Paper deep dive
Machine Learning and Time-Series Foundation Models for Day-Ahead Electricity Forecasting in a University Building
Amey Vasulkar, Samuel Kernan Freire, Arie-Willem de Leeuw, Tadeo-Baldiri Salcedo-Rahola
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 10/12/2026, 6:11:17 AM
Summary
This study benchmarks statistical, machine learning, recurrent neural network, and pretrained TimesFM foundation models for day-ahead electricity forecasting in a university building. Using chronological validation on 57 test days, TimesFM 3.0 with covariates achieved the lowest overall mean absolute error (MAE), reducing it by approximately 43% compared to weekly seasonal-naive forecasting. XGBoost remained a competitive alternative with rapid local fitting, while LSTM performed poorly. The research highlights the trade-off between forecast accuracy, computational requirements, and the utility of external covariates, particularly for holiday periods.
Entities (10)
Relation Signals (7)
XGBoost → offers → rapid local fitting
confidence 95% · XGBoost offers rapid local fitting and inference.
TimesFM 3.0 → outperforms → weekly seasonal-naive forecasting
confidence 95% · TimesFM 3.0 with covariates achieves the lowest overall errors, reducing mean absolute error by approximately 43% relative to weekly seasonal-naive forecasting.
TimesFM 3.0 → requires → no building-specific weight training
confidence 95% · TimesFM requires no building-specific weight training, while XGBoost offers rapid local fitting and inference.
TimesFM 3.0 → uses → pretrained weights
confidence 95% · We use the pretrained TimesFM 2.5 and 3.0 weights supplied by Google Research
Covariates → improves → holiday forecasts
confidence 90% · Adding covariates improves its holiday forecasts but slightly worsens workday and weekend forecasts
LSTM → underperforms → TimesFM 3.0
confidence 90% · The evaluated LSTM takes longer to fit without outperforming either approach.
TimesFM 3.0 → iscompetitivewith → XGBoost
confidence 85% · XGBoost remains competitive, although paired daily comparisons do not establish a clear mean-absolute-error advantage for TimesFM.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Day-ahead electricity forecasts can help building operators schedule heating, ventilation, and other services around anticipated demand. Choosing a forecasting model also requires balancing accuracy with its forecast-time inputs, local training, and computational requirements. We benchmark statistical, machine-learning, recurrent neural-network, and pretrained TimesFM models for a university building. We use chronological validation and evaluate all configurations on the same 57 test days, comparing accuracy across workdays, weekends, and holidays alongside local fitting and forecasting time. Weather covariates use retrospective observations. TimesFM 3.0 with covariates achieves the lowest overall errors, reducing mean absolute error by approximately 43\% relative to weekly seasonal-naive forecasting. XGBoost remains competitive, although paired daily comparisons do not establish a clear mean-absolute-error advantage for TimesFM. TimesFM also performs strongly using only past electricity consumption. Adding covariates improves its holiday forecasts but slightly worsens workday and weekend forecasts, leaving overall mean absolute error almost unchanged. TimesFM requires no building-specific weight training, while XGBoost offers rapid local fitting and inference. The evaluated LSTM takes longer to fit without outperforming either approach. An exploratory historical evaluation reveals seasonal variation, while a separate weather-source comparison produces only small changes in the selected TimesFM configuration's accuracy. These results support pretrained TimesFM as a strong initial forecasting baseline for this building, with XGBoost as a competitive locally fitted alternative. Covariates are particularly useful when improved holiday forecasts justify the additional input preparation.
Tags
Links
- Source: https://arxiv.org/abs/2610.12638v1
- Canonical: https://arxiv.org/abs/2610.12638v1
Trouble viewing inline? Open PDF directly →
Full Text
81,841 characters extracted from source content.
Expand or collapse full text
captionstableheading Machine Learning and Time-Series Foundation Models for Day-Ahead Electricity Forecasting in a University Building Amey N. Vasulkar Samuel Kernan Freire Arie-Willem de Leeuw Tadeo-Baldiri Salcedo-Rahola 2026-10-08 Abstract Day-ahead electricity forecasts can help building operators schedule heating, ventilation, and other services around anticipated demand. Choosing a forecasting model also requires balancing accuracy with its forecast-time inputs, local training, and computational requirements. We benchmark statistical, machine-learning, recurrent neural-network, and pretrained TimesFM models for a university building. We use chronological validation and evaluate all configurations on the same 57 test days, comparing accuracy across workdays, weekends, and holidays alongside local fitting and forecasting time. Weather covariates use retrospective observations. TimesFM 3.0 with covariates achieves the lowest overall errors, reducing mean absolute error by approximately 43% relative to weekly seasonal-naive forecasting. XGBoost remains competitive, although paired daily comparisons do not establish a clear mean-absolute-error advantage for TimesFM. TimesFM also performs strongly using only past electricity consumption. Adding covariates improves its holiday forecasts but slightly worsens workday and weekend forecasts, leaving overall mean absolute error almost unchanged. TimesFM requires no building-specific weight training, while XGBoost offers rapid local fitting and inference. The evaluated LSTM takes longer to fit without outperforming either approach. An exploratory historical evaluation reveals seasonal variation, while a separate weather-source comparison produces only small changes in the selected TimesFM configuration’s accuracy. These results support pretrained TimesFM as a strong initial forecasting baseline for this building, with XGBoost as a competitive locally fitted alternative. Covariates are particularly useful when improved holiday forecasts justify the additional input preparation. Introduction Buildings have been estimated to account for 36% of global energy use and 40% of associated carbon dioxide emissions (Khalil et al., 2022). Improving their operation complements improvements to the building fabric and installed equipment. In particular, scheduling heating, ventilation, and air conditioning (HVAC) and other building services around anticipated demand provides an opportunity to align operation with building activity. Reviews of data-driven building energy modelling identify forecasting as a component of such management and control (Bourdeau et al., 2019; Khalil et al., 2022). Day-ahead electricity forecasting supplies the next 24 hourly electricity-consumption values before the forecast day begins. The hourly profile can inform service scheduling and identify periods of expected high demand, whereas a daily total cannot distinguish when that demand occurs. Producing such forecasts routinely requires a model that can capture changing consumption patterns from the available data. Building operators must therefore choose both a forecasting method and a procedure for preparing its inputs and keeping it up to date. The practical value of that choice depends on forecast accuracy, covariate requirements, local fitting, and inference resources. Statistical models, tree ensembles, and recurrent neural networks offer different combinations of these requirements (Khalil et al., 2022). Pretrained time-series foundation models introduce another possibility by forecasting new series without building-specific weight training (Das et al., 2024). Whether that transfer provides a useful alternative to a well-tuned local model remains an empirical question for the building concerned. University buildings provide an informative setting for this comparison because teaching days, weekends, examinations, and academic holidays produce changes in operating schedules. Studies of academic calendars, occupancy, and campus electricity consumption motivate assessing these periods separately (Alfalah et al., 2023; Ghanavati et al., 2023; Gui et al., 2021). A forecast can reproduce routine teaching-day variation while still overstating consumption during a closure. These schedule changes make average performance across all hours an incomplete basis for choosing a model. In this paper, we examine a modern, well-insulated teaching building at the Delft campus of The Hague University of Applied Sciences (THUAS), extending our preceding study of the same building (Stienstra et al., 2026). For models using weather covariates, practical application also requires a suitable weather source. Historical evaluations can use station observations or reanalysis, while next-day weather must be supplied by a forecast issued beforehand. Differences between these products may affect the electricity-consumption forecast, making weather-source sensitivity relevant alongside model accuracy. This study contributes a temporally validated comparison of locally fitted and pretrained forecasting configurations, together with evidence on when their accuracy differences matter. We compare models on the same forecast days, assess differences across workdays and closures, and quantify uncertainty in their mean daily errors. An extended historical TimesFM evaluation examines seasonal behaviour beyond the winter-weighted test; a controlled weather-source comparison examines whether its forecasts depend strongly on the supplied weather product. Together with measured fitting and forecast times, these analyses support model selection according to accuracy, input requirements, and local modelling effort. Our target is electricity consumption: tariff-dependent costs and savings from subsequent control decisions are not evaluated. Related Work The reviews by Bourdeau et al. (2019) and Khalil et al. (2022) document a broad range of statistical, machine-learning, and deep-learning approaches to building energy forecasting. They organise the evidence by building characteristics, available data, preprocessing, forecast horizon, and evaluation method. These distinctions are important when interpreting reported performance: a method evaluated on daily consumption for one building is not directly comparable with a method predicting the hourly profile of another. The literature therefore motivates comparison under explicit experimental conditions rather than selection of an algorithm solely from a published accuracy value. University-building studies illustrate the importance of both weather and operating schedules. Amber et al. (2017) examined five years of daily electricity consumption for two buildings at London South Bank University, using multiple regression with weather, weekday, and building-type covariates. Their 2011 evaluation reported normalised RMSE of 12% for the administrative building and 13% for the academic building; weekday index, temperature, and building type were significant predictors. This provides an applied precedent for a comparatively simple model with contextual covariates. However, its daily resolution does not establish performance for the 24-hour profile required by day-ahead scheduling. Studies of academic calendars and occupancy further motivate distinguishing teaching activity from closures, without implying that realised attendance is known in advance (Alfalah et al., 2023; Gui et al., 2021). Our preceding THUAS study provides a directly relevant comparison of multiple linear regression, Holt–Winters, Prophet, XGBoost, and LSTM for the same building (Stienstra et al., 2026). Its reported results cover 51 workdays: Prophet had MAE 17.78 kWh and R2=0.92R^2=0.92, compared with MAE 22.53 kWh for XGBoost and 29.21 kWh for LSTM. XGBoost nevertheless had the lowest MAPE and mean squared error. Thus, even within one study, the preferred model depended on the error measure, and the recurrent neural network did not provide the best reported accuracy. The study also identified difficulties on non-workdays, when lower consumption and different temporal patterns complicated forecasting. Its workday-only results leave open how the comparison changes when closures are included and how performance varies over a longer seasonal record. Hybrid approaches combine models to improve different parts of a forecast. For day-ahead forecasting in a German energy community, Semmelmann et al. (2022) used a bidirectional LSTM to predict the general consumption pattern and XGBoost to improve peak timing and magnitude. Adding the peak correction reduced MAPE by 0.14 percentage points overall and 3.55 percentage points for peaks. Furthermore, the larger peak-specific gain also shows why an overall score can hide improvements during the periods a model component is designed to address. In university buildings, the preceding THUAS study identifies closures as another period requiring separate attention (Stienstra et al., 2026). These studies motivate reporting errors for distinct operating periods alongside the pooled score. Pretrained forecasting studies introduce a different route to accuracy. Das et al. (2024) reported zero-shot TimesFM performance approaching supervised, dataset-specific methods across public benchmark collections. In an applied electricity-demand comparison, Llanes-Guilarte et al. (2026) evaluated TimesFM 2.0 and Lag-Llama on small-community consumption data and reported that their relative performance depended on temporal resolution. These studies support investigating pretrained transfer, but they do not establish its accuracy for a particular university building or for later TimesFM versions. Building-level evaluation must still examine context requirements, the value of external covariates, and performance during schedule changes. The remaining gap is a clear account of how locally fitted and pretrained methods compare when the building, forecast hours, and model-selection procedure are held consistent. Their performance across routine operation, closures, and seasons also needs to be considered alongside the effort of preparing covariates and fitting models. Weather introduces an additional uncertainty: the station observations or reanalysis used in a historical evaluation differ from the forecasts available for next-day operation. Establishing how strongly those source differences affect electricity-consumption forecasts would make the model comparison more useful in practice. Research Questions We address four questions: 1. How do the selected statistical, machine-learning, LSTM, and TimesFM configurations compare with daily and weekly seasonal-naive forecasts on the same 24-hour targets? 2. Are differences in forecast accuracy consistent across workdays, weekends, and holidays, and what does an extended historical evaluation of TimesFM reveal about seasonal performance? 3. How sensitive is a fixed TimesFM 3.0 model with covariates to KNMI observations (Royal Netherlands Meteorological Institute, n.d.), ERA5 reanalysis (Hersbach et al., 2020), and historical ECMWF IFS weather (Open-Meteo, n.d.-a)? 4. What local fitting time, forecast time, memory use, and historical context are required by the evaluated configurations? The comparison among all models uses the same 57-day test set. Two additional analyses address the interpretation of this winter-weighted benchmark: the extended historical evaluation examines TimesFM across more seasons and closure periods, while the weather-source comparison tests its sensitivity to the supplied weather product. Neither analysis changes the selected model settings. We report the resource quantities measured during execution; total pretraining cost and complete model-storage costs are outside this comparison. Methods and Data Building, Electricity Consumption, and Time Convention We use hourly whole-building electricity consumption for the THUAS Delft campus building. The institutional dataset was assembled for the preceding site study, which identifies electricity measurements from the Joulz portal and scheduling data from THUAS’s planning systems (Stienstra et al., 2026). Electricity consumption is the response variable, measured in kWh for each hourly interval. The dataset contains 13,752 hourly observations arranged in 573 consecutive 24-hour forecast windows. It was supplied as a training/validation file with 12,384 observations and a test file with 1,368 observations. Both files contain a timestamp, electricity consumption, and 47 candidate covariates. We use the source files together for descriptive statistics while preserving their chronological separation for model selection and testing. All timestamps are reported in UTC. The complete series spans 4 September 2022 at 23:00 UTC through 30 March 2024 at 22:00 UTC; forecasts are issued daily at 23:00 UTC for the next 24 hourly targets. We retain the institutional operating-day labels for calendar comparisons. Appendix Section Time Convention and Weather Alignment documents the time conversion and weather alignment. Figure 1: Whole-building hourly electricity consumption from 4 September 2022 at 23:00 UTC to 30 March 2024 at 22:00 UTC. The upper panel marks the validation and test periods; the lower panel expands the first complete Monday–Sunday week in UTC to show the hourly variation. Both panels retain the original hourly resolution. Weather Data The main benchmark uses hourly KNMI observations from Rotterdam Airport station 344, the station included in the supplied building dataset (Royal Netherlands Meteorological Institute, n.d.). The preceding site study describes the selected station as approximately 4 km from the building (Stienstra et al., 2026). These observations provide a regional weather reference rather than measurements at the building itself. To examine whether the choice of this reference affects electricity-consumption forecasts, we also use ERA5 reanalysis and historical ECMWF IFS forecasts obtained through Open-Meteo (Hersbach et al., 2020; Open-Meteo, n.d.-b, n.d.-a). Table 1 distinguishes their roles. Table 1: Weather sources and their role. The main 11-model benchmark uses KNMI for weather covariates; the additional source comparison uses a fixed TimesFM 3.0 configuration. Source Product Role in the comparison KNMI Observations at Rotterdam Airport station 344 Observed-weather reference ERA5 Gridded reanalysis Retrospective weather reconstruction Historical IFS Archived ECMWF IFS forecasts from Open-Meteo Forecast-product substitution The supplied dataset contains 21 candidate weather quantities. We retain the seven with corresponding definitions across all three products and harmonise their units: air temperature, dew-point temperature, relative humidity, mean-sea-level pressure, hourly precipitation, wind speed, and wind direction. The other 14 weather quantities are excluded, as described in the next section. The retained quantities’ spatial and temporal sampling are not identical; Appendix Section Time Convention and Weather Alignment documents the alignment and remaining differences. Reanalysis describes past conditions using observations and a numerical model; it is not a weather forecast issued before the target day. Historical IFS is an archived forecast product, although the downloaded records do not identify when each individual value became available. These distinctions let us test weather-source sensitivity without assuming that all three products were accessible at the historical forecast origin. Forecast-Time Inputs We reduce the 47 candidate covariates to 30 in two steps. First, we remove 14 of the 21 candidate weather quantities because they lack consistently defined counterparts across the three sources, retaining the seven listed in Section Weather Data. The removed quantities include weather codes and additional daily aggregates. Second, we exclude three further covariates: realised people counts, cumulative room bookings, and the one-hour electricity-consumption lag, whose target-period values are unavailable for the complete day-ahead forecast. Thus, 47−14−3=3047-14-3=30 covariates remain: 21 calendar and schedule quantities, seven weather quantities, and two historical electricity-consumption lags (Table 2). Models supporting external covariates use these same quantities; seasonal-naive forecasts and target-only LSTM and TimesFM models use past electricity consumption, while Prophet base uses its trend, seasonal, and holiday components. Table 2: The 30 retained covariates, grouped into three domains. Availability refers to the forecast origin. Appendix Section Covariate Catalogue gives the individual stored variables. Domain Covariates (number) Units Availability Calendar and schedule (21) Hour, year, month, week, day, day of week (6) Calendar indices Known in advance Sine/cosine pairs for hour, day, day of week, week, month (10) Dimensionless Known in advance Work-week, weekend, holiday, opening-hours, examination-week indicators (5) Binary Calendar known; schedules assumed known Weather (7) Air and dew-point temperatures (2) °C Retrospective Relative humidity; sea-level pressure (2) %; hPa Retrospective Hourly precipitation (1) m Retrospective Wind speed; wind direction (2) m/s; degrees Retrospective Historical consumption (2) Electricity consumption 24 and 168 hours earlier (2) kWh Observed before the forecast origin We assume that the supplied opening-hours and examination schedules were available before the forecast day; their historical publication times were not retained. Realised people counts and cumulative room bookings describe activity during the target period. The one-hour lag would require unobserved consumption after the first forecast step, whereas the 24- and 168-hour lags are available for every target hour. Models The benchmark contains 11 configurations, listed in Table 3. Ridge is multiple linear regression with a penalty that shrinks its coefficients (Hoerl & Kennard, 1970). XGBoost combines regression trees and can represent nonlinear relationships and interactions between covariates (Chen & Guestrin, 2016). Prophet combines trend, daily and weekly seasonality, and holiday effects; we test it with and without external covariates (Taylor & Letham, 2018). The LSTM uses recurrent memory cells to learn temporal relationships (Hochreiter & Schmidhuber, 1997); our implementation receives sequences of past electricity consumption and produces the next 24 hourly values. Its hybrid adds a ridge prediction of the LSTM residual. TimesFM is a decoder-only time-series foundation model pretrained on collections of real and synthetic series (Das et al., 2024). Zero-shot forecasting means that its network weights are not fitted to this building; past electricity consumption is still required as context. We use the pretrained TimesFM 2.5 and 3.0 weights supplied by Google Research, retaining 2.5 as a version reference (Google Research, 2025, 2026). TimesFM 3.0, released in August 2026, supports native multivariate forecasting with historical and future covariates. We evaluate both versions using target history alone, and additionally evaluate TimesFM 3.0 with the same 30 covariates used by the regression-based methods. This native interface differs from the external ridge residual modelling used in the TimesFM 2.5 hybrids. Context length and inference settings are selected by validation, without changing the pretrained weights. Table 3: Models and their forecast-time inputs. Prophet represents the holiday indicator through its native holiday component, leaving 29 external regressors. Configuration Forecast-time inputs Building-specific fitting Seasonal-naive 24 h Previous day’s consumption at the corresponding hour None Seasonal-naive 168 h Previous week’s consumption at the corresponding hour None Ridge 30 covariates Standardisation and regression fitted at each origin XGBoost 30 covariates Regression trees fitted at each origin Prophet base Past consumption and native calendar/holiday components Trend, seasonality, and holiday effects fitted at each origin Prophet with regressors Past consumption, 29 external covariates, and native holiday effects Prophet and its regressors fitted jointly at each origin TimesFM 2.5 target-only Past consumption Pretrained weights remain fixed LSTM base Past consumption Network and scaling fitted at each origin LSTM + ridge LSTM forecast and 30 covariates for residual correction LSTM and residual regression fitted from prior observations TimesFM 3.0 target-only Past consumption Pretrained weights remain fixed TimesFM 3.0 with covariates Past consumption and 30 covariates Pretrained weights remain fixed For a forecast origin t, let yt+h∈ℝy_t+h denote electricity consumption at target hour t+ht+h, with h=0,…,23h=0,…,23. At each of these 24 target hours, t+h∈ℝ30x_t+h ^30 contains the 30 covariates; y<ty_<t denotes past consumption. The hour index h and the number of covariates therefore describe different dimensions. A model M with parameters θ and external covariates produces the forecast in Equation 1: y^t+h=Mθ(y<t,t+h),h=0,…,23. y_t+h=M_θ\! (y_<t,x_t+h ), h=0,…,23. (1) The seasonal-naive forecasts in Equation 2 copy an observed value from the corresponding hour one day or one week earlier: y^t+h(k)=yt+h−k,k∈24,168. y_t+h^(k)=y_t+h-k, k∈\24,168\. (2) For the LSTM hybrid, rjr_j in Equation 3 is the observed error of an earlier base forecast at target hour j. Ridge models that error from the covariates, and its prediction is added to the LSTM forecast: rj=yj−y^jLSTM,y^t+hhybrid=y^t+hLSTM+r^t+h.r_j=y_j- y^LSTM_j, y^hybrid_t+h= y^LSTM_t+h+ r_t+h. (3) Only residuals whose actual consumption is known before t enter the correction fit. The correction is added once. This follows the general residual-modelling principle of Zhang (2003). Appendix Section Additional TimesFM 2.5 Hybrid Experiments compares the TimesFM 2.5 target-only model with two jointly tuned residual hybrids: TimesFM followed by ridge correction, and ridge followed by TimesFM forecasting of residuals. These additional experiments examine whether external residual modelling improves on the base forecast. TimesFM 2.5 target-only also remains in the main benchmark as a version reference. Training, Validation, and Test Design We use sliding-window, rolling-origin cross-validation, following the chronological evaluation principle described by Tashman (2000). At each forecast origin, refitted models use the immediately preceding 8,760 hourly observations, equivalent to 365 days, and predict the following 24 hours. Both ends of the training window then advance by 24 hours. Scalers and imputers are fitted only to that origin’s history. XGBoost reserves the final 14 days of its training window for early stopping, while the LSTM uses the final 10% of its training sequences. Prophet’s yearly seasonality is disabled because each fit contains only one annual cycle. TimesFM keeps its pretrained weights fixed and receives a context length selected by validation. Table 4: Chronological evaluation periods. Each forecast day begins at 23:00 UTC and contains 24 consecutive targets. Stage Target interval in UTC Forecast days Hourly targets Purpose Initial history 4 September 2022, 23:00–4 November 2023, 22:00 426 10,224 Supply past observations for subsequent forecasts Validation 4 November 2023, 23:00–2 February 2024, 22:00 90 2,160 Select hyperparameters Test 2 February 2024, 23:00–30 March 2024, 22:00 57 1,368 Evaluate the selected configurations A hyperparameter candidate is one complete combination of settings. Each candidate is evaluated over the same 90 consecutive daily folds, producing 2,160 validation predictions. We select the lowest pooled RMSE, using MAE and then candidate order to break ties. We then hold those settings fixed during the 57-day test and continue advancing the forecast origin by one day; models whose policy requires daily fitting are refitted on the available history. The test is not used to tune these settings. Ridge uses a grid of 25 log-spaced penalties. Optuna’s tree-structured Parzen estimator proposes 96 trials for XGBoost, 96 for each Prophet configuration, 72 for the LSTM base, and 48 for its ridge residual correction (Akiba et al., 2019). All five Optuna searches use seed 20260729 and disable trial pruning, so every candidate is scored over the complete 90-day validation period. This differs from early stopping within an individual XGBoost or LSTM fit, which limits fitting iterations without dropping forecast days. TimesFM 2.5 uses a grid of eight combinations of context length and input normalisation. Each TimesFM 3.0 configuration uses a 64-combination grid: eight context lengths from 168 to 8,760 hours, symmetric averaging on/off, a non-negative-output option on/off, and input z-normalisation on/off. The forecast horizon remains 24 hours throughout. The selected TimesFM 2.5 context is 672 hours. Both TimesFM 3.0 configurations select 8,760 hours, symmetric averaging, no non-negative-output transform, and no optional z-normalisation; the point prediction is the median forecast quantile. The selected LSTM uses a 168-hour lookback and one 64-unit layer. Complete validation scores and selected settings are reported in Appendix Section Validation and Selected Settings. Descriptive Analysis We summarise the distribution of hourly electricity consumption, mean hourly profiles, and differences between workdays, weekends, and holidays. A workday is a day labelled as neither a holiday nor a weekend; a holiday label takes precedence over a weekend label when forming the three separate groups. We report counts with every comparison. Pearson correlations describe the direction and strength of linear associations between consumption and individual covariates. Full-period correlations characterise the dataset; they are not used to select models from the test outcomes. Daily and weekly autocorrelation and stationarity diagnostics provide supporting descriptions in the appendix. The main Results section focuses on consumption levels, calendar patterns, and the three covariate domains. Additional TimesFM Analyses We use two supplementary analyses to examine conditions not resolved by the main model comparison. The extended historical evaluation asks how forecast errors change across seasons and closure periods beyond the short winter test. The weather-source comparison asks whether a TimesFM forecast changes materially when the weather description changes. Both use the selected TimesFM 3.0 configuration with covariates; neither involves further hyperparameter selection or building-specific weight training. For the historical evaluation, we start with 168 hours of context at 11 September 2022, 23:00 UTC, and advance the forecast origin by 24 hours for 566 daily forecasts. Context grows by 24 hours per origin until it reaches 8,760 hours, after which the window slides at that fixed length. Every forecast covers only the next 24 hours and receives only earlier electricity-consumption observations. The pretrained model can therefore be examined during the first year of this short record without reserving that entire year for local weight fitting. The weekly seasonal-naive forecast uses the corresponding hours from the preceding week and provides a reference for changing forecasting difficulty. This is an advantage of the selected pretrained evaluation procedure, not a claim that shorter-history evaluation is impossible with other methods. We aggregate the historical forecast errors by workday, weekend, holiday, UTC calendar month, and meteorological season. Winter comprises December–February, spring March–May, summer June–August, and autumn September–November. Monthly and seasonal groups follow target timestamps and may divide a forecast window; we report their hourly counts and context ranges. Because the fixed settings were selected using later validation data, this extended historical evaluation is exploratory rather than an independent all-season test. The weather comparison uses the three products described in Section Weather Data. It keeps electricity-consumption history, the other 23 covariates, forecast windows, and all TimesFM 3.0 settings fixed, replacing only the seven weather covariates in both the 8,760-hour context and target hours. We harmonise units, use mean-sea-level pressure throughout, and assign weather to matching hourly intervals in UTC. This tests substitution of one complete weather source for another; it does not test removing weather or independently changing only the future weather values. Appendix Section Time Convention and Weather Alignment describes the interval mapping and treatment of missing wind directions. This comparison uses an explicit interval assignment across all three weather sources. Its KNMI forecasts are therefore reported separately from the original benchmark, which retains the supplied weather join. The archived IFS data allow a forecast-product comparison, but the absence of per-value issue times prevents reconstructing exactly which forecast an operator could have obtained at each historical origin. Accuracy and Day-Level Uncertainty We compute pooled hourly MAE, RMSE, R2R^2, MAPE, and sMAPE, combining absolute and percentage error measures to describe forecast accuracy (Hyndman & Koehler, 2006). For actual consumption yiy_i, forecast y^i y_i, error ei=yi−y^ie_i=y_i- y_i, n target hours, and mean observed consumption y¯ y, Equation 4 defines these measures: MAE =1n∑i|ei|, = 1n _i|e_i|, RMSE =1n∑iei2, = 1n _ie_i^2, (4) R2 R^2 =1−∑iei2∑i(yi−y¯)2, =1- _ie_i^2 _i(y_i- y)^2, MAPE =100n∑i|eiyi|, = 100n _i | e_iy_i |, sMAPE =200n∑i|ei||yi|+|y^i|. = 200n _i |e_i||y_i|+| y_i|. MAE gives the average size of an hourly error in kWh. RMSE gives larger errors more weight because it squares errors before averaging. Reporting both reveals the influence of large misses on overall accuracy. Median absolute error would describe the middle error but would give little information about that upper tail. MAPE expresses errors relative to measured consumption; its denominator is nonzero throughout this dataset, whose minimum consumption is 13.9 kWh. We calculate R2R^2 from the pooled target hours of the reported group, not by averaging daily R2R^2 values. Its denominator measures variation around that group’s mean. During a relatively flat closure period, even modest errors can therefore produce a low or negative R2R^2; for a constant series it is undefined. A negative value means that squared forecast error exceeds the error of the group’s mean-value reference, not that consumption has no temporal pattern. We interpret R2R^2 alongside MAE and RMSE, which retain a directly interpretable hourly kWh scale across workdays and closures (Hyndman & Koehler, 2006). To assess whether the pooled differences depend on a few forecast days, we calculate a daily MAE from each 24-hour forecast and compare models on the same days. XGBoost provides a common fitted-model reference; this choice sets the direction of the reported differences and does not determine which model wins. We also compare the TimesFM versions directly. For model m and reference q, the daily difference in Equation 5 is negative when model m is more accurate: d(m,q)=124∑h=023(|yd,h−y^d,h(m)|−|yd,h−y^d,h(q)|).d_d^(m,q)= 124 _h=0^23 (|y_d,h- y^(m)_d,h|-|y_d,h- y^(q)_d,h| ). (5) We resample the 57 paired days with replacement 10,000 times, using seed 20260727, and report the 2.5th and 97.5th percentiles of the mean difference as a percentile bootstrap interval (Davison & Hinkley, 1997). Keeping each day together preserves dependence among its 24 hourly errors. We apply the same procedure to the three pairwise weather-source comparisons, using the fixed TimesFM 3.0 configuration throughout. An interval containing zero indicates that these days do not clearly establish the direction of the mean difference; it does not demonstrate equal performance. The intervals are not adjusted for multiple comparisons and do not preserve dependence between successive days. Day-type scores are additional descriptive comparisons and do not select hyperparameters. To illustrate the hourly forecast profiles, we select the first chronological test forecast origin in each of the workday, weekend, and holiday groups. We display the observed consumption and saved forecasts from TimesFM 3.0 with and without covariates and XGBoost for those windows. This rule is independent of forecast errors; the aggregate comparisons use all 57 test days. Computational Measurements We measure fitting or updating time separately from the time needed by an already-loaded model to forecast 24 hours. We report the median and 95th percentile across test origins, peak process memory, and peak GPU memory where recorded. For the LSTM hybrid, the correction time is reported separately from the LSTM training time it requires. These measurements describe the executed configurations on their recorded hardware. They exclude the foundation model’s original pretraining cost and are not a comparison of total model-development costs. Results Electricity Consumption and Calendar Patterns The dataset has no missing or duplicated hourly timestamps and no missing or negative electricity-consumption values. Mean hourly consumption is 140.31 kWh, the median is 112.15 kWh, and values range from 13.9 to 452.0 kWh (Table 5). The time series in Figure 1 shows changes across the study period, while the expanded week makes the repeated daytime rises and overnight consumption visible. The marked validation and test periods locate the model comparison within this longer record. Table 5: Electricity-consumption summary over the complete dataset. Consumption values are kWh per hourly interval. Workdays exclude weekends and holidays; holiday labels take precedence. Group Hourly observations Mean (kWh) Median (kWh) SD (kWh) Minimum (kWh) Maximum (kWh) Full series 13752.0 140.31 112.15 76.86 13.90 452.00 Holiday 2760 87.77 77.10 33.57 13.90 374.30 Weekday 8040 176.23 161.25 79.16 56.30 452.00 Weekend 2952 91.58 83.00 30.47 46.10 306.10 Mean hourly consumption is substantially higher on workdays (176.23 kWh) than on weekends (91.58 kWh) or holidays (87.77 kWh). The separation is concentrated in daytime operation (Figure 2). Figure 2: Mean hourly electricity-consumption profiles for workdays, weekends, and institutional holidays over the complete dataset. Direct labels and distinct line styles identify the three groups. Hours are reported in UTC; the original operating-day memberships are retained. The grouped correlations in Figure 3 show the relative associations of calendar, weather, and past-consumption covariates with electricity consumption. Opening hours and the daily and weekly lags have strong positive associations, with full-period Pearson correlations of approximately 0.79, 0.69, and 0.80, respectively. The daily and weekly autocorrelations are 0.687 and 0.788 over the full series (Appendix Section Supporting Descriptive Diagnostics). Figure 3: Pearson correlations between the 30 retained covariates and electricity consumption over the complete dataset. The calendar/schedule, weather, and historical-consumption panels use a common correlation scale and consistent bar thickness; the zero line separates positive and negative associations. Target-day weather is retrospective. Benchmark Accuracy All 11 configurations in Table 3 forecast the same 57 test windows and 1,368 hourly targets. Table 6 reports their pooled errors and R2R^2. Figure 4 places the configurations in the same MAE order in both panels, allowing their MAE and RMSE positions to be compared directly. Table 6: Test accuracy for the 11 model configurations. MAE and RMSE are in kWh; lower errors and higher R2R^2 are better. Weather covariates use KNMI observations as supplied in the original benchmark. Model Forecast-time inputs MAE (kWh) RMSE (kWh) R² MAPE (%) sMAPE (%) TimesFM 3.0 with covariates Calendar, lags and observed weather 19.97 30.26 0.798 16.94 15.27 TimesFM 3.0 target-only Past consumption 20.01 32.45 0.768 16.99 14.94 TimesFM 2.5 target-only Past consumption 21.07 31.85 0.777 18.50 16.09 XGBoost Calendar, lags and observed weather 22.45 31.11 0.787 18.98 16.86 Prophet with regressors Calendar, lags and observed weather 24.91 34.22 0.742 21.39 19.73 Ridge Calendar, lags and observed weather 27.18 37.87 0.684 22.93 20.43 LSTM + ridge residual Calendar, lags and observed weather 30.07 43.53 0.583 24.85 22.36 Prophet base Calendar and holidays 30.22 38.19 0.679 28.83 24.99 LSTM base Past consumption 30.84 46.22 0.530 25.99 22.56 Seasonal-naive 168 h Past consumption 35.27 53.26 0.376 29.29 24.80 Seasonal-naive 24 h Past consumption 37.57 60.45 0.196 31.93 26.87 Figure 4: Pooled MAE and RMSE on the 1,368 common test targets, with configurations ordered by increasing MAE from top to bottom in both panels. Points further left indicate lower error. Colours and marker shapes distinguish TimesFM configurations, XGBoost, and the supporting comparisons; every configuration is labelled. TimesFM 3.0 with covariates has the lowest MAE (19.97 kWh), RMSE (30.26 kWh), and MAPE (16.94%), with R2=0.798R^2=0.798. Its MAE is almost identical to the target-only version’s 20.01 kWh, but its RMSE is lower by 2.19 kWh. TimesFM 3.0 target-only has the lowest sMAPE (14.94%). XGBoost has MAE 22.45 kWh and RMSE 31.11 kWh; it therefore has higher MAE but slightly lower RMSE than either target-only TimesFM version. The weekly seasonal-naive forecast improves on the daily forecast, with MAE 35.27 versus 37.57 kWh. Relative to the weekly reference, TimesFM 3.0 with covariates reduces MAE by 15.30 kWh, or 43.4%. Prophet’s external regressors reduce MAE from 30.22 to 24.91 kWh. The LSTM residual correction gives a smaller reduction, from 30.84 to 30.07 kWh. Differences Across Forecast Days The paired comparisons in Figure 5 show how the mean daily-MAE differences vary when forecast days are resampled. The numerical intervals are given in Table 7. Table 7: Mean daily-MAE differences from XGBoost and 95% paired bootstrap intervals, in kWh. Negative differences favour the listed model; positive differences favour XGBoost. Model Reference Mean daily-MAE difference (kWh) 95% percentile CI (kWh) Paired days Seasonal-naive 24 h XGBoost 15.12 [8.338, 22.252] 57 Seasonal-naive 168 h XGBoost 12.82 [8.421, 17.742] 57 Ridge XGBoost 4.73 [2.841, 6.671] 57 Prophet base XGBoost 7.76 [4.260, 10.958] 57 Prophet with regressors XGBoost 2.46 [0.422, 4.622] 57 TimesFM 2.5 target-only XGBoost -1.38 [-4.636, 2.218] 57 TimesFM 3.0 target-only XGBoost -2.44 [-5.825, 1.342] 57 TimesFM 3.0 with covariates XGBoost -2.48 [-5.140, 0.132] 57 LSTM base XGBoost 8.39 [3.526, 13.461] 57 LSTM + ridge residual XGBoost 7.62 [3.427, 11.903] 57 Figure 5: Mean differences in daily MAE and their 95% bootstrap intervals. The all-model comparison uses XGBoost as the reference; the additional panel compares TimesFM configurations directly. Differences are the named model minus its reference, so negative values favour the named model. An interval crossing zero does not establish a direction of improvement. TimesFM 3.0 with covariates has a mean daily-MAE difference of -2.48 kWh from XGBoost, but its interval spans -5.14 to 0.13 kWh. The target-only TimesFM 3.0 and TimesFM 2.5 comparisons also include zero. Direct comparisons between TimesFM 3.0 and 2.5 are similarly inconclusive. The native-covariate minus target-only TimesFM 3.0 difference is only -0.04 kWh, with an interval from -2.735 to 2.090 kWh. Thus, the lowest pooled point estimate does not establish a clear daily-MAE advantage among these configurations over the 57 days. The intervals comparing both seasonal-naive forecasts, ridge, both Prophet configurations, and both LSTM configurations with XGBoost are wholly positive. Within this test, these comparisons provide clearer evidence of lower mean daily MAE for XGBoost. They remain comparisons of the particular fitted configurations and covariates used here. Workdays, Weekends, and Holidays The test contains 34 workdays (816 hours) and 23 non-workdays (552 hours). Non-workdays comprise 12 weekend days and 11 holidays. Table 8 reports the complete model comparison. Figure 6 separates workdays, weekends, and holidays for the three TimesFM configurations and XGBoost, showing how their relative performance changes with day type and error measure. The corresponding figure for all 11 configurations is provided in Appendix Section Complete Day-Type Comparison. Table 8: Pooled test errors by day type, in kWh. Subgroup rankings describe the observed errors and have not been tested for statistical significance. Model Workday MAE (kWh) Workday RMSE (kWh) Non-workday MAE (kWh) Non-workday RMSE (kWh) TimesFM 3.0 with covariates 18.85 26.00 21.63 35.63 TimesFM 3.0 target-only 16.65 23.00 24.98 42.75 TimesFM 2.5 target-only 17.71 23.25 26.03 41.41 XGBoost 22.88 32.15 21.81 29.50 Prophet with regressors 23.18 31.56 27.47 37.80 Ridge 27.41 37.88 26.84 37.85 LSTM + ridge residual 30.26 44.31 29.80 42.35 Prophet base 25.88 31.97 36.62 45.87 LSTM base 30.26 46.33 31.70 46.07 Seasonal-naive 168 h 32.29 47.59 39.68 60.67 Seasonal-naive 24 h 36.27 57.75 39.50 64.25 Figure 6: MAE and RMSE for TimesFM 2.5, TimesFM 3.0 with and without covariates, and XGBoost, evaluated separately on 34 workdays, 12 weekend days, and 11 holidays. Model settings are identical across groups. Each metric uses a common scale across the three day types; subgroup differences are descriptive. TimesFM 3.0 target-only has the lowest workday MAE and RMSE, 16.65 and 23.00 kWh. On non-workdays, TimesFM 3.0 with covariates has slightly lower MAE than XGBoost (21.63 versus 21.81 kWh), but higher RMSE (35.63 versus 29.50 kWh). Within TimesFM 3.0, adding covariates reduces holiday RMSE from 53.74 to 39.47 kWh. In contrast, workday RMSE increases from 23.00 to 26.00 kWh and weekend RMSE from 29.25 to 31.71 kWh. Figure 7 complements these aggregate errors with the hourly profiles from the first test forecast in each day-type group. The observed and forecast curves show differences in predicted consumption level and timing within each window. The complete group scores above provide the basis for comparing model performance. Figure 7: Observed electricity consumption and frozen TimesFM 3.0 and XGBoost forecasts for the first chronological test forecast origin in each day-type group. The origin labels are in UTC, and each panel contains the corresponding 24 hourly targets on a common consumption scale. Black identifies observations; colour and line style identify the forecasts. The selection rule does not use forecast errors. Extended Historical Evaluation The extended historical evaluation covers 566 forecast origins and 13,584 hourly targets, from 11 September 2022 at 23:00 UTC through 30 March 2024 at 22:00 UTC. The first seven days supply the initial context. TimesFM 3.0 with covariates gives MAE 16.60 kWh and RMSE 26.74 kWh, compared with 29.16 and 49.03 kWh for the weekly seasonal-naive forecast (Table 9). The 95th percentile of absolute error is 48.52 kWh for TimesFM and 115.38 kWh for the seasonal-naive forecast. Table 9: Extended historical evaluation on 13,584 common target hours. Error values are in kWh. The TimesFM settings were selected using later validation data; these scores are exploratory. Scenario n MAE RMSE R² p95 absolute error TimesFM 3.0 with covariates 13584 16.6 26.74 0.879 48.52 Seasonal-naive 168 h 13584 29.16 49.03 0.593 115.38 Seasonal errors vary substantially. TimesFM MAE is 20.25 kWh in winter, 19.92 kWh in spring, 9.89 kWh in summer, and 13.98 kWh in autumn. The chronological monthly comparison in Figure 8 shows this variation alongside the seasonal-naive reference. These groups have unequal coverage, and the available TimesFM context grows during the first year. Figure 8: Monthly pooled MAE and RMSE for TimesFM 3.0 with covariates and the weekly seasonal-naive forecast during the exploratory historical evaluation. The aligned panels share a chronological month axis, and the two models use different line styles and markers. Months refer to UTC target timestamps; September 2022 and March 2024 are partial months. The historical evaluation contains 330 workdays, 121 weekend days, and 115 holiday days. TimesFM MAE is 16.86, 14.65, and 17.89 kWh, respectively. Holiday RMSE remains highest at 34.97 kWh, compared with 23.81 kWh on workdays and 25.23 kWh on weekends. The corresponding pooled R2R^2 is 0.910 on workdays but -0.086 on holidays. Appendix Section Historical Evaluation and Weather Breakdowns provides the counts, monthly context ranges, and season/day-type tables. Weather-Source Sensitivity All three weather sources produce complete forecasts for the same 57 windows and 1,368 target hours. Table 10 reports their absolute errors, while Figure 9 isolates the paired differences between sources. MAE is 19.968 kWh with KNMI observations, 19.968 kWh with ERA5, and 20.137 kWh with historical IFS. Corresponding RMSE values are 30.288, 30.221, and 30.409 kWh. ERA5 has the lowest point estimates, although its MAE differs from KNMI by less than 0.001 kWh. Table 10: Fixed TimesFM 3.0 weather-source comparison using the same UTC hourly intervals. Context, other covariates, and model settings are held constant. This KNMI result uses the supplementary interval assignment and is distinct from the original benchmark row. Source Target hours MAE (kWh) RMSE (kWh) R² MAPE (%) sMAPE (%) KNMI station 344 1368 19.9682 30.2878 0.7981 16.9134 15.2548 ERA5 reanalysis 1368 19.968 30.2208 0.799 16.9045 15.2468 Historical IFS 1368 20.1373 30.4094 0.7965 17.1393 15.389 Figure 9: Paired differences in daily MAE between weather sources, with 95% day-block bootstrap intervals. The zero line denotes no difference; a positive value indicates larger error for the first source named. All comparisons use the same 57 forecast windows and fixed TimesFM 3.0 settings. Absolute error values are reported in Table 10. The ERA5-minus-KNMI mean daily-MAE difference is effectively zero, with a 95% interval of -0.167 to 0.171 kWh. Historical IFS has a difference of 0.169 kWh from KNMI, with an interval of 0.007 to 0.331 kWh, and 0.169 kWh from ERA5, with an interval of 0.003 to 0.325 kWh. Thus the unadjusted intervals indicate a small increase for historical IFS, approximately 0.85% of KNMI MAE. Their lower bounds lie close to zero, and the analysis does not adjust for multiple comparisons or dependence between successive days. The day-type pattern is consistent across sources: holiday MAE is approximately 24.74–25.29 kWh, compared with 18.84–18.97 kWh on workdays. Appendix Section Historical Evaluation and Weather Breakdowns reports the day-type figure, monthly/seasonal errors, and numerical paired intervals. Fitting and Forecasting Time Table 11: Measured time and memory during test forecasting. Times are seconds per 24-hour forecast origin, summarised by median and 95th percentile. The LSTM ridge-correction row is additional to the LSTM base cost. Model Fit/update median / p95 (s) Warm 24-h inference median / p95 (s) Peak RSS (MiB) Peak GPU allocated (MiB) TimesFM 3.0 with covariates 0.000000 / 0.000000 0.9215 / 0.9680 4402.73 2470.39 TimesFM 3.0 target-only 0.000000 / 0.000000 0.4047 / 0.4566 4402.73 1308.73 TimesFM 2.5 target-only 0.000000 / 0.000000 0.1197 / 0.1394 2435.05 898.29 XGBoost 0.8418 / 3.43 0.001232 / 0.004213 378.00 not applicable Prophet with regressors 4.19 / 5.97 0.3201 / 0.3582 254.04 not applicable Ridge 0.0302 / 0.0359 0.002373 / 0.003514 191.07 not applicable LSTM + ridge (correction only) 0.007708 / 0.009089 0.001936 / 0.002448 3443.93 not recorded Prophet base 2.09 / 2.84 0.2969 / 0.3550 235.48 not applicable LSTM base 99.05 / 204.05 0.1497 / 0.1789 3443.93 not recorded Seasonal-naive 168 h 0.000000 / 0.000000 0.000080 / 0.000104 188.10 not applicable Seasonal-naive 24 h 0.000000 / 0.000000 0.000081 / 0.000135 184.08 not applicable TimesFM’s model weights were trained externally and reused for this building. In our execution, the median forecast time was 0.12 s for TimesFM 2.5, 0.41 s for target-only TimesFM 3.0, and 0.92 s for TimesFM 3.0 with covariates. Loading TimesFM 3.0 took approximately two seconds before forecasting. These measurements exclude its original pretraining. Ridge, XGBoost, Prophet, and the LSTM required building-specific fitting at each origin under the evaluated policy. Median fitting times were approximately 0.03 s for ridge, 0.84 s for XGBoost, 2.10 s for Prophet base, 4.19 s for Prophet with regressors, and 99.05 s for the LSTM. The ridge correction added approximately 0.008 s to the LSTM fit. Discussion Accuracy in Relation to Previous Studies TimesFM provides a strong starting point for building electricity forecasting without local neural-network training. TimesFM 3.0 with covariates ranks first in pooled MAE and RMSE, while the target-only 3.0 and 2.5 models use only past electricity consumption and also achieve lower pooled MAE than all locally fitted configurations. XGBoost is the closest locally fitted alternative, with pooled RMSE only 0.85 kWh above TimesFM 3.0 with covariates. These results demonstrate that pretrained forecasting can deliver leading accuracy without fitting model weights to the building. The paired daily-MAE intervals leave the ranking of TimesFM and XGBoost uncertain, while the LSTM’s longer local fitting brings no accuracy advantage under the evaluated policy. The preceding THUAS evaluation found the lowest workday MAE with Prophet and the lowest MAPE and mean squared error with XGBoost (Stienstra et al., 2026). The present comparison adds pretrained forecasting and evaluates the selected configurations over 57 days including closures. TimesFM’s strong performance extends the evidence for zero-shot transfer reported by Das et al. (2024) and the electrical-demand comparison by Llanes-Guilarte et al. (2026) to this university building. XGBoost retains an advantage when larger non-workday errors are the priority: its non-workday RMSE is 6.13 kWh lower than that of TimesFM 3.0 with covariates, despite their MAEs differing by only 0.18 kWh. The preferred error profile therefore depends on the operating period and the consequences of larger misses. TimesFM 2.5 already provides a strong target-only forecast, with MAE 21.07 kWh. Version 3.0 reduces target-only MAE to 20.01 kWh, although its RMSE is slightly higher, at 32.45 versus 31.85 kWh. The native-covariate 3.0 configuration gives the lowest pooled values for both measures. The extended historical evaluation and weather-source comparison further assess this selected configuration across seasons and alternative inputs. Together, these results support using pretrained TimesFM as an initial forecasting baseline alongside seasonal-naive references, before investing in building-specific model training. What the Covariates Add The benefit of covariates depends on how effectively a model already uses consumption history. External regressors reduce Prophet’s pooled MAE by 5.31 kWh, ridge residual correction reduces LSTM MAE by 0.77 kWh, and native covariates reduce TimesFM 3.0 MAE by only 0.04 kWh. The TimesFM result indicates two complementary strengths: the consumption series contains recurring daily and weekly operating patterns, and the pretrained model can use those patterns effectively with its weights held fixed. Together, these provide a plausible explanation for the small overall gain from explicitly supplying calendar, weather, and historical-consumption covariates. For TimesFM, the main benefit appears during holidays: adding the covariate set reduces MAE from 34.40 to 24.76 kWh and RMSE from 53.74 to 39.47 kWh. The 14.27 kWh reduction in holiday RMSE accompanies smaller increases of 3.00 kWh on workdays and 2.46 kWh on weekends, yielding a lower pooled RMSE. These gains were obtained with the complete calendar, weather, and historical-consumption covariate set. The results favour this configuration when reducing holiday errors is a priority, despite its smaller loss of accuracy during routine operation. The practical choice is therefore straightforward: target-only TimesFM provides strong forecasts from consumption history alone, while its covariate version improves closure-period forecasts at the cost of additional input preparation. This distinction separates the value of a simple initial forecast from the value of refining performance during specific operating periods. Performance Across the Year The extended historical evaluation uses a practical advantage of pretrained forecasting: TimesFM can forecast after a seven-day context warm-up without fitting new network weights. This permits examination of periods otherwise reserved for the benchmark’s 365-day local fitting window. Its MAE varies from 9.89 kWh in summer to 20.25 kWh in winter, showing the value of examining the full operating year when assessing forecasting difficulty. The comparison with weekly seasonal-naive forecasts adds a reference for each period. Seasonal differences should be interpreted alongside the unequal coverage and growing context during the first year; the analysis extends the assessment of TimesFM rather than the ranking of all 11 configurations. The historical holiday results also clarify the role of accuracy metrics. Holiday MAE is close to workday MAE, but holiday R2R^2 is slightly negative because squared forecast errors exceed the variation around the holiday group’s mean. Lower consumption variability makes this reference harder to improve upon, while occasional large errors increase the squared-error total. MAE and RMSE remain directly interpretable in hourly kWh, whereas R2R^2 describes performance relative to variation within the group. Reporting them together makes the closure-period errors easier to assess. Weather Inputs and Practical Model Choice The weather experiment addresses a practical concern about the benchmark’s use of station observations: would a different weather product substantially change the TimesFM forecasts? For the tested configuration, replacing KNMI with ERA5 changes MAE by less than 0.001 kWh, and replacing it with historical IFS increases MAE by approximately 0.17 kWh, or 0.85%. The IFS increase is small in magnitude even though its unadjusted paired interval excludes zero. This provides evidence of limited sensitivity to the three tested weather products, rather than evidence of identical predictions or exactly equal accuracy. This finding strengthens the case for investigating pretrained forecasting at THUAS: the covariate configuration’s accuracy is not strongly dependent on choosing one of these weather sources. The small source differences also persist within day types, including holidays, where covariates provide the clearest improvement over target-only TimesFM. However, weather-source substitution and covariate inclusion answer different questions. All three source variants contain weather, and the other 23 covariates remain unchanged. The experiment does not show that weather is unimportant, that reanalysis is necessarily more accurate at the building, or that another forecasting model would have the same sensitivity. Our comparison replaces both historical and future weather together and does not reconstruct issue-time-specific inputs. Target-only TimesFM avoids the weather-data requirement entirely, providing a simpler input pipeline; the covariate configuration offers the holiday-error improvement discussed above. Any deployment must also respect the checkpoint’s restrictions noted below. Local Modelling Effort Pretrained forecasting replaces local weight fitting with historical-context preparation and inference. In this benchmark, that change does not entail an observed pooled-accuracy penalty relative to XGBoost: TimesFM 3.0 with covariates has lower point estimates, with uncertainty that prevents a clear daily-MAE ranking. The target-only version further reduces the number of data sources needed. Nevertheless, both TimesFM configurations underwent validation of context and inference settings; zero-shot forecasting is not equivalent to an untuned or zero-history application. XGBoost offers a different practical advantage. After a median fit of 0.84 s, its measured 24-hour inference took approximately 0.0012 s, compared with 0.92 s for TimesFM 3.0 with covariates in the recorded CPU/GPU environments. Both inference times are short relative to a daily forecasting cycle, making their difference less consequential for one building than for a larger forecasting service. Under the evaluated policy, XGBoost is refitted each day as new observations arrive, requiring an ongoing fitting procedure and maintenance of its covariate inputs. TimesFM incorporates new observations by advancing its context without refitting network weights. Both approaches still require checks on input quality and forecast performance. The LSTM’s approximately 99 s median fit is appreciably longer, without lower errors in this benchmark. Limitations The study is designed to support a site-specific model comparison, not a universal ranking of forecasting methods. Using the same target hours, chronological model selection, seasonal-naive references, and paired daily comparisons makes the 57-day test informative for this building despite its limited duration. The test is weighted toward late winter and early spring, however, and contains only 11 holiday days. Its bootstrap intervals are not adjusted for multiple comparisons and do not model dependence between successive days. The extended historical evaluation adds seasonal and closure-period evidence, but uses shorter contexts at early origins and settings selected from later validation data; summer is represented only once. It therefore complements rather than replaces an independent all-season test. Rotterdam Airport observations were used because they are the regional weather source in the supplied dataset. Comparing them with two gridded products directly tests whether dependence on that source substantially changes the selected TimesFM results; the small measured differences reduce that concern for the evaluated period. They do not establish exact weather conditions at the building or test other stations. The products differ in temporal sampling, and the supplementary comparison uses a consistent endpoint assignment while the original benchmark retains the supplied weather join. Independent documentation of the original meter interval-label convention was unavailable (Appendix Section Time Convention and Weather Alignment). Historical IFS identifies a forecast product, but its downloaded values lack the issue times needed to reconstruct a fully operational forecast sequence. Treating opening hours and examination weeks as known one day ahead is a reasonable planning assumption for scheduled university activities. We exclude realised attendance and cumulative room bookings rather than treating tomorrow’s observed activity as an available input. Nevertheless, the historical publication times of the supplied schedules were not retained, and calendar labels do not measure actual attendance. The results therefore apply to the stated schedule-availability assumption. The resource measurements distinguish recurring fitting from forecasting, which is sufficient to compare those tasks within the recorded executions. They do not establish total resource efficiency: LSTM GPU peaks and some model-storage measurements are missing, runs use different CPU/GPU settings, and development and pretraining costs are excluded. Finally, the evaluated TimesFM 3.0 weights are released for non-commercial, non-production research use (Google Research, 2026). The results can motivate operational assessment, but do not establish permission to deploy this checkpoint in production. Conclusion Pretrained TimesFM and locally fitted XGBoost provide the strongest alternatives among the 11 configurations evaluated for this university building. TimesFM 3.0 with covariates achieved the lowest pooled MAE and RMSE, 19.97 and 30.26 kWh per hourly target, after chronological validation and evaluation on 57 test days. Its MAE was 43.4% lower than the weekly seasonal-naive reference. XGBoost remained close in RMSE at 31.11 kWh, with rapid inference after fitting and no clear separation from TimesFM in the paired daily-MAE comparisons. Workday and non-workday errors lead to different choices: target-only TimesFM 3.0 had the lowest workday errors, whereas XGBoost had the lowest non-workday RMSE. Adding covariates to TimesFM substantially reduced holiday errors but provided almost no pooled MAE improvement. The exploratory 566-day historical evaluation extended the analysis beyond the winter test and revealed substantial seasonal variation without additional network training. Replacing KNMI observations with ERA5 or historical IFS changed MAE by at most 0.17 kWh, showing limited weather-source sensitivity within the selected covariate configuration. Target-only TimesFM therefore offers a strong initial forecast for this building from past electricity consumption alone, without local weight training or external covariate preparation. Adding covariates substantially improves holiday forecasts, with smaller increases in workday and weekend errors. XGBoost remains an accurate, fast alternative when local fitting and ongoing covariate maintenance are acceptable. The practical starting point is a pretrained target-only forecast, followed by comparison with a locally fitted model and the addition of covariates where closure-period accuracy justifies the extra inputs. Future Work Future work should evaluate the selected approaches on independent buildings and complete seasonal holdouts, and examine the separate contributions of calendar, weather, and historical-consumption covariates. An operational pilot could connect forecast errors to service-scheduling decisions and measure the resulting energy and cost outcomes. Data Availability and Reproducibility The electricity-consumption and institutional scheduling data were supplied for the THUAS building study. The data and analysis code will be publicly released through 4TU. The research archive retains data checksums, ordered covariates, validation selections, hourly prediction ledgers, software and hardware records, analysis scripts, and generated tables and figures. A GitHub repository will accompany publication. Supplementary Materials Covariate Catalogue Table 12: Stored names, scientific meanings, and units for the 30 retained covariates. The manuscript uses the scientific names; the stored names enable reproduction. Source identifier(s) Scientific meaning Domain Unit or encoding Exam_weeks Exam-week indicator calendar/schedule binary Opening_hours Opening-hours indicator calendar/schedule binary Hour Hour label calendar/schedule hour Year Calendar year calendar/schedule year Month Calendar month calendar/schedule month Week ISO week number calendar/schedule week Day Day of month calendar/schedule day DayofWeek Day of week calendar/schedule day Month_sin, Month_cos Cyclic month encoding calendar/schedule unitless Week_sin, Week_cos Cyclic week encoding calendar/schedule unitless Day_sin, Day_cos Cyclic day encoding calendar/schedule unitless DayOfWeek_sin, DayOfWeek_cos Cyclic weekday encoding calendar/schedule unitless Hour_sin, Hour_cos Cyclic hour encoding calendar/schedule unitless Work_week Work-week indicator calendar/schedule binary Weekend Weekend indicator calendar/schedule binary Holiday Institutional holiday/closure indicator calendar/schedule binary Temperature Air temperature weather °C Temperature_dew Dew-point temperature weather °C Relative_humidity Relative humidity weather % Total_precipitation Total precipitation weather m per hourly interval Pressure Sea-level pressure weather hPa Wind_speed Wind speed weather m s-1 Wind_direction Wind direction weather degrees lag_24_hours Consumption lag, 24 h observed consumption kWh lag_168_hours Consumption lag, 168 h observed consumption kWh Validation and Selected Settings Table 13: Validation scores and selected settings for the benchmark configurations. Every candidate was evaluated on 2,160 hourly validation targets. Model Validation trials Validation MAE (kWh) Validation RMSE (kWh) Validation R² Selected parameter details TimesFM 3.0 with covariates 64 19.98 28.90 0.861 context_values=8760; make_positive=False; point_forecast=median_quantile_index_4; return_quantiles=True; use_symmetric_averaging=True; use_znorm=False TimesFM 3.0 target-only 64 19.37 28.64 0.863 context_values=8760; make_positive=False; point_forecast=median_quantile_index_4; return_quantiles=True; use_symmetric_averaging=True; use_znorm=False TimesFM 2.5 target-only 8 23.38 36.26 0.781 context_values=672; normalize_inputs=False XGBoost 96 21.27 28.76 0.862 colsample_bytree=0.6239; learning_rate=0.01401; max_depth=7; min_child_weight=0.999; n_estimators=1350; reg_alpha=0.4845; reg_lambda=0.05573; subsample=0.7332 Prophet with regressors 96 24.13 32.16 0.827 changepoint_prior_scale=0.01784; changepoint_range=0.8846; daily_fourier_order=10; daily_mode=additive; holidays_prior_scale=4.843; regressor_prior_scale=0.02489; seasonality_prior_scale=0.1609; weekly_fourier_order=7; weekly_mode=additive Ridge 25 26.82 35.69 0.787 alpha=215.4 LSTM + ridge residual 48 30.32 42.75 0.695 residual_feature_scaling=standard; residual_history_rows=1440; ridge_alpha=655.8 Prophet base 96 35.40 45.61 0.653 changepoint_prior_scale=0.0144; changepoint_range=0.8424; daily_fourier_order=12; daily_mode=additive; holidays_prior_scale=0.854; seasonality_prior_scale=4.112; weekly_fourier_order=7; weekly_mode=additive LSTM base 72 32.08 45.91 0.648 architecture_profile=‘dropout’: 0.05, ‘l2’: 1e-06, ‘layers’: 1, ‘units’: 64; lookback_hours=168; optimizer_profile=‘batch_size’: 32, ‘learning_rate’: 0.0006, ‘optimizer’: ‘rmsprop’ Seasonal-naive 168 h 1 38.03 58.46 0.430 lag_hours=168 Seasonal-naive 24 h 1 40.19 63.29 0.332 lag_hours=24 Complete Day-Type Comparison Figure 10 extends the focused comparison in Figure 6 to all 11 configurations. The workday, weekend, and holiday groups use the same forecast windows and fixed model settings as the main benchmark. Figure 10: MAE and RMSE for all 11 configurations on workdays, weekends, and holidays. Model ordering is identical across panels, and each metric uses the same scale across day types. Counts are 34 workdays, 12 weekend days, and 11 holidays. Additional TimesFM 2.5 Hybrid Experiments We compare two jointly tuned TimesFM 2.5 hybrids with the target-only base. In the first, ridge predicts residuals of TimesFM forecasts; in the second, TimesFM forecasts residuals of ridge predictions. Context and residual-regression settings are selected together using the 90 validation origins. The joint-search design was finalised after initial benchmark results had been inspected, so these additional comparisons are exploratory. Table 14: Validation and test errors for TimesFM 2.5 target-only and the two jointly tuned hybrids. Validation covers 2,160 target hours and testing 1,368; all errors are in kWh. Configuration Validation MAE (kWh) Validation RMSE (kWh) Test MAE (kWh) Test RMSE (kWh) TimesFM 2.5 target-only 23.38 36.26 21.07 31.85 TimesFM 2.5 + ridge residual (jointly tuned) 23.77 36.30 20.88 31.44 Ridge + TimesFM 2.5 residual (jointly tuned) 22.54 31.66 25.14 34.91 Adding ridge correction to TimesFM gives a small test improvement over the base: MAE decreases from 21.07 to 20.88 kWh and RMSE from 31.85 to 31.44 kWh. Reversing the order produces the lowest validation RMSE of the three configurations but higher test errors, with MAE 25.14 kWh and RMSE 34.91 kWh. Residual modelling therefore provides no consistent improvement across the two hybrid constructions. Time Convention and Weather Alignment We convert the supplied fixed UTC+1 timestamp labels to UTC by subtracting one hour, preserving all rows and 24-hour forecast windows. Calendar covariates and workday/holiday membership retain the institutional operating-clock labels. Thus, a forecast beginning at 23:00 UTC covers portions of two UTC dates, while monthly and seasonal error groups follow the individual UTC target timestamps. For the supplementary weather comparison, a target labelled u is assigned to the interval [u,u+1h)[u,u+1\,h). Temperature, humidity, pressure, and wind are represented at the interval endpoint; precipitation is accumulated over that interval. The KNMI workbook key marks the start of the observation interval, so ERA5 and IFS use the next valid-time row relative to that key. This mapping was checked across all 57 context and forecast windows. Station wind averages over a short observation period, while the gridded wind field is instantaneous. The original benchmark and extended historical evaluation preserve the inherited source join. Under the adopted fixed UTC+1 and interval-start interpretation, that join selects KNMI observations one hourly row later than the supplementary endpoint mapping. This is a difference between timestamp assignments, not evidence of incorrect KNMI measurements. Independent documentation of whether the original meter timestamp denotes an interval start or end was unavailable. We therefore keep the two result sets distinct and state their conventions explicitly. Wind direction is a numerical covariate in degrees, not a binary weather category. The KNMI value 990 denotes variable direction rather than an angle; preprocessing treats it as missing and substitutes the median from observations preceding the forecast origin. The separate weather-code covariate is excluded from all 30-covariate configurations. In the supplementary comparison, pressure is mean-sea-level pressure for all sources. Open-Meteo documents the historical IFS forecast archive (Open-Meteo, n.d.-a). Our downloaded records retain weather valid times but not complete model-initialisation and publication times for each value. This is sufficient to identify the supplied forecast product, but not to reconstruct exactly which run an operator could have accessed at every historical forecast origin. Historical Evaluation and Weather Breakdowns The historical tables report target-hour groups. Workday, weekend, and holiday membership follows the source operating day. Calendar months and meteorological seasons follow UTC target timestamps, so a forecast window may contribute hours to two adjacent groups. Seasonal coverage is unequal; in particular, summer is observed only in 2023. Table 15: Historical evaluation by day type and meteorological season. MAE and RMSE are in kWh; target-hour counts identify the coverage of each group. Group Level Model Target hours MAE (kWh) RMSE (kWh) Day type holiday TimesFM 3.0 with covariates 2760 17.887 34.969 Day type weekend TimesFM 3.0 with covariates 2904 14.648 25.227 Day type workday TimesFM 3.0 with covariates 7920 16.863 23.81 Day type holiday Seasonal-naive 168 h 2760 40.62 69.059 Day type weekend Seasonal-naive 168 h 2904 20.574 35.139 Day type workday Seasonal-naive 168 h 7920 28.317 44.806 UTC season Winter (DJF) TimesFM 3.0 with covariates 4344 20.247 29.47 UTC season Summer (JJA) TimesFM 3.0 with covariates 2208 9.893 14.243 UTC season Spring (MAM) TimesFM 3.0 with covariates 2927 19.915 33.975 UTC season Autumn (SON) TimesFM 3.0 with covariates 4105 13.976 22.692 UTC season Winter (DJF) Seasonal-naive 168 h 4344 41.681 63.009 UTC season Summer (JJA) Seasonal-naive 168 h 2208 15.006 24.204 UTC season Spring (MAM) Seasonal-naive 168 h 2927 31.717 53.167 UTC season Autumn (SON) Seasonal-naive 168 h 4105 21.704 37.702 Table 16: Chronological monthly errors and context coverage. Context ranges are the minimum and maximum numbers of prior hourly observations supplied to TimesFM for forecasts contributing to the UTC month. UTC month Target hours Context min Context max MAE (kWh) RMSE (kWh) Naive MAE (kWh) Naive RMSE (kWh) 2022-09 457 168 624 10.519 14.674 10.82 14.428 2022-10 744 624 1368 10.169 19.31 18.163 35.337 2022-11 720 1368 2088 19.474 31.703 30.046 45.45 2022-12 744 2088 2832 17.181 23.717 44.034 68.687 2023-01 744 2832 3576 19.842 26.994 40.608 59.109 2023-02 672 3576 4248 18.398 27.194 30.404 46.698 2023-03 744 4248 4992 17.274 23.133 36.743 54.228 2023-04 720 4992 5712 27.103 47.829 32.074 58.871 2023-05 744 5712 6456 18.187 34.065 32.2 59.184 2023-06 720 6456 7176 11.936 16.583 12.607 24.28 2023-07 744 7176 7920 8.779 12.166 16.732 24.899 2023-08 744 7920 8664 9.029 13.709 15.601 23.411 2023-09 720 8664 8760 12.192 18.383 12.672 20.581 2023-10 744 8760 8760 13.081 21.26 30.099 53.956 2023-11 720 8760 8760 17.313 24.573 24.285 34.409 2023-12 744 8760 8760 20.47 28.851 38.871 59.88 2024-01 744 8760 8760 22.593 34.369 50.282 73.69 2024-02 696 8760 8760 22.999 34.261 45.011 65.255 2024-03 719 8760 8760 17.239 25.571 25.658 37.003 Table 17: Paired daily-MAE differences between weather sources, in kWh, with 95% percentile intervals from 10,000 resamples of the 57 forecast-day blocks. Positive values favour the second source named. Contrast Origin days Estimate (kWh) 95% CI low (kWh) 95% CI high (kWh) ERA5 reanalysis − KNMI station 344 57 -0.0002 -0.1672 0.1706 Historical IFS − KNMI station 344 57 0.1691 0.0071 0.3305 Historical IFS − ERA5 reanalysis 57 0.1693 0.0027 0.3249 Table 18: Weather-source errors by operating day type. Each source is evaluated on 816 workday, 288 weekend, and 264 holiday target hours. Source Day type Target hours MAE (kWh) RMSE (kWh) KNMI station 344 holiday 264 24.741 39.332 KNMI station 344 weekend 288 18.783 31.84 KNMI station 344 workday 816 18.842 26.069 ERA5 reanalysis holiday 264 24.965 39.621 ERA5 reanalysis weekend 288 18.461 31.488 ERA5 reanalysis workday 816 18.883 25.948 Historical IFS holiday 264 25.291 40.118 Historical IFS weekend 288 18.717 31.532 Historical IFS workday 816 18.971 26.051 Figure 11: MAE and RMSE for each weather source on workdays, weekends, and holidays. Within each group, distinct markers identify KNMI, ERA5, and historical IFS on the same error scale. All sources use the same target hours: 816 workday, 288 weekend, and 264 holiday hours. Table 19: Weather-source errors by UTC month and meteorological season. This test contains only February 2024 (649 winter target hours) and March 2024 (719 spring target hours); month and season therefore identify the same two groups. Source UTC month / season Target hours MAE (kWh) RMSE (kWh) KNMI station 344 February 2024 / DJF 649 23.0108 34.7614 KNMI station 344 March 2024 / MAM 719 17.2218 25.5866 ERA5 reanalysis February 2024 / DJF 649 23.061 34.7231 ERA5 reanalysis March 2024 / MAM 719 17.1761 25.4828 Historical IFS February 2024 / DJF 649 23.1571 34.7774 Historical IFS March 2024 / MAM 719 17.4115 25.8402 Supporting Descriptive Diagnostics The full-series autocorrelations are 0.687 at 24 hours and 0.788 at 168 hours (Figure 12). ADF and KPSS tests give different conclusions for the raw level because their null hypotheses differ; the KPSS statistic falls from 1.55 to 0.0046 after first differencing. A linear trend in daily mean consumption explains little variation (R2=0.0026R^2=0.0026). We retain these as supporting descriptions rather than criteria for selecting the final forecast model. Figure 12: Autocorrelation of hourly electricity consumption through 336 hours. The highlighted 24- and 168-hour lags identify daily and weekly recurrence, supporting the seasonal-naive references. Figure 13: Mean hourly electricity consumption over the complete dataset. References References Akiba, T., Sano, S., Yanase, T., Ohta, T., & Koyama, M. (2019). Optuna: A next-generation hyperparameter optimization framework. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2623–2631. https://doi.org/10.1145/3292500.3330701 Alfalah, B., Shahrestani, M., & Shao, L. (2023). Identifying occupancy patterns and profiles in higher education institution buildings with high occupancy density—a case study. Intelligent Buildings International, 15(2), 45–61. https://doi.org/10.1080/17508975.2022.2137451 Amber, K. P., Aslam, M. W., Mahmood, A., Kousar, A., Younis, M. Y., Akbar, B., Chaudhary, G. Q., & Hussain, S. K. (2017). Energy consumption forecasting for university sector buildings. Energies, 10(10), 1579. https://doi.org/10.3390/en10101579 Bourdeau, M., Zhai, X. qiang, Nefzaoui, E., Guo, X., & Chatellier, P. (2019). Modeling and forecasting building energy consumption: A review of data-driven techniques. Sustainable Cities and Society, 48, 101533. https://doi.org/10.1016/j.scs.2019.101533 Chen, T., & Guestrin, C. (2016). XGBoost: A scalable tree boosting system. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 785–794. https://doi.org/10.1145/2939672.2939785 Das, A., Kong, W., Sen, R., & Zhou, Y. (2024). A decoder-only foundation model for time-series forecasting. Proceedings of the 41st International Conference on Machine Learning, Proceedings of Machine Learning Research, 235, 10148–10167. https://proceedings.mlr.press/v235/das24c.html Davison, A. C., & Hinkley, D. V. (1997). Bootstrap methods and their application. Cambridge University Press. https://doi.org/10.1017/CBO9780511802843 Ghanavati, A., Rawlins, M., Dow, D., Sweeny, C., & Smith, J. (2023). Interpretation of the power consumption characterization of an urban university campus toward power system planning. Sustainability. https://doi.org/10.3390/su152014835 Google Research. (2025). TimesFM 2.5 200M PyTorch checkpoint and model card. https://huggingface.co/google/timesfm-2.5-200m-pytorch Google Research. (2026). TimesFM: Version 3.0 documentation and pretrained models. https://github.com/google-research/timesfm Gui, X., Gou, Z., & Lu, Y. (2021). Reducing university energy use beyond energy retrofitting: The academic calendar impacts. Energy and Buildings, 231, 110647. https://doi.org/10.1016/j.enbuild.2020.110647 Hersbach, H., Bell, B., Berrisford, P., Hirahara, S., Horányi, A., Muñoz-Sabater, J., Nicolas, J., Peubey, C., Radu, R., Schepers, D., Simmons, A., Soci, C., Abdalla, S., Abellan, X., Balsamo, G., Bechtold, P., Biavati, G., Bidlot, J., Bonavita, M., … Thépaut, J.-N. (2020). The ERA5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146(730), 1999–2049. https://doi.org/10.1002/qj.3803 Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8), 1735–1780. https://doi.org/10.1162/neco.1997.9.8.1735 Hoerl, A. E., & Kennard, R. W. (1970). Ridge regression: Biased estimation for nonorthogonal problems. Technometrics, 12(1), 55–67. https://doi.org/10.1080/00401706.1970.10488634 Hyndman, R. J., & Koehler, A. B. (2006). Another look at measures of forecast accuracy. International Journal of Forecasting, 22(4), 679–688. https://doi.org/10.1016/j.ijforecast.2006.03.001 Khalil, M., McGough, A. S., Pourmirza, Z., Pazhoohesh, M., & Walker, S. (2022). Machine learning, deep learning and statistical analysis for forecasting building energy consumption—a systematic review. Engineering Applications of Artificial Intelligence, 115, 105287. https://doi.org/10.1016/j.engappai.2022.105287 Llanes-Guilarte, D. S., Herrera-Semenets, V., Bustio-Martínez, L., G. Ordiano, J. Á., & Santos-Moreno, M. (2026). Forecasting electrical demand with zero-shot lag llama and TimesFM V2. In Progress in artificial intelligence and pattern recognition. IWAIPR 2025 (Vol. 16328, p. 323–335). Springer Nature Switzerland. https://doi.org/10.1007/978-3-032-11358-0_27 Open-Meteo. (n.d.-a). Historical forecast API. Retrieved https://open-meteo.com/en/docs/historical-forecast-api Open-Meteo. (n.d.-b). Historical weather API. Retrieved https://open-meteo.com/en/docs/historical-weather-api Royal Netherlands Meteorological Institute. (n.d.). Uurwaarden van weerstations. Retrieved https://w.daggegevens.knmi.nl/klimatologie/uurgegevens Semmelmann, L., Henni, S., & Weinhardt, C. (2022). Load forecasting for energy communities: A novel LSTM-XGBoost hybrid model based on smart meter data. Energy Informatics. https://doi.org/10.1186/s42162-022-00212-9 Stienstra, J., Vasulkar, A. N., Leeuw, A.-W. de, Freire, S. K., & Salcedo-Rahola, T.-B. (2026). Data-driven approaches to forecasting building energy consumption: Case study of a modern, well-insulated university building. In CLIMA 2025 (Vol. 788, p. 1533–1543). Springer. https://doi.org/10.1007/978-3-032-10546-2_147 Tashman, L. J. (2000). Out-of-sample tests of forecasting accuracy: An analysis and review. International Journal of Forecasting, 16(4), 437–450. https://doi.org/10.1016/S0169-2070(00)00065-0 Taylor, S. J., & Letham, B. (2018). Forecasting at scale. The American Statistician, 72(1), 37–45. https://doi.org/10.1080/00031305.2017.1380080 Zhang, G. P. (2003). Time series forecasting using a hybrid ARIMA and neural network model. Neurocomputing, 50, 159–175. https://doi.org/10.1016/S0925-2312(01)00702-0