Comparative review of hybrid forecasting models for short-term prediction of building thermal load

Nikolaos A. Efkarpidis; Despoina Kothona; Georgios C. Christoforidis

cs.LGSep 19, 2026

Comparative review of hybrid forecasting models for short-term prediction of building thermal load

Authors: Nikolaos A. Efkarpidis, Despoina Kothona, Georgios C. Christoforidis

Organizations: VIVAVIS Schweiz AG - 5405 Baden, Switzerland · Department of Electrical and Computer Engineering, University of Western Macedonia, 50100 Kozani, Greece

Abstract

In this paper, a comparative review of different hybrid models for short-term forecasting of building thermal demand is carried out. Particularly, the assessment tackles the comparison of data-driven models enhanced with other state-of-the-art techniques. At the first step, the existing techniques reported in the literature are analysed. It is concluded that Metaheuristics or a data-driven model are used to identify the parameters of the basic model. The qualitative evaluation includes for each method the input and output features, main advantages and drawbacks. At the second step, an existing dataset of historical thermal demand from Scottish households, as well as historical weather forecasts are utilized to assess additionally the performance of existing hybrid methods. From the assessment of 13 hybrid methods, the Empirical Modal Decomposition - long short-term memory - Markov (EMD-LSTM-Markov) model can predict with the highest accuracy the day-ahead power pattern of heating and domestic hot water (DHW) demands. Though local power peaks are also accurately predicted, high power swells and spikes are underestimated. Other methods, such as Support Vector Machine - Simulated Annealing (SVM-SA) and Random Forest - Improved Sparrow Search Algorithm - LSTM (RF-ISSA-LSTM) predict a smooth pattern of heating and DHW demand profiles with rapid changes underestimating most power peaks.

Figures & tables

Figure 1: Classification of short-term building thermal load forecasting methods.

Figure 2: Classification of the techniques applied for short-term thermal load forecasting in buildings.

Ref.	Inputs	Outputs
Cen and Lim (2024)	• Historical outputs: Electricity demand of lighting, HVAC units and plug loads for each of the 33 building zones. • Building and indoor environment: Historical indoor ambient light (IAL), indoor ambient temperature (IAT), and indoor relative humidity (RH).	• Electricity demand of lighting, HVAC units and plug loads. • Indoor ambient light. • Indoor ambient temperature. • Indoor relative humidity.
Dong et al. (2022)	• Weather: Solar radiation, outdoor temperature, outdoor humidity. • Historical outputs: Cooling load during the previous six hours.	• Hourly cooling load.
Ren et al. (2022)	• Weather: Outdoor temperature, dew point temperature, humidity, air pressure, wind speed. • Calendar: Working day type. • Historical outputs: Electricity, heating and cooling loads (previous 4 h and previous 7 h), together with the corresponding values at $T$ , $T-1$ and $T+1$ of the previous week.	• Electricity load. • Heating load. • Cooling load.
Liu et al. (2023a)	• Weather: Outdoor dry-bulb temperature, outdoor relative humidity, solar radiation intensity (time-shifted). • Building and indoor environment: Indoor CO 2 concentration, indoor dry-bulb temperature, indoor relative humidity (time-shifted). • Historical outputs: HVAC energy consumption (1 h lag).	• HVAC energy consumption.
Yan et al. (2023)	• Weather: Air pressure, dew point temperature, air temperature, wind speed, wind direction (4 h lag). • Calendar: Month, date, day of week, day type, hour (4 h lag). • Historical outputs: Electricity, cooling and heating loads (1–24 h historical values).	• Electricity load. • Cooling load. • Heating load.
Yan et al. (2024)	• Weather: Outdoor temperature, relative humidity, wind speed, wind direction, global horizontal radiation, direct normal radiation, diffuse horizontal radiation. • Calendar: Month, day, hour, day of week, day type. • Historical outputs: Electricity, cooling, heating and gas loads (1–24 h historical values).	• Electricity load. • Cooling load. • Heating load. • Gas load.

Table 1: Input and output features of the selected hybrid data-driven forecasting models.

Model	Advantages	Drawbacks
PatchTCN–TST Cen and Lim (2024)	• Channel-independent forecasting enables scalable multivariate prediction. • Patching improves local dependency extraction and reduces attention complexity. • TCN enhances temporal feature representation. • Superior accuracy across multiple metrics and scenarios.	• Inter-channel dependencies (e.g., temperature and AC power correlation) are neglected.
DwdAdam–ILSTM Dong et al. (2022)	• Captures long-term dependencies while avoiding gradient issues. • Superior convergence, stability, and accuracy over SVR, BPNN, and LSTM variants. • Strong generalization with limited training samples.	• Generalization validated mainly for commercial buildings.
CLSTM–AR–MFFCLA Ren et al. (2022)	• Combines linear (AR) and nonlinear (LSTM/CLSTM) modeling. • Multi-dimensional feature fusion improves spatial–temporal learning. • Stable performance for peak and non-workday loads.	• Increased architectural and computational complexity.
PCC–GRU Liu et al. (2023a)	• PCC-based time shifting enhances feature correlation and model stability. • GRU models nonlinear dynamics and peak-shaving behavior. • Mitigates energy storage interference.	• Limited performance improvement for SVR-based models.
FTTrans–E–BL Yan et al. (2023)	• Feature-time attention captures temporal and global dependencies. • Pareto-based multi-task learning improves adaptability and stability. • Probabilistic forecasting quantifies uncertainty.	• Higher computational complexity despite reduced training time.
TE–BiSRU–MTL Yan et al. (2024)	• Transfer entropy clustering captures nonlinear multi-feature dependencies. • Temporal attention and Bi-SRU enhance accuracy and efficiency. • Multi-task learning improves generalization across horizons.	• Increased complexity due to ensemble and multi-task frameworks.

Table 2: Advantages and drawbacks of selected hybrid data-driven load forecasting models with references

Panel of Figure 3 — Figure 3: Typical clusters of DHW and heating demand for two households: (a) household 90. (b) household 92.

Figure 4: Adjustment of provided weather forecasts to the prediction horizon of next day with time resolution of 15 minutes.

Panel of Figure 5 — Figure 5: Inputs (blue) for home: (a) 68, (b) 81, (c) 82, (d) 88, (e) 90, (f) 92, (g) 101, (h) 107, (i) 109 and (j) all collectively.

Panel of Figure 6 — Figure 6: MAE boxplots from the application of all heat forecasting models at home: (a) 68, (b) 81, (c) 82, (d) 88, (e) 90, (f) 92, (g) 101, (h) 107, and (i) 109.

Panel of Figure 7 — Figure 7: RMSE boxplots from the application of all heat forecasting models at home: (a) 68, (b) 81, (c) 82, (d) 88, (e) 90, (f) 92, (g) 101, (h) 107, and (i) 109.

Panel of Figure 8 — Figure 8: P95AE boxplots from the application of all heat forecasting models at home: (a) 68, (b) 81, (c) 82, (d) 88, (e) 90, (f) 92, (g) 101, (h) 107, and (i) 109.

Panel of Figure 9 — Figure 9: Highest heating demand day with most rapid changes for: (a) home 92 - 22/03/2018, and (b) home 109 - 10/05/2018.

Panel of Figure 10 — Figure 10: MAE boxplots from the application of all DHW forecasting models at home: (a) 68, (b) 81, (c) 82, (d) 88, (e) 90, (f) 92, (g) 101, (h) 107, and (i) 109.

Panel of Figure 11 — Figure 11: RMSE boxplots from the application of all DHW forecasting models at home: (a) 68, (b) 81, (c) 82, (d) 88, (e) 90, (f) 92, (g) 101, (h) 107, and (i) 109.

Panel of Figure 12 — Figure 12: P95AE boxplots from the application of all DHW forecasting models at home: (a) 68, (b) 81, (c) 82, (d) 88, (e) 90, (f) 92, (g) 101, (h) 107, and (i) 109.

Panel of Figure 13 — Figure 13: Day with most DHW spikes for: (a) home 81 - 01/01/2018, and (b) home 109 - 21/04/2018, as well as most stochastic DHW day for: (c) home 81 - 22/12/2017, and (d) home 109 - 01/05/2018.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter	Value
Patch length	$8$
Number of patches	$12$
Transformer dimension	$128$
Attention heads	$4$
Batch size	$8$
Training epochs	$40$

Appendix

Table 3: PatchTCN-TST parameters

Parameter	Value
LSTM batch size	$8$
LSTM hidden layers	$1$
LSTM hidden units	$64$
Training epochs	$25$
Learning rate	$10^{-3}$
Activation function	Softplus

Appendix

Table 4: DwdAdam-ILSTM parameters

Parameter	Value
Batch size	$32$
Training epochs	$30$
Learning rate	$10^{-3}$
Encoder input size	Selected features
Encoder hidden size	$64$
Encoder layers	$1$

Appendix

Table 5: CLSTM-AR forecasting model parameters

Parameter	Value
Training batch size	$32$
Training epochs	$8$
GRU layers	$1$
GRU hidden units	$64$
Output dimension	$96$
Activation function	Softplus()

Appendix

Table 6: PCC-GRU parameters

Parameter	Value
Training epochs	$30$
Training batch size	$32$
Learning rate	$5\times 10^{-4}$
Input feature embedding dimension	$64$
Transformer encoder layers	$2$
Transformer attention heads	$4$

Appendix

Table 7: FTTrans-E-BL parameters

Parameter	Value
Copula entropy bins	$20$
Relative entropy tolerance	$0.15$
Bi-GRU hidden size	$128$
Bi-GRU output size	$256$
Training epochs	$20$
Training batch size	$32$

Appendix

Table 8: TE-BiSRU-MTL parameters

Parameter	Value
Training epochs	$100$
Training batch size	$32$
Learning rate	$10^{-3}$
Wavelet family	bior2.8
WTD decomposition level	$3$
CNN layers	$2$

Appendix

Table 9: WTD-CNN-LSTM parameters

Parameter	Value
Correlation subsample	$200$
Training batch size	$32$
Inference batch size	$16$
Training epochs	$40$
Early stopping patience	$6$
Learning rate	$10^{-3}$

Appendix

Table 10: aTCN-LSTM parameters

Parameter	Value
Training epochs	$30$
Learning rate	$10^{-3}$
Training batch size	$32$
Inference batch size	$32$
Optimizer	Adam
Teacher forcing ratio	$0.5$

Appendix

Table 11: EMD-LSTM-Markov parameters

Parameter	Value
SA cycles	$2$
SA trials per cycle	$3$
SA subsample	$1500$
LinearSVR maximum iterations	$5000$
LinearSVR tolerance	$10^{-3}$
Search space - $C$ range	$[0.01,10]$

Appendix

Table 12: SVM-SA parameters

Parameter	Value
PCA number of components	$150$
ANN hidden layers	$1$
ANN hidden neurons	$20$
ANN hidden activation function	ReLU
ANN output activation function	Softplus
ANN output layer	$96$ neurons

Appendix

Table 13: GA/iPSO-ANN parameters

Parameter	Value
Conv1 filters	$32$
Conv2 filters	$32$
Parallel convolution kernels	[3, 7]
Residual connection	Enabled
Dropout rate	$0.2$
Adaptive pooling output length	$24$

Appendix

Table 14: SSA-CNN-SVM parameters

Parameter	Value
ISSA population size	$4$
ISSA candidates evaluated	$4$
LSTM batch size	$128$
Evaluation batch size	$512$
Early stopping patience	$8$
RF trees	$120$

Appendix

Table 15: RF-ISSA-LSTM parameters

Explore similar work

Benchmarking Transformer and xLSTM for Time-Series Forecasting of Heat Consumption
May 10, 2026Marja Wahl, Daniel R. Bayer, Sven Rausch +1Forecasting Backbone Time Series Forecasting
RAUSCH Technology GmbH, Würzburg, Germany · Modeling and Simulation, University of Würzburg, Würzburg, Germany
XGBoost "is all you need": the case of forecasting transmitted heat energy in District Heating Systems
Aug 11, 2026Milan ZdravkovićElectricity Price Forecasting Battery Energy Storage Systems
Faculty of Mechanical Engineering University of Niš ul. Aleksandra Medvedeva 14, 18000 Niš, Serbia
Systematic Evaluation of TabPFN-TS and Chronos-2 for Zero-Shot Heat Load Forecasting in District Heating Networks
Aug 20, 2026Ben Spoek, Karim K. Ben Hicham, Kai Derzsi +3Accurate Short-Term Electricity Load Forecasting Zero-Shot Forecasting
Chair of Energy Efficient Buildings and Indoor Climate, RWTH Aachen University, Mathieustr. 10, 52074 Aachen, Germany · Process Systems Engineering, RWTH Aachen University, Forckenbeckstr. 51, 52074 Aachen, Germany

May 10, 2026, cs.LGJ/K move · Enter open · S save

Benchmarking Transformer and xLSTM for Time-Series Forecasting of Heat Consumption

Marja Wahl, Daniel R. Bayer, Sven Rausch, Marco Pruckner

RAUSCH Technology GmbH, Würzburg, Germany · Modeling and Simulation, University of Würzburg, Würzburg, Germany

Forecasting Backbone Time Series Forecasting Transformer Architectures Energy Resources Offshore Wind Turbine

Open paper AskPDF arXiv

Obtaining an accurate short-term forecasting for heat demand is an essential part of operating district heating networks cost-efficient and reliable. Heat consumption time series at the building level are highly dependent on exogenous variables such as outdoor temperature and individual usage patterns, making forecasting in this context a challenging task. Thus, this paper benchmarks novel Transformer-based and xLSTM architectures for short-term heat-demand forecasting. Using hourly data from 25 German buildings (2017-2025), we compare three-hour and 24-hour forecasting horizons relevant for intraday control and day-ahead scheduling. We establish a multi-building benchmark that tests whether models trained on pooled, heterogeneous building data are able to generalize across diverse building stock. The results show that the xLSTM achieves the lowest RMSE (19.88 kWh for three-hour, 21.47 kWh for 24-hour forecasts), while the Temporal Fusion Transformer attains the best MAE (9.16 kWh for three-hour forecasts). As xLSTMs and Transformers require long training times and have a huge number of trainable parameters, their sustainability remains questionable. Therefore, this paper further investigates the trade-off between predictive accuracy and computational resource demand of the evaluated forecasting models. The findings indicate that also low-parameter models like a traditional fully-connected network achieve good predictive results, highlighting that marginal accuracy gains of the novel prediction models come at substantial resource expense for this use case.

Abstract

Explore similar work

Benchmarking Transformer and xLSTM for Time-Series Forecasting of Heat Consumption

XGBoost "is all you need": the case of forecasting transmitted heat energy in District Heating Systems

Systematic Evaluation of TabPFN-TS and Chronos-2 for Zero-Shot Heat Load Forecasting in District Heating Networks