From physics to machine learning in forecasting thermal load of a district-heated residential building: Assessment of different methods and influence of input data

Accurate thermal load forecasting is crucial for optimizing the operation of heating systems, particularly under predictive control strategies. The choice of the forecasting method is strongly influenced by data availability and compatibility with control logic. However, there is still a lack in scientific literature on the detailed evaluation of forecast methods for the operation of buildings in weather-based control mode and in forecast control mode of building heating system. This study presents a structured multi-method approach to model space heating demand of a residential building in Poland, which has been monitored since 2018 and is being operated in forecasting mode since 2021 through the forHEAT system developed by the Lublin University of Technology. The analysis is articulated in three methods. Method 1 uses a calibrated EnergyPlus model to simulate building behavior under both weather-based (2018–2021) and forecasting-based control (2021–2024), offering insights into seasonal load patterns and about the limitations of static modeling under dynamic control. Method 2 develops 26 nonlinear autoregressive neural networks with exogenous inputs (NARX), exploring combinations of seasonal subsets, training durations, and control paradigms. Results show that recent, seasonally aligned training data enable the best performance ( R 2 = 0.937, cvRMSE = 9.96%), while shorter or misaligned datasets lead to higher error and reduced generalization. Method 3 introduces a simplified empirical model based on equivalent outdoor temperature ( T eq ) for real-time implementation. While T eq maintains low bias ( NMBE within ± 0.6%), it suffers from higher variability ( cvRMSE up to 27%) and limited adaptability. Overall, NARX networks consistently outperform both EnergyPlus and T eq across metrics, particularly in scenarios with transitional or mismatched regimes, making them the most robust tool for accurate heat load forecasting. Future developments can expand the model to other building typologies and climatic locations.