Enhancing short-term building energy forecasting with photovoltaic thermal effects and seasonal heat transfer dynamics using deep learning models

Photovoltaic (PV) systems in industrial buildings affect both electricity generation and the thermal environment, particularly through shading and heat retention effects. These PV thermal effects influence rooftop heat transfer, which can significantly impact building energy consumption, especially in low-insulation structures. However, most short-term forecasting models neglect this interaction, reducing predictive accuracy. This study proposes a short-term energy forecasting framework that integrates PV thermal dynamics using Long Short-Term Memory (LSTM) networks. A PV-roof thermal model is established via the heat balance method to compute hourly heat exchange. Correlation analysis confirms that PV module temperature and heat transfer rates strongly influence load profiles. To capture seasonal variability, seasonal waveform-based clustering is applied to group daily load patterns. A Bidirectional LSTM (BiLSTM) model is then trained with thermal and meteorological features. The framework is validated using high-resolution data from an industrial building in Jingzhou, China—a city with a humid subtropical climate. Results show that models incorporating PV thermal effects significantly improve accuracy. The BiLSTM model achieves superior performance, with RMSE of 152.41 kW and MAPE of 1.84 % in summer, and RMSE of 97.59 kW and MAPE of 1.31 % in winter.