Weekly periodicity feature-driven CNN-GRU and LightGBM hybrid model for short-term multi-step building load forecasting

Accurate building load forecasting is essential for the efficient scheduling and operation of energy systems. However, due to the pronounced non-stationarity and volatility of building energy consumption, short-term load forecasting remains challenging. To address this issue, this paper proposes a short-term multi-step load forecasting method based on a convolutional neural network (CNN)-gated recurrent unit (GRU) and light gradient boosting machine (LightGBM) hybrid model, termed CGLHM. First, a data-driven approach is used to verify the weekly periodicity of building load, based on which weekly periodic features are constructed. Then, CNN-GRU is employed to extract short-term dynamic dependencies, while LightGBM captures cross-week repetitive patterns enhanced by periodic features. Finally, an adaptive weighted fusion mechanism is used to combine the outputs of the two branches across different forecasting steps, thereby improving overall prediction accuracy. The proposed model is compared with multiple benchmark models in terms of forecasting accuracy and computational cost. Experimental results show that CGLHM achieves the best overall forecasting accuracy across different prediction horizons and seasonal conditions while maintaining acceptable computational cost. The results further demonstrate the robustness and practical applicability of the proposed method across different building types.