Research on heat load forecasting for public buildings: Multi-step feature selection and discrete wavelet transform of time series data

Heat load forecasting is of great significance for energy conservation and carbon reduction in public buildings. However, its usage patterns differ from those of residential buildings, making the long-term trends and short-term fluctuations of the heat load more difficult to capture. This study proposes a strategy combining multi-step feature selection with discrete wavelet transform for time series data to enhance heat load forecasting accuracy. First, the heat load time series data undergoes discrete wavelet transformation to reveal the heat load variation patterns and high and low frequency characteristics. The forecast period is determined by threshold, sliding time window and cross-correlation function. The prediction time points are determined by detecting the extreme values of the high-frequency components of the DWT. A two-step feature selection process is used to identify an optimal feature set, and the DWT components are then added to the feature set to optimize the prediction. The proposed method are verified in a commercial building. The regulation period (or forecast period) for the initial cold period and the final cold period is 4 h, while the regulation period for the high cold period is 6 h. The proportion of forecast errors within the range of 0 ≤ pMAPE <5% has increased to 93.3%, with an MAPE of 1.64% and an R2 of 0.991. And through generalization verification, further validation of the above process and method is carried out.