Research on heat load forecasting methods for heating stations based on adaptive multi-stage wavelet denoising and multi-technology integration

This paper addresses the challenges of data noise interference and model performance optimization in heat load forecasting for district heating stations (DHS). It proposes an innovative multi-technology fusion forecasting method. First, sliding box plots are employed to identify anomalous data. Following data repair using PCHIP interpolation, an adaptive multi-stage wavelet denoising (AMSWD) method is designed. This method effectively suppresses noise while preserving key features of the original signal with high fidelity. A hybrid heating station thermal load prediction model integrating DOA-T-BiTCN-BiGRU-ThermaFocus-Mamba is constructed. The DOA optimization algorithm enables adaptive hyperparameter configuration. An improved bidirectional temporal convolutional network (T-BiTCN) with TeLU activation function extracts local spatial features, while bidirectional gated recurrent units (BiGRU) capture long-term temporal dependencies. The ThermaFocus-Mamba module dynamically adjusts information flow, significantly enhancing the model's prediction accuracy and generalization capability. To validate the proposed model, thermal load data and related metrics from three distinct heating stations in an urban area were collected for experimental evaluation. Experimental results demonstrate that the proposed model exhibits outstanding prediction accuracy and generalization capabilities across diverse heating scenarios and multi-timescale forecasting tasks. Its overall performance significantly outperforms baseline models and existing state-of-the-art prediction methods, providing a novel and effective solution for energy-optimized management in intelligent heating systems.