Short-term Load Forecasting Method of IES Based on TCN Enhanced by Attention Mechanism with Decomposition Strategy

Introduction: Constructing integrated energy systems (IES) represents a significant trend in energy structure transformation and system development. Due to the inherent characteristics of IES, such as load variation, fluctuation, and complex energy coupling, accurate multi-energy load forecasting is essential to ensure economic dispatch and optimal operation. Methods: In this study, the coupling relationships between energy sources are quantified using the Maximal Information Coefficient (MIC), and the energy load influencing factors are identified through Grey Relation Analysis (GRA). Furthermore, to enhance frequency separation and reduce data noise, the time-varying filter approach for empirical mode decomposition (TVF-EMD) is applied. Additionally, an Attention mechanism is integrated with Temporal Convolutional Networks (TCN) to develop a multi-task hybrid model, which effectively extracts temporal features and assigns varying degrees of attention to shared features. The proposed TVF-EMD-ATCN method is validated using publicly available IES datasets from Arizona State University. Results: Compared to traditional single-task and data decomposition methods, the proposed model achieved Mean Absolute Percentage Errors (MAPE) of 0.0181, 0.0209, and 0.0221 for cooling, heating, and electricity load forecasting, respectively. Other evaluation metrics also demonstrated superior performance. Conclusion: The experimental results indicate that the proposed method offers excellent forecasting accuracy and strong reliability for short-term IES load prediction.