Short-term power load forecasting for estate-level buildings considering multilevel feature extraction and adaptive fusion

Accurate short-term power load forecasting (STLF) for estate-level building clusters is a core technology achieving efficient energy management, optimal dispatch and sustainable operations in smart power systems. To address the limitations of existing methods, including the failure to adequately obtain features across different time scales, and to dynamically adapt the weights of features under diverse load patterns, this paper proposes a novel multilevel feature extraction and adaptive fusion framework for STLF for estate-level building clusters. Firstly, in an effort to simultaneously and adequately obtain local and global feature information, and retain the inherent chaotic characteristics of historical load data, a dual-branch parallel multilevel feature extraction network based on temporal convolutional networks and Transformer (TT) is specifically designed to extract local and global feature information through their complementary functionalities. Subsequently, an adaptive feature fusion mechanism (AFFM) is designed to dynamically adjust the weights of local and global features under diverse load patterns. Finally, the fused features are fed into the multi-layer perceptron (MLP) to generate the final forecasting. Building on this framework, a novel model, termed TT-AFFM-MLP, is designed for load forecasting in building clusters. Experiments on three actual datasets from different regions with varying temporal granularity show that the proposed model consistently outperforms several other state-of-the-art models, proving its high accuracy and strong robustness.