Research on Short-Term Power Load Forecasting Method based on Temporal Convolutional Network

Under the smart grid paradigm, the surging penetration of distributed energy resources and deep integration of demand response mechanisms have led to complex characteristics in power load sequences, including strong nonlinearity, multi-scale coupling, and stochastic abrupt changes, posing significant challenges to traditional forecasting methods. This paper proposes an enhanced Temporal Convolutional Network (TCN) framework featuring an innovative multi-channel spatiotemporal feature decoupling architecture. Through synergistic optimization of dynamic feature weighting modules and temporal attention mechanisms, effective fusion of multi-source heterogeneous features is achieved. The method employs a hierarchical dilated causal convolution structure to preserve temporal causality while enhancing long-range dependency capture capabilities. A gated mechanism-based dynamic allocation strategy for feature contribution levels is designed to precisely quantify the spatiotemporal coupling relationships among meteorological factors, date types, and historical load data. Empirical studies demonstrate that compared with baseline models like LSTM and GRU, the proposed model achieves 23.7% and 18.4% improvements in MAE (15.8MW) and RMSE (21.3MW) metrics respectively, with 67.5% higher training efficiency and 32.1% reduction in prediction error for load mutations. This research provides a reliable solution for high-precision load forecasting in smart grid environments.