Research on Multi-Algorithm Fusion and Deep Learning Models for Power Consumption Forecasting

Accurate power load forecasting is a critical element in ensuring the safe and economic operation of power grids. However, the nonlinear characteristics of load data and abrupt operational conditions pose significant challenges to forecasting. This study proposes a forecasting method based on a Long Short-Term Memory-Gated Recurrent Unit (LSTMGRU) hybrid deep learning architecture, designed with a multiscale spatiotemporal feature extraction module. Through parallel convolutional structures and graph convolutional networks, effective load feature extraction is achieved. An adaptive weighted combined loss function is constructed to optimize the model training process, and a Stacking strategy is employed to integrate various base algorithms to improve generalization performance. The feature extraction module incorporates attention mechanisms to enhance the capture of key spatiotemporal features, while residual connections are utilized to mitigate the vanishing gradient problem. Experimental results demonstrate that this method exhibits excellent forecasting performance under various load conditions, particularly during load fluctuations, with strong adaptability. Through comparative analysis with traditional methods and other deep learning models, the advantages of this method in terms of prediction accuracy and model stability are validated, providing a new technological solution for power system load forecasting.