A short-term load forecasting hybrid model based on hierarchical optimization and ensemble networks

Accurate short-term load forecasting (STLF) is crucial for ensuring the stable operation of power systems and promoting the development of electricity markets. However, current research for STLF overlooks model optimization methods and the need for more flexible weighted ensemble methods, which leads to less research on excavating the prediction performance of the models and improving their utilization efficiency. To address these issues, this paper proposes a novel STLF model based on Hierarchical Optimization and Ensemble Networks (HOEN), which sequentially applies a deep learning model optimization method, a multi-model ensemble method, and a residual correction method to perform hierarchical optimization and ensemble of LSTM, GRU and MTS-Mixers. Firstly, a Spiral Fitting Optimization (SFO) method is proposed based on the theory of Stochastic Weight Averaging (SWA). This method explores the optimal model parameter matrix by fitting the logarithmic spiral curve between the parameter weights and loss values of models, thereby excavating and optimizing the predictive performance of the model. Secondly, an Optimal Model Classification Ensemble (OMCE) method is proposed, which dynamically assigns the weights of each model by constructing a data-driven dynamic weighting mechanism, improving the utilization efficiency of the optimized models. Finally, residual correction method is applied to optimize the forecast residuals of the ensemble model, further enhancing its predictive performance. The proposed methods and models are validated using load data from Hubei Province, China, and Belgium in this paper. The results demonstrate that our methods and models exhibit clear advantages in terms of accuracy, stability, and generalization ability in STLF.