Load Forecasting Method based on Multi Loss Function Collaborative Optimization

Abstract

Accurate load forecasting can help the power sector to formulate a reasonable power generation scheme, which can ensure the reliability of power supply while minimizing resource waste. However, most of the existing prediction methods based on deep learning only regard the minimum loss function of the training dataset under laboratory conditions as the optimal model, resulting in low generalization of the model and poor performance of the model in solving practical engineering problems with universality. To solve the above problems, this paper proposes a load forecasting model based on multi loss function collaborative optimization, considering the constraint relationship between the variance, deviation and model generalization error of forecasting results. Considering the different physical meanings of different loss functions, the model calculates the weighted sum of multiple loss functions, and then optimizes the weight value of combined loss functions by using genetic algorithm. The results show that the prediction error of combined loss function is smaller than that of single loss function under the premise of selecting appropriate weight parameters.