Hyperparameter Tuning of a Correntropy based ANN for Daily Electric Power Peak Load Forecasting by Modified Brain Storm Optimization

Abstract

Listen

Translate article

This paper proposes a hyperparameter tuning method of a Correntropy based artificial neural network (ANN) for daily electric power peak load forecasting by modified brain storm optimization (MBSO). When the conventional least mean square (LMS) based ANN is utilized and outliers are included in training data, outliers must be removed by engineers and it is a time-consuming job. Although a Correntropy based ANN can ignore the outliers during learning automatically, ignored training data rate must be tuned appropriately. Namely, hyperparameters of the Correntropy based ANN including the ignored training data rate must be tuned appropriately for constructing an appropriate model. Conventionally, a grid search and particle swarm optimization (PSO) have been utilized to tune hyperparameters. However, the grid search takes a long time and PSO suffers from premature convergence. Therefore, the proposed method utilizes MBSO, which has a good balance between intensification and diversification, in order to tune hyperparameters more efficiently. Using the proposed method, electric power utilities can reduce engineering jobs for tuning hyperparameters of a Correntropy based ANN. Moreover, the utilities can reduce engineering jobs for removing outliers included in training data. Namely, the proposed method can drastically advance automatic model generation of daily electric power peak load forecasting. Effectiveness of the proposed method is verified by comparison with a conventional LMS based ANN, a hyperparameter tuning method of a Correntropy based ANN by a grid search, and a hyperparameter tuning method of a Correntropy based ANN by PSO. The results are verified by the t-test and the Wilcoxon signed-rank sum test.