Multi-dimensional data-based medium- and long-term power-load forecasting using double-layer CatBoost

Abstract

In this study, a medium- and long-term power load prediction method is proposed based on the two-layer categorical boosting (CatBoost) algorithm with multi-dimensional feature considerations. Simultaneously, the influences of economic fluctuation, power generation disruption, and meteorological data on power load are considered, whereby the dimension of power-load forecasting data characteristics is broadened. A randomised search cross-validation (CV) regression model is also applied to model parameter optimisation. Real data from a province in northeast China were used for the training and test sets. Compared with nine advanced load prediction models, including eXtreme gradient boosting and adaptive boosting, the coefficient of determination (R2) of the proposed method was 0.925, mean average percentage error (MAPE) was 0.0158, and root-mean-square error (RMSE) was 274.2036. In this study, a popular, viable artificial intelligence technology, two-layer CatBoost, was explored, and multi-dimensional external variables of power generation were added for the first time for load prediction. Finally, a higher accuracy load forecasting tree model is presented. The method has good potential for use in medium- and long-term power-load forecasting applications.

Keywords

• Load forecasting
• Machine learning
• CatBoost
• Randomised search CV