Enhancing Efficiency of Ensemble Machine Learning Models for Short-Term Load Forecasting through Feature Selection

Electric load forecasting directly affects management planning and is necessary for the reliable functioning of electric power systems. Furthermore, accurate forecasts are required by power companies for electricity generation and distribution planning. This paper compares different approaches to the feature selection process for solving the problem of short-term forecasting of industrial enterprise power consumption using machine learning methods. A comparative analysis of the following machine learning methods was carried out: adaptive boosting over decision trees (AdaBoost), gradient boosting over decision trees (Gradient Boosting), and random forest (Random Forest). The choice of such models is due to the presence of binary features in the original data, which are best handled by decision trees. The selection of the most significant features for these models is especially important as they are prone to overfitting when the input data set is large. The most significant features of the original time series were determined by calculating the correlation coefficients between the predicted hour and the preceding hours. Moreover, the repair schedule is used as an additional feature. Several approaches to feature selection were considered. The results showed that using data only on selected hours instead of the entire time series improves the accuracy of the model and the speed of its training. It was found that the use of the maintenance schedule also significantly increases the accuracy of the forecast. The Gradient Boosting model showed the best result with an average error of 5.48 %.