Improving Accuracy of Machine Learning Based Short-Term Load Forecasting Models with Correlation Analysis and Feature Engineering

Short-term load forecasting is an integral part of the electric power system management and is necessary to ensure the electricity market operation. Load forecasting of individual market participants, especially industrial enterprises, can be particularly challenging due to many external factors that influence the load pattern. These factors may include a change in weather conditions, holidays and days off, economic trends, etc. Therefore, there is a need to develop new and more advanced forecasting models that are able to produce adequate predictions of future power consumption under such conditions. This study considers the task of day-ahead load forecasting of a mining and processing plant in Yakutia. The analysis of the exogenous data use impact on its load pattern was carried out. The forecasting models were built using two machine learning methods: decision tree ensembles (Adaptive Boosting, Gradient Boosting and Random Forest) and artificial neural networks (Deep Feedforward Neural Network). To improve the model performance, the following data preprocessing techniques were used: correlation analysis, feature engineering, and intelligent feature selection. The experiments show that the appropriate feature selection can significantly increase the accuracy of the forecasting models. In this case, taking into account the weather data allowed to reduce the average forecasting error by 22% for the neural network model, yielding the best result among the considered models.