Medium-term load forecasting in isolated power systems based on ensemble machine learning models

Over the past decades, power companies have been implementing load forecasting to determine trends in the electric power system (EPS); therefore, load forecasting is applied to solve the problems of management and development of power systems. This paper considers the issue of building a model of medium-term forecasting of load graphs for EPS with specific properties, based on the use of ensemble machine learning methods. This paper implements the approach of identification of the most significant features to apply machine learning models for medium-term load forecasting in an isolated power system. A comparative study of the following models was carried out: linear regression, support vector regression (SVR), decision tree regression, random forest (Random Forest), gradient boosting over decision trees (XGBoost), adaptive boosting over decision trees (AdaBoost), AdaBoost over linear regression. Isolation of features from a time series allows for the implementation of simpler and more overfitting-resistant models. All the above makes it possible to increase the reliability of forecasts and expand the use of information technologies in the planning, management, and operation of isolated EPSs. Calculations of the total forecast error have proved that the characteristics of the proposed models are high quality and accurate, and thus they can be used to forecast the real load of a power system.