Implementation of Neural Networks and Feature Selection for Short Term Load Forecast

Forecasting power demand is a significant factor in the planning and secure operation of power systems. Forecasting is based on considering all the factors that affect power demand. Reducing of the forecasting error can be achieved by identifying the factors that affect power demand, understanding their influence on the forecasting and adequate usage in the forecasting model. The selection of the optimal number and combination of features that will lead to a prediction model with less error is performed using the Recursive Feature Elimination with Cross-Validation (RFECV) elimination method. The features are selected according to a statistical importance measure. Then, the Artificial Neural Networks (ANN) are used to build load forecasting models, for a specified hour in future, and then, by combining those models, an one day-ahead model was created. The proposed approach combines the abilities of the neural networks to learn the nonlinear relationship between features and the optimization capability of the RFECV elimination method to find the best model for forecasting.