Optimized Multi-model Machine Learning Approach for Residential Load Prediction

Accurate prediction of the residential load is important to increase the utilization of available local energy resources, to address over/under voltage problems, and overloading issues in low voltage (LV) distribution grids. The need for an optimized prediction model has been observed that performs well even when the dataset has characteristic deviations. This paper aims to present an optimized model for residential electrical load prediction. Linear regression, regression tree, ensemble, and gaussian process regression (GPR) methods are considered, and based on the error metrics evaluation, a prediction model with minimum error is selected for load prediction. Calendar and weather-based features are selected based on the principal component analysis. A case study has been performed, considering the measured smart meter data from a multi-storied residential house in Houston, Texas, USA. A comparison of predicted and measured residential electrical load shows that the suggested approach performs satisfactorily.