Machine learning estimation of prosumer net metering using optimal polynomial basis functions for dimensionality reduction

Abstract Solar energy has become an essential renewable resource for residential prosumers engaging in bidirectional electricity exchange with the grid. However, accurately estimating net consumption remains a challenge due to the prevalence of limited yet redundant local Photovoltaic (PV) data. This paper designs an input dimension reduction approach using optimal polynomial basis functions to enhance machine learning estimation. Optimal higher-order powers are determined via an L 2 -norm performance function based on matrix conditioning. Additionally, Pearson correlation and the Least Absolute Shrinkage and Selection Operator (LASSO) are employed to eliminate redundancies and quantify key input–output relationships. The refined data is used to train Batch Least Squares (BLS), Neural Network (NN), and Artificial Bee Colony (ABC) algorithms. Results demonstrate that the BLS model, integrated with Pearson-based feature selection, achieves superior precision with a Mean Absolute Error (MAE) of 0.0037 and a Root Mean Square Error (RMSE) of 0.0048. Validation on unseen test data confirms the model's robustness, maintaining a low MAE of 0.0036 and RMSE of 0.0046.