Bayesian Optimized Supervised Learning for Short-Term Net Load Forecasting in Distribution Grids with Net-Metering

As power systems evolve into more intelligent, flexible, and interactive networks with a higher penetration of Renewable Energy Sources (RES) on the demand side, net load forecasting has become a cost-effective and essential technique for planning, stability, and reliability of modern power grids. This study explores both direct and indirect methodologies for short-term net load forecasting (STNLF) in distribution grids considering solar net-metering using deep learning (DL) and machine learning (ML) techniques. The proposed models begin with feature engineering and filtering using Pearson Correlation and Mutual Information (MI) indices, followed by developing a forecasting models employing a Bayesian Optimized Long Short-Term Memory (BO-LSTM) network and Bayesian Optimized Random Forest (BO-RF). The model’s effectiveness was validated using historical net load data from the Lahore Electric Supply Company (LESCO) spanning January 2022 to January 2024, as LESCO has the highest number of net-metering customers among all Distribution Companies (DISCOS) in Pakistan. Results demonstrate the BO-RF model’s capability to deliver accurate and robust STNLF, achieving a Mean Absolute Percentage Error (MAPE) of $\mathbf{3. 5 6 \%}$ for the test data and $\mathbf{2. 2 1 \%}$ for the train data. This study also compares the efficiency of indirect and direct methods for STNLF, demonstrating that the indirect method generally performs better than direct method in distribution grids.