AI-Based Energy Management for Renewable-Integrated Smart Grids using Load Forecasting and Battery Optimization

Abstract

Renewable-integrated smart grids and commercial microgrids need operating strategies that keep photovoltaic (PV) surplus, time-of-use (TOU) prices and peak-demand charges under control. This paper develops and evaluates a forecast-driven energy management framework that combines (i) day-ahead load and PV forecasting, (ii) a linear-programming battery scheduler with incremental demand-charge accounting, hourly re-optimisation and calibrated uncertainty margins, and (iii) a real-time peak guard that protects execution from forecast errors. It is tested on one year of public hourly hospital-load data and NREL TMY3 irradiance processed through a physical PV model (1.95 MWp PV, 1 MW / 2 MWh battery) under an explicit TOU-plus-demand-charge tariff, using rolling-origin evaluation over 214 days. An ensemble of Ridge regression, random forest, LightGBM and a multilayer perceptron lowers day-ahead load RMSE by 19.8% relative to a seasonal-naïve benchmark (51.9 vs 64.7 kW; MAPE 2.61%), whereas the learned PV forecast improves on smart persistence by only 1.6%. The AI-driven scheduler reduces operating cost by 4.52% ($26,243) and mean monthly peak import by 12.0% relative to PV-only operation, capturing 60.9% of a perfect-foresight bound (7.42%); a persistence-driven scheduler attains a comparable 4.98%, so forecast accuracy gains do not translate one-for-one into savings. Removing the real-time guard turns the saving into a 2.34% loss, and error-scaling experiments show that PV forecast error, rather than load-forecast error, governs the value of information