AI-Driven Forecasting For Optimizing Energy Storage and Reducing Carbon Emissions in Renewable-Heavy Smart Grids

Abstract

Distributed generation is more intermittent, and intelligent storage control systems are essential to match this variability across the grid and keep carbon emissions to a minimum. This study compares the actual performance of a forecast-driven, carbon-aware battery storage strategy to the baseline without battery storage and with conventional self-consumption and involves the actual performance of a carbon-aware battery storage strategy based on forecasts, uses operational data from a TX-123BT Texas test feeder over five years (2017-2021) of data with the same standardized operating conditions. The Day-ahead (DA) carbon intensity was predicted using a RF and DNN model, both with a persistence baseline. The best performance was provided by the RF (16.3% MAPE vs. 18.5% for DNN). Ablation and SHAP analysis of a climate feature showed that real exogenous features were more beneficial to the RF than they were to the DNN. A tabular Q-learning storage scheduling agent using a novel marginal emissions reward was deployed. With the default battery rating (4,000 kW / 15,000 kWh), the carbon-aware agent decreased emissions by 0.0074%, and the self-consumption strategy yielded no benefit because residual load never fell below zero. A sweep of the capacity sensitivity, however, did show that the reduction of carbon scales perfectly linearly with battery power rating. These results show that effectiveness is limited by storage scale and the feeder's ~23% renewable penetration rather than control methodology.