AI-Driven Energy Management for a Solar-Integrated Smart Building: Forecast-Informed Battery Scheduling under Time-of-Use and Demand Charges

Abstract

Commercial buildings with rooftop photovoltaic (PV) generation and battery storage face a scheduling problem: solar output and building demand are both uncertain, while electricity prices vary by hour and by the monthly peak. This paper develops and evaluates an AI-driven energy management framework in which gradient-boosting models forecast day-ahead building load and PV output, and a linear-programming optimiser converts those forecasts into a battery schedule that minimises energy cost, demand charges and degradation cost. The framework is tested on a simulated 5,000 m² office building with a 200 kWp rooftop array and a 300 kWh / 100 kW battery in a hot semi-arid climate, using a synthetic three-year hourly data set (two years for training, one for testing). On the held-out year, the gradient-boosting load forecaster reached an RMSE of 5.1 kW (R² = 0.989), against 15.0 kW for a persistence benchmark. The forecast-driven schedule cut annual electricity cost by 10.0% relative to PV-only operation and by 6.6% relative to a rule-based battery controller, and reduced evening peak-tariff imports by 82%. It did not reduce the monthly maximum demand, which a perfect-foresight benchmark could cut by almost half; this gap quantifies the cost of forecast error. A sensitivity analysis over battery size and peak tariff shows that the value of storage depends strongly on the tariff spread, with simple payback ranging from about six to more than forty years under assumed capital costs. Building load, tariffs, costs and roof-structure parameters are simulated or assumed, so the results show relative performance of control strategies rather than predictions for a specific site.