The exponential growth of digital infrastructures and connected devices has made energy demand increasingly variable and difficult to anticipate. In 2023, smart buildings accounted for nearly 20% of urban energy consumption, underscoring the urgency of optimized management. This paper investigates how artificial intelligence (AI) can improve real time optimization of energy consumption in smart grids.We collect and pre process IoT sensor time series and evaluate two neural approachesMultilayer Perceptron(MLP) and Long Short Term Memory (LSTM)against a seasonal ARIMA baseline. On a simulated campus scale testbed inspired by ouruniversity infrastructure, LSTM improves next hour demand forecasting accuracy by 18.6% over ARIMA and by 5.8% over MLP, achieving an RMSE of 0.218 kWh. A redistribution simulation driven by predictions yields an average 14.7% reduction in energy losses and a 9.3% net energy gain in office buildings. We discuss robustness to miss data (≤5%), abrupt load changes, and operational disturbances, and situate our findings with respect to recent literature including LSTM based building forecasting, deep reinforcement learning for grid control, and IoT enabled management frameworks. We conclude with actionable deployment considerations for African campuses and municipal facilities.
Kra et al. (Sat,) studied this question.