Short-Term Electricity Load Forecasting for Campus Buildings Using LSTM Networks

Abstract

Accurate short-term load forecasts help campus facility managers schedule air conditioning and avoid peak-demand charges. This paper evaluates long short-term memory (LSTM) networks for hourly load forecasting across five academic buildings using two years of smart-meter data enriched with calendar and weather features. The LSTM is compared with seasonal ARIMA and gradient-boosted trees over a rolling-origin evaluation. For a 24-hour horizon, the LSTM achieved a mean absolute percentage error of 6.4%, compared with 9.1% for seasonal ARIMA and 7.2% for gradient boosting. Performance dropped during semester breaks, when occupancy patterns change abruptly, which suggests that academic-calendar features are essential. The model has been integrated into a dashboard that alerts operators to expected peaks a day in advance.