Comparative Analysis of Sequential and Non-Sequential Machine Learning Models for Electrical Load Forecasting in Medical Facilities

Abstract

Medical buildings require a reliable power supply, yet their electrical demand is difficult to predict because equipment use, occupancy and air-conditioning loads change rapidly, and many accurate forecasting models are too demanding for low-cost edge devices. This study compares a sequential model, long short-term memory (LSTM), with a non-sequential model, extreme gradient boosting (XGBoost), for hour-ahead load forecasting in a medical faculty building. Public circuit-level measurements were averaged to hourly values and split chronologically into training, validation and test sets. Both models used the same 24-hour look-back window, were tuned by grid search on the validation set and were run on the same computing platform. Accuracy was assessed with the root mean square error (RMSE), the mean absolute error and the coefficient of determination, and cost with training and inference time. The tuned LSTM gave the lowest test error (RMSE 6.12 kW), ahead of XGBoost (7.51 kW), and both outperformed a persistence baseline (8.30 kW). However, XGBoost trained about 220 times faster and produced forecasts about 180 times faster. The results reveal a clear accuracy-versus-cost trade-off that can guide model selection for edge-based energy management in medical buildings.