Machine learning framework for short- and medium-term building energy consumption forecasting

Accurate forecasting of building energy consumption is crucial for increasing energy efficiency, enabling demand-side management, and improving decarbonization efforts. However, building load profiles are highly variable, impacted by operational schedules, tenant behavior, and weather conditions, limiting the accuracy of conventional forecasting systems. This study rigorously compares six forecasting approaches—Long Short-Term Memory (LSTM), Artificial Neural Networks (ANN), k-Nearest Neighbors (KNN), Extreme Gradient Boosting (XGBoost), Autoregressive Integrated Moving Average (ARIMA), and Seasonal ARIMA (SARIMA)—applied to 2 months of hourly consumption data from a commercial building. Short-term forecasting performance was evaluated at an hourly temporal resolution using MAE, MSE, RMSE, and R 2 to enable consistent comparison among the six models. Results show that LSTM achieved the lowest MAE and RMSE, while SARIMA achieved the highest test-set R 2 . ANN and KNN performed reliably under low-variance conditions, whereas XGBoost was more effective at lower consumption levels than at peak loads. ARIMA and SARIMA capture seasonality but struggle with outliers and near-zero loads. A separate supplementary synthetic-data analysis examined ANN forecasting over a 14-day daily horizon. The novelty of this work lies in systematic comparative benchmarking within a unified experimental framework that clarifies the forecasting horizon and quantifies performance differences. The findings provide actionable insights for the design of smart building energy management systems and inform policy initiatives aimed at improving demand-side flexibility and sustainable energy use.