Department

Construction Management

Additional Department

Civil and Environmental Engineering

Document Type

Article

Publication Date

Fall 9-2025

Embargo Period

12-12-2025

Abstract

University buildings are energy-intensive and operate on complex schedules, making electricity demand forecasting particularly challenging. This study develops and evaluates monthly forecasting models for a public university campus in Georgia using six years of data (January 2019–December 2024) that integrate weather variables and academic calendar indicators. Three modeling approaches are compared: Seasonal Autoregressive Integrated Moving Average (SARIMA), SARIMA with exogenous variables (SARIMAX), and a hybrid SARIMAX–Long Short-Term Memory (LSTM) model. Feature selection methods, correlation analysis, Granger causality, Random Forest importance, Recursive Feature Elimination (RFE), and Least Absolute Shrinkage and Selection Operator (LASSO) regression, were applied to optimize input relevance. The hybrid SARIMAX–LSTM model outperformed all others, achieving a Mean Absolute Percentage Error (MAPE) of 8.10%, compared to 15.26% for SARIMA, 10.62% for SARIMAX, and 8.43% for LSTM. It reduced error by 23.7% in MAPE, 23.3% in Mean Absolute Error (MAE), and 22.7% in Root Mean Square Error (RMSE) relative to the best standalone models. Statistical diagnostics confirmed the model’s residual validity. These findings highlight the advantage of combining linear and nonlinear modeling techniques to improve forecasting accuracy in institutional settings. The proposed framework supports proactive energy planning and aligns with broader sustainability and operational goals.

Journal Title

Energy and Buildings

Journal ISSN

0378-7788

Volume

347

Digital Object Identifier (DOI)

10.1016/j.enbuild.2025.116400

Share

COinS