High-resolution energy consumption forecasting of a university campus power plant based on advanced machine learning techniques

Effective long-term energy forecasting is essential for efficient management of large institutions like university campuses, yet traditional forecasting methods frequently fall short in capturing complex consumption behaviors. To bridge this gap, this study introduces an advanced machine learning (ML) framework leveraging an extensive hourly energy consumption dataset from the University of Missouri campus over a period of six years from 2017 to 2022. The dataset uniquely integrates energy demand data from the university's Combined Heat and Power Plant (CHPP) alongside critical environmental parameters, such as air temperature, humidity, wind speed/direction, atmospheric pressure, and solar intensity, capturing distinctive consumption patterns across pre-pandemic, pandemic, and post-pandemic periods. Several ML algorithms - Decision Tree (DT), Random Forest (RF), Support Vector Regressor (SVR), K-Nearest Neighbor (KNN), and eXtreme Gradient Boosting (XGBoost) - were rigorously trained, validated, and benchmarked. The XGBoost model evidently emerged as superior, achieving impressive forecasting accuracy with MAE of 0.8680, RMSE of 1.2078, MAPE of 3.12 %, and R2 of 0.94. Additionally, the model's probabilistic forecasts were validated using the negative log likelihood (NLL = 1.5924), confirming robust performance and reliable uncertainty quantification. Furthermore, SHapley Additive exPlanations (SHAP) were employed to interpret the model predictions, highlighting the critical roles of air temperature, daily temporal cycles, and seasonal factors in driving energy usage. Finally, during the deployment phase, the optimal model was employed to forecast energy demand for the full year 2023-the primary objective of this study-exhibiting high robustness through close adherence to actual demand patterns.