Short-Term Load Forecasting for a Campus Building: Regression with Residual Correction and Gaussian Smoothing versus a Feedforward Neural Network

Abstract Accurate short-term load forecasting supports efficient operation and energy planning for large buildings, yet many campus facilities lack the data volume and ex-pertise needed for complex models. This paper compares two hourly forecasting approaches for Shepard Hall at The City College of New York using metered load data from 2022 to January 2026. The first is a multiple linear regres-sion model with daily and weekly lag features, an hourly mean-load feature, and weekend indicators, enhanced by hour- and day-type residual correction and a five-point Gaussian smoothing filter. The second is a two-hidden-layer feedforward neural network with 14 engineered tem-poral, calendar, and academic-schedule features, trained with recency-weighted loss and the Adam optimizer. On the test week of February 1–7, 2026, Gaussian smooth-ing reduced the regression mean squared error (MSE) by 46%, from 91.68 to 49.26 kW2 (RMSE 7.02 kW), while the neural network achieved an MSE of 219.53 kW2 (RMSE 14.82 kW). The results show that a simple, interpretable regression with targeted post-processing can outperform a deeper neural network for single-building forecasting with limited data.