At the University of Missouri College of Engineering, researchers Sanjeev Khanna and Saad Alsamraee tackled this problem. Using six years’ worth of hourly data—over 52,000 data points—from the Combined Heat and Power Plant, the study covers energy usage trends before, during, and after the COVID-19 pandemic. Surprisingly, campus-wide consumption held steady even in 2020, falling only slightly from 28.26 MWh in 2019 to 26.37 MWh. Reduced classroom and office usage was offset by the continued operation of research and essential facilities.
Older energy forecasting models—like ARIMA (Autoregressive integrated moving average)—were decent at handling basic trends but struggled with the kind of messy, real-world data that university campuses produce. To tackle this, the University of Missouri research team tested several machine learning tools that are better at spotting complex patterns in data. These included Decision Trees, Random Forest, Support Vector Regression, K-Nearest Neighbors, and one called XGBoost.
XGBoost stood out from the rest. It handled the huge dataset with ease and delivered the most accurate forecasts—cutting prediction errors by 46%. That kind of improvement isn’t just impressive on paper—it has real-world benefits, like helping facilities teams better plan energy use and cut costs.
XGBoost not only performed well with past data—it held up when predicting the university’s energy use a year in advance. For 2023, it forecasted 241,236 megawatt hours—just a hair off the actual number, 239,217. That’s less than a 1% difference, which gives campus managers confidence in planning everything from maintenance schedules to energy purchases.
Even when the researchers introduced problems—like removing 10% of the data to mimic real-world outages—the model stayed solid. They also used a tool called SHAP (Shapley Additive Explanations) which helped to understand what influenced the predictions most. Turns out, outside temperature had the biggest impact, followed by what time and day it was—patterns that line up with how buildings actually operate.
This study lays out a clear path for other large campuses or facilities—like hospitals or government buildings—to do the same. With tighter budgets and climate goals to meet, using smarter forecasting tools like this isn’t just a nice upgrade—it’s becoming a must-have.
A professor of mechanical engineering and director of the Midwest Industrial Assessment Center, Khanna has garnered support from the DOE over the last 18 years. As he explains, "By knowing when there are going to be peaks and valleys and how much energy will be needed, even on an hour-by-hour basis, we can ultimately help power plants better plan ahead so they can be as efficient as possible with energy use. This research can help universities and industries reduce carbon emissions and save money."
