Data Mining and Feature Selection in Large Hospital Central Heating Plant

<p dir="ltr">In the face of anthropogenic climate change, with buildings contributing to roughly one-third of CO2 emissions, optimizing boiler operations is essential. This is particularly pertinent in hospitals in cold climates, where high heating loads are driven by the need for 100% outdoor air in ventilation systems, coupled with strict temperature and humidity controls. A major challenge lies in existing facilities, where older equipment often lacks integration with the Building Automation System (BAS), having few accessible measurement points outside their proprietary boiler controllers. Addressing this, our paper introduces a data mining and feature selection approach to aid the development of grey-box emulators and optimization algorithms for complex, existing central heating plants. We utilize a real-world building to develop a dataset reflective of typical plants. This includes analyzing BAS points, weather data, and variables from physical models, leading to a dataset for feature selection and engineering. The study finds Random Forest Regression slightly superior to Lasso Regression for feature selection, effectively capturing the dynamics of the central heating plant and boiler system. We discuss the key features for predicting Return Water Temperature and forecasting Thermal Power, vital for assessing heating system efficiency and performance. Our results show that Random Forest Regression adeptly predicts Thermal Power and Return Water Temperature with the selected features, despite some challenges in predicting extreme temperatures. Testing the model's accuracy without significant predictors like VFD Speed and Flow Rate highlights the predictive value of Hour of the Day and Day-type for occupancy patterns. Future research will leverage these features to develop online optimization models for the central plant system. The goal is to simultaneously reduce CO2 emissions and energy costs in the studied hospital.</p>