Estimating cooling loads of Arizona State University buildings using microclimate data and machine learning

In hot arid urban climates, building cooling equipment consumes the greatest share of energy out of all building end-use equipment. For existing buildings, dynamic system-based building energy simulation tools have provided valuable information on the impact of microclimates on cooling energy use. However, such models require in-depth information on building model parameters and often suffer from modeler bias even when sufficient calibration indices are satisfied. This study presents a data-driven approach for predicting the cooling loads of three university buildings in Arizona using simulated microclimate data. A microclimate model ENVI-met generated the input micro-scale weather data for each building. The ENVI-met simulation was validated using in-situ observations during the summer of 2018. Multiple machine learning algorithms were implemented. A final model was selected and used as baseline to predict cooling loads for each building in the dataset. The model predicts chill water tons per square meter using microclimate variables that include mean air temperature, mean absolute humidity, shading levels, and direct shortwave radiation. The black-box model was explained using an advanced machine learning model interpretation library in Python: SHAP. The baseline model predicted cooling loads with a prediction accuracy score of 0.98 using the tree-based algorithm Random Forest. Sensitivity analyses and scenario results showed that cooler microclimates reduced cooling loads for the modeled buildings. The developed framework will be used in future study extensions to explore the impacts of simulated microclimate scenarios generated by ENVI-met.