Limitations of machine learning for building energy prediction: ASHRAE Great Energy Predictor III Kaggle competition error analysis

Research is needed to explore the limitations and potential for improvement of machine learning building energy prediction. With this aim, ASHRAE Great Energy Predictor III (GEPIII) Kaggle competition was launched in 2019. This effort largest meter its kind, with 4370 participants who submitted 39,403 predictions. The test dataset included two years hourly whole readings from 2380 meters 1448 buildings at 16 locations. paper analyzes various sources types residual model error an aggregation competition’s top 50 solutions. analysis reveals using standard inputs historical meter, weather, basic metadata. errors are classified according timeframe, behavior, magnitude, incidence single or across a campus. results show models have within range acceptability (RMSLEscaled = < 0.1) on 79.1% data. Lower magnitude (in-range) (0.1 RMSLEscaled 0.3) occur 16.1% These could be remedied innovative training data onsite Web-based sources. Higher (out-of-range) > 4.8% unlikely accurately predicted.