How Data can Mislead Machine Learning Prediction Process: Case Study of Building Cooling and Heating Loads

Building heating and cooling power consumption optimization is an important field to reduce and optimize power consumption and to reach green building. To optimize the loads, the building can be constructed utilizing simulation software with its real materials, diminutions, glazing and orientation. Subsequently, the cooling and heating consumptions can be calculated. This method is a complex process. To reduce its complexity, machine learning (ML) has emerged to predict the cooling and heating loads of buildings. However, these algorithms require datasets to train, validate and to be tested. In this work, we attempted to work with a building dataset to predict its heating and cooling loads. Five different supervised machine learning regression algorithms are leveraged. In addition, a new hybrid multi-layer machine learning algorithm (HMLM) is proposed. The dataset utilized is a dataset constructed from simulation software. Our main purpose in this work is to show that simulation datasets may mislead the ML algorithms since their data has been extracted utilizing different assumptions and different preconfigured values. Moreover, simulation datasets should be verified with any real datasets before utilizing them. We have reduced the features in the dataset to only the building height. Our results show that with only this variable an acceptable model with accuracy exceeds 91% can be trained. This is misleading information obtained from the model.