Comparative study of machine learning-based multi-objective prediction framework for multiple building energy loads

Buildings are one of the significant sources of energy consumption and greenhouse gas emission in urban areas all over the world. Lighting control and building integrated photovoltaic (BIPV) are two effective measures in reducing overall primary energy consumption and carbon emission during building operation. Due to the complex energy nature of the building, accurate day-ahead prediction of heating, cooling, lighting loads and BIPV electrical power production is essential in building energy management. Owing to the changing metrological conditions, diversity and complexity of buildings, building energy load demands and BIPV electrical power production is highly variable. This may lead to poor building energy management, extra primary energy consumption or thermal discomfort. In this study, three machine learning-based multi-objective prediction frameworks are proposed for simultaneous prediction of multiple energy loads. The three machine learning techniques are artificial neural network, support vector regression and long-short-term-memory neural network. Since heating, cooling, lighting loads and BIPV electrical power production share similar affecting factors, it is computational time saving to adopt the proposed multi-objective prediction framework to predict multiple building energy loads and BIPV power production. The ANN-based predictive model results in the smallest mean absolute percentage error while SVM-based one cost the shortest computation time.