Online HVAC system modeling with BMS data using recurrent neural networks

This paper presents a method to develop and tune HVAC system models with BMS data using recurrent neural networks (RNN). To update RNN models automatically online, we eliminate ad hoc adjustment technique in three points. One is in selection of inputs variables. We propose a method of calculating the necessity of input variables with using connections' weights of neurons. Second is in the selection of hidden layers unit numbers. We introduce "Cross-validation" method and it gives criterion to evaluate parametric model. Last is in selection of training data. We apply cluster analysis to BMS data to evaluate scarcity of data. Proposed approaches are tested real measurements of actual buildings in Japan.