Insight into predictive models: On the joint use of clustering and classification by association (CBA) on building time series

Abstract

Data-driven, black box machine learning models have received a lot of attention in the field of building control. They have been used successfully to predict building behaviour given information like weather forecasts and real time sensor information. In these models, the occupant behaviour is considered to act exogenously on the building. We consider the users as active elements of the building operation control loop. To make educated control decisions they have to be informed about how the building will behave. Therefore, we propose a prediction model which explains to occupants the day-ahead building behaviour using a clustering and classification by association model. We benchmark this approach to a neural network regression model and only observed a small loss of accuracy. Knowing the upcoming building behaviour, occupants can adjust their behaviour (e.g. putting on clothes) or the building systems settings (e.g. set points) accordingly. The proposed method is a promising way to decode complex regression models into readable rules, which in future may be useful in conjunction with for example voice-based virtual assistants.