Feature assessment in data-driven models for unlocking building energy flexibility

Data-driven approaches are playing an increased role in building automation. This can, in part, be attributed to building operation and energy management system data becoming more readily accessible. A particular application is models to allow predictive control harnessing building energy flexibility, which is of interest to different stakeholders including; energy utilities, aggregators and end-users. Given the possibility of thousands of data features, feature selection becomes a critical part of the model development process. This paper considers various filter, wrapper and embedded methods applied in conjunction with three predictors in addressing the problem of constructing a suitable data-driven model to facilitate predictive control and provision of energy flexibility in a large commercial building. The feature selection algorithms are generally shown to significantly reduce model evaluation time and, in some cases, increase model accuracy. A random forest model with embedded feature selection was found to be the optimal solution in terms of model accuracy.