Learning energy demand domain knowledge via feature transformation

Domain knowledge is an essential factor for forecasting energy demand. This paper introduces a method that incorporates machine learning techniques to learn domain knowledge by transforming the input features. Our approach divides the inputs into subsets and then searches for the best machine learning technique for transforming each subset of inputs. Preprocessing of the inputs is not required in our approach because the machine learning techniques appropriately transform the inputs. Hence, this technique is capable of learning where nonlinear transformations of the inputs are needed. We show that the learned data transformations correspond to energy forecasting domain knowledge. Transformed subsets of the inputs are combined using ensemble regression, and the final forecasted value is obtained. Our approach is tested with natural gas and electricity demand signals. Experimental results show how this method can learn domain knowledge, which yields improved forecasts.