Sensitivity Based Feature Selection for Recurrent Neural Network Applied to Forecasting of Heating Gas Consumption

The paper demonstrates the importance of feature selection for recurrent neural network applied to problem of one hour ahead forecasting of gas consumption for office building heating. Although the accuracy of the forecasting is similar for both the feed-forward and the recurrent network, the removal of features leads to accuracy reduction much earlier for the feed-forward network. The recurrent network can perform well even with 50% of features. This brings significant benefits in scenarios, where the neural network is used as a blackbox model of building consumption, which is called by an optimizer that minimizes the consumption. The reduction of input dimensionality leads to reduction of costs related to measurement equipment, but also costs related to data transfer.