Selecting input variables for HPP reservoir water inflow forecasting using mutual information

In this paper a new algorithm for selecting input variables for artificial neural network (ANN) based water inflow forecasting model using the information theoretic concept of mutual information is presented. Operation of hydropower plant (HPP) is directly connected with current water inflow, especially in case of run of river HPP with relatively small accumulation capacity. Water inflow forecasting depends on considerable amount of factors thus creating a very large set of potential input variables. Unknown time delay between precipitation time and related influence on water inflow makes the input variable set even bigger. While it is essential that the information contained in the input set is sufficient to determine the neural network output, the presence of too many input features can lead to poor generalization performance. Algorithm presented in this paper reduces the dimension of input variable set by eliminating variables with low information content. The performance of proposed algorithm is tested on the specific case of Soca river CHS and some results are presented as well