Learning hydrologic flow separation algorithm and local ANN committee modeling

Learning models are used in hydrologic forecasting more and more often. Natural phenomena are, however, multi-stationary and are composed of a number of interacting processes. Their modeling often assumes the existence of one single model handling all processes, which often suffers from inaccuracies. A solution is to model various processes separately by different models optimized to represent every single process, and to merge them in a committee. In this paper two approaches are considered. In the first one the flow separation algorithm based on hydrologic domain knowledge is used to partition the data and train separate ANN models. To be used in operation, such algorithm has to be also learned by a classifier. In the second approach the data is partitioned by a clustering algorithm. The validity of the approach is demonstrated on realistic case studies. In order to obtain optimal modularity, the models accuracy is evaluated based on several parameters. The modular models showed better performance than models trained on the whole data set almost in all cases.