Natural phenomena are multistationary and are composed of a number of interacting processes, so one single model handling all processes often suffers from inaccuracies. A solution is to partition data in relation to such processes using the available domain knowledge or expert judgment, to train separate models for each of the processes, and to merge them in a modular model (committee). In this paper a problem of water flow forecast in watershed hydrology is considered where the flow process can be presented as consisting of two subprocesses -- base flow and excess flow, so that these two processes can be separated. Several approaches to data separation techniques are studied. Two case studies with different forecast horizons are considered. Parameters of the algorithms responsible for data partitioning are optimized using genetic algorithms and global pattern search. It was found that modularization of ANN models using domain knowledge makes models more accurate, if compared with a global model trained on the whole data set, especially when forecast horizon (and hence the complexity of the modelled processes) is increased.