Training artificial neural networks to perform rainfall disaggregation

Hydrologists and engineers need methods to disaggregate hourly rainfall data into subhourly increments for many hydrologic and hydraulic engineering applications. In the present engineering environment where time efficiency and cost effectiveness are paramount characteristics of engineering tools, disaggregation techniques must be practical and accurate. One particularly attractive technique for disaggregating long-term hourly rainfall records into subhourly increments involves the use of artificial neural networks (ANNs). A past investigation of ANN rainfall disaggregation models indicated that although ANNs can be applied effectively there are several considerations concerning the characteristics of the ANN model and the training methods employed. The research presented in this paper evaluated the influence on performance of several ANN model characteristics and training issues including data standardization, geographic location of training data, quantity of training data, number of training iterations, and the number of hidden neurons in the ANN. Results from this study suggest that data from rainfall-gauging stations within several hundred kilometers of the station to be disaggregated are adequate for training the ANN rainfall disaggregation model. Further, we found the number of training iterations, the limits of data standardization, the number of training data sets, and the number of hidden neurons in the ANN to exhibit varying degrees of influence over the ANN model performance.