Neural networks for water systems analysis: from fundamentals to complex pattern recognition

摘要

Accurate river flows are crucial for effective water resource management. However, estimating flows in ungauged rivers, particularly those in difficult to access terrains, is a challenging problem for water scientists and managers. As a solution, hydrological regionalisation (HR) has been proposed to estimate river flows based on proxy-basin, interpolation and regression methods. Recently, neural networks have been shown to produce improved estimates. In this study, HR-based artificial neural networks (ANN) models were developed for estimating monthly flows in ungauged rivers in New Zealand using hydrological and geomorphological attributes. After rigorous input selection, multilayer perceptron (MLP) networks were first developed by trial and error. Then, a new MLP method, not involving trial and error, was developed by clustering the correlated hidden neurons in a trained MLP to simplify the model structure; this produced overall better results than the trial-and-error MLP and a genetic algorithm optimised MLP. Results show that accurate and parsimonious MLP models can be developed for flow estimation based on HR using the new method. Therefore, the study presents the hydrological community with improved neural networks tools based on HR to estimate flows in ungauged rivers for more effective water management.