Using pruning algorithms and genetic algorithms to optimise network architectures and forecasting inputs in a neural network rainfall-runoff model

Abstract

Four design tool procedures are examined to create improved neural network architectures for forecasting runoff from a small catchment. Different algorithms are used to remove nodes and connections so as to produce an optimised forecasting model, thereby reducing computational expense without loss in performance. The results also highlight issues in selecting analytical methods to compare outputs from different forecasting procedures.