A comprehensive comparison of four input variable selection methods for artificial neural network flow forecasting models

Abstract

Artificial neural networks (ANNs) are increasingly used for flood forecasting. The performance of these models relies on the selection of appropriate inputs. However, Input Variable Selection (IVS) is typically performed using expert knowledge or simple linear methods. This research compares and evaluates four IVS methods including two model-free methods: partial correlation (PC), partial mutual information (PMI), and two novel model-based methods: an improved input omission (IO), and improved combined neural pathway strength (CNPS). A comprehensive comparison of performance efficacy for multiple IVS methods has not been published in literature before. Each method is used for daily and hourly lead times in the Bow and Don Rivers (both in Canada), respectively. These watersheds represent different hydrological systems and were selected to highlight the performance of the IVS methods under differing conditions. This research determines that the proposed CNPS produces the strongest performing ANNs based on the robustness of the inputs selected, comparison to other IVS methods, and models developed without IVS. Additionally, this research demonstrates that standard termination criteria do not reliably identify the optimum number of inputs for the ANNs and using a model-based optimization of inputs is recommended. As a result, it is recommended that the number of inputs be determined using a systematic approach, where each input selection is informed by an IVS-based input ranking, rather than a predefined termination criterion. Lastly, this research demonstrates that input usefulness is not binary concept; the correct number of selected inputs is dependant on the desired model complexity, instead of an arbitrarily selected IVS termination criteria.