River Flow Forecasting Using an Improved Artificial Neural Network

Artificial neural network (ANN) is a popular data-driven modelling technique that has found application in river flow forecasting over the last two decades. This can be attributed to its ability to assimilate complex and nonlinear input-output relationships inherent in hydrological processes within a river catchment. However despite its prominence, ANNs are still prone to certain problems such as overfitting and over-parameterization, especially when used under limited availability of datasets. These problems often influence the predictive ability of ANN-derived models, with inaccurate and unreliable results as resultant effects. This paper presents a study aimed at finding a solution to the aforementioned problems. Two evolutionary computational techniques namely differential evolution (DE) and genetic programming (GP) were applied to forecast monthly flow in the upper Mkomazi River, South Africa using a 19-year baseline record. Two case studies were considered. Case study 1 involved the use of correlation analysis in selecting input variables during model development while using DE algorithm for optimization purposes. However in the second case study, GP was incorporated as a screening tool for determining the dimensionality of the ANN models, while the learning process was subjected to sensitivity analysis using the DE-algorithm. Results from the two case studies were evaluated comparatively using three standard model evaluation criteria. It was found that results from case study 1 were considerably plagued by the problems of overfitting and over-parameterization, as significant differences were observed in the error estimates and R2 values between the training and validation phases. However, results from case study 2 showed great improvement, as the overfitting and memorization problems were significantly minimized, thus leading to improved forecast accuracy of the ANN models. It was concluded that the conjunctive use of GP and DE can be used to improve the performance of ANNs, especially when availability of datasets is limited.