Prediction of flow characteristics using multiple regression and neural networks: A case study in Zimbabwe

Abstract

The feasibility of predicting flow characteristics from basin descriptors using multiple regression and neural networks has been investigated on 52 basins in Zimbabwe. Flow characteristics considered were average annual runoff, base flow index, flow duration curve, and average monthly runoff . Mean annual runoff is predicted using linear equations from mean annual precipitation, basin slope, and proportion of a basin underlain by granite and gneiss. A multiple regression equation is derived to predict the base flow index from mean annual precipitation, slope, and proportion of a basin with grasslands. Findings indicate that a neural network predicts the base flow index with comparable accuracy to multiple regression. Differences in lithology and land cover type were not significant in explaining the base flow index. An exponential model was able to describe flow duration curves, and coefficients of this model could be predicted from the base flow index. Best predictions of flow duration curves were made by a neural network from base flow index, slope, and proportion of a basin with grasslands. The distribution of mean annual runoff into monthly flows was predicted by a neural network from base flow index, slope, and proportion of a basin with grasslands. The study found the base flow index to be important for predicting flow characteristics, and recommends studies aimed at improving prediction of the base flow index.