Flood prediction for the Vaal River Basin (South Africa) using machine learning techniques

Abstract

Flood prediction is a critical tool for disaster management and risk mitigation. Machine learning models are viable alternatives to the traditional techniques of flood prediction and analysis, which often fail in capturing the complex nonlinear relationship among meteorological parameters. This study evaluated the performance of an artificial neural network (ANN) to predict the flooding indicator (surface volume) in the Vaal River Basin using the key meteorological parameters: historical records of rainfall, wind speed, humidity, and maximum temperature. A 30-year (1994–2024) dataset was collected from the South African Weather Service and preprocessed using standard techniques. Hyper-parameter optimisation of the models was carried out using a grid-search method. The ANN model was developed by testing different topologies, training algorithms and activation functions at both the hidden and output layers. The performance of the models was evaluated using relevant statistical metrics, namely root mean square error (RMSE), mean absolute percentage error (MAPE), mean absolute error (MAE), value accounted for (VAE) and R-value. The ANN model with tansigtansig activation function and Levenberg Marquardt training algorithms outperformed other architectures with RMSE of 6.245, MAPE of 25.95%, MAE of 4.656, VAE of 7.843 and R-value of 0.823 at the training. This research demonstrated the viability of machine learning-based flooding predictions based on weather variables, contributing to flood risk management strategies.