Forecasting Flood Magnitudes and Annual Exceedance Probabilities (2015–2052) Using ANN and Gumbel Distribution

Abstract

The purpose of this article is to forecast flood magnitudes and their respective P-per cent annual exceedance probabilities in the Gucha-Migori River Basin for the period 2015–2052 to support flood risk management and early warning planning. Historical river discharge and precipitation data were used to develop and validate forecasting models. An artificial neural network (ANN) with a 1-20-1 network topology, where the single input represents hydrological data, the twenty hidden neurons process patterns, and the single output represents predicted flood magnitude, was trained and validated to predict future flood events, while the Gumbel distribution model was applied to estimate annual exceedance probabilities. Model performance was evaluated using correlation coefficients and error metrics. The ANN effectively captured non-linear flood dynamics, with minimum R values of 9.20×10?¹, 9.27×10?¹, and 9.15×10?¹ for training, validation, and testing, respectively. Forecasted maximum flood magnitudes for successive five-year periods from 2015 to 2052 ranged from 299 to 502 m³/s. Corresponding annual non-exceedance probabilities derived from Gumbel’s distribution varied between 64 per cent and 99.6 per cent. These results indicate a significant variability in future flood magnitudes, emphasising the need for proactive flood management strategies. The study provides critical information for decision-makers, enabling the development of flood response plans, early warning systems, and preparedness measures in the Gucha-Migori River Basin.