Interpretable Analysis of Rainfall-Runoff Forecasting Using MLP and Perturbation-based Approach in the Sisaony River, Madagascar

Abstract

This study examines the performance of a rainfall-runoff model based on a Multi-Layer Perceptron (MLP), supplemented by an Explainable Artificial Intelligence (XAI) analysis using a perturbation-based approach. The MLP model predicts the hourly instantaneous discharge of the Sisaony River in Madagascar based on precipitation, two estimates of potential evapotranspiration (PET), and past discharge data, for forecast horizons of 24 hours, 18 hours, 12 hours, 9 hours, 6 hours, or 3 hours. The results are compared to forecasts obtained from a conventional GRP (Génie Rural pour la Prévision de crue)-type rainfall-runoff model and show that the MLP performs well, particularly for longer horizons. To address the issue of interpretability, a perturbation-based explainability approach was applied by replacing input variables with zero or their mean. The analysis reveals that discharge is the most influential variable, confirming the strong autoregressive behavior of the system. Finally, the study demonstrates that combining deep learning models with explainability techniques provides both strong predictive performance and improved understanding of model behavior, offering a promising approach for flood forecasting and risk management in data-limited regions.

Keywords: Multi-Layer Perceptron (MLP), Explainable Artificial Intelligence (XAI), Perturbation-Based Approach, Rainfall–Runoff, Sisaony River

DOI: 10.54941/ahfe1007279