PERFORMANCE EVALUATION OF MLP AND CNN FOR RESERVOIR OUTFLOW FORECASTING

Abstract

Flood forecasting is critical for downstream disaster risk mitigation. Traditional hydrological models often fail to capture the nonlinear, dynamic interactions between inflows, storage levels, and operational reservoir releases during extreme events. This study evaluates Multilayer Perceptron (MLP) and Convolutional Neural Network (CNN) deep learning architectures for hourly reservoir outflow forecasts during the catastrophic 2006 Ukai flood. Using chronological data splits and sliding-window input sequences, model performance was rigorously assessed via MAE, RMSE, and R². The MLP consistently outperformed the CNN across training, validation, and testing phases. On unseen test data, the MLP achieved MAE = 49,575 cusecs, RMSE = 96,193 cusecs, and R² = 0.635, compared to CNN values of 57,386 cusecs, 103,645 cusecs, and 0.576, respectively. Validation metrics similarly favored MLP (MAE 35,880 versus 42,402 cusecs; R² 0.67 versus 0.61). Time-series and scatter analyses confirmed MLP's superior stability and peak-capture capability, whereas CNN exhibited systematic underestimation and lagged responses during rapid flow variations. These findings suggest that simpler fully connected architectures provide robust, operationally practical solutions for real-time flood forecasts under limited datasets, while CNN may require expanded data or hybrid frameworks for improved reliability.