This paper proposes a novel approach to enhance flood forecasting models by using Generative Adversarial Networks (GANs) to generate synthetic flood events. A modified time-series GAN incorporates mass conservation, energy balance, and hydraulic principles through regularization terms in the loss function and mass conservative LSTMs. PCA and t-SNE analyze the synthetic data. The generated synthetic data augments the original data to train a probabilistic neural runoff model for multi-step ahead flood event forecasting. Results show statistically significant improvement (9.8% NSE) in multistep-ahead predictions using the augmented dataset compared to using only original data. While reducing Prediction Interval Normalized Average Width (PINAW), the model also trades off Prediction Interval Coverage Probability (PICP).

Divas Karimanzira