Time-series generative adversarial networks for flood forecasting

Abstract

Reliable flood forecasting serves as one of the fundamental tasks in flood management. However, the forecast performance of modern data-driven models depends heavily on the quantity and quality of training data. The lack of field data of flood events highly underpins the development of machine learning (ML) flood forecasting techniques. With consideration of the rareness of flood events and the high dimensionality of flood time series, two latest variants of generative adversarial networks (GANs), Time-series Generative Adversarial Network (TimeGAN) and Real-world Time Series Generative Adversarial Network (RTSGAN), were applied to generate synthetic flood timeseries. In the example analysis of the Xijiang River Basin in southern China, results show that time-series GANs can accurately and efficiently mimic the spatiotemporal correlations between flood series from multiple sites, and RTSGAN outperforms TimeGAN especially when the lengths of time series are long. Additionally, gradient boosting regression tree (GBRT), long-short term memory (LSTM) and quantile regression method integrated LSTM networks (QRLSTM) flood forecasting models were trained with real sequences and an extra amount of synthetic sequences. It shows that the expansion of training datasets in this way is promising in reducing prediction errors of classical machine learning models. In the 3-day-ahead flood forecast, the introduction of synthetic training datasets has little positive effect on both GBRT and LSTM. In 24-hour-ahead flood forecast, as the lead period increases, GBRT-RTSGAN outperforms GBRT evidenced by a 6.7% increase in Nash-Sutcliffe efficiency (NSE) metric, a 56.5% reduction in average absolute relative error (AARE) metric and a 33.1% reduction in root-mean square error (RMSE) metric. For interval forecast, the use of synthetic training datasets helps QRLSTM obtain lower prediction interval normalized averaged width (PINAW), however, at the cost of prediction interval coverage probability (PICP).