A novel multi-step ahead forecasting model for flood based on time residual LSTM

Abstract

Climate change has significantly impacted hydrology, including extreme precipitation, and changing precipitation patterns that could lead to an increase in flooding. Life and property benefit from accurate and reliable multi-step flood forecasting. Recently, Recurrent Neural Network (RNN) have become increasingly popular among hydrology researchers for their ability to capture historical dependencies, simplify computations by ignoring intermediate hydrological processes, and provide higher prediction accuracy than traditional models. However, RNN-based flood prediction models face two significant challenges. Firstly, due to their strict time-serial, RNN suffer from gradient issues such as vanishing and exploding gradients, which can make training RNN models difficult. To address this issue, we propose a Residual Long Short-Term Memory (ResLSTM) model that incorporates time residual connections into the time connections of LSTM. Secondly, most flood prediction models output a deterministic value, but the natural hydrological characteristics of the basin are a nonlinear and complex system with many influencing factors that have some randomness. This requires the use of probabilistic methods to modeling. Thus, we introduce the probabilistic forecasting model Autoregressive Recurrent Networks (DeepAR) into our flood prediction model, which outputs a prediction interval rather than a deterministic value. Then, we build four flood probability prediction models by combining DeepAR and four enhanced RNN, including ResLSTM (ours), Long Short-Term Memory (LSTM), Gate Recurrent Unit (GRU), and Time Feedforward Connections Simple Gate Recurrent Unit (TFC-SGRU). The performance of these models is evaluated by the long-term hydrologic data of the Passaic and Ramapo River basins in the United States. The results demonstrate that the prediction interval of the four models is more adaptive to flood uncertainties. And the accuracy of peak flow prediction is nearly 100% within a 90% prediction probability interval. The temporal residual-based model is more accurate and robust than the original LSTM and GRU. We believe this study fills a research gap in multi-step-ahead flood probability prediction and improves the accuracy and reliability of flood prediction models.