Probabilistic multi-step ahead streamflow forecast based on deep learning

Abstract The use of deep learning methods for fluvial flood forecasting is rapidly gaining traction, offering a promising solution to the challenges associated with accurate yet time-consuming numerical models. This paper presents two physics-inspired deep learning approaches specifically designed for fluvial flood forecasting, each embracing different learning principles: centralized and federated learning. The centralized model utilizes an Encoder-Decoder technique to handle input data of varying types and scales, while the federated model is based on a node-link graph with a seq2seq internal model. Both models are enhanced with a probabilistic forecasting head to account for the inherent uncertainty in streamflow forecasts. The objective of these approaches is to address the limitations of traditional numerical models while leveraging the potential of deep learning to improve the speed, accuracy, and scalability of flood forecasting. To validate their effectiveness, the models were tested across different use cases. The findings from the federated learning approach emphasize the importance of catchment clustering before model utilization and demonstrate the models’ ability to generalize effectively in catchments with similar properties. On the other hand, the results of the centralized method highlight the model’s reliance on the test set falling within the data range of the training set (Average NSE and KGE for the sixth hour ahead of 0.88 and 0.78, respectively). To address this limitation, the paper suggests the development of a method for the future, such as leveraging a numerical model or using Generative Adversarial Networks, to generate highly extreme events, particularly in the context of a changing climate. The models are implemented in a flexible operational framework based on open standards, ensuring their adaptability and usability in various settings.