Flood Forecast and Control for Urban Rivers Using LSTM Neural-Network

To make better prediction and control of river water navigation in a Danish city - Vejle, the Long-Short-Term-Memory (LSTM) neural-network model is adopted to predict the water-level nearby a high flooding-risk area using correlated historical data. A set of feedback control solutions are developed based on the extension of the obtained LSTM model to automatically regulate a distribution-gate system, which guides the coming stream-flow into separated urban rivers. The proposed control solutions are tested in simulation based on four historic events, and it can be observed that two floods at the critical areas since 2017 could have been prevented by balancing flow-splits using automatic feedback control, which was manually controlled in the past. This study demonstrates a clear and promising potential to use machine learning methods for supporting development of smart cities and their climate adaption strategies.