Search

Efficient and accurate streamflow predictions are important for urban water management. Data-driven models, especially neural network (NN) models can predict streamflow fast, while the results are uncertain in some complex river systems. Physically based models can reveal the underlying physics, but it is relatively slow and computationally costly. This work focuses on evaluating the reliability of three NN models (artificial neural networks (ANN), long short-term memory networks (LSTM), adaptive neuro-fuzzy inference system (ANFIS)) and one physically based model (SOBEK) in terms of efficiency and accuracy for average and peak streamflow simulation. All the models are applied for a tidal river and a mountainous river in Shenzhen. The results show that, the ANN model calculates fastest since the hidden layer's structure is simple. The LSTM model is reliable in average streamflow simulation in tidal river with the lowest bias while the ANFIS model has the best accuracy for peak streamflow simulation. Furthermore, the SOBEK model shows reliability in simulating average and peak streamflow in mountainous river due to its ability to capture uneven spatial rainfall in the area. Overall, the results indicate that the LSTM model can be a helpful supplementary to physically based models in streamflow simulation of complex urban river systems, by giving fast streamflow predictions with usually acceptable accuracy. Our results can provide helpful information for hydrological engineers in the application of flooding early warning and emergency preparedness in the context of flooding risk management. • Accurate and efficient streamflow simulations are important for urban flooding management. • Neural network models have better performance than physically based models in terms of urban flooding prediction time. • The LSTM model has the best performance in tidal river basin. • The SOBEK model performs better than neural network models in mountainous river basin. • The ANFIS model is a better choice for quick flood events prediction. [ABSTRACT FROM AUTHOR]

• Machine learning (ML) models can determine the rainfall flooding threshold as a line. projected in a plane spanned by two principal components, thereby providing a binary result (flood or no flood). • Compared to the conventional critical rainfall curve, the proposed models, especially the subspace discriminant analysis, greatly raise the accuracy (ACC) to 96.5% and lowering the false alert rate to 25%. • Rainfall threshold based flood prediction can be executed rapidly and simply, this method allows decision makers time for a high-level assessment of flood risk, providing valuable lead time for citizens in the flood-prone areas to be warned. Urban pluvial flooding is a threatening natural hazard in urban areas all over the world, especially in recent years given its increasing frequency of occurrence. In order to prevent flood occurrence and mitigate the subsequent aftermath, urban water managers aim to predict precipitation characteristics, including peak intensity, arrival time and duration, so that they can further warn inhabitants in risky areas and take emergency actions when forecasting a pluvial flood. Previous studies that dealt with the prediction of urban pluvial flooding are mainly based on hydrological or hydraulic models, requiring a large volume of data for simulation accuracy. These methods are computationally expensive. Using a rainfall threshold to predict flooding based on a data-driven approach can decrease the computational complexity to a great extent. In order to prepare cities for frequent pluvial flood events – especially in the future climate – this paper uses a rainfall threshold for classifying flood vs. non-flood events, based on machine learning (ML) approaches, applied to a case study of Shenzhen city in China. In doing so, ML models can determine several rainfall threshold lines projected in a plane spanned by two principal components, which provides a binary result (flood or no flood). Compared to the conventional critical rainfall curve, the proposed models, especially the subspace discriminant analysis, can classify flooding and non-flooding by different combinations of multiple-resolution rainfall intensities, greatly raising the accuracy to 96.5% and lowering the false alert rate to 25%. Compared to the conventional model, the critical indices of accuracy and true positive rate (TPR) were 5%-15% higher in ML models. Such models are applicable to other urban catchments as well. The results are expected to be used to assist early warning systems and provide rational information for contingency and emergency planning. [ABSTRACT FROM AUTHOR]