Rapid prediction method for human-vehicle risk based on numerical simulation and Deep Learning algorithm

The safety of people and property is seriously threatened by urban floods brought on by severe rains. Flood disaster prevention and mitigation greatly depend on quick and accurate risk forecasts and early warning. However, because of their computational complexity and lengthy computation durations, physically based numerical models struggle to meet the timeliness requirements of flood forecasting. The present study provides a swift predictive strategy for human-vehicle risk utilizing a numerical model and deep learning techniques. The method initially employs a numerical model to simulate water depth and flow velocity across various rainfall scenarios. Subsequently, it assesses risk levels with a human-vehicle instability algorithm to provide associated flood risk maps. A Fully Connected Neural Network-Convolutional Neural Network (FCNN-CNN) model is developed and trained utilizing these risk maps. The architecture and hyperparameters of the FCNN-CNN model are continuously refined to satisfy forecasting demands, yielding an expedited forecasting model. The proposed model possesses the following advantages: The training data for the deep learning model originate from numerical simulations of physical processes, ensuring high dependability, purity, and a enough volume of data. The model thoroughly evaluates the effects of water depth and flow velocity on danger levels, explicitly delineating the prediction targets and risk assessment processes. Conversely, converting the prediction goal from regression to classification diminishes the complexity and hardware requirements of the deep learning model, thus improving prediction efficiency. The FCNN-CNN model considers the two-dimensional characteristics of the risk maps, hence enhancing the model's predictive accuracy and efficiency. The findings indicate that the developed prediction model can precisely forecast flood-induced human-vehicle instability risks for adults, children, SUVs, and automobiles across different rainfall intensities, achieving prediction accuracies of 0.9213, 0.9366, 0.9837, and 0.9730, respectively. The mean prediction duration for each risk map is 0.27 seconds, illustrating the model's robust accuracy and promptness.