Research And Application Of Urban Waterlogging Depth Prediction Method Based On Stacking Ensemble Learning

With the acceleration of urbanization,some cities’ drainage system and underground pipe networks construction have failed to keep up with the pace of urban development in a timely and effective manner.Once it is hit by heavy rainfall or continuous rainfall,it is easy to cause urban waterlogging disasters.This paper conducts an in-depth analysis of the problem of urban waterlogging depth prediction,and proposes a solution based on the historical rainfall,water accumulation data and urban topographic hydrological data in Ningbo City,combined with deep learning and machine learning.(1)This thesis proposes a waterlogging depth forecasting method based on multilayer stacking ensemble structure and quasi-attention,to solve the insufficient feature extraction and multi-collinearity problems among multiple data sources in existing flood scenarios.This method is based on a multi-layer stacking ensemble learning structure,in which three heterogeneous models with high accuracy are used as base models in the first layer to extract features of the original data through K-fold crossvalidation.Then,the generated new features are added to the second layer model for feature fusion and final prediction.Meanwhile,a weight scaling feature processing method is designed based on the quasi-attention idea during the feature generation process,enhancing the features with high prediction accuracy and time-series weight,and weakening the features that are relatively weak,thus reducing the collinearity degree of new features.Experimental results demonstrate that this method can achieve good performance in flood prediction tasks.(2)This thesis proposes a spatiotemporal prediction method based on spatial attention mechanism and STL-LSTM,to mitigate the spatial influence among multivariable time-series data and the impact of periodic and fluctuating factors of flood time-series data on prediction results.This method uses the STL time-series decomposition algorithm to separate the seasonal sequence,trend sequence and residual sequence from the flood time-series data.Then,the Attention-LSTM network predicts the three decomposition sequences respectively and eventually weighted sum to obtain the final prediction result.In this method,the Attention mechanism is used to allocate spatial weights to multivariable time-series data in the spatial dimension,dynamically allocating space weights to the input features in a single time step of LSTM.Experimental results demonstrate that this method achieves good performance in flood prediction tasks based on multivariable time-series data.(3)Design and implementation of intelligent forecasting and warning system for urban waterlogging depth.Based on the prediction model of urban waterlogging depth proposed in this paper,an urban waterlogging depth prediction and early warning system is designed and implemented.The system takes the prediction model as the core,provides real-time monitoring and prediction of the waterlogging depth in the city,and realizes visualization on the Web.