Urban inland inundation rapid prediction method based on deep learning

Abstract

The invention relates to the technical field of urban inland inundation rapid prediction, and discloses an urban inland inundation rapid prediction method based on deep learning, and the method comprises the following steps: S1, collecting historical meteorological data, landform data, urban drainage system data and historical inland inundation event data; s2, performing data cleaning on the collected data, removing noise, filling missing values, and processing abnormal values; s3, constructing a deep learning model architecture; when urban inland inundation risk prediction is carried out, multi-source heterogeneous data are integrated and standardized, and a unified spatial-temporal characteristic analysis framework is constructed, so that the system can eliminate magnitude differences of weather, terrain and drainage system data, and data comparability of different regions is ensured; and meanwhile, dynamic feature extraction is performed on real-time rainfall data by using a deep learning model, an abnormal fluctuation rule of meteorological elements is identified, the waterlogging risk pre-judgment capability of extreme weather events is improved, and the stability and credibility of a prediction result are enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of urban waterlogging rapid prediction, in particular to a kind of urban waterlogging rapid prediction method based on deep learning. BACKGROUND

[0002] With the acceleration of global climate change and urbanization, urban waterlogging problem is increasingly serious.Waterlogging not only affects urban traffic and residents' life, but also can cause huge economic losses and casualties.Therefore, it is very important to predict and warn urban waterlogging quickly and accurately for ensuring urban safety and residents' life.

[0003] Currently, since urban waterlogging prediction involves multi-source heterogeneous data, when real-time waterlogging risk assessment is carried out, the deep learning model that depends on meteorological data and topographic data cannot dynamically perceive the change of drainage system topology caused by urban road construction, and when the key drainage node is migrated, the model spatial feature extraction will fail, resulting in the decrease of waterlogging prediction accuracy.

[0004] Therefore, the present application provides a kind of urban waterlogging rapid prediction method based on deep learning to solve the above problems. SUMMARY

[0005] In view of the shortcomings of the prior art, the present application provides an urban waterlogging rapid prediction method based on deep learning, which solves the problem of model spatial feature extraction failure in the background art, resulting in the decrease of waterlogging prediction accuracy.

[0006] To achieve the above purpose, the present application provides the following technical solutions: An urban waterlogging rapid prediction method based on deep learning, the method comprising the following steps: S1, collect historical meteorological data, topographic data, urban drainage system data and historical waterlogging event data; S2, data cleaning is performed on the collected data to remove noise, fill in missing values and process abnormal values; S3, a deep learning model architecture is constructed, including an input layer, a hidden layer and an output layer, wherein the hidden layer includes an LSTM network branch for processing time series data to capture the change rule of meteorological elements over time, a CNN network branch for processing spatial data to extract the spatial features of topography and urban drainage system, and a fusion layer for integrating the outputs of LSTM and CNN branches; S4, initialize model parameters and perform forward propagation calculation; S5, calculate the loss between model output and true value; S6, perform back propagation, calculate gradient and update model parameters; S7. Iteratively train the model by repeating the forward propagation, loss calculation, and back propagation steps until the model converges. S8. Input real-time meteorological data, topographic data, and urban drainage system data into the trained model, and output the probability and water depth range of urban flooding in the future.

[0007] Preferably, step S1 includes the following steps: S11. Connect the meteorological bureau and meteorological station data sources through the meteorological data collection component to obtain historical meteorological data, including rainfall, wind speed, wind direction and temperature, and store the extracted meteorological elements in the local database; S12. Connect to the GIS system through the topographic data collection component to obtain topographic data, including elevation, slope and aspect, and process the extracted topographic features before storing them in the local database. S13. Connect to the drainage management department's database through the urban drainage system data collection component to obtain drainage system data, including drainage pipe distribution and drainage capacity, and perform consistency processing on the extracted drainage elements. S14. Connect to the emergency management department's database through the historical flooding event data collection component to obtain historical flooding event data, including the time, location, extent, and water depth of flooding, and form a historical sample database; S15. Through data integration and management components, meteorological data, topographic data, drainage system data, and historical flooding event datasets are transformed into a structured dataset in a unified format, and data backup and access control are performed.

[0008] Preferably, step S2 includes the following steps: S21. Fill in the missing data using the mean fill, median fill, and forward fill methods to ensure data continuity; S22. Identify and process outliers in the data, use the 3σ principle and box plot method to detect outliers, and correct and delete them; S23. Smooth random noise by using moving average filtering and Gaussian filtering algorithms to eliminate data fluctuations. S24. Normalize the data to scale the features of different dimensions to a uniform numerical range using the minimum-maximum normalization formula. S25. Convert text data into numerical vectors and use one-hot encoding and embedding layer technology to ensure that the data is suitable for model input.

[0009] Preferably, step S3 includes the following steps: S31. The design input layer receives preprocessed multidimensional data, including time-series meteorological data and spatial topographic data; S32. Construct an LSTM network branch containing at least two layers of LSTM units, with 50-100 units per layer, to capture the changing patterns of rainfall and wind speed over time. S33. Construct a CNN network branch containing convolutional layers, pooling layers, and activation layers to extract spatial distribution features of elevation, slope, and drainage pipes. S34. Design a fusion layer, using a fully connected layer and attention mechanism to weightedly fuse the temporal feature vector output by the LSTM branch and the spatial feature vector output by the CNN branch. S35. Construct an output layer to generate the probability value of water flooding and the predicted water depth range. The probability value is output through the Sigmoid function, and the water depth range is output through the linear regression layer.

[0010] Preferably, step S4 includes the following steps: S41. Initialize the model weight parameters using the Xavier and He initialization methods, and set the weight matrix to a random normal distribution. S42. Initialize the bias parameters to a zero vector to ensure the stability of the initial model state; S43. Process the input data in batches, with each batch consisting of 32-128 bytes, and load them into memory; S44. Perform forward propagation computation, processing the time series data through LSTM branches. The calculation formula is as follows: ; in For time step The hidden state, Input data; S45. Spatial data is processed through CNN branches, and the calculation formula is as follows: ; in For feature maps, For elevation data, For convolution kernel, For bias.

[0011] Preferably, step S5 includes the following steps: S51. Calculate the predicted output of the model, using the following formula: ; in For predicted values, For input data, For model parameters, For model functions; S52. Using the mean squared error (MSE) as the loss function, the calculation formula is as follows: ; in This is the loss value. For the sample size, Let be the predicted value for the i-th sample. This represents the true value of the i-th sample. S53. Use cross-entropy loss for classification tasks to help calculate the confidence level of the probability of waterlogging; S54. Use the mean absolute error (MAE) for the regression task to help optimize the water depth range prediction. S55. Calculate the total loss as the weighted sum of MSE and auxiliary loss, with a weight ratio of 0.7:0.3.

[0012] Preferably, step S6 includes the following steps: S61. Calculate the gradient of the loss function with respect to the model parameters using automatic differentiation techniques; S62. Update the parameters using the Adam optimization algorithm with adaptive moment estimation. The formula is: ; in For learning rate, For gradient operators, The value of the loss function. These are model parameters; S63. Set the initial learning rate to 0.001 and adopt a learning rate decay strategy, reducing it by 10% every 10 epochs; S64. Apply gradient clipping techniques to limit the gradient norm to below 5.0 to prevent training instability; S65. Update the parameters of the fusion layer and backpropagate the gradients of the LSTM and CNN branches using the chain rule.

[0013] Preferably, step S7 includes the following steps: S71. Set the maximum number of iterations T to 100-500 and define an early stopping mechanism to terminate training when the validation set loss does not decrease for 5 consecutive iterations. S72. After each batch of training, calculate the loss of the training set and the validation set, and record the loss curve; S73. Apply Dropout regularization technique, setting the dropout rate to 0.2-0.5 in LSTM and CNN layers; S74. Use batch normalization layers to accelerate convergence and reduce internal covariate shift. S75. When the loss function converges to the preset threshold and reaches the maximum number of iterations, save the optimal model parameters.

[0014] Preferably, step S8 includes the following steps: S81. Receive real-time meteorological data streams, including radar rainfall forecasts and real-time wind speed monitoring data; S82. Load the preprocessing module to normalize and filter noise in real-time data; S83. Input real-time data into the trained model and perform inference calculations; S84. Output the probability value of waterlogging. The probability threshold is set to 0.7. If the probability is exceeded, an early warning will be triggered. S85 outputs predicted water depth range with an accuracy of ±0.1 meters and generates a risk heat map based on a geographic grid.

[0015] Preferably, the method further includes the following steps: S101: Deploy the model to the cloud server to support high-concurrency real-time prediction; S102. Regularly update training data and incrementally train the model by combining data from new urban flooding events. S103. Apply transfer learning techniques to adapt the pre-trained model to different urban areas; S104. Integrate a multi-source data feedback mechanism to dynamically adjust the weights of the fusion layer based on the prediction error; S105. Generate a prediction report, including a probability distribution map of urban flooding, a spatiotemporal variation curve of water depth, and a confidence analysis.

[0016] Beneficial effects Compared with existing technologies, this invention provides a rapid urban flooding prediction method based on deep learning, which has the following beneficial effects: 1. In this invention, when predicting urban flooding risk, a unified spatiotemporal feature analysis framework is constructed by integrating multi-source heterogeneous data and standardizing them. This enables the system to eliminate the magnitude differences between meteorological, topographical, and drainage system data, ensuring the comparability of data from different regions. At the same time, a deep learning model is used to dynamically extract features from real-time rainfall data, identify abnormal fluctuation patterns of meteorological elements, improve the ability to predict flooding risk during extreme weather events, and enhance the stability and reliability of the prediction results.

[0017] 2. In this invention, during model training and optimization, a dual-branch network architecture that integrates spatiotemporal features is designed to collaboratively extract the temporal variation patterns of meteorological data and the spatial distribution characteristics of terrain data. This enables the model to avoid the risk of misjudgment caused by the fragmentation of spatiotemporal features in traditional methods. Combined with an adaptive parameter update mechanism, the feature fusion weights are dynamically adjusted to ensure that high-risk areas for urban flooding can still be located in complex urban environments, reducing the spatial bias of prediction results. Furthermore, the weight matrix is ​​initialized using a truncated normal distribution based on Xavier and He strategies. This initialization strategy ensures the reliability and efficiency of model training, enabling the deep neural network to quickly converge to a high-performance state, ensuring that the prediction system can immediately provide stable and reliable output results after deployment.

[0018] 3. In this invention, when deploying the prediction system, a cloud-based incremental learning and migration adaptation mechanism is established to continuously integrate new urban flooding event data and optimize model parameters, enabling the system to adapt to the differences in drainage capacity in different urban areas; at the same time, multi-dimensional risk visualization reports are generated, and the emergency management platform is coordinated to carry out graded early warning response, improve the protection efficiency of key areas in old urban areas and low-lying areas, and strengthen the proactive prevention and control capabilities of urban disaster prevention and mitigation. Attached Figure Description

[0019] Figure 1 This is a flowchart of a rapid urban flooding prediction method based on deep learning according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] For specific implementation examples, please refer to: Figure 1 A rapid urban flooding prediction method based on deep learning, comprising the following steps: S1. Collect historical meteorological data, topographic data, urban drainage system data, and historical flooding event data; S2. Perform data cleaning on the collected data, removing noise, filling in missing values, and handling outliers; S3. Construct a deep learning model architecture, including an input layer, a hidden layer, and an output layer. The hidden layer includes an LSTM network branch for processing time series data to capture the changing patterns of meteorological elements over time, a CNN network branch for processing spatial data to extract spatial features of topography and urban drainage systems, and a fusion layer to integrate the outputs of the LSTM and CNN branches. S4. Initialize model parameters and perform forward propagation calculations; S5. Calculate the loss between the model output and the true value; S6. Perform backpropagation, calculate the gradient, and update the model parameters; S7. Iteratively train the model by repeating the forward propagation, loss calculation, and back propagation steps until the model converges. S8. Input real-time meteorological data, topographic data, and urban drainage system data into the trained model, and output the probability and water depth range of urban flooding in the future.

[0022] S1 includes the following steps: S11. Connect the meteorological bureau and meteorological station data sources through the meteorological data collection component to obtain historical meteorological data, including rainfall, wind speed, wind direction and temperature, and store the extracted meteorological elements in the local database; S12. Connect to the GIS system through the topographic data collection component to obtain topographic data, including elevation, slope and aspect, and process the extracted topographic features before storing them in the local database. S13. Connect to the drainage management department's database through the urban drainage system data collection component to obtain drainage system data, including drainage pipe distribution and drainage capacity, and perform consistency processing on the extracted drainage elements. S14. Connect to the emergency management department's database through the historical flooding event data collection component to obtain historical flooding event data, including the time, location, extent, and water depth of flooding, and form a historical sample database; S15. Through data integration and management components, meteorological data, topographic data, drainage system data, and historical flooding event datasets are transformed into a structured dataset in a unified format, and data backup and access control are performed.

[0023] S2 includes the following steps: S21. Fill in the missing data using the mean fill, median fill, and forward fill methods to ensure data continuity; S22. Identify and process outliers in the data, use the 3σ principle and box plot method to detect outliers, and correct and delete them; S23. Smooth random noise by using moving average filtering and Gaussian filtering algorithms to eliminate data fluctuations. S24. Normalize the data, scaling features of different dimensions to a uniform numerical range using the min-max normalization formula, where the normalization formula is: ; in The original data values, These are the normalized data values. The minimum value of the data. The maximum value of the data; S25. Convert text data into numerical vectors, using one-hot encoding and embedding layer techniques to ensure the data is suitable for model input, including the following steps: S251. Identify the category features in text data, including discrete variables such as the location of flooding events and the type of drainage pipes, and perform word segmentation and binning preprocessing to form a standardized category list; S252. Apply one-hot coding technology to convert each category value into a binary vector, where the vector length is equal to the total number of categories, the position corresponding to a specific category is 1, and the rest are 0, to ensure data discretization. S253. For high-dimensional sparse one-hot encoded vectors, embedding layer technology is used to map them to a low-dimensional dense vector space. The dimensionality is reduced by a trainable weight matrix, which improves the computational efficiency of the model. S254. The embedded numerical vector is concatenated and fused with other numerical features to form an input tensor of a unified dimension, ensuring compatibility with the input layer of the deep learning model. S255. Verify the dimension and distribution of the numerical vectors, and ensure that there are no missing or outliers through sampling checks. Finally, output a standardized numerical data stream for the model to use.

[0024] S3 includes the following steps: S31. The design of the input layer to receive preprocessed multidimensional data, including time-series meteorological data and spatial topographic data, is as follows: The input layer uses a multi-channel tensor structure to receive heterogeneous data: Time series data, including rainfall and wind speed, are organized into a two-dimensional matrix using a sliding time window format. ,in The time step is set to 60 minutes by default. The characteristic number; Spatial data, including elevation and slope, is converted into a 100m×100m geographic grid. Each grid contains feature maps with C channels. Data is then accessed via a data interface to obtain radar rainfall data and GIS topographic data from the meteorological bureau, ensuring standardized data input while supporting real-time streaming processing. S32. Construct an LSTM network branch containing at least two layers of LSTM units, with each layer containing 50-100 units, to capture the changing patterns of rainfall and wind speed over time. The implementation method, tailored to practical application scenarios, is as follows: The LSTM branch uses a two-layer stacked structure with 64 and 32 cells respectively, and the time step is... =60, with minute-level rainfall sequences as the core input; For urban flooding prediction scenarios, the model focuses on learning the cumulative effect of short-term heavy rainfall. The hidden state transfer mechanism can remember the temporal dependency between the peak rainfall delay and the drainage response. During training, historical rainstorm event data is used to optimize parameters and enhance the prediction robustness under extreme weather conditions. S33. Construct a CNN network branch containing convolutional layers, pooling layers, and activation layers to extract spatial distribution features of elevation, slope, and drainage pipes. The implementation method, combined with practical application scenarios, is as follows: The CNN branch adopts an encoder structure: the first layer uses 32 3×3 convolutional kernels to extract local terrain features, including depressions and steep slopes; the second layer uses 64 convolutional kernels to capture the topological relationship of the drainage network, including pipe connectivity and catchment area boundaries. For the scenario of insufficient drainage capacity in old urban areas, the model focuses on learning grid areas with elevations below 3m and low pipe density. The ReLU activation function is used to enhance the expression of spatial nonlinear features. The pooling layer uses max pooling 2×2 window to compress the data dimension and retain key features. S34. Design a fusion layer, employing a fully connected layer and attention mechanism, to weightedly fuse the temporal feature vector output from the LSTM branch with the spatial feature vector output from the CNN branch. The specific implementation is as follows: The fusion layer employs a multi-head attention mechanism to dynamically integrate spatiotemporal features. Specifically, four attention heads are set up: two focus on meteorological time-series features, and the other two focus on topographic spatial features. Attention weights are calculated using an additive model, mapping the temporal feature vector output by the LSTM and the spatial feature vector output by the CNN to the same latent space through a trainable parameter matrix, and then calculating the similarity score. ; in The time feature vector output by the LSTM branch. The spatial feature vector output by the CNN branch. For the transpose of the trainable weight vector, To score attention, , is a trainable weight matrix; S35. Construct an output layer to generate the probability value of water flooding and the predicted water depth range. The probability value is output through the Sigmoid function, and the water depth range is output through the linear regression layer.

[0025] S4 includes the following steps: S41. Initialize the model weight parameters using the Xavier and He initialization methods, setting the weight matrix to a random normal distribution. This initialization strategy ensures the reliability and efficiency of model training, enabling the deep neural network to quickly converge to a high-performance state. The specific implementation includes the following steps: S411. Determine the structural parameters of each layer of the neural network, including the input dimension. and output dimensions And define the shape of the weight matrix according to the layer type; S412. For layers using the Siqmoid activation function, apply the Xavier initialization method, starting the weight matrix with a mean of 0 and a standard deviation of... Random sampling is performed within a normal distribution to ensure that the weight variance adapts to the linear interval of the activation function; S413. For layers using the ReLU activation function, apply the He initialization method, starting the weight matrix with a mean of 0 and a standard deviation of... Random sampling is performed within a normal distribution to maintain an appropriate variance range in the weight distribution, matching the characteristics of the ReLU activation function; S414. Set a global random seed to fix the initial state of the random number generator, so that the results of all random processes, including weight initialization, are predictable. S415. Verify the initialization results by sampling and calculating the mean and variance of the weight matrix to ensure that its distribution is within the expected range and to avoid unstable training caused by initial values ​​that are too large or too small. S42. Initialize the bias parameters to a zero vector to ensure the stability of the initial model state; Specifically, this is implemented by inputting dimensions into the forward propagation layer. and output dimensions Using the baseline, calculate the standard deviation: ; And in the interval The sample is truncated and sampled to generate initial weight values; S43. Process the input data in batches, with each batch consisting of 32-128 bytes, and load them into memory; S44. Perform forward propagation computation, processing the time series data through LSTM branches. The calculation formula is as follows: ; in For time step The hidden state, Input data; S45. Spatial data is processed through CNN branches, and the calculation formula is as follows: ; in For feature maps, For elevation data, For convolution kernel, For bias.

[0026] S5 includes the following steps: S51. Calculate the predicted output of the model, using the following formula: ; in For predicted values, For input data, For model parameters, For model functions; S52. Using the mean squared error (MSE) as the loss function, the calculation formula is as follows: ; in This is the loss value. For the sample size, Let be the predicted value for the i-th sample. This represents the true value of the i-th sample. S53. Apply cross-entropy loss to the classification task to help calculate the confidence level of the probability of flooding, including the following steps: S531. Define a binary classification task for waterlogging occurrence, and encode the waterlogging occurrence status in historical waterlogging event data into binary labels, where 1 indicates waterlogging has occurred and 0 indicates it has not occurred, forming the basic truth values ​​for supervised learning. S532. Apply the cross-entropy loss function to calculate the difference between the model's predicted probability and the true label. The formula is: ; in For classification loss value, The first output of the model via the Sigmoid function The probability of flooding for each sample. For the corresponding real tags, The number of samples; S533. By optimizing the model parameters through backpropagation and minimizing the cross-entropy loss value, the predicted probability distribution is made closer to the true distribution, thereby improving the calibration of the probability output and indirectly enhancing the confidence of the probability of urban flooding. S534. During training, monitor the curve of cross-entropy loss decreasing with the number of iterations. When the loss converges stably, determine that the probability value output by the model has high confidence and can be used for subsequent early warning decisions. S535. When applying the trained model to new data, the output probability value is evaluated in conjunction with the loss history to assist the system in performing reliable risk classification, including the following steps: S5351. Perform model inference on the new input data, obtain the output probability value, and extract the distribution characteristics of the probability value; S5352. Retrieve historical loss data from the model training phase, including the validation set loss curve and the final convergence value, and calculate the correlation between the current prediction and the historical loss. S5353. Assess the confidence level of the probability value: When the probability value is close to 0 and the loss history is at a low level, it is judged as a low-risk state with high confidence; when the probability value is close to 1 and the loss history is at a low level, it is judged as a high-risk state with high confidence. S5354. Based on the confidence assessment results, the risk classification is dynamically weighted. High confidence probability directly triggers an early warning, while low confidence probability requires additional verification. S5355: Output the final risk level to assist the emergency management platform in decision-making and response; S54. Use the mean absolute error (MAE) for the regression task to help optimize the water depth range prediction. S55. Calculate the total loss as the weighted sum of MSE and auxiliary loss, with a weight ratio of 0.7:0.3.

[0027] S6 includes the following steps: S61. Calculate the gradient of the loss function with respect to the model parameters using automatic differentiation techniques, including the following steps: S611. Define the loss function L as the difference between the model output and the true value, where L depends on the model parameters θ, including weights and biases, and ensure that the loss function is differentiable to support gradient calculation. S612. Construct a computation graph to trace all the operation sequences from the input data through the model forward propagation to the loss output, including linear transformations and basic activation function operations; S613. Apply automatic differentiation technology, traverse the computation graph through reverse pattern differentiation, and recursively calculate the partial derivatives ∂L / ∂θ of the loss function L with respect to each parameter θ using the chain rule; S614. In the deep learning framework, the built-in automatic differentiation engine is called to perform the gradient calculation process; S615, Output gradient vector This includes the gradient value of each parameter, which is used in subsequent optimization algorithms to update the model parameters; S62. Update the parameters using the Adam optimization algorithm with adaptive moment estimation. The formula is: ; in For learning rate, For gradient operators, The value of the loss function. These are model parameters; S63. Set the initial learning rate to 0.001 and adopt a learning rate decay strategy, reducing it by 10% every 10 epochs; S64. Apply gradient clipping techniques to limit the gradient norm to below 5.0 to prevent training instability. This includes the following steps: S641. Calculate the L2 norm of the gradients of all model parameters, using the following formula: in For the total number of parameters, The gradient vector, For loss function, For the first One model parameter; S642. Compare the gradient norm with the preset threshold 5.0. If the condition is met, the trimming operation will be performed; otherwise, the parameter update will proceed directly. S643. Scale the gradient vector to ensure that the norm after clipping does not exceed the threshold, as shown in the following formula: ; in The gradient after clipping; S644, trim the gradient The parameters are passed to the optimization algorithm for parameter updates to prevent training oscillations caused by gradient explosion. S645. Monitor the pruning frequency. When pruning is triggered frequently, it is recorded as a training instability signal to help adjust the learning rate and model architecture. S65. Update the parameters of the fusion layer and backpropagate the gradients of the LSTM and CNN branches using the chain rule, including the following steps: S651. Starting from the fusion layer, calculate the loss function. Regarding the gradient of the output features of the fusion layer ,in This is the output vector of the fusion layer; S652. Apply the chain rule to propagate the gradient to the LSTM branch, as shown in the following formula: ; in The output features of the LSTM branch; S653. Similarly, applying the chain rule, propagate the gradient to the CNN branch, as shown in the following formula: ; in The output features of the CNN branch; S654. Continue backpropagation, calculate the parameter gradient layer by layer through the inner layers of LSTM and CNN branches, and complete the gradient update of the entire network.

[0028] S7 includes the following steps: S71. Set the maximum number of iterations T to 100-500 and define an early stopping mechanism to terminate training when the validation set loss does not decrease for 5 consecutive iterations. S72. After each batch of training, calculate the loss of the training set and the validation set, and record the loss curve; S73. Apply Dropout regularization, setting the dropout rate to 0.2-0.5 in the LSTM and CNN layers. The Dropout operation formula is as follows: Where h is the input feature vector. This is the feature vector after Dropout processing. For the mask vector, For discard rate; S74. Use batch normalization layers to accelerate convergence and reduce internal covariate shift. S75. When the loss function converges to the preset threshold and reaches the maximum number of iterations, save the optimal model parameters.

[0029] S8 includes the following steps: S81. Receive real-time meteorological data streams, including radar rainfall forecasts and real-time wind speed monitoring data; S82. Load the preprocessing module to normalize and filter noise in real-time data; S83. Input real-time data into the trained model and perform inference calculations; S84. Output the probability value of waterlogging occurrence, with a probability threshold set to 0.7. A warning is triggered if this threshold is exceeded. The Sigmoid activation function formula is: ; in This represents the probability value of waterlogging occurring. is the linear combination value output by the fusion layer, where e is the natural constant; S85 outputs predicted water depth ranges with an accuracy of ±0.1 meters, and generates risk heat maps based on geographic grids. The specific implementation process and technical means are as follows: The predicted water depth is generated by the linear activation function of the regression output layer and optimized using the Huber loss function. This function combines the advantages of mean squared error and mean absolute error, and is robust to outliers, ensuring an accuracy requirement of ±0.1 meters. The risk heat map is generated based on GIS gridding technology, dividing the urban area into 100m×100m grid units. Each grid is filled with the predicted water depth value, and the water depth value is mapped to a red-yellow-green gradient color scheme through a color mapping algorithm: green for water depth <0.3m, yellow for 0.3-0.5m, and red for >0.5m. The heat map is output in GeoJSON format and integrated with Leaflet and ArcGIS platforms for visualization, supporting zooming and clicking to query details.

[0030] The method also includes the following steps: S101: Deploy the model to the cloud server to support high-concurrency real-time prediction; S102. Regularly update training data and incrementally train the model by combining data from new urban flooding events. S103. Apply transfer learning techniques to adapt the pre-trained model to different urban areas. The formula for adjusting the transfer learning loss is as follows: ; in For the total loss, Losses due to new missions, For the current parameter, For pre-trained parameters, The regularization coefficient is used. S104. Integrate a multi-source data feedback mechanism to dynamically adjust the weights of the fusion layer based on the prediction error; S105. Generate a prediction report, including a probability distribution map of urban flooding, a spatiotemporal variation curve of water depth, and a confidence analysis.

[0031] The operation steps of this deep learning-based rapid urban flooding prediction method are as follows: Step 1: Data Fusion and Preprocessing Stage First, a distributed data acquisition system aggregates multi-source heterogeneous data in real time: rainfall and wind speed time-series monitoring data provided by the meteorological department, spatial topological data of elevation and slope from the Geographic Information System (GIS), pipeline distribution and drainage capacity parameters from the drainage network database, and location and water depth records of historical flooding events. The raw data undergoes standardization processing via a cleaning module, missing values ​​are filled using sliding window interpolation, outlier detection algorithms are used to eliminate sensor anomalies, and finally, normalization techniques are used to uniformly map meteorological and topographic features across different dimensions to the [0,1] interval, forming a spatiotemporally aligned structured training sample library.

[0032] Step 2: Spatiotemporal Feature Collaborative Modeling Stage A dual-branch deep neural network architecture is constructed to process the fused data: the Long Short-Term Memory (LSTM) branch extracts the temporal dependencies of meteorological data and learns the dynamic evolution of rainfall intensity and duration through a gating mechanism; the Convolutional Neural Network (CNN) branch captures the spatial correlation between topography and drainage networks and uses convolutional kernels to scan elevation raster data to identify low-lying and flood-prone areas. The outputs of the two branches are dynamically weighted and integrated through an attention fusion layer, giving higher weight to areas experiencing sudden rainstorms, generating a joint feature vector that combines sensitivity to temporal changes with spatial distribution resolution.

[0033] Step 3: Adaptive Training Optimization Phase A multi-objective loss function is employed to supervise model training: the cross-entropy loss function optimizes the classification task for the probability of urban flooding, while the mean squared error loss function constrains the regression accuracy for water depth prediction. During training, a random dropout technique is introduced, randomly masking the outputs of some neurons in the LSTM and CNN layers to enhance the model's robustness to data noise. A gradient pruning strategy is simultaneously implemented to limit the gradient magnitude during backpropagation, preventing oscillations and divergence during parameter updates. An early stopping mechanism is automatically triggered when the prediction error on the validation set fails to decrease over consecutive iterations, preserving the optimal weight parameters.

[0034] Step 4: Real-time Prediction and Feedback Phase During the deployment phase, real-time rainfall forecasts from meteorological radar and data from IoT water level sensors are received and preprocessed before being input into the trained model. The inference engine generates two key outputs based on the fused features: a gridded probability map of urban flooding and an estimated water depth range for the corresponding geographic unit. When the probability of a target area exceeds a preset threshold, the system automatically triggers a tiered early warning signal and pushes a risk heat map to the emergency management platform. The prediction results are overlaid onto a 3D urban geographic base map, visually indicating the depth of road flooding and the affected area of ​​critical infrastructure.

[0035] Step 5: Dynamic Evolution and Migration Adaptation Stage A continuous model optimization mechanism is established: Actual water depth data from new urban flooding events is collected periodically and compared with concurrent prediction results to form an error feedback dataset. An incremental learning strategy is employed to fine-tune model parameters, adapting to environmental changes caused by pipeline network upgrades while retaining existing knowledge. When deploying across regions, a transfer learning protocol is initiated, keeping the spatial feature extraction layer parameters unchanged and adjusting only the time series analysis module. This enables rapid knowledge transfer from high-data-density areas to newly developed urban areas, ensuring consistent prediction capabilities across the entire region.

[0036] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0037] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A rapid urban flooding prediction method based on deep learning, characterized in that, The method includes the following steps: S1. Collect historical meteorological data, topographic data, urban drainage system data, and historical flooding event data; S2. Perform data cleaning on the collected data, removing noise, filling in missing values, and handling outliers; S3. Construct a deep learning model architecture, including an input layer, a hidden layer and an output layer. The hidden layer includes an LSTM network branch for processing time series data to capture the changing patterns of meteorological elements over time, a CNN network branch for processing spatial data to extract spatial features of topography and urban drainage systems, and a fusion layer to integrate the outputs of the LSTM and CNN branches. S4. Initialize model parameters and perform forward propagation calculations; S5. Calculate the loss between the model output and the true value; S6. Perform backpropagation, calculate the gradient, and update the model parameters; S7. Iteratively train the model by repeating the forward propagation, loss calculation, and back propagation steps until the model converges. S8. Input real-time meteorological data, topographic data, and urban drainage system data into the trained model, and output the probability and water depth range of urban flooding in the future.

2. The method for rapid prediction of urban flooding based on deep learning according to claim 1, characterized in that: S1 includes the following steps: S11. Connect the meteorological bureau and meteorological station data sources through the meteorological data collection component to obtain historical meteorological data, including rainfall, wind speed, wind direction and temperature, and store the extracted meteorological elements in the local database; S12. Connect to the GIS system through the topographic data collection component to obtain topographic data, including elevation, slope and aspect, and process the extracted topographic features before storing them in the local database. S13. Connect to the drainage management department's database through the urban drainage system data collection component to obtain drainage system data, including drainage pipe distribution and drainage capacity, and perform consistency processing on the extracted drainage elements. S14. Connect to the emergency management department's database through the historical flooding event data collection component to obtain historical flooding event data, including the time, location, extent, and water depth of flooding, and form a historical sample database; S15. Through data integration and management components, meteorological data, topographic data, drainage system data, and historical flooding event datasets are transformed into a structured dataset in a unified format, and data backup and access control are performed.

3. The method for rapid prediction of urban flooding based on deep learning according to claim 1, characterized in that: S2 includes the following steps: S21. Fill in the missing data using the mean fill, median fill, and forward fill methods to ensure data continuity; S22. Identify and process outliers in the data, use the 3σ principle and box plot method to detect outliers, and correct and delete them; S23. Smooth random noise by using moving average filtering and Gaussian filtering algorithms to eliminate data fluctuations. S24. Normalize the data to scale the features of different dimensions to a uniform numerical range using the minimum-maximum normalization formula. S25. Convert text data into numerical vectors and use one-hot encoding and embedding layer technology to ensure that the data is suitable for model input.

4. The method for rapid prediction of urban flooding based on deep learning according to claim 1, characterized in that: S3 includes the following steps: S31. The design input layer receives preprocessed multidimensional data, including time-series meteorological data and spatial topographic data; S32. Construct an LSTM network branch containing at least two layers of LSTM units, with 50-100 units per layer, to capture the changing patterns of rainfall and wind speed over time. S33. Construct a CNN network branch containing convolutional layers, pooling layers, and activation layers to extract spatial distribution features of elevation, slope, and drainage pipes. S34. Design a fusion layer, using a fully connected layer and attention mechanism to weightedly fuse the temporal feature vector output by the LSTM branch and the spatial feature vector output by the CNN branch. S35. Construct an output layer to generate the probability value of water flooding and the predicted water depth range. The probability value is output through the Sigmoid function, and the water depth range is output through the linear regression layer.

5. The method for rapid prediction of urban flooding based on deep learning according to claim 1, characterized in that: S4 includes the following steps: S41. Initialize the model weight parameters using the Xavier and He initialization methods, and set the weight matrix to a random normal distribution. S42. Initialize the bias parameters to a zero vector to ensure the stability of the initial model state; S43. Process the input data in batches, with each batch consisting of 32-128 bytes, and load them into memory; S44. Perform forward propagation computation, processing the time series data through LSTM branches. The calculation formula is as follows: ； in For time steps The hidden state, Input data; S45. Spatial data is processed through CNN branches, and the calculation formula is as follows: ； in For feature maps, For elevation data, For convolution kernel, For bias.

6. The method for rapid prediction of urban flooding based on deep learning according to claim 1, characterized in that: S5 includes the following steps: S51. Calculate the predicted output of the model, using the following formula: ； in For predicted values, For input data, For model parameters, For model functions; S52. Using the mean squared error (MSE) as the loss function, the calculation formula is as follows: ； in The loss value. For the sample size, Let be the predicted value for the i-th sample. This represents the true value of the i-th sample. S53. Use cross-entropy loss for classification tasks to help calculate the confidence level of the probability of waterlogging; S54. Use the mean absolute error (MAE) for the regression task to help optimize the water depth range prediction. S55. Calculate the total loss as the weighted sum of MSE and auxiliary loss, with a weight ratio of 0.7:0.

3.

7. The method for rapid prediction of urban flooding based on deep learning according to claim 1, characterized in that: S6 includes the following steps: S61. Calculate the gradient of the loss function with respect to the model parameters using automatic differentiation techniques; S62. Update the parameters using the Adam optimization algorithm with adaptive moment estimation. The formula is: ； in For learning rate, For gradient operators, The value of the loss function. These are model parameters; S63. Set the initial learning rate to 0.001 and adopt a learning rate decay strategy, reducing it by 10% every 10 epochs; S64. Apply gradient clipping techniques to limit the gradient norm to below 5.0 to prevent training instability; S65. Update the parameters of the fusion layer and backpropagate the gradients of the LSTM and CNN branches using the chain rule.

8. The method for rapid prediction of urban flooding based on deep learning according to claim 1, characterized in that: S7 includes the following steps: S71. Set the maximum number of iterations T to 100-500 and define an early stopping mechanism to terminate training when the validation set loss does not decrease for 5 consecutive iterations. S72. After each batch of training, calculate the loss of the training set and the validation set, and record the loss curve; S73. Apply Dropout regularization technique, setting the dropout rate to 0.2-0.5 in LSTM and CNN layers; S74. Use batch normalization layers to accelerate convergence and reduce internal covariate shift. S75. When the loss function converges to the preset threshold and reaches the maximum number of iterations, save the optimal model parameters.

9. The method for rapid prediction of urban flooding based on deep learning according to claim 1, characterized in that: S8 includes the following steps: S81. Receive real-time meteorological data streams, including radar rainfall forecasts and real-time wind speed monitoring data; S82. Load the preprocessing module to normalize and filter noise in real-time data; S83. Input real-time data into the trained model and perform inference calculations; S84. Output the probability value of waterlogging. The probability threshold is set to 0.

7. If the probability is exceeded, an early warning will be triggered. S85 outputs predicted water depth range with an accuracy of ±0.1 meters and generates a risk heat map based on the geographic grid.

10. The method for rapid prediction of urban flooding based on deep learning according to claim 1, characterized in that: The method further includes the following steps: S101: Deploy the model to the cloud server to support high-concurrency real-time prediction; S102. Regularly update training data and incrementally train the model by combining data from new urban flooding events. S103. Apply transfer learning techniques to adapt the pre-trained model to different urban areas; S104. Integrate a multi-source data feedback mechanism to dynamically adjust the weights of the fusion layer based on the prediction error; S105. Generate a prediction report, including a probability distribution map of urban flooding, a spatiotemporal variation curve of water depth, and a confidence analysis.

Images