A three-dimensional virtual simulation flood forecasting and early warning method and system

Abstract

The application discloses a three-dimensional virtual simulation flood prediction and early warning method and system, obtains terrain data and rainfall time series data of a target area; generates a flood evolution multi-time step prediction result containing water depth distribution based on the terrain data and the rainfall time series data; generates a flood peak value prediction result based on the rainfall time series data and the multi-time step prediction result; performs three-dimensional presentation on the multi-time step prediction result and the flood peak value prediction result in a three-dimensional visualization platform, and outputs early warning information based on a preset threshold rule. The application reduces the dependence on complex numerical simulation modeling and iterative solution by simultaneously using terrain constraints and rainfall driving for evolution prediction and peak value prediction and linking three-dimensional presentation and threshold early warning output, and improves the rapid early warning support capability of a sudden rainfall scene.

Description

Technical Field

[0001] This application relates to the field of time series data prediction technology, and in particular to a three-dimensional virtual simulation flood prediction and early warning method and system. Background Technology

[0002] With the increasing frequency of extreme rainfall events and the acceleration of urbanization, the suddenness and destructiveness of floods are constantly increasing. To improve the efficiency of flood prevention and disaster reduction response, relevant industries are gradually building flood forecasting, early warning, and emergency command systems for river basins, cities, and key areas. Rapid prediction of flood inundation range, water depth, and their evolution over time, as well as advance estimation of peak water levels / peak values ​​at key sections or regions, are crucial foundations for personnel evacuation, road control, risk assessment, and emergency dispatch. Simultaneously, to facilitate commanders' intuitive understanding of disaster development and rapid decision-making, it is often necessary to dynamically display the forecast results using 3D virtual simulation and combine them with threshold rules to achieve tiered early warning and decision support.

[0003] In related technologies, flood inundation prediction and peak forecasting mainly rely on hydrodynamic numerical simulation methods. For example, hydrological-hydrodynamic models are established based on governing equations such as the shallow water equation, and discrete solutions are performed by combining inputs such as topographic elevation, rainfall, and boundary conditions to obtain spatiotemporal results such as water depth, flow velocity, and inundation range. Although these methods have clear physical meaning, they often require strong professional modeling experience and preparation of multi-source data. Moreover, the calculation process usually involves iterative solutions of differential or partial differential equations. When dealing with high-resolution terrain and long-term simulations, the computational load is large and the time consumption is long, making it difficult to meet the needs of rapid early warning and real-time interactive display under sudden rainstorms. To improve computational efficiency, existing technologies sometimes use methods such as simplifying models, reducing grid resolution, or shortening the calculation step size for approximation. However, this may cause errors to gradually accumulate during the time-series propagation, resulting in deviations from key moments (such as the inflection point of inundation expansion and peak arrival time) and key indicators (such as peak water level and maximum water depth) in the flood evolution process. In addition, while some traditional machine learning methods can accelerate predictions to some extent, they have limited ability to characterize complex nonlinear hydrological relationships and strong spatiotemporal correlations, making it difficult to simultaneously guarantee prediction accuracy, stability, and physical consistency. In particular, in high-uncertainty scenarios such as extreme rainfall, prediction reliability remains insufficient.

[0004] Therefore, in flood inundation evolution prediction and flood peak forecasting, the high complexity of modeling and calculation, the time-consuming numerical simulation calculation and the easy accumulation of errors, as well as the insufficient reliability of nonlinear prediction under extreme rainfall scenarios, have become urgent problems to be solved. Summary of the Invention

[0005] This application provides a three-dimensional virtual simulation flood prediction and early warning method and system, which aims to solve the problems of high modeling and calculation complexity, time-consuming numerical simulation calculation and easy error accumulation, and insufficient reliability of nonlinear prediction in extreme rainfall scenarios in the existing technology for flood inundation evolution prediction and flood peak forecast.

[0006] In a first aspect, a three-dimensional virtual simulation method for flood prediction and early warning, the method comprising: Acquire topographic data and rainfall time series data for the target area; Based on the terrain data and the rainfall time series data, a multi-time step prediction result of flood evolution is generated, and the multi-time step prediction result includes water depth distribution; Based on the rainfall time series data and the multi-time step prediction results, a flood peak prediction result is generated; The multi-time-step prediction results and the flood peak prediction results are presented in three dimensions in a three-dimensional visualization platform, and early warning information is output based on preset threshold rules.

[0007] Optionally, in the above scheme, generating multi-time-step prediction results of flood evolution based on the terrain data and the rainfall time series data includes: Feature extraction is performed on the terrain data to obtain a terrain feature representation; The terrain feature representation is constructed into a terrain feature sequence that is consistent with the time step of the rainfall time series data; The rainfall time series data is represented in a spatiotemporal manner to obtain a rainfall feature sequence; The terrain feature sequence and the rainfall feature sequence are weighted and fused to obtain a spatiotemporal feature sequence; The spatiotemporal feature sequence is input into a convolutional recurrent neural network to obtain the multi-time-step prediction result.

[0008] Optionally, in the above scheme, the step of extracting features from the terrain data to obtain a terrain feature representation includes: The terrain data is identified as digital elevation model data; The digital elevation model data is input into a terrain feature extraction network to obtain a terrain feature map, wherein the terrain feature extraction network includes convolutional layers, nonlinear activation layers, and downsampling layers; The terrain feature map is determined as the terrain feature representation.

[0009] Optionally, in the above scheme, constructing the terrain feature representation into a terrain feature sequence consistent with the time step of the rainfall time series data includes: The target time step length is determined based on the number of time steps in the rainfall time series data; The terrain feature representation is copied or broadcast along the time dimension to obtain a terrain feature sequence of the target time step length.

[0010] Optionally, in the above scheme, the step of performing spatiotemporal representation of the rainfall time series data to obtain a rainfall feature sequence includes: The rainfall values ​​at each time step are expanded into a rainfall feature map that is consistent with the spatial grid corresponding to the terrain feature representation; Based on the rainfall feature maps corresponding to each of the time steps, a multi-time-step rainfall feature sequence is obtained.

[0011] Optionally, in the above scheme, inputting the spatiotemporal feature sequence into a convolutional recurrent neural network to obtain the multi-time-step prediction result includes: The spatiotemporal feature sequence is input into a convolutional recurrent neural network in time step order. The convolutional recurrent neural network includes at least two layers of convolutional long short-term memory network (ConvLSTM) to update the hidden state at each time step and output the spatiotemporal feature representation at each time step. The spatiotemporal feature representations of each time step are decoded to obtain the water depth prediction map corresponding to each time step. The water depth prediction maps at each time step are combined to form a multi-time step water depth distribution corresponding to the spatial grid of the target area, which is used as the multi-time step prediction result.

[0012] Optionally, in the above scheme, decoding the spatiotemporal feature representations of each time step to obtain the water depth prediction map corresponding to each time step includes: Spatial resolution restoration processing is performed on the spatiotemporal feature representations of each time step to obtain the processed features corresponding to each time step; wherein, the spatial resolution restoration processing includes deconvolution upsampling and / or interpolation upsampling; Channel mapping is performed on the processed features corresponding to each time step to obtain the water depth prediction map corresponding to each time step, as well as the first direction velocity component prediction map and the second direction velocity component prediction map corresponding to the water depth prediction map. The first direction velocity component and the second direction velocity component are orthogonal to each other.

[0013] Optionally, in the above scheme, generating flood peak prediction results based on the rainfall time series data and the multi-time step prediction results includes: Based on the rainfall time series data and the multi-time step prediction results, peak prediction input features are obtained; The peak prediction input features are used to generate a model, which outputs flood peak prediction results, including peak water level and / or peak water depth. Specifically, when the conditional generation model is a conditional diffusion model, the flood peak prediction result is obtained by performing stepwise denoising iterations based on the peak prediction input features.

[0014] Optionally, in the above scheme, the step of outputting early warning information based on preset threshold rules includes: Obtain the preset threshold rules corresponding to the target area, wherein the preset threshold rules include at least one type of threshold and the warning level corresponding to the threshold; The water depth statistics for each time step are extracted from the multi-time step prediction results. The water depth statistics include the maximum water depth, average water depth, and the percentage of grid cells exceeding the water depth threshold in the target area. Extract peak indexes from the flood peak prediction results; The water depth statistics and / or the peak index are compared with the preset threshold rule to obtain the comparison result and the warning level corresponding to the comparison result; Output the warning information corresponding to the warning level in the three-dimensional visualization platform, and associate the warning information with the multi-time step prediction results and / or the flood peak prediction results for display.

[0015] Secondly, a three-dimensional virtual simulation flood prediction and early warning system, the system comprising: The data acquisition module is used to acquire topographic data and rainfall time series data of the target area; An evolution prediction module is used to generate multi-time-step prediction results of flood evolution based on the terrain data and the rainfall time series data, wherein the multi-time-step prediction results include water depth distribution; The peak prediction module is used to generate flood peak prediction results based on the rainfall time series data and the multi-time step prediction results; The 3D visualization and early warning module is used to present the multi-time step prediction results and the flood peak prediction results in 3D on a 3D visualization platform, and output early warning information based on preset threshold rules.

[0016] Thirdly, a computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to perform the following steps: Compared with the prior art, this application has at least the following beneficial effects: This application, based on further analysis and research of existing technical problems, recognizes the issues of high modeling and computational complexity, time-consuming numerical simulation calculations and easy error accumulation, and insufficient reliability of nonlinear predictions under extreme rainfall scenarios in flood inundation evolution prediction and flood peak forecasting. By acquiring topographic data and rainfall time-series data of the target area, a multi-time-step flood evolution prediction result including water depth distribution is first generated based on topographic conditions and rainfall-driven factors within the same processing link. Then, under the common constraints of the rainfall time-series data and the multi-time-step prediction results, a flood peak prediction result is generated. The above evolution prediction results and peak prediction results are then presented in three dimensions on a three-dimensional visualization platform, and early warning information is output according to preset threshold rules. Because topographic data provides spatial constraints on flood diffusion and confluence, and rainfall time-series data provides spatial constraints on flood diffusion and confluence, the application addresses these issues. The time-varying sequence provides external input drivers, and the multi-time-step water depth distribution can characterize the continuous evolution of floods from occurrence to expansion. Based on this, the flood peak prediction results can be generated to centrally represent key risk indicators. At the same time, the three-dimensional presentation displays the time-series evolution and peak information in an intuitive way. The comparison output of threshold rules transforms the prediction results into actionable risk level prompts. Therefore, the prediction calculation does not have to rely on complex physical numerical simulation and high-cost iterative modeling to form usable time-series evolution and peak output. It can be directly used for business visualization and early warning release, thereby solving the problems of high modeling and calculation complexity, time-consuming numerical simulation calculation and easy error accumulation, and difficulty in timely support of early warning decisions by nonlinear prediction in extreme rainfall scenarios. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a three-dimensional virtual simulation flood prediction and early warning method provided in one embodiment of this application; Figure 2 This is a schematic block diagram of a network structure provided in one embodiment of this application; Figure 3 A geographic digital elevation information map used in a network provided in one embodiment of this application; Figure 4 A simulation data diagram of a two-dimensional hydrodynamic model used in a network provided in one embodiment of this application; Figure 5 This application provides a 256×256 water depth map obtained by interpolation in a network according to one embodiment of the present application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] In the description of this application: unless otherwise stated, "a plurality of" means two or more. The terms "first," "second," "third," etc., in this application are intended to distinguish the objects referred to and do not have any special meaning in terms of technical connotation (e.g., they should not be construed as an emphasis on importance or order). Expressions such as "including," "comprising," and "having" also mean "not limited to" (certain units, components, materials, steps, etc.).

[0020] In one embodiment, such as Figure 1 As shown, a three-dimensional virtual simulation flood prediction and early warning method is provided, including the following steps: Acquire topographic data and rainfall time series data for the target area; Based on the terrain data and the rainfall time series data, a multi-time step prediction result of flood evolution is generated, and the multi-time step prediction result includes water depth distribution; Based on the rainfall time series data and the multi-time step prediction results, a flood peak prediction result is generated; The multi-time-step prediction results and the flood peak prediction results are presented in three dimensions in a three-dimensional visualization platform, and early warning information is output based on preset threshold rules.

[0021] This embodiment of the three-dimensional virtual simulation flood prediction and early warning method focuses on simulating flood evolution and forecasting peak values ​​for a target area. The target area can be a watershed, an urban area, or a specified grid range. Topographic data can be a digital elevation model (DEM), and its spatial resolution can be set according to application requirements, such as a high-resolution raster covering details like buildings and roads. Rainfall time series data can be a time-step sequence of rainfall intensity or amount, with time steps ranging from minutes to hours. This data can be derived from historical rainfall data or real-time weather forecasts / radar extrapolation data.

[0022] When generating multi-time-step flood evolution prediction results, the flood evolution prediction model can be input with the DEM and rainfall sequence to obtain a water depth distribution sequence output by time step. The water depth distribution can be output in a raster format consistent with the DEM, with each raster corresponding to a water depth value, thus forming a water depth field that evolves over time. Furthermore, the flood evolution prediction results can include lateral flow velocity information corresponding to the water depth for three-dimensional visualization and subsequent calculation of physical consistency constraints or statistical indicators.

[0023] When generating flood peak prediction results, rainfall time series data and multi-time step prediction results of flood evolution are used together as conditional information and input into the peak prediction model to obtain peak water level and / or peak water depth. The peak prediction results can output a peak index for the entire target area, or output peak indices separately for preset sub-areas, key sections, or key points.

[0024] When presented on a 3D visualization platform, the system maps multi-timestep water depth fields to 3D water surface height fields or volume rendering parameters, and overlays peak prediction results as labels, legends, or curves, thus providing a time-varying view of the inundation evolution and peak information. Warning information is generated based on preset threshold rules, which can define warning levels according to water depth thresholds, peak water level thresholds, or a combination of thresholds. The system compares the prediction results with the thresholds and outputs the corresponding warning level.

[0025] This embodiment completes multi-time step prediction of flood evolution, flood peak prediction, three-dimensional presentation, and threshold early warning output in the same process. It can realize the display of disaster evolution and risk level prompts under a unified data and model framework, which is convenient for providing interpretable prediction and early warning results for emergency command or risk assessment scenarios.

[0026] In this embodiment, generating multi-time-step prediction results of flood evolution based on the terrain data and the rainfall time-series data includes: Feature extraction is performed on the terrain data to obtain a terrain feature representation; The terrain feature representation is constructed into a terrain feature sequence that is consistent with the time step of the rainfall time series data; The rainfall time series data is represented in a spatiotemporal manner to obtain a rainfall feature sequence; The terrain feature sequence and the rainfall feature sequence are weighted and fused to obtain a spatiotemporal feature sequence; The spatiotemporal feature sequence is input into a convolutional recurrent neural network to obtain the multi-time-step prediction result.

[0027] In the terrain feature extraction stage, the DEM can be input into the terrain feature extraction network, which outputs a terrain feature representation. This representation can be a multi-channel feature map, where channels characterize spatial patterns related to water accumulation evolution, such as slope changes, depression structures, and runoff paths. This terrain feature representation serves as a static spatial condition, providing consistent terrain constraint information for each time step in subsequent time-series prediction.

[0028] In the terrain feature sequence construction stage, the terrain feature representation can be copied or broadcast along the time dimension according to the time step length of the rainfall time series, resulting in a terrain feature sequence consistent with the rainfall time step. This construction method avoids repeating terrain feature extraction calculations at each time step, while ensuring that each time step has usable terrain condition input.

[0029] The spatialization stage of rainfall expands the rainfall values ​​at each time step into a two-dimensional rainfall feature map consistent with the spatial grid representing the terrain features, transforming rainfall from a "time series" into a "spatial-temporal series." In implementation, the rainfall intensity at each time step can be spatially broadcast to form a constant-value map of the same size as the terrain grid; alternatively, if spatial rainfall distribution is available, a rainfall raster map for each time step can be directly generated, thus constructing a multi-time-step rainfall feature sequence.

[0030] The weighted fusion stage fuses terrain features and rainfall features at each time step to obtain a spatiotemporal feature sequence. Weighted fusion can be implemented in several ways: one approach is to apply weight coefficients to the terrain and rainfall feature maps separately and then sum them; another approach is to first concatenate the channels and then use a learnable weight generation network to obtain the fusion weights, and then weight the two types of features accordingly; alternatively, features can be grouped and weighted according to the channel dimension to achieve different responses of different terrain patterns to rainfall input.

[0031] In the convolutional recurrent neural network (RNN) stage, the aforementioned spatiotemporal feature sequences are input into the RNN (e.g., ConvLSTM). The model recursively updates the hidden states in the time dimension and outputs the prediction results for each time step, thus obtaining multi-time-step prediction results of flood evolution. The output includes at least the water depth distribution and may also include the flow velocity component for subsequent three-dimensional rendering, physical consistency assessment, or statistical indicator extraction.

[0032] This embodiment achieves multi-time-step water depth evolution prediction on a unified spatiotemporal feature representation by performing spatiotemporal alignment and weighted fusion of DEM and rainfall sequence, and using convolutional recurrent neural network for temporal recursion. This is beneficial for improving the ability to model the temporal correlation of flood evolution and supporting the simulation output of continuous time steps.

[0033] In this embodiment, the step of extracting features from the terrain data to obtain a terrain feature representation includes: The terrain data is identified as digital elevation model data; The digital elevation model data is input into a terrain feature extraction network to obtain a terrain feature map, wherein the terrain feature extraction network includes convolutional layers, nonlinear activation layers, and downsampling layers; The terrain feature map is determined as the terrain feature representation.

[0034] In this embodiment, the terrain data can be identified as a Digital Elevation Model (DEM). The DEM can be a raster matrix, with each pixel corresponding to the ground elevation. To extract terrain features suitable for time-series prediction, the DEM can be input into a terrain feature extraction network. The terrain feature extraction network may include convolutional layers, nonlinear activation layers, and downsampling layers. Convolutional layers are used to extract spatial features such as slope structure, depression boundaries, and valley orientation from the local neighborhood; nonlinear activation layers are used to enhance the representation of nonlinear terrain patterns; and downsampling layers are used to expand the receptive field, aggregate multi-scale terrain information, and reduce the computational burden on subsequent time-series networks.

[0035] In implementation, the terrain feature extraction network can be a multi-stage convolutional encoder structure, with each stage consisting of several convolutional layers and activation layers. The spatial size is progressively reduced and the number of channels is increased through downsampling. The terrain feature map output by the network can serve as a terrain feature representation for subsequent fusion with rainfall features and as input to a ConvLSTM. The spatial size of the terrain feature map can be smaller than the original DEM size, and then restored to the target resolution in the subsequent decoding stage.

[0036] This embodiment converts the DEM into a multi-channel topographic feature map, which can express key topographic structure information in a way that is more suitable for neural network spatiotemporal modeling, thereby providing stable spatial condition input for subsequent flood evolution prediction.

[0037] In this embodiment, constructing the terrain feature representation into a terrain feature sequence consistent with the time step of the rainfall time series data includes: The target time step length is determined based on the number of time steps in the rainfall time series data; The terrain feature representation is copied or broadcast along the time dimension to obtain a terrain feature sequence of the target time step length.

[0038] After obtaining the terrain feature representation, the target time step length can be determined based on the number of time steps in the rainfall time series data. For example, if the rainfall series contains N time steps, the target time step length is N. Subsequently, the terrain feature representation is copied or broadcast along the time dimension to construct a terrain feature sequence of length N, so that each time step corresponds to the same terrain feature map.

[0039] In implementation, copying or broadcasting can employ tensor expansion operations to expand two-dimensional or three-dimensional terrain feature maps into four-dimensional tensors (time × channel × height × width) with a time dimension, thereby aligning them with the rainfall feature sequence in the time dimension. If the terrain remains unchanged over short timescales, this construction method can be used directly; if terrain changes need to be introduced in specific applications (e.g., temporary embankments, road closures), terrain features at certain time steps can be updated after broadcasting.

[0040] This embodiment constructs a static terrain feature representation as a terrain feature sequence consistent with the rainfall time step, which ensures that the time series prediction model obtains consistent spatial condition input at each time step, facilitating the synchronous fusion and recursive prediction of terrain and rainfall in the time dimension.

[0041] In this embodiment, the step of performing spatiotemporal representation of the rainfall time series data to obtain a rainfall feature sequence includes: The rainfall values ​​at each time step are expanded into a rainfall feature map that is consistent with the spatial grid corresponding to the terrain feature representation; Based on the rainfall feature maps corresponding to each of the time steps, a multi-time-step rainfall feature sequence is obtained.

[0042] Rainfall time series data can be either a single rainfall value (e.g., average rainfall intensity) for each time step, or a spatial distribution map of rainfall for each time step. For the former, the rainfall value at each time step can be expanded into a rainfall feature map with a spatial grid consistent with the topographic feature representation. This expansion can be achieved by filling the rainfall value into a two-dimensional matrix of the same size as the topographic feature map, forming a constant-value rainfall map; or, if spatial differences need to be considered, a pre-defined spatial weight template can be introduced to map the rainfall value into a non-uniform rainfall map.

[0043] Based on the rainfall feature maps corresponding to each time step, a multi-time-step rainfall feature sequence can be obtained by combining them in chronological order. This rainfall feature sequence has the same number of time steps as the terrain feature sequence, and its spatial dimensions are consistent with the terrain feature representation, thus providing an alignment basis for subsequent weighted fusion.

[0044] This embodiment transforms the rainfall sequence into a spatiotemporal feature sequence consistent with the terrain grid, enabling the aligned expression of rainfall input and spatial terrain conditions on the same grid, which facilitates subsequent time-step fusion and recursive prediction.

[0045] In this embodiment, the step of inputting the spatiotemporal feature sequence into a convolutional recurrent neural network to obtain the multi-time-step prediction result includes: The spatiotemporal feature sequence is input into a convolutional recurrent neural network in time step order. The convolutional recurrent neural network includes at least two layers of convolutional long short-term memory network (ConvLSTM) to update the hidden state at each time step and output the spatiotemporal feature representation at each time step. The spatiotemporal feature representations of each time step are decoded to obtain the water depth prediction map corresponding to each time step. The water depth prediction maps at each time step are combined to form a multi-time step water depth distribution corresponding to the spatial grid of the target area, which is used as the multi-time step prediction result.

[0046] The spatiotemporal feature sequences are input into a convolutional recurrent neural network (RNN) in time-step order. The RNN consists of at least two ConvLSTM layers. The first ConvLSTM layer establishes a temporal recursive relationship based on local spatial convolutions, outputting the spatiotemporal hidden state of the first layer. The second ConvLSTM layer further models the spatiotemporal dependencies of higher layers based on the output of the first layer, obtaining the spatiotemporal feature representations for each time step. This structure allows the network to preserve spatial locality while accumulating historical information in the temporal dimension.

[0047] The spatiotemporal feature representations at each time step are decoded to obtain the water depth prediction map for that time step. The decoding process may include restoring the features from the encoded resolution to the target spatial resolution and mapping multi-channel features to a single-channel water depth output. Restoring to the target spatial resolution can be accomplished through upsampling or deconvolution operations, and channel mapping can be achieved through convolutional layers. Each time step outputs a water depth prediction map corresponding to the spatial grid of the target region.

[0048] The water depth prediction maps at each time step are combined in chronological order to form a multi-time-step water depth distribution sequence, which serves as the multi-time-step prediction result. This sequence can be directly used for playback and display on a 3D platform, or it can be used as input for subsequent peak prediction and threshold warning statistics.

[0049] This embodiment uses ConvLSTM to recursively extrapolate spatiotemporal features and outputs water depth prediction maps at each time step, which can form a continuous water depth evolution sequence consistent with the time step, making it convenient to support dynamic prediction and subsequent visualization and early warning calculation of flood inundation processes.

[0050] In this embodiment, the decoding process of the spatiotemporal feature representations of each time step to obtain the water depth prediction map corresponding to each time step includes: Spatial resolution restoration processing is performed on the spatiotemporal feature representations of each time step to obtain the processed features corresponding to each time step; wherein, the spatial resolution restoration processing includes deconvolution upsampling and / or interpolation upsampling; Channel mapping is performed on the processed features corresponding to each time step to obtain the water depth prediction map corresponding to each time step, as well as the first direction velocity component prediction map and the second direction velocity component prediction map corresponding to the water depth prediction map. The first direction velocity component and the second direction velocity component are orthogonal to each other.

[0051] Spatial resolution restoration processing is used to restore the spatiotemporal feature representations at each time step from a lower resolution to the target spatial grid resolution. This can be achieved through deconvolutional upsampling or interpolation upsampling followed by convolutional thinning. Deconvolutional upsampling can reconstruct spatial details using learnable parameters; interpolation upsampling can employ bilinear interpolation or nearest-neighbor interpolation, and then restore boundary and structural information through convolutional layers after upsampling. These processes yield the processed features corresponding to each time step.

[0052] Channel mapping is used to map processed features to output physical quantity channels. Convolutional layers can be used to map multi-channel features to a water depth prediction map; simultaneously, within the same channel mapping process, prediction maps of the first and second directional velocity components corresponding to the water depth prediction map are output. The first and second directions can be defined in the specification as two mutually orthogonal directions consistent with the spatial grid coordinate system, such as the x and y directions, or east and north. The output format can be consistent with the water depth prediction... Figure 1 All of these are grid fields aligned with the spatial grid of the target area, so that each time step corresponds to three grid images: water depth, the velocity component in the first direction, and the velocity component in the second direction.

[0053] In the 3D visualization platform, the velocity component prediction map can be further converted into a vector field display or a flow direction arrow display. The vector direction is synthesized by two orthogonal components, and the vector magnitude is determined by the combination of component amplitudes, thus presenting it in conjunction with the water depth evolution image.

[0054] This embodiment, by simultaneously outputting a water depth prediction map and a mutually orthogonal flow velocity component prediction map during the decoding stage, can provide a more complete spatiotemporal field representation of flood evolution results, facilitating the display of flow direction in a three-dimensional scene and further analysis and calculations related to water depth.

[0055] In this embodiment, generating flood peak prediction results based on the rainfall time series data and the multi-time step prediction results includes: Based on the rainfall time series data and the multi-time step prediction results, peak prediction input features are obtained; The peak prediction input features are used to generate a model, which outputs flood peak prediction results, including peak water level and / or peak water depth. Specifically, when the conditional generation model is a conditional diffusion model, the flood peak prediction result is obtained by performing stepwise denoising iterations based on the peak prediction input features.

[0056] Peak prediction input features are constructed from rainfall time series data and multi-time-step prediction results. Construction methods may include extracting temporal and statistical features from the rainfall sequence, such as cumulative rainfall, maximum rainfall intensity, peak arrival time index, and rate of change within the time window; simultaneously, temporal and spatial aggregation of multi-time-step water depth distribution sequences are performed, for example, extracting the maximum water depth sequence, average water depth sequence, and area percentage sequence exceeding the threshold for each time step for the target area or preset sub-region, or compressing the water depth field into a fixed-dimensional temporal feature vector through an encoding network. These rainfall features and evolution features can be concatenated to form the peak prediction input features.

[0057] The peak value prediction is input into a conditional generative model, which outputs the flood peak value prediction result. The conditional generative model can be a regression network, a sequence-to-scalar prediction network, or a generative model; the output can be the peak water level and / or the peak water depth. If it is necessary to output the peak value for multiple sub-regions separately, the output can be a set of peak value indicators corresponding to multiple regions.

[0058] When the conditional generation model is a conditional diffusion model, the inference phase can start from the initial noise state and perform progressive denoising iterations under the conditional constraints of the peak prediction input features. During the iteration process, the noise samples are guided by the conditional input to gradually approximate the peak prediction result representation until the final peak water level and / or peak water depth output is obtained. In the training phase, the peak label data can be progressively noisy to form training samples at different noise stages, and the denoising network can be trained by combining time step information, so that it can predict the denoising direction or denoising residual under the conditional input constraints, thereby supporting the progressive denoising iterations in the inference phase.

[0059] This embodiment constructs the peak prediction input features by combining the rainfall time series and the multi-time step prediction results of flood evolution, and uses a conditional generation model (especially the stepwise denoising generation mechanism of the conditional diffusion model) to output the flood peak index. It can simultaneously utilize rainfall-driven information and evolution process information in the peak prediction, making it easier to obtain peak prediction results that match the conditions of the target area.

[0060] In this embodiment, the step of outputting early warning information based on preset threshold rules includes: Obtain the preset threshold rules corresponding to the target area, wherein the preset threshold rules include at least one type of threshold and the warning level corresponding to the threshold; The water depth statistics for each time step are extracted from the multi-time step prediction results. The water depth statistics include the maximum water depth, average water depth, and the percentage of grid cells exceeding the water depth threshold in the target area. Extract peak indexes from the flood peak prediction results; The water depth statistics and / or the peak index are compared with the preset threshold rule to obtain the comparison result and the warning level corresponding to the comparison result; Output the warning information corresponding to the warning level in the three-dimensional visualization platform, and associate the warning information with the multi-time step prediction results and / or the flood peak prediction results for display.

[0061] Pre-configured threshold rules can be configured by flood control departments or business systems. Each threshold rule must include at least one type of threshold and its corresponding warning level. Thresholds can be set as a single water depth threshold, multiple water depth threshold segments, or simultaneously include peak water level thresholds, peak water depth thresholds, and combined rules. Warning levels can be multi-level, such as alert, blue, yellow, orange, and red, or can be defined according to the business system.

[0062] Water depth statistics are extracted from multi-time-step prediction results. Extraction methods include calculating the maximum and average water depth of the target area for each time-step water depth grid field; the percentage of grid cells exceeding the water depth threshold can be obtained by statistically analyzing the ratio of the number of grid cells exceeding the threshold to the total number of grid cells. If preset sub-regions or key facility areas are set, the above statistics can also be calculated separately for each sub-region to form a regional-level statistical set. Peak value indicators are extracted from flood peak prediction results; peak value indicators can be peak water level, peak water depth, or a combination thereof.

[0063] The system compares water depth statistics and / or peak indices with preset threshold rules to obtain comparison results and determine the corresponding warning level. The comparison logic can include single-threshold comparison, multi-threshold segmented comparison, or combined rule comparison. For example, the warning level can be raised when both the water depth statistics exceed a first threshold and the peak indices exceed a second threshold. The comparison results can record the triggered threshold item, the trigger time step, and the trigger area, providing a basis for subsequent display and traceability.

[0064] Output and display early warning information in a 3D visualization platform. This display can include overlaying warning level indicators while playing flood evolution footage, highlighting areas triggering thresholds on a map, showing comparison curves of corresponding statistics with thresholds, and displaying peak indicator labels. If the early warning information changes over time, the points where the warning level changes can be marked on the timeline to help users pinpoint stages of escalating risk.

[0065] This embodiment extracts statistics from the multi-time-step water depth prediction results and compares them with the peak prediction index and preset threshold rules in a hierarchical manner. Then, it presents the warning information in a 3D platform, which can form a traceable threshold triggering logic and an intuitive display method, making it easy to output hierarchical warning prompts for business scenarios.

[0066] In one embodiment, a deep learning-based method for simulating flood inundation evolution is provided. Traditional numerical simulation-based hydrological prediction uses mathematical and physical equations to build complex models and predict hydrological elements such as watershed runoff and velocity, with a focus on two key indicators: peak flood level prediction and flood inundation process prediction. However, such methods involve complex model building, and a trade-off between computational efficiency and accuracy in numerical simulation is difficult. With the rapid development of computer hardware technology, the application of machine learning and deep learning in traditional fields has attracted considerable attention from researchers.

[0067] like Figure 2 , Figure 3 , Figure 4 and Figure 5 As shown, this embodiment considers the changes in lateral hydrodynamic elements during neural network training, decoupling the water flow into orthogonal flows. , The direction is determined, and the quasi-linear one-dimensional shallow water equation is given by equation (1.1). (1.1); In the formula, It's traffic. It is time. It specifies the direction. It is the cross-sectional area of ​​the flow path. It is gravitational acceleration. It is the horizontal plane height. It is the underwater elevation. It is the hydraulic radius. It is the Manning coefficient of friction.

[0068] The implementation primarily comprises three core modules: First, a flood evolution time-series prediction module based on the ConvLSTM architecture, which integrates multiple information sources such as geographic elevation, real-time inundation depth, and shallow water equations to achieve dynamic simulation of the flood evolution process. Second, an innovative flood peak prediction module based on the Diffusion model is constructed, which can effectively handle nonlinear prediction problems under extreme rainfall events. Finally, a 3D visualization platform is developed based on the Unity engine, possessing three main functional characteristics: supporting the visualization of multi-dimensional prediction results, including 3D dynamic simulation of the flood evolution process and interactive display of peak water level prediction results; providing a user-friendly interface that supports flexible switching of prediction models and parameter adjustment; and integrating a real-time data interface to dynamically acquire meteorological and hydrological data and update prediction results, while automatically triggering a graded early warning mechanism based on preset thresholds.

[0069] The ConvLSTM flood prediction model employs an end-to-end deep learning architecture to achieve time-series prediction of the evolution of dam-break floods through four key stages.

[0070] First, the terrain feature extraction stage uses an encoder network (containing convolutional layers, ReLU activation function and max pooling layer) to convert the input 256×256 high-resolution elevation map (DEM) into a 128×128 feature map with 64 channels, and then uses a dimension expansion operation to copy the static terrain features to 25 time steps.

[0071] Secondly, in the multimodal data fusion stage, the model expands the rainfall data of 25 time steps (1 value per time step) to a 128×128 spatial dimension through a broadcasting mechanism, and splices it with the terrain features in the channel dimension to form a 65-channel spatiotemporal feature cube.

[0072] Then, in the spatiotemporal evolution modeling stage, a two-layer ConvLSTM network is used to process the spatiotemporal feature sequence. While maintaining the 128×128 spatial dimension, the network generates 128 channels of spatiotemporal features through hidden layers and adjusts the dimensions to adapt to subsequent processing.

[0073] Finally, in the prediction and decoding stage, the LSTM output features at each time step are deconvolved and upsampled, and the 128-channel features are expanded into 3-channel outputs (corresponding to water depth, x-direction flow velocity and y-direction flow velocity, respectively) through 1×1 convolution, and finally outputting a 256×256 resolution multi-parameter prediction sequence for 25 time steps.

[0074] This embodiment addresses the problem that traditional machine learning models, which either rely on expert experience and statistical features or are shallow machine learning models with only one hidden layer or no hidden layer nodes, struggle to characterize the complex nonlinear relationships between hydrological elements. It also solves the problem of traditional numerical simulation models struggling to balance prediction accuracy and speed. A loss function based on shallow water equations was designed in the training of the convolutional short-time network to ensure the realism and reliability of the network output structure, making the simulation results more consistent with physical reality.

[0075] This embodiment employs a ConvLSTM architecture, mitigating the problems of gradient explosion and vanishing gradients, and effectively catering to the strongly relevant contextual information in the application scenario. In data processing, multi-source information is integrated to design a flood evolution time-series prediction module, incorporating multi-dimensional information such as geographical elevation and real-time inundation depth. In the network training scheme used in this embodiment, the shallow water equation is designed as a constraint in the loss function, ensuring the physical realism of the dynamic simulation of the flood evolution process.

[0076] In one embodiment, a three-dimensional virtual simulation flood prediction and early warning system is provided, comprising the following modules, wherein: The data acquisition module is used to acquire topographic data and rainfall time series data of the target area; An evolution prediction module is used to generate multi-time-step prediction results of flood evolution based on the terrain data and the rainfall time series data, wherein the multi-time-step prediction results include water depth distribution; The peak prediction module is used to generate flood peak prediction results based on the rainfall time series data and the multi-time step prediction results; The 3D visualization and early warning module is used to present the multi-time step prediction results and the flood peak prediction results in 3D on a 3D visualization platform, and output early warning information based on preset threshold rules.

[0077] In this embodiment, the 3D virtual simulation flood prediction and early warning system includes a data acquisition module, an evolution prediction module, a peak prediction module, and a 3D visualization and early warning module. The data acquisition module acquires topographic data and rainfall time-series data for the target area. The topographic data can be a DEM file or raster data, and the rainfall time series can be generated from historical data, weather forecasts, or real-time observation data. The data acquisition module can also provide data preprocessing functions, such as coordinate unification, grid resampling, missing value imputation, and time step alignment.

[0078] The evolution prediction module is used to generate multi-timestep water depth distributions based on terrain data and rainfall time series. The module can further include sub-modules for terrain feature extraction, terrain feature sequence construction, rainfall spatialization, weighted fusion, ConvLSTM temporal recursion, and decoding output. The output of the evolution prediction module can be a water depth distribution sequence and, optionally, a positive alternating current velocity component sequence.

[0079] The peak prediction module generates flood peak prediction results based on rainfall time series and multi-time step prediction results. Internally, the module may include a peak input feature construction submodule and a conditional generation model submodule. The peak input feature construction submodule extracts rainfall statistical features and convergent features of the evolution process to form an input vector; the conditional generation model submodule performs conditional generation inference, and the conditional generation model can be a conditional diffusion model, performing progressive denoising iterations to output peak water level and / or peak water depth.

[0080] The 3D visualization and early warning module is used to present multi-time-step prediction results and peak prediction results in 3D and output early warning information. Internally, the module may include a 3D mapping and rendering submodule, a timeline playback submodule, a threshold rule management submodule, a statistical extraction submodule, a threshold comparison and grading submodule, and an early warning display association submodule. The threshold rule management submodule stores and updates the mapping between thresholds and early warning levels; the statistical extraction and threshold comparison submodule corresponds to the extraction and comparison process of statistical / peak indicators; and the early warning display association submodule associates and displays early warning levels with 3D visuals, statistical curves, or labels.

[0081] This embodiment integrates data acquisition, flood evolution prediction, flood peak prediction, and 3D visualization and threshold early warning functions into the same system in a modular manner. It can complete the entire process from input data to prediction, presentation and early warning output under a unified system architecture, which is convenient for engineering deployment and business integration.

[0082] The specific implementation details of each module can be found in the above description of the limitations of the three-dimensional virtual simulation flood prediction and early warning method, and will not be repeated here.

[0083] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A three-dimensional virtual simulation flood prediction and early warning method, characterized in that, The method includes: Acquire topographic data and rainfall time-series data for the target area; Based on the terrain data and the rainfall time series data, a multi-time step prediction result of flood evolution is generated, and the multi-time step prediction result includes water depth distribution; Based on the rainfall time series data and the multi-time step prediction results, a flood peak prediction result is generated; The multi-time-step prediction results and the flood peak prediction results are presented in three dimensions in a three-dimensional visualization platform, and early warning information is output based on preset threshold rules.

2. The method according to claim 1, characterized in that, The generation of multi-time-step flood evolution prediction results based on the terrain data and the rainfall time series data includes: Feature extraction is performed on the terrain data to obtain a terrain feature representation; The terrain feature representation is constructed into a terrain feature sequence that is consistent with the time step of the rainfall time series data; The rainfall time series data is represented in a spatiotemporal manner to obtain a rainfall feature sequence; The terrain feature sequence and the rainfall feature sequence are weighted and fused to obtain a spatiotemporal feature sequence; The spatiotemporal feature sequence is input into a convolutional recurrent neural network to obtain the multi-time-step prediction result.

3. The method according to claim 2, characterized in that, The step of extracting features from the terrain data to obtain a terrain feature representation includes: The terrain data is identified as digital elevation model data; The digital elevation model data is input into a terrain feature extraction network to obtain a terrain feature map, wherein the terrain feature extraction network includes convolutional layers, nonlinear activation layers, and downsampling layers; The terrain feature map is determined as the terrain feature representation.

4. The method according to claim 2, characterized in that, The step of constructing the terrain feature representation into a terrain feature sequence consistent with the time step of the rainfall time series data includes: The target time step length is determined based on the number of time steps in the rainfall time series data; The terrain feature representation is copied or broadcast along the time dimension to obtain a terrain feature sequence of the target time step length.

5. The method according to claim 2, characterized in that, The process of performing spatiotemporal representation on the rainfall time series data to obtain a rainfall feature sequence includes: The rainfall values ​​at each time step are expanded into a rainfall feature map that is consistent with the spatial grid corresponding to the terrain feature representation; Based on the rainfall feature maps corresponding to each of the time steps, a multi-time-step rainfall feature sequence is obtained.

6. The method according to claim 2, characterized in that, The step of inputting the spatiotemporal feature sequence into a convolutional recurrent neural network to obtain the multi-time-step prediction result includes: The spatiotemporal feature sequence is input into a convolutional recurrent neural network in time step order. The convolutional recurrent neural network includes at least two layers of convolutional long short-term memory network (ConvLSTM) to update the hidden state at each time step and output the spatiotemporal feature representation at each time step. The spatiotemporal feature representations of each time step are decoded to obtain the water depth prediction map corresponding to each time step. The water depth prediction maps at each time step are combined to form a multi-time step water depth distribution corresponding to the spatial grid of the target area, which is used as the multi-time step prediction result.

7. The method according to claim 6, characterized in that, The decoding process of the spatiotemporal feature representations at each time step to obtain the water depth prediction map corresponding to each time step includes: Spatial resolution restoration processing is performed on the spatiotemporal feature representations of each time step to obtain the processed features corresponding to each time step; wherein, the spatial resolution restoration processing includes deconvolution upsampling and / or interpolation upsampling; Channel mapping is performed on the processed features corresponding to each time step to obtain the water depth prediction map corresponding to each time step, as well as the first direction velocity component prediction map and the second direction velocity component prediction map corresponding to the water depth prediction map. The first direction velocity component and the second direction velocity component are orthogonal to each other.

8. The method according to claim 1, characterized in that, The process of generating flood peak prediction results based on the rainfall time series data and the multi-time step prediction results includes: Based on the rainfall time series data and the multi-time step prediction results, peak prediction input features are obtained; The peak prediction input features are used to generate a model, which outputs flood peak prediction results, including peak water level and / or peak water depth. Specifically, when the conditional generation model is a conditional diffusion model, the flood peak prediction result is obtained by performing stepwise denoising iterations based on the peak prediction input features.

9. The method according to claim 1, characterized in that, The output of early warning information based on preset threshold rules includes: Obtain the preset threshold rules corresponding to the target area, wherein the preset threshold rules include at least one type of threshold and the warning level corresponding to the threshold; The water depth statistics for each time step are extracted from the multi-time step prediction results. The water depth statistics include the maximum water depth, average water depth, and the percentage of grid cells exceeding the water depth threshold in the target area. Extract peak indexes from the flood peak prediction results; The water depth statistics and / or the peak index are compared with the preset threshold rule to obtain the comparison result and the warning level corresponding to the comparison result; Output the warning information corresponding to the warning level in the three-dimensional visualization platform, and associate the warning information with the multi-time step prediction results and / or the flood peak prediction results for display.

10. A three-dimensional virtual simulation flood prediction and early warning system, characterized in that, The system includes: The data acquisition module is used to acquire topographic data and rainfall time series data of the target area; An evolution prediction module is used to generate multi-time-step prediction results of flood evolution based on the terrain data and the rainfall time series data, wherein the multi-time-step prediction results include water depth distribution; The peak prediction module is used to generate flood peak prediction results based on the rainfall time series data and the multi-time step prediction results; The 3D visualization and early warning module is used to present the multi-time step prediction results and the flood peak prediction results in 3D on a 3D visualization platform, and output early warning information based on preset threshold rules.

Images