A multi-source data dynamic risk early warning method and system based on ST-GAN

Abstract

The application discloses a kind of based on ST-GAN's multi-source data dynamic risk early warning method and system, the method includes the following steps: S1. constructing the spatiotemporal generation confrontation network suitable for spatiotemporal data characteristics, utilize the network to generate extreme precipitation data, realize spatiotemporal unbalanced data reduction, obtain the equalization data of final output;S2. the equalization data of final output with high-dimensional space variable is carried out multi-source data fusion, and nonlinear spatiotemporal information conversion equation, and by local linearization obtains linear approximation model, to predict future time series;S3. under the present situation that extreme rainfall data amount is relatively insufficient, neural network based on dual learning theory accurately learns the parameter of nonlinear spatiotemporal conversion, estimates extreme weather event.The application is through spatiotemporal generation confrontation network (ST-GAN) and multi-source data fusion engine, significantly improves the precision and efficiency of natural disaster warning.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of natural disaster monitoring and early warning, and particularly relates to a multi-source data dynamic risk early warning method and system based on ST-GAN. BACKGROUND

[0002] In the prior art, natural disaster prediction and early warning mainly faces the following problems: data imbalance: extreme weather and geological disaster data are seriously uneven in space-time distribution, resulting in insufficient prediction accuracy of traditional models. Difficulties in multi-source data fusion: multi-source data such as radar, remote sensing and ground observation have strong heterogeneity, which is difficult to effectively integrate. Poor real-time performance: the existing model has high computational complexity, which is difficult to meet the rapid response demand after an earthquake or extreme weather. Lack of physical mechanism: pure data-driven models lack explainability for disaster physical processes, affecting the reliability of early warning. SUMMARY

[0003] The purpose of the present application is to overcome the shortcomings of the prior art and provide a multi-source data dynamic risk early warning method and system based on ST-GAN, which significantly improves the accuracy and efficiency of natural disaster early warning through a spatio-temporal generative adversarial network (ST-GAN) and a multi-source data fusion engine.

[0004] The purpose of the present application is achieved by the following technical solution: a multi-source data dynamic risk early warning method based on ST-GAN, comprising the following steps:

[0005] S1. Construct a spatio-temporal generative adversarial network suitable for spatio-temporal data characteristics, generate extreme precipitation data using the network, realize spatio-temporal unbalanced data reduction, and obtain the final output of balanced data;

[0006] S2. Perform multi-source data fusion on the final output of balanced data and high-dimensional space variables, and perform non-linear spatio-temporal information conversion equation, and obtain a linear approximation model through local linearization to predict future time series;

[0007] S3. In the current situation of relatively insufficient extreme rainfall data, a neural network based on dual learning theory accurately learns the parameters of nonlinear spatio-temporal conversion to estimate extreme weather events.

[0008] A multi-source data dynamic risk early warning system based on ST-GAN, comprising:

[0009] The balanced processing module is configured to construct a spatio-temporal generative adversarial network suitable for spatio-temporal data characteristics, generate extreme precipitation data using the network, realize spatio-temporal unbalanced data reduction, and obtain the final output of balanced data;

[0010] The multi-source data fusion module is used to fuse the final output equalized data with high-dimensional spatial variables. It is a nonlinear spatiotemporal information transformation equation, and obtains a linear approximation model through local linearization to predict future time series.

[0011] The dual training module is used to accurately learn the parameters of nonlinear spatiotemporal transformation based on the dual learning theory of neural networks, given the relatively insufficient amount of extreme rainfall data, and to predict extreme weather events and provide risk warnings.

[0012] The beneficial effects of this invention are: improved prediction accuracy: by using a spatiotemporal generative adversarial network (ST-GAN) to equalize extreme rainfall and geological disaster data, the problem of uneven spatiotemporal distribution is effectively solved, and the prediction reliability of the model under sparse data conditions is significantly improved.

[0013] 2. Enhance multi-source data fusion capabilities: Utilize dynamic equations (STI equations) and attention mechanisms to effectively integrate heterogeneous data from multiple sources such as radar, remote sensing, and ground monitoring, thereby improving the comprehensiveness of feature extraction.

[0014] 3. Strong generalization ability: It exhibits stable predictive performance in test areas with different geological conditions, and is applicable to multiple disaster scenarios such as flash floods, landslides, and debris flows.

[0015] 4. Significantly improves real-time performance: By adopting a physical information neural network proxy model to replace traditional numerical simulation, the prediction time for geological disasters is greatly shortened, meeting the needs of rapid post-disaster response. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention;

[0017] Figure 2 This is a schematic diagram of a multi-source data dynamic risk early warning system. Detailed Implementation

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0019] This invention constructs a multi-source data dynamic early warning model using a spatiotemporal generative adversarial network (ST-GAN), combining physical mechanisms and data-driven approaches to achieve full-chain processing of "extreme rainfall forecasting - geological disaster prediction - comprehensive early warning," specifically:

[0020] like Figure 1 As shown, a multi-source data dynamic risk early warning method based on ST-GAN includes the following steps:

[0021] S1. Construct a spatio-temporal generative adversarial network adaptive to the characteristics of spatio-temporal data, generate extreme precipitation data by using the network, realize spatio-temporal unbalanced data reduction, and obtain the final output of balanced data;

[0022] S101. Given the time series data of extreme rainfall as the original unbalanced data:

[0023]

[0024] S102. Adopt an encoder-decoder architecture for sequence reconstruction:

[0025] The encoder compresses the input data into hidden space features :

[0026] ;

[0027] wherein is the historical spatio-temporal sequence data, each contains observation values of multiple spatial positions;

[0028] The decoder reconstructs the data based on the hidden space features

[0029] wherein , are the encoder and the decoder, respectively; and are the learnable parameters of the encoder and the decoder, respectively; for spatio-temporal sequence reconstruction, the loss function is , the first term represents the mean square error loss, which ensures that the reconstructed data is consistent with the true data in the time dimension, and the second term is a spatial smoothing regular term, which forces the generated data to maintain geographical continuity in the spatial dimension; is the number of spatial positions, is the regularization coefficient;

[0030] The loss function is used to pre-train the encoder-decoder to learn the reconstruction of spatio-temporal sequences; during pre-training, multiple training samples are constructed from the complete extreme rainfall time series data, each sample being a continuous historical sequence and its corresponding real data at the next time ; the parameters of the encoder-decoder obtained by pre-training are used to initialize the generative adversarial network;

[0031] , represent the observation values of the i-th spatial position contained in , ​

[0032] S103. Construct and train the generative adversarial model:

[0033] A1. Design of the generator:

[0034] Given random noise ;

[0035] Generate historical hidden state from historical time series data by encoder ;

[0036] Give the signal generation process of the generator:

[0037] ;

[0038] where, : generated hidden space feature, : network parameters of the generator, g is a neural network function;

[0039] Output synthetic rainfall data:

[0040] Reconstruct by decoder, output synthetic rainfall data ;

[0041] A2. Design of the discriminator:

[0042] Input: real data or generated data ;

[0043] Discrimination process: D( ; )→[0,1] : parameters of the discriminator; where, is the input of the discriminator, taking real data or generated data ;

[0044] Output probability value: 1 = real, 0 = generated;

[0045] A3. Design of the loss function:

[0046] Generator loss: ;

[0047] where, is the synthetic rainfall data generated by the generator and reconstructed by the decoder, is the probability that the discriminator judges the generated data as "real"; the goal of the generator is to maximize this probability, so it is embodied in the loss function as minimizing its negative logarithm;

[0048] Discriminator loss: ​

[0049] where, is the rainfall data from the real data distribution, is the synthetic rainfall data generated by the generator and reconstructed by the decoder, is the output function of the discriminator; the goal of the discriminator is to maximize the accuracy of judging real data as real and generated data as false , so it is embodied in the loss function as minimizing its negative log-likelihood;

[0050] A4, construct a set of time series data composed of the data set, and perform generative adversarial training. In each training process, the generative adversarial training model is updated based on the loss function, and a trained generative adversarial model is obtained.

[0051] S104. Through the above generative adversarial training, the decoder finally outputs the synthetic t-time rainfall data , which together with the original data at t-time forms an equalization data set, and after preprocessing, an equalization data is obtained.

[0052] Output synthetic rainfall data and original data together form an equalization data set, and the specific implementation process includes: spatio-temporal blank filling: in the sparse area of historical data (such as rare rainstorm area), the generator generates synthetic rainfall events in accordance with the physical correlation between terrain slope and radar reflectivity, which conforms to meteorological rules; extreme event enhancement: for low-frequency extreme events (such as once-in-a-century rainstorm), generate samples of multiple intensity levels by adjusting noise distribution ; physical consistency constraint: the discriminator forces to meet the reality constraints (such as slope > 25° area rainfall > 30mm / h), and rejects the generated results that violate the laws of fluid mechanics; the final output of the equalization data will be used as the input of the multi-source fusion engine.

[0053] S2. Perform multi-source data fusion on the final output of the equalization data and high-dimensional spatial variables, and construct a differential equation;

[0054] Input: equalization data and high-dimensional spatial variables, where the high-dimensional spatial variables include atmospheric pressure, humidity;

[0055] Processing:

[0056] Organize the equalized extreme rainfall data into a space-time data matrix, and construct the spatial vector as where, represents the equalized rainfall at time point t and spatial position i, and D is the total number of spatial positions;

[0057] The time vector is constructed as wherein represents the spatial aggregated rainfall at time point , which is obtained by averaging the equalized rainfall at time point for all spatial locations, i.e. , and L is the size of the time window.

[0058] The high-dimensional spatial variables are fused with the equalized rainfall data to construct an enhanced spatial vector:

[0059]

[0060] wherein is the atmospheric pressure spatial vector, is the humidity spatial vector.

[0061] Based on the above data matrix, a nonlinear spatiotemporal information conversion equation is constructed:

[0062]

[0063] wherein Φ and Ψ are both nonlinear differentiable functions, the first equation is the elementary form of the STI equation, and the second equation is the conjugate form of the STI equation. represents the spatial vector of multiple variables at time point t, and represents the time vector of variable y at multiple time points t, t + 1, …, t + L−1; I represents the unit function.

[0064] Φ and Ψ in the spatiotemporal information conversion equation are both nonlinear differentiable functions, which are locally linearized to obtain a linear approximation model:

[0065]

[0066] The linearization approximation model approximates the mapping functions Φ and Ψ by fitting matrices A and B, specifically:

[0067] Using the generated equalized dataset , matrices A and B are solved by least squares method or the like, so that and ; by solving A and B, a linear transformation is used to approximately predict the future time series .

[0068] wherein AB = I, A and B are L×D and D×L matrices respectively, L is the time dimension, D is the spatial dimension, and I represents the L×L unit matrix.

[0069] S3. In the current situation of relatively insufficient extreme rainfall data, the neural network based on the dual learning theory can accurately learn the differentiable functions Φ and Ψ of nonlinear space-time transformation, and predict extreme weather events.

[0070] Dual learning is a machine learning method that aims to improve model performance by learning from two related tasks. In the current situation of relatively insufficient extreme rainfall data, the neural network based on the dual learning theory can accurately learn the differentiable functions Φ and Ψ of nonlinear space-time transformation, and help to reasonably and deeply explain the formation mechanism and influencing factors of extreme rainfall, and accurately predict extreme weather events. As shown in the figure, the space-time converter CTS and the time-space converter CST are both implemented in the form of neural networks. The goal of the time-space converter is to convert the input spatial information X into corresponding temporal information Y, while the space-time converter aims to convert the temporal information Y back into corresponding spatial information X. According to the space-time information conversion equation, ideally, the two conversion methods are reciprocal. That is, the output obtained by applying the space-time converter after applying the time-space converter should remain consistent with the original input, specifically:

[0071] S301. Constructing space-time converter CTS and time-space converter CST:

[0072] The input of the time-space converter CST is spatial information , and the output is corresponding temporal information , that is where is the enhanced spatial vector constructed in step S2 , is the corresponding temporal vector ;

[0073] The input of the time-space converter CTS is temporal information , and the output is corresponding spatial information ;

[0074] According to the space-time information conversion equation, ideally, the two conversion methods are reciprocal, that is:

[0075]

[0076]

[0077] S302. Improve model performance by mutual learning from two related tasks through dual learning:

[0078] The training data consists of the equalized data set obtained by steps S1 and S2, and each training sample includes: for the time-space converter CST, the output is X, and the label is Y; for the space-time converter CTS, the input is Y, and the label is X;

[0079] The two converters are jointly trained by minimizing the following loss function:

[0080]

[0081] wherein is the temporal transformer reconstruction loss, is the spatio-temporal transformer reconstruction loss; is the cycle consistency loss, is the weight coefficient of the cycle consistency loss;

[0082] Under the current situation of relatively insufficient extreme rainfall data, based on the dual learning theory, the two converters provide mutual supervision signals, so as to accurately learn the differentiable functions Φ and Ψ of nonlinear spatio-temporal transformation, and obtain a nonlinear spatio-temporal information conversion equation based on the differentiable functions, which is used for predicting extreme weather events and conducting risk warning.

[0083] A multi-source data dynamic risk warning system based on ST-GAN, comprising:

[0084] A balanced processing module is configured to construct a spatio-temporal generative adversarial network suitable for the characteristics of spatio-temporal data, generate extreme precipitation data by using the network, realize spatio-temporal unbalanced data reduction, and obtain balanced data as a final output.

[0085] A multi-source data fusion module is configured to perform multi-source data fusion on the balanced data as the final output and high-dimensional spatial variables, and convert nonlinear spatio-temporal information by using a nonlinear spatio-temporal information conversion equation, and obtain a linear approximation model by local linearization to predict future time series.

[0086] A dual training module is configured to accurately learn parameters (differentiable functions Φ and Ψ) of nonlinear spatio-temporal transformation by using a neural network based on the dual learning theory under the current situation of relatively insufficient extreme rainfall data, and predict extreme weather events and conduct risk warning.

[0087] In the embodiments of the present application, the above model can be integrated into a GIS platform to publish warning information in real time.

[0088] As Figure 2 shown, the multi-source data dynamic risk warning system of the present application adopts a three-layer architecture (as shown in the figure) of “data acquisition→intelligent processing→warning publication”, and the specific working principle is as follows:

[0089] 2.1 Data acquisition layer: input source: satellite remote sensing, weather radar, ground sensors (such as seismographs, rain gauges), historical disaster database. Function: real-time acquisition of multi-source heterogeneous data (such as rainfall time series, terrain raster, lithology distribution), and transmission to the data center through the Internet of Things (IoT).

[0090] 2.2 Intelligent processing layer: core module: ST-GAN model: solve the problem of spatio-temporal imbalance of data (generate synthetic data to fill in missing areas).

[0091] 2.3 Multi-source fusion engine: combine dynamic equation (such as formula AX'=Y') and physical information neural network, associate meteorological, geological and topographic data.

[0092] 2.4 Early warning model: based on Mamba algorithm (small sample optimization) and double attention mechanism (strong earthquake prediction), output disaster probability heat map.

[0093] 2.5 Early warning release layer: function: through the GIS three-dimensional visualization platform, real-time push early warning information to the government end (emergency command system) and the public end (mobile phone APP).

[0094] Dynamic update: according to the real-time monitoring data, the prediction results are corrected.

[0095] The application significantly improves the accuracy and efficiency of natural disaster warning through spatio-temporal generative adversarial network (ST-GAN) and multi-source data fusion engine. Compared with the prior art, the traditional method relies on historical statistical data and isolated model, which leads to problems such as inaccurate prediction in small sample area, difficulty in multi-source data fusion, slow response after earthquake, etc. The application innovatively introduces ST-GAN to generate synthetic data, filling the data gap of extreme events; through the spatio-temporal differential equation, the physical consistency of meteorological and geological data is forced, avoiding the subjectivity of artificial threshold setting; and the physical information neural network proxy model is used to replace the high-cost numerical simulation, shortening the prediction time of strong earthquake-induced geological disasters from hours to minutes. At the same time, the double attention mechanism visualizes the key disaster-causing factors, enhancing the model's explainability and providing scientific basis for disaster prevention and control.

Claims

1. A method for dynamic risk early warning of multi-source data based on ST-GAN, characterized in that: The method comprises the following steps: S1. Constructing a spatio-temporal generative adversarial network adaptive to the characteristics of spatio-temporal data, generating extreme precipitation data by using the network, reducing spatio-temporal unbalanced data, and obtaining balanced data as final output; The step S1 comprises: S101. Given time series data of extreme rainfall , as raw non-equilibrium data: = , ,…., ； S102. Adopting an encoder-decoder architecture for sequence reconstruction: An encoder compresses input data into latent space features : ； wherein, = , ,…., is historical spatio-temporal sequence data, each contains observation values for multiple spatial locations;​ Decoder based on latent space features Reconstruction data ; In the formula , are respectively an encoder and a decoder; and are respectively a learnable parameter of the encoder and the decoder; for spatio-temporal sequence reconstruction, the loss function is , the first term represents a mean square error loss, ensuring that the reconstructed data is consistent with the real data in the time dimension, and the second term is a spatial smoothing regular term, forcing the generated data to maintain geographical continuity in the spatial dimension; is the number of spatial positions, is a regularization coefficient; Using a loss function The encoder-decoder is pre-trained to learn the reconstruction of spatio-temporal sequences; during pre-training, multiple training samples are constructed from the complete extreme rainfall time series data, each sample being a continuous historical sequence and its corresponding real data of the next moment The parameters of the encoder-decoder obtained by pre-training are used to initialize the generative adversarial network; , denotes , the observation value of the i-th spatial position comprised in the vector S103. Constructing and training a generative adversarial model: A1. Design of the generator: Given random noise N(0, 1); by the encoder from the historical time series data generating a historical hidden state: ; The signal generation process of the generator is given: ； wherein, : generated latent space feature, : network parameters of the generator, g is a neural network function; Output synthetic rainfall data: reconstructing, by the decoder, output synthetic rainfall data ; A2. Design of the discriminator: Input: real data or generated data ; Discrimination process: D (x) → [0, 1] ; ) where x is the input of the discriminator, taking real data or generated data ; and is the parameter of the discriminator. ​ Output probability value: 1 = real, 0 = generated; A3. Design of the loss function: Generator loss: ; wherein, is synthetic rainfall data produced by the generator and reconstructed by the decoder, is the probability that the discriminator judges the generated data as "real"; the goal of the generator is to maximize this probability, thus embodied in the loss function as minimizing the negative logarithm thereof; Discriminator loss: ; in, It is rainfall data derived from real data distribution. It is synthetic rainfall data generated by a generator and reconstructed by a decoder. It is the output function of the discriminator; the goal of the discriminator is to maximize the accuracy of judging true data. And judging the falsity of generated data The accuracy is therefore reflected in the loss function as minimizing its negative log-likelihood; A4、constructing multiple sets of time series data The generated data set is subjected to generative adversarial training, and the generative adversarial training model is updated based on a loss function in each training process to obtain a trained generative adversarial model. S104. Through the above generative adversarial training, the decoder finally outputs the synthesized rainfall data at time t , which together with the original data at time t comprises the balanced data set, and after preprocessing, the balanced data is obtained; S2. Multi-source data fusion of the balanced data as final output and high-dimensional spatial variables, construction of a nonlinear spatio-temporal information conversion equation, and obtaining a linear approximation model through local linearization to predict future time series; The step S2 comprises: Input: balanced data and high-dimensional spatial variables, wherein the high-dimensional spatial variables include atmospheric pressure, humidity; Processing: The equalized extreme rainfall data is organized into a spatial-temporal data matrix, with the spatial vector constructed as wherein, denotes the equalized rainfall at spatial location i at time point t, and D is the total number of spatial locations. The time vector is constructed as wherein denotes the spatial aggregated rainfall at the time point is obtained by averaging the homogenized rainfall at the time point over all spatial locations, i.e. L is the size of the time window. Fuse the high-dimensional spatial variables and the balanced rainfall data to construct an enhanced spatial vector: ； wherein is the atmospheric pressure space vector, is the humidity space vector; Based on the above data matrix, a nonlinear spatio-temporal information conversion equation is constructed: ； where Φ and Ψ are both non-linear differentiable functions, the first equation is the elementary form of the STI equation, and the second equation is the conjugate form of the STI equation; denotes a spatial vector of multiple variables at time point t, and denotes a temporal vector of variable y at multiple time points t, t + 1, …, t + L−1; I denotes an identity function; Both Φ and Ψ in the spatio-temporal information conversion equation are nonlinear differentiable functions, which are locally linearized to obtain a linear approximation model: ； The linearization approximation model approximates the mapping functions Φ and Ψ by fitting matrices A and B, specifically: Using the generated equalized dataset , solve matrix A and B by least square method etc. so that and ; using linear transformation to approximate the prediction of future time series ; Where, AB = I, A and B are LxD and DxL matrices respectively, L is the time dimension, D is the spatial dimension, and I represents the LxL unit matrix; S3. Under the current situation of relatively insufficient extreme rainfall data, a neural network based on dual learning theory accurately learns the parameters of nonlinear spatio-temporal conversion to estimate extreme weather events.

2. The method of claim 1, wherein the method is based on a ST-GAN. The balanced data obtained after preprocessing comprises: Spatio-temporal blank filling: In the sparse area of historical data, the generator generates synthetic rainfall events in accordance with the physical correlation between terrain slope and radar reflectivity, in accordance with meteorological rules; Extreme event enhancement: for low frequency extreme events, by adjusting the noise distribution N( , ) generate multiple intensity levels samples; Physical consistency constraints: enforcer forces Satisfy reality constraints, reject generated results that violate laws of fluid mechanics.

3. The method of claim 1, wherein the method is based on a ST-GAN. The step S3 comprises: S301. Constructing a spatio-temporal converter CTS and an astro-temporal converter CST: The input of the space-time converter CST is a spatial information , the output is a corresponding time information , i.e. wherein is the enhanced spatial vector constructed in step S2 , is a corresponding time vector ; The input of the space-time converter CTS is temporal information and the output is corresponding spatial information ; According to the space-time information conversion equation, in the ideal case, the two conversion methods are reciprocal, that is: ； ； S302. Improve the model performance by mutual learning from two related tasks through dual learning: The training data consists of the balanced data set obtained by steps S1 and S2, and each training sample includes: for the astro-temporal converter CST, the output is X and the label is Y; for the spatio-temporal converter CTS, the input is Y and the label is X; The two converters are jointly trained by minimizing the following loss function: ； wherein is a temporal transformer reconstruction loss, is a temporal transformer reconstruction loss; is a cycle consistency loss, is a weight coefficient of the cycle consistency loss; Under the current situation of relatively insufficient extreme rainfall data, based on the dual learning theory, the two converters provide supervision signals to each other, thereby accurately learning the differentiable functions Φ and Ψ of nonlinear spatio-temporal conversion, obtaining a nonlinear spatio-temporal information conversion equation based on the differentiable function, and used for estimating extreme weather events and conducting risk warning.

4. A multi-source data dynamic risk early warning system based on ST-GAN, adopting the method of any one of claims 1-3. It comprises: A balancing processing module for constructing a spatio-temporal generative adversarial network adaptive to the characteristics of spatio-temporal data, generating extreme precipitation data by using the network, reducing spatio-temporal unbalanced data, and obtaining balanced data as final output; A multi-source data fusion module is configured to perform multi-source data fusion on the equalized data and high-dimensional spatial variables of the final output, construct a nonlinear spatiotemporal information conversion equation, and obtain a linear approximation model through local linearization to predict future time series; A dual training module is configured to accurately learn parameters of nonlinear spatiotemporal conversion based on a neural network of a dual learning theory in the current situation of relatively insufficient extreme rainfall data, estimate an extreme weather event, and perform risk early warning.

Images