Anomaly warning method of sparse spatial wind vector field based on multi-source heterogeneous spatiotemporal data

Abstract

The present invention provides a sparse space wind vector field anomaly warning method based on multi-source heterogeneous spatiotemporal data. It includes: step S1: acquisition and spatiotemporal standardization of multi-source heterogeneous wind vector data; step S2: construction and weight update of multi-scale geographic spatiotemporal weighted autoregressive model; step S3: decomposition of vector field and construction of regional warning pilot feature set; step S4: boundary flux calculation and PINN-GPR fitting of sparse spatial connected domain; step S5: four-dimensional spatiotemporal data variational assimilation and update of sparse space boundary flux; step S6: abnormal wind field warning based on regional warning pilot feature matching. The present invention can realize the abnormal warning of sparse space wind vector field and adaptive update of warning library.

Description

Technical Field

[0001] The present invention belongs to the field of meteorological science and technology and data science and technology, and in particular relates to a sparse spatial wind vector field anomaly early warning method based on multi-source heterogeneous spatiotemporal data. Background Art

[0002] In the field of meteorological science, especially in the development of wind vector field analysis and early warning systems, there are some key challenges, which mainly stem from the difficulty of data acquisition, the inconsistency of data quality and the complexity of data processing. With the advancement of technology, the use of multi-source heterogeneous spatiotemporal data has become an important way to solve these problems.

[0003] In vast geographic areas or inaccessible areas, traditional meteorological observation equipment may not have comprehensive coverage, resulting in sparse wind vector data. This sparsity limits the accuracy and reliability of meteorological models in these areas.

[0004] Modern weather forecasting involves collecting data from multiple data sources (such as ground stations, satellites, aircraft, sounding balloons, etc.). These data sources have significant differences in temporal resolution, spatial resolution, measurement technology, and data format, and integrating these heterogeneous data has become a major challenge. Rapidly processing large amounts of real-time meteorological data and generating accurate warning information is the key to improving the efficiency of meteorological services and reducing the risks of weather-related disasters. However, existing technologies are often limited in data processing speed, warning response time, and application space. Summary of the invention

[0005] In view of this, the present invention aims to overcome the shortcomings of the above-mentioned problems in the prior art and proposes a sparse spatial wind vector field anomaly warning method based on multi-source heterogeneous spatiotemporal data. The method can integrate multi-source heterogeneous spatiotemporal data and enhance the wind vector field anomaly warning capability in data sparse space.

[0006] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0007] The first aspect of the present invention provides a sparse spatial wind vector field anomaly warning method based on multi-source heterogeneous spatiotemporal data, comprising the following steps:

[0008] Step S1: Acquire multi-source heterogeneous spatiotemporal wind vector data, and perform scale division and multi-scale data grid standardization according to the spatial resolution and temporal resolution characteristics of these data;

[0009] Step S2: construct a multi-scale geographic spatiotemporal weighted autoregressive model based on geographic distribution characteristics, combine static characteristics, periodic characteristics and time-varying characteristics to construct a benchmark model and its periodic fluctuation characteristics and opportunity data attenuation characteristics, and determine the weight update model;

[0010] Step S3: Decomposing the vector field into a divergence field and a curl field using the Helmholtz theorem, and constructing a regional wind vector field abnormal warning leading feature set based on the features of historical abnormal events in the divergence field and the curl field in the target area;

[0011] Step S4: Calculate the wind field flux at the boundary grid of the sparse spatial connected domain, use the physical information neural network to construct fluid dynamics constraints and atmospheric science prior knowledge constraints, and use the Gaussian process regression to fit the divergence and curl features in the sparse spatial connected domain;

[0012] Step S5: performing variational assimilation of four-dimensional space-time data and updating of boundary fluxes in sparse space. In the variational assimilation of four-dimensional space-time data, based on the observation data at the current moment and the response of the model parameters at the previous moment at the current moment, combined with the fluid dynamics constraints of the physical information neural network, the boundary fluxes of each connected domain in the sparse space are corrected;

[0013] Step S6: Perform abnormal wind field warning based on regional warning leading feature matching, match the current wind field features with the warning leading feature set, and according to the matching degree threshold, when the abnormal probability of the wind field sample exceeds the threshold, the system issues a warning.

[0014] Furthermore, the step S1 specifically includes:

[0015] Acquisition of multi-source heterogeneous spatiotemporal wind vector data including data provided by land-based observation equipment, airborne observation equipment, meteorological satellites, sounding balloons and other observation equipment capable of measuring wind speed and direction;

[0016] According to the spatial resolution characteristics of various types of data, the data are divided into multiple levels of high resolution, medium resolution and low resolution based on the observation accuracy;

[0017] In terms of time resolution, the data is divided into high-frequency sampling data, low-frequency sampling data, and opportunity data according to the sampling frequency and the length of the opportunity window. The opportunity data refers to the non-periodic observation data collected within the event window when the opportunity event occurs, including the observation data obtained by the airborne observation equipment on the route in the target area and the observation data collected by the sounding balloon.

[0018] Within each scale level, the spatial and temporal resolutions are rasterized and the observation data are redistributed according to preset grid cells to achieve the fusion of different data sources on a unified spatial and temporal scale, to standardize multi-scale data and to generate multi-scale data grids with a unified format.

[0019] Furthermore, the step S2 specifically includes:

[0020] Based on the geographical distribution characteristics of various types of observation equipment in the target area, including the geographical location distribution of land-based observation equipment, the orbital characteristics of meteorological satellites and their relative position relationship with the ground observation area, and the spatial layout of routes in the region, a multi-scale geographic spatiotemporal weighted autoregressive model is constructed;

[0021] The benchmark model is designed by combining multiple features, including static features, periodic features and time-varying features; the static features refer to the spatial features reflected by the observation performance related to the geographical location of the observation equipment and its observation range, which are relatively static; the periodic features refer to the periodic changes of the observation data over time; and the time-varying features refer to the dynamic characteristics of the observation data over time;

[0022] Based on the above characteristics, the periodic fluctuation characteristics and opportunity data decay characteristics of the benchmark model are constructed.

[0023] Furthermore, the step S3 specifically includes:

[0024] Based on the Helmholtz theorem, the wind vector field is decomposed into two independent components, the divergence field and the curl field, to achieve a refined analysis of the wind field characteristics. Any smooth vector field can be uniquely decomposed into an irrotational divergence field and a passive curl field, and the divergence and convergence of the airflow in the wind field, as well as the rotation and vortex characteristics of the airflow are studied respectively:

[0025]

[0026] Among them, U(X) is the wind vector field, which is smooth in three-dimensional space; is the gradient operator, which represents the total differential in all directions of space; is the divergence field, representing the divergence of the scalar potential function φ(X); is the curl field, expressed as a vector potential function A(X);

[0027] Based on the performance of historical abnormal wind events in the target area, the features in the divergence field and the curl field are extracted to construct a leading feature set for regional wind vector field abnormal warning. Specifically, by analyzing the divergence field features and curl field features related to abnormal wind events in historical data, the performance patterns of these events after wind field decomposition are identified.

[0028] Furthermore, the step S4 specifically includes:

[0029] Based on the reconstructed regional wind vector field, the wind field flux is calculated at the boundary grid of the sparse spatial connected domain, wherein the sparse spatial connected domain refers to a specific area formed after the multi-source heterogeneous spatiotemporal data is reconstructed based on the multi-scale geographic spatiotemporal weighted aggregation;

[0030] Physical constraints are constructed using physical information neural networks to ensure that the network output is consistent with the observed data and satisfies the physical laws and atmospheric science prior knowledge used for constraints. The physical constraints include but are not limited to the Navier-Stokes equations of fluid dynamics and the continuity equation:

[0031]

[0032] Among them, v is the fluid velocity vector field; ρ is the fluid density; p is the pressure field; μ is the dynamic viscosity coefficient; f is the external force density;

[0033] In rectangular coordinates it is:

[0034]

[0035] Among them, u, v, w are the velocity components of the fluid at point (x, y, z) at time t;

[0036] Use Gaussian process regression to fit the divergence and curl features in the sparse spatial connected domain to achieve data prediction and interpolation in data sparse areas;

[0037] Iterate using the Adam optimizer:

[0038]

[0039] Among them, θ (t) is the parameter value at the tth iteration; η is the learning rate; m is the first-order momentum estimate; v is the second-order momentum estimate; β 1 and β 2 is the momentum parameter, and ∈ is the insurance number to prevent the system from dividing by zero.

[0040] Furthermore, the step S5 specifically includes:

[0041] In the process of variational assimilation of four-dimensional space-time data, based on the model output at the previous moment and the observed data at the current moment, under the constraints of the physical information neural network, the model parameters are optimized by variational methods to ensure that the model response error is minimized. The objective function J(x) in the process of variational assimilation of four-dimensional space-time data is:

[0042]

[0043] Among them, x i is at time t i The model state, x i =M(t i ,t 0 )x 0 , where M(t i ,t 0 ) is from t 0to i Model evolution operator of b is the prior state vector; B is the background error covariance matrix; y i,j is the observation vector; H i,j is the observation operator, which transforms the model space into the observation space; R i,j is the observation error covariance matrix;

[0044]

[0045] Among them, J b (x) is the background term, which represents the difference between the updated physical field and the background field, weighted by the covariance; J o (x) is the observation term, which represents the difference between the model prediction value and the observed value, weighted by the covariance;

[0046] For the boundary areas of sparsely connected domains, the wind field flux is calculated and corrected based on the assimilated model results.

[0047] Furthermore, the step S6 specifically includes:

[0048] Match the current wind farm features with the warning leading feature set. According to the matching degree threshold, when the abnormal probability of the wind farm sample exceeds the threshold, the system issues a warning. The matching strategy is based on pattern matching and classification of a support vector machine.

[0049]

[0050] subject to you i (w T x i +b)≥1-ξ i ,ξ i ≥0;

[0051] Where w is the normal vector of the decision hyperplane, b is the bias, ξ i is the slack variable, C is the penalty parameter, y i is the class label, x i is the eigenvector.

[0052] The second aspect of the present invention provides a sparse spatial wind vector field anomaly warning device based on multi-source heterogeneous spatiotemporal data, comprising:

[0053] Data acquisition unit: used to acquire multi-source heterogeneous spatiotemporal wind vector data, and perform scale division and multi-scale data grid standardization according to the spatial resolution and temporal resolution characteristics of these data;

[0054] Weight determination unit: used to construct a multi-scale geographic spatiotemporal weighted autoregressive model based on geographic distribution characteristics, combine static characteristics, periodic characteristics and time-varying characteristics to build a benchmark model and its periodic fluctuation characteristics and opportunity data attenuation characteristics, and determine the weight update model;

[0055] Model building unit: Use the Helmholtz theorem to decompose the vector field into a divergence field and a curl field, and build a regional wind vector field anomaly warning leading feature set based on the characteristics of historical abnormal events in the target area in the divergence field and the curl field;

[0056] Data processing unit: used to calculate wind field flux at the boundary grid of sparse spatial connected domain, construct fluid dynamics constraints and atmospheric science prior knowledge constraints using physical information neural network, and use Gaussian process regression to fit the divergence and curl characteristics in sparse spatial connected domain;

[0057] Data updating unit: used for performing variational assimilation of four-dimensional space-time data and updating boundary flux of sparse space. In the variational assimilation of four-dimensional space-time data, based on the observation data at the current moment and the response of the model parameters at the previous moment at the current moment, combined with the fluid dynamics constraints of the physical information neural network, the boundary flux of each connected domain in the sparse space is corrected;

[0058] Early warning unit: used to carry out abnormal wind field warning based on regional early warning leading feature matching, match the current wind field characteristics with the early warning leading feature set, and according to the matching degree threshold, when the abnormal probability of the wind field sample exceeds the threshold, the system issues an early warning.

[0059] A third aspect of the present invention provides an electronic device, comprising a processor and a memory connected to the processor for storing executable instructions for the processor, wherein the processor is used to execute the above-mentioned sparse spatial wind vector field anomaly warning method based on multi-source heterogeneous spatiotemporal data.

[0060] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned sparse spatial wind vector field anomaly warning method based on multi-source heterogeneous spatiotemporal data.

[0061] Compared with the prior art, the sparse spatial wind vector field anomaly warning method based on multi-source heterogeneous spatiotemporal data described in the present invention has the following advantages:

[0062] Based on the geographic spatiotemporal characteristics of multi-source heterogeneous observation sources in the region, the present invention constructs a spatiotemporal multi-scale adaptive weight matrix of multi-source heterogeneous observation data. By fully considering the spatiotemporal distribution characteristics of different observation sources and their information contributions at different spatiotemporal scales, the matrix can dynamically adjust the weight of each observation source in the data fusion process, thereby improving the accuracy and reliability of data fusion. This method effectively integrates the information of different observation sources in the region and enhances the detection and analysis capabilities of complex meteorological phenomena.

[0063] In view of the complexity of calculating the prior weight matrix of the geographic spatiotemporal characteristics of different target areas, the present invention allows the use of uniform isotropic initial parameters for startup. In the subsequent system operation process, by utilizing the update of the multi-scale spatiotemporal weight matrix, the sensitivity difference of the observed data is adjusted, and the system adaptively updates and gradually optimizes the prior weight matrix of the target area. This adaptive update mechanism effectively reduces the dependence on the initial weight setting, and gradually improves the ability to accurately model specific areas in the continuous accumulation and analysis of observed data, thereby enhancing the adaptability and early warning accuracy of the system under different geographic spatiotemporal conditions.

[0064] The present invention uses dense spatial observation data to enhance adjacent sparse spatial data under physical constraints, thereby achieving abnormal warning of sparse spatial wind vector fields. Based on the existing observation equipment network, this method increases the warning range of wind vector field anomalies, improves the time lead time and spatial coverage of warnings, thereby helping to respond to potential wind field anomalies more promptly and effectively, improve the ability to respond to sudden meteorological phenomena during flight, reduce flight risks caused by wind shear and turbulence, and thus ensure the safety and punctuality of flight operations. This will not only help improve the safety management level of the civil aviation system, but will also bring a more stable operating environment and reduce flight delays and passenger inconveniences caused by abnormal weather. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings:

[0066] Figure 1 The present invention is a flow chart of a sparse spatial wind vector field anomaly warning method based on multi-source heterogeneous spatiotemporal data. DETAILED DESCRIPTION

[0067] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0068] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and the like are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, features defined as "first", "second", and the like may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0069] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood by specific circumstances.

[0070] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0071] Embodiment 1:

[0072] like Figure 1 The present invention provides a sparse spatial wind vector field anomaly warning method based on multi-source heterogeneous spatiotemporal data, and the steps are as follows:

[0073] Step S1: Acquisition and spatiotemporal standardization of multi-source heterogeneous wind vector data;

[0074] Acquire multi-source heterogeneous spatiotemporal wind vector data in the target area, where data sources include but are not limited to data provided by land-based observation equipment, airborne observation equipment, meteorological satellites, sounding balloons, and other observation equipment capable of measuring wind speed and direction.

[0075] Scale multi-source heterogeneous data and standardize multi-scale data grids.

[0076] According to the spatial resolution characteristics of various types of data, the data are divided into spatial scales based on the observation accuracy, including but not limited to dividing the data into multiple levels such as high resolution, medium resolution and low resolution.

[0077] In terms of time resolution, the data is divided into time scales according to the sampling frequency and the length of the opportunity window, specifically including but not limited to dividing the data into high-frequency sampling data, low-frequency sampling data, and opportunity data. The opportunity data refers to the non-periodic observation data collected within the event window when the opportunity event occurs, usually but not limited to the observation data obtained by airborne observation equipment on the route within the target area, the observation data collected by the sounding balloon, etc.

[0078] Within each scale level, the spatial and temporal resolutions are rasterized and the observation data are redistributed according to preset grid cells to achieve the fusion of different data sources on a unified spatial and temporal scale, to standardize multi-scale data and to generate multi-scale data grids with a unified format.

[0079] Step S2: Construction and weight update of multi-scale geographic spatiotemporal weighted autoregressive model;

[0080] Based on the geographical distribution characteristics of various types of observation equipment in the target area, including but not limited to the geographical location distribution of land-based observation equipment, the orbital characteristics of meteorological satellites and their relative position relationship with the ground observation area, and the spatial layout of routes in the region, a multi-scale geographic spatiotemporal weighted autoregressive model is constructed.

[0081] The benchmark model is designed by combining multiple features, including static features, periodic features, and time-varying features.

[0082] The static characteristics refer to the spatial characteristics reflected by the observation performance related to the geographical location of the observation equipment and its observation range, etc., which are relatively static;

[0083] The periodic characteristics refer to the periodic changes of the observed data over time, including but not limited to diurnal changes, regional circulation, seasonal changes, etc.;

[0084] The time-varying characteristics refer to the dynamic characteristics of the observed data changing over time, such as the evolution of the weather system, and the characteristic that the confidence of the observed data gradually decays over time during the sampling period of the observation equipment.

[0085] Based on the above characteristics, the periodic fluctuation characteristics and opportunity data decay characteristics of the benchmark model are constructed. The periodic fluctuation characteristics are used to describe the periodic changes of the observed data. The opportunity data decay characteristics are used to deal with the uncertainty and decay law of the opportunity data in time.

[0086] On this basis, the method of the present invention determines the dynamic adjustment mechanism of the multi-scale geographic spatiotemporal weighted autoregressive model through a weight update model. The weight update model takes into account the real-time nature of the observation data, the reliability of the data source and its correlation with the benchmark model, and can adaptively adjust the model when new observation data arrives.

[0087] Step S3: Decomposition of vector field and construction of regional early warning pilot feature set;

[0088] Based on the Helmholtz theorem, the wind vector field is decomposed into two independent components: divergence field and curl field.

[0089] Helmholtz's theorem states that any smooth vector field can be uniquely decomposed into an irrotational divergence field and a sourceless curl field. This decomposition process allows us to study the divergence and convergence of airflow in the wind field, as well as the rotation and vortex characteristics of airflow:

[0090]

[0091] Among them, U(X) is the wind vector field, which is smooth in three-dimensional space; is the gradient operator, which represents the total differential in all directions of space; is the divergence field, representing the divergence of the scalar potential function φ(X); is the curl field, expressed as the curl of the vector potential function A(X).

[0092] Further, based on the performance of historical wind field anomaly events in the target area, the features in the divergence field and the curl field are extracted to construct a leading feature set for regional wind vector field anomaly warning.

[0093] Specifically, by analyzing the divergence and curl field features associated with abnormal wind events in historical data, we can identify the manifestation patterns of these events after wind field decomposition. Such features include but are not limited to abnormal divergence or convergence areas in the divergence field, high curl vortex areas in the curl field, and drastic gradient changes.

[0094] Step S4: boundary flux calculation and PINN-GPR fitting of sparse spatial connected domains;

[0095] Specifically, the wind field flux is calculated at the boundary grid of the sparse spatial connected domain based on the reconstructed regional wind vector field. The sparse spatial connected domain refers to a specific area formed after the multi-source heterogeneous spatiotemporal data is reconstructed based on the multi-scale geographic spatiotemporal weighted aggregation. The characteristics of this area are that it mainly relies on observation data with wide coverage, low spatial resolution and long sampling time interval, such as satellite observation data.

[0096] Based on the fitting results of the wind vector field in dense space, the wind field flux is calculated at the boundary grid of the sparse spatial connected domain. The fluid dynamics constraints and prior knowledge constraints of atmospheric science are constructed using physical information neural networks. The divergence and curl characteristics in the sparse spatial connected domain are fitted using Gaussian process regression.

[0097] For the divergence field Use the scalar potential function φ(X) to describe it. Assume that φ(X) is a stationary Gaussian process:

[0098]

[0099] Among them, κ s (X, X′) are the Mattern kernels for the divergence part. The Mattern kernel function is expressed as:

[0100]

[0101] Where v is the smoothing parameter; Γ(v) is the gamma function of v; K v is the modified Bessel function of the second kind; l is the length scale parameter.

[0102] The second type of modified Bessel function is expressed as:

[0103]

[0104] Among them, I v (z) is the modified Bessel function of the first kind, expressed as:

[0105]

[0106] Among them, the gamma function is expressed as:

[0107]

[0108] The regression of the divergence field is:

[0109]

[0110] For The curl field is described by the vector potential function A(X).

[0111] Assume A(X) is a multivariate Gaussian process:

[0112]

[0113] Among them, κ v (X,X′) are the Mattern kernels for the curl part respectively.

[0114] The regression of the curl field is:

[0115]

[0116] Then the flow field U(X) is expressed as:

[0117]

[0118] Physical constraints are constructed using physical information neural networks to ensure that the network output is consistent with the observed data and satisfies the physical laws and atmospheric science prior knowledge used for constraints. The physical constraints include but are not limited to the Navier-Stokes equations of fluid dynamics and the continuity equation:

[0119]

[0120] Among them, v is the fluid velocity vector field; ρ is the fluid density; p is the pressure field; μ is the dynamic viscosity coefficient; f is the external force density;

[0121] In rectangular coordinates it is:

[0122]

[0123] Among them, u, v, w are the velocity components of the fluid at the point (x, y, z) at time t.

[0124] Then the loss function corresponding to the physical information is:

[0125]

[0126] Among them, Loss div is the loss function of the divergence field; Loss NS is the loss function of the curl field; N is the number of observed data points, x i is the corresponding position of the observation data.

[0127] Then the total loss function Loss all It is expressed as:

[0128] Loss all =Loss data +λ div Loss div +λ NS Loss NS ;

[0129] Among them, λ NS and λ div is the penalty coefficient that controls the intensity of the corresponding penalty term; Loss data is the error term based on the observed data and is expressed as:

[0130]

[0131] Among them, y i is x i The observed value at .

[0132] Then the total loss function Loss all The gradient with respect to the model parameters θ is:

[0133]

[0134] in, represents the differential operator with respect to the model parameters θ.

[0135] The iterative process uses the Adam optimizer for iteration:

[0136]

[0137] Among them, θ (t) is the parameter value at the tth iteration; η is the learning rate; m is the first-order momentum estimate; v is the second-order momentum estimate; β 1 and β 2 is the momentum parameter, and ∈ is the insurance number to prevent the system from dividing by zero.

[0138] Step S5: Variational assimilation of four-dimensional space-time data and updating of sparse space boundary flux;

[0139] Four-dimensional space-time data variational assimilation and updating of sparse space boundary flux. In the process of four-dimensional space-time data variational assimilation, based on the model output at the previous moment and the observation data at the current moment, under the constraints of the physical information neural network, the model parameters are optimized by variational methods to ensure that the model response error is minimized. The objective function J(x) in the process of four-dimensional space-time data variational assimilation is:

[0140]

[0141] Among them, x s,i is at time t i The model state, that is, x s,i =M(t i ,t 0 )x s,0 , where M(t i ,t 0 ) is from t 0 to i Model evolution operator of s,b is the prior state vector; B is the background error covariance matrix; y i,j is the observation vector; H i,j is the observation operator, which transforms the model space into the observation space; R i,j is the observation error covariance matrix;

[0142] The components of the objective function are:

[0143]

[0144] Among them, J b (x) is the background term, which represents the difference between the updated physical field and the background field, weighted by the covariance; Jo (x) is the observation term, which represents the difference between the model predicted value and the observed value, weighted by the covariance.

[0145] The minimum of the objective function is the optimal model state estimation in the space-time dimension. n+1 The observed data at time t n The model parameters at time t n+1 The boundary flux of each connected domain in the sparse space is corrected by combining the fluid dynamics constraints of the physical information neural network.

[0146] Step S6: abnormal wind field warning based on regional warning leading feature matching.

[0147] The feature matching uses support vector machine as a classification algorithm to match the current wind field features with the warning pilot feature set:

[0148]

[0149] subject to you tag,i (w T x eig,i +b)≥1-ξ i ,ξ i ≥0;

[0150] Where w is the normal vector of the decision hyperplane, b is the bias, ξ i is the slack variable, C is the penalty parameter, y tag,i is the class label, x eig,i is the eigenvector.

[0151] The classifier outputs the probability that each wind farm sample belongs to the abnormal class. According to the preset threshold, when the abnormal probability of the wind farm sample exceeds the threshold, the system issues an early warning.

[0152] Embodiment 2:

[0153] A sparse spatial wind vector field anomaly early warning device based on multi-source heterogeneous spatiotemporal data, comprising:

[0154] Data acquisition unit: used to acquire multi-source heterogeneous spatiotemporal wind vector data, and perform scale division and multi-scale data grid standardization according to the spatial resolution and temporal resolution characteristics of these data;

[0155] Weight determination unit: used to construct a multi-scale geographic spatiotemporal weighted autoregressive model based on geographic distribution characteristics, combine static characteristics, periodic characteristics and time-varying characteristics to build a benchmark model and its periodic fluctuation characteristics and opportunity data attenuation characteristics, and determine the weight update model;

[0156] Model building unit: Use the Helmholtz theorem to decompose the vector field into a divergence field and a curl field, and build a regional wind vector field anomaly warning leading feature set based on the characteristics of historical abnormal events in the target area in the divergence field and the curl field;

[0157] Data processing unit: used to calculate wind field flux at the boundary grid of sparse spatial connected domain, construct fluid dynamics constraints and atmospheric science prior knowledge constraints using physical information neural network, and use Gaussian process regression to fit the divergence and curl characteristics in sparse spatial connected domain;

[0158] Data updating unit: used for performing variational assimilation of four-dimensional space-time data and updating boundary flux of sparse space. In the variational assimilation of four-dimensional space-time data, based on the observation data at the current moment and the response of the model parameters at the previous moment at the current moment, combined with the fluid dynamics constraints of the physical information neural network, the boundary flux of each connected domain in the sparse space is corrected;

[0159] Early warning unit: used to carry out abnormal wind field warning based on regional early warning leading feature matching, match the current wind field characteristics with the early warning leading feature set, and according to the matching degree threshold, when the abnormal probability of the wind field sample exceeds the threshold, the system issues an early warning.

[0160] Embodiment three:

[0161] An electronic device includes a processor and a memory connected to the processor for storing instructions executable by the processor, wherein the processor is used to execute the above-mentioned sparse spatial wind vector field anomaly warning method based on multi-source heterogeneous spatiotemporal data.

[0162] Embodiment 4:

[0163] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned sparse spatial wind vector field anomaly warning method based on multi-source heterogeneous spatiotemporal data.

[0164] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A sparse spatial wind vector field anomaly warning method based on multi-source heterogeneous spatiotemporal data, characterized by: The steps include: Step S1: Acquire multi-source heterogeneous spatiotemporal wind vector data, and perform scale division and multi-scale data grid standardization according to the spatial resolution and temporal resolution characteristics of these data; Step S2: construct a multi-scale geographic spatiotemporal weighted autoregressive model based on geographic distribution characteristics, combine static characteristics, periodic characteristics and time-varying characteristics to construct a benchmark model and its periodic fluctuation characteristics and opportunity data attenuation characteristics, and determine the weight update model; Step S3: Decomposing the vector field into a divergence field and a curl field using the Helmholtz theorem, and constructing a regional wind vector field abnormal warning leading feature set based on the features of historical abnormal events in the divergence field and the curl field in the target area; Step S4: Calculate the wind field flux at the boundary grid of the sparse spatial connected domain, use the physical information neural network to construct fluid dynamics constraints and atmospheric science prior knowledge constraints, and use the Gaussian process regression to fit the divergence and curl features in the sparse spatial connected domain; Step S5: performing variational assimilation of four-dimensional space-time data and updating of boundary fluxes in sparse space. In the variational assimilation of four-dimensional space-time data, based on the observation data at the current moment and the response of the model parameters at the previous moment at the current moment, combined with the fluid dynamics constraints of the physical information neural network, the boundary fluxes of each connected domain in the sparse space are corrected; Step S6: Perform abnormal wind field warning based on regional warning leading feature matching, match the current wind field features with the warning leading feature set, and according to the matching degree threshold, when the abnormal probability of the wind field sample exceeds the threshold, the system issues a warning.

2. The sparse spatial wind vector field anomaly early warning method based on multi-source heterogeneous spatiotemporal data according to claim 1 is characterized by: The step S1 specifically includes: Acquisition of multi-source heterogeneous spatiotemporal wind vector data including data provided by land-based observation equipment, airborne observation equipment, meteorological satellites, sounding balloons and other observation equipment capable of measuring wind speed and direction; According to the spatial resolution characteristics of various types of data, the data are divided into multiple levels of high resolution, medium resolution and low resolution based on the observation accuracy; In terms of time resolution, the data is divided into high-frequency sampling data, low-frequency sampling data, and opportunity data according to the sampling frequency and the length of the opportunity window. The opportunity data refers to the non-periodic observation data collected within the event window when the opportunity event occurs, including the observation data obtained by the airborne observation equipment on the route in the target area and the observation data collected by the sounding balloon. Within each scale level, the spatial and temporal resolutions are rasterized and the observation data are redistributed according to preset grid cells to achieve the fusion of different data sources on a unified spatial and temporal scale, to standardize multi-scale data and to generate multi-scale data grids with a unified format.

3. The sparse spatial wind vector field anomaly warning method based on multi-source heterogeneous spatiotemporal data according to claim 1 is characterized by: The step S2 specifically includes: Based on the geographical distribution characteristics of various types of observation equipment in the target area, including the geographical location distribution of land-based observation equipment, the orbital characteristics of meteorological satellites and their relative position relationship with the ground observation area, and the spatial layout of routes in the region, a multi-scale geographic spatiotemporal weighted autoregressive model is constructed; The benchmark model is designed by combining multiple features, including static features, periodic features and time-varying features; the static features refer to the spatial features reflected by the observation performance related to the geographical location of the observation equipment and its observation range, which are relatively static; the periodic features refer to the periodic changes of the observation data over time; and the time-varying features refer to the dynamic characteristics of the observation data over time; Based on the above characteristics, the periodic fluctuation characteristics and opportunity data decay characteristics of the benchmark model are constructed.

4. The sparse spatial wind vector field anomaly early warning method based on multi-source heterogeneous spatiotemporal data according to claim 1 is characterized by: The step S3 specifically includes: Based on the Helmholtz theorem, the wind vector field is decomposed into two independent components, the divergence field and the curl field, to achieve a refined analysis of the wind field characteristics. Any smooth vector field can be uniquely decomposed into an irrotational divergence field and a passive curl field, and the divergence and convergence of the airflow in the wind field, as well as the rotation and vortex characteristics of the airflow are studied respectively: Among them, U(X) is the wind vector field, which is smooth in three-dimensional space; is the gradient operator, which represents the total differential in all directions of space; is the divergence field, representing the divergence of the scalar potential function φ(X); is the curl field, expressed as a vector potential function A(X); Based on the performance of historical abnormal wind events in the target area, the features in the divergence field and the curl field are extracted to construct a leading feature set for regional wind vector field abnormal warning. Specifically, by analyzing the divergence field features and curl field features related to abnormal wind events in historical data, the performance patterns of these events after wind field decomposition are identified.

5. The sparse spatial wind vector field anomaly early warning method based on multi-source heterogeneous spatiotemporal data according to claim 1 is characterized by: The step S4 specifically includes: Based on the reconstructed regional wind vector field, the wind field flux is calculated at the boundary grid of the sparse spatial connected domain, wherein the sparse spatial connected domain refers to a specific area formed after the multi-source heterogeneous spatiotemporal data is reconstructed based on the multi-scale geographic spatiotemporal weighted aggregation; Physical constraints are constructed using physical information neural networks to ensure that the network output is consistent with not only the observed data, but also the physical laws used for constraints and prior knowledge of atmospheric science. The physical constraints include the Navier-Stokes equations of fluid dynamics and the continuity equation: Among them, v is the fluid velocity vector field; ρ is the fluid density; p is the pressure field; μ is the dynamic viscosity coefficient; f is the external force density; In rectangular coordinates it is: Among them, u, v, w are the velocity components of the fluid at point (x, y, z) at time t; Use Gaussian process regression to fit the divergence and curl features in the sparse spatial connected domain to achieve data prediction and interpolation in data sparse areas; Iterate using the Adam optimizer: Among them, θ (t) is the parameter value at the tth iteration; η is the learning rate; m is the first-order momentum estimate; v is the second-order momentum estimate; β1 and β2 are momentum parameters, and ∈ is the insurance number to avoid system division by zero.

6. The method for anomaly warning of sparse spatial wind vector field based on multi-source heterogeneous spatiotemporal data according to claim 1 is characterized by: The step S5 specifically includes: In the process of variational assimilation of four-dimensional space-time data, based on the model output at the previous moment and the observed data at the current moment, under the constraints of the physical information neural network, the model parameters are optimized by variational methods to ensure that the model response error is minimized. The objective function J(x) in the process of variational assimilation of four-dimensional space-time data is: Among them, x i is at time t i The model state, x i =M(t i ,t0)x0, where M(t i ,t0) is from t0 to t i Model evolution operator of b is the prior state vector; B is the background error covariance matrix; y i,j is the observation vector; H i,j is the observation operator, which transforms the model space into the observation space; R i,j is the observation error covariance matrix; Among them, J b (x) is the background term, which represents the difference between the updated physical field and the background field, weighted by the covariance; J o (x) is the observation term, which represents the difference between the model prediction value and the observed value, weighted by the covariance; For the boundary areas of sparsely connected domains, the wind field flux is calculated and corrected based on the assimilated model results.

7. The method for anomaly warning of sparse spatial wind vector field based on multi-source heterogeneous spatiotemporal data according to claim 1 is characterized by: The step S6 specifically includes: Match the current wind farm features with the warning leading feature set. According to the matching degree threshold, when the abnormal probability of the wind farm sample exceeds the threshold, the system issues a warning. The matching strategy is based on pattern matching and classification of a support vector machine. subject to y i (w T x i +b)≥1-ξ i ,ξ i ≥0; Where w is the normal vector of the decision hyperplane, b is the bias, ξ i is the slack variable, C is the penalty parameter, y i is the class label, x i is the eigenvector.

8. A sparse spatial wind vector field anomaly warning device based on multi-source heterogeneous spatiotemporal data, characterized by: include: Data acquisition unit: used to acquire multi-source heterogeneous spatiotemporal wind vector data, and perform scale division and multi-scale data grid standardization according to the spatial resolution and temporal resolution characteristics of these data; Weight determination unit: used to construct a multi-scale geographic spatiotemporal weighted autoregressive model based on geographic distribution characteristics, combine static characteristics, periodic characteristics and time-varying characteristics to build a benchmark model and its periodic fluctuation characteristics and opportunity data attenuation characteristics, and determine the weight update model; Model building unit: Use the Helmholtz theorem to decompose the vector field into a divergence field and a curl field, and build a regional wind vector field anomaly warning leading feature set based on the characteristics of historical abnormal events in the target area in the divergence field and the curl field; Data processing unit: used to calculate wind field flux at the boundary grid of sparse spatial connected domain, construct fluid dynamics constraints and atmospheric science prior knowledge constraints using physical information neural network, and use Gaussian process regression to fit the divergence and curl characteristics in sparse spatial connected domain; Data updating unit: used for performing variational assimilation of four-dimensional space-time data and updating boundary flux of sparse space. In the variational assimilation of four-dimensional space-time data, based on the observation data at the current moment and the response of the model parameters at the previous moment at the current moment, combined with the fluid dynamics constraints of the physical information neural network, the boundary flux of each connected domain in the sparse space is corrected; Early warning unit: used to carry out abnormal wind field warning based on regional early warning leading feature matching, match the current wind field characteristics with the early warning leading feature set, and according to the matching degree threshold, when the abnormal probability of the wind field sample exceeds the threshold, the system issues an early warning.

9. An electronic device, comprising a processor and a memory connected to the processor for storing instructions executable by the processor, characterized in that: The processor is used to execute the sparse spatial wind vector field anomaly warning method based on multi-source heterogeneous spatiotemporal data as described in any one of claims 1 to 7 above.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the sparse spatial wind vector field anomaly warning method based on multi-source heterogeneous spatiotemporal data described in any one of claims 1 to 7 is implemented.

Images