Method for determining regional cold air process index based on spatio-temporal graph neural network

Abstract

This invention discloses a method for determining regional cold air process indices based on spatiotemporal graph neural networks. The method includes: acquiring a multivariate time-series dataset of historical meteorological observations and a target cold air process index sequence; selecting multiple key variables from the multivariate time-series dataset that have the strongest time-lag causal driving relationship with the target cold air process index sequence; constructing a dynamic directed spatiotemporal graph using meteorological stations as nodes and multiple key variables as node features; inputting the dynamic directed spatiotemporal graph into a spatiotemporal graph attention network to obtain an initial node feature matrix; inputting the initial node feature matrix into a content-aware gating unit to obtain adaptively filtered node features; and inputting the adaptively filtered node features into a lightweight prediction module to obtain the predicted regional cold air process index. This invention can significantly improve computational and inference efficiency while ensuring prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, and in particular to a method, system, storage medium, and electronic device for determining regional cold air process indices based on spatiotemporal neural networks. Background Technology

[0002] Accurate forecasting of regional cold air processes is a critical requirement for meteorological forecasting, disaster prevention and mitigation, and energy dispatch. The core challenge lies in effectively modeling the dynamic propagation mechanisms and multi-scale interactions of cold air systems from high-dimensional, nonlinear, and strongly spatiotemporally coupled meteorological observation data. Traditional numerical weather prediction models (such as WRF) rely on complex physical parameterization schemes, resulting in high computational costs and sensitivity to initial value errors, with limited forecasting capabilities for sudden and localized processes. While classical statistical and machine learning methods (such as regression and SVR) are computationally efficient, they struggle to capture the complex nonlinear relationships and spatiotemporal dependencies among meteorological elements, leading to poor model interpretability.

[0003] In recent years, deep learning has provided a new paradigm for meteorological time-series forecasting. Recurrent Neural Networks (RNNs, LSTMs) and their variants handle temporal dependencies through gating mechanisms, but they are insufficient in handling spatial correlations among multiple stations. Spatiotemporal graph neural networks effectively improve the performance of multivariate spatiotemporal forecasting by modeling meteorological stations as graph nodes and combining graph convolution with temporal modeling. However, existing methods still have significant limitations: First, the graph structure is mostly based on static geographical distances, which cannot reflect the physical correlations driven by atmospheric circulation (such as wind fields) and dynamically evolving with weather conditions; second, the model input usually contains a large number of meteorological variables, lacking a dynamic feature selection mechanism based on physical mechanisms, resulting in high input noise and low model efficiency; finally, the complex spatiotemporal attention mechanism is susceptible to noise interference and lacks an adaptive information filtering module for the physical characteristics of meteorological processes.

[0004] Furthermore, in the field of dynamic graph structure learning, frameworks such as MTGNN attempt to learn implicit spatial relationships through node embedding, but their learning process lacks explicit physical guidance and exhibits weak generalization ability in extreme meteorological events with scarce data. Regarding feature selection, while methods based on mutual information or LASSO can measure dependencies between variables, they are mostly static or synchronous analyses, failing to effectively identify causal driving relationships with time-lag effects (such as the impact of upstream pressure increases on downstream cooling over the next 24 hours). For example, while the MetNet model proposed by Tsinghua University achieves end-to-end precipitation nowcasting, its input radar reflectivity map is a highly processed product and does not address the feature selection problem of the original multivariate time series. The GNN-MLP hybrid model developed by the Institute of Atmospheric Physics, Chinese Academy of Sciences, aggregates spatial information through graph networks, but its graph structure is based on fixed administrative divisions and fails to incorporate meteorological dynamic constraints, resulting in significant prediction lags during rapidly changing processes such as the passage of cold fronts.

[0005] To improve efficiency, existing research has also explored lightweight LSTM variants (such as minLSTM) to reduce computational complexity by simplifying the gating structure. However, when these lightweight models are directly applied to high-dimensional multivariate weather forecasting, their performance often degrades due to redundant and noisy input features, and they do not achieve synergistic optimization with the upstream spatiotemporal feature extraction module.

[0006] In summary, the limitations of existing solutions can be attributed to two contradictions: First, an imbalance between model expressiveness and physical interpretability. While complex black-box models can improve fitting ability, their internal decision-making mechanisms lack meteorological explanations, making them difficult for operational forecasters to trust and adopt. Second, a disconnect between data-driven flexibility and the constraints of physical laws. Purely data-driven methods are prone to overfitting historical noise, while relying solely on physical formulas struggles to handle complex nonlinearities. Summary of the Invention

[0007] This invention provides a method, system, storage medium, and electronic device for determining regional cold air process indices based on spatiotemporal graph neural networks, which can significantly improve computation and inference efficiency while ensuring prediction accuracy.

[0008] This invention provides a method for determining regional cold air process indices based on spatiotemporal graph neural networks, comprising: A multivariate time series dataset of historical meteorological observations and a target cold air process index sequence are obtained. Several key variables with the strongest time-delay causal driving relationship to the target cold air process index sequence are selected from the multivariate time series dataset. A dynamic directed spatiotemporal graph is constructed using meteorological stations as nodes and the aforementioned key variables as node features. The dynamic directed spatiotemporal graph is input into the spatiotemporal graph attention network to obtain an initial node feature matrix; the initial node feature matrix is ​​input into the content-aware gating unit to obtain adaptively filtered node features; the adaptively filtered node features are input into the lightweight prediction module to obtain the predicted regional cold air process index.

[0009] Furthermore, according to the above method for determining the regional cold air process index based on spatiotemporal graph neural network, the multivariate time series dataset includes multiple candidate variables; Several key variables with the strongest time-delay causal driving relationship to the target cold air process index sequence were selected from the multivariate time series dataset, including: For each candidate variable and objective, calculate the causal strength under different time lags; For each candidate variable, the maximum causal strength is taken as the global causal contribution of the candidate variable. The top K candidate variables with the highest global causal contribution are selected as key variables, and a set of key variables is formed.

[0010] Furthermore, according to the above-mentioned method for determining the regional cold air process index based on a spatiotemporal graph neural network, the step of calculating the causal strength under different time lags for each candidate variable and target includes: Calculate the causality strength under a single time delay: By embedding dimension m and delay step The one-dimensional time series of candidate vectors is transformed into an m-dimensional phase space trajectory; Select the nearest neighbor points, and in the phase space of the candidate vectors, find the k historical states that are closest to the current state at time t. Using the phase space information of these k neighbors, we can obtain the target's position in the phase space. The state at any given time is measured by the prediction error or the correlation coefficient between the predicted value and the true value, which measures the ability to predict the target using the candidate vector phase space, i.e., the causal strength. Traversal delay Repeat the above steps to calculate the candidate vector under all candidate time delays. The causal strength.

[0011] Furthermore, according to the above method for determining the regional cold air process index based on a spatiotemporal graph neural network, the edges of the dynamic directed spatiotemporal graph include static basic edges and dynamic meteorological flow edges. The static basic edge is represented by the following formula:

[0012] in, For static base edges, Indicates site i With the site j Spatial distance between them It is the bandwidth parameter of the Gaussian kernel function; The dynamic meteorological flow edge is represented by the following formula:

[0013] in, Let be the dynamic meteorological flow edge at time t. For the site i exist t Wind vector at any moment From i point to j The position vector, Scaling factor It is the dot product.

[0014] Furthermore, according to the above method for determining the regional cold air process index based on a spatiotemporal graph neural network, the dynamic directed spatiotemporal graph is represented by the following formula:

[0015] in, It is a dynamic directed spacetime graph sequence. This represents all stations in a dynamic directed spatiotemporal graph sequence. Let represent the set of edges at time t. It is a set representation of connection relationships. Denotes the edge at time t. It is a matrix representation of the connection weights. This represents the matrix consisting of the eigenvectors of all stations at time t; in,

[0016] For static base edges, For dynamic meteorological flow edges, It is an adjustable hyperparameter or a learnable scalar.

[0017] Furthermore, according to the above-mentioned method for determining the regional cold air process index based on a spatiotemporal graph neural network, the initial node feature matrix is ​​input into a content-aware gating unit to obtain adaptively filtered node features, including: The initial node feature matrix is ​​linearly transformed, and a channel-node fine-grained vector gate is generated through an activation function; The vector gate and the initial node feature matrix are multiplied element-wise to obtain the adaptively filtered node features.

[0018] Furthermore, according to the above-described method for determining the regional cold air process index based on a spatiotemporal graph neural network, the processing procedure of the lightweight prediction module is expressed by the following formula:

[0019]

[0020]

[0021]

[0022]

[0023] in, For the Gate of Oblivion These are the weight matrix and bias term of the forget gate, respectively. The input to the lightweight prediction module is the adaptively filtered node features. For input gate, These are the weight matrix and bias term of the input gate, respectively. In the candidate hidden state, , To generate the weight matrix and bias terms for candidate states, For the normalized forgetting gate, For the normalized input gate, This represents the final hidden state at the current moment. To hide the time for discussion. This indicates element-wise multiplication.

[0024] The present invention also provides a system for determining regional cold air process indices based on spatiotemporal graph neural networks, comprising: The acquisition and filtering module is used to acquire a multivariate time series dataset of historical meteorological observations and a target cold air process index sequence, and to filter out several key variables with the strongest time-delay causal driving relationship to the target cold air process index sequence from the multivariate time series dataset. The dynamic directed spatiotemporal graph construction module is used to construct a dynamic directed spatiotemporal graph with meteorological stations as nodes and the aforementioned key variables as node features. The regional cold air process index prediction module is used to input the dynamic directed spatiotemporal graph into a spatiotemporal graph attention network to obtain an initial node feature matrix; input the initial node feature matrix into a content-aware gating unit to obtain adaptively filtered node features; and input the adaptively filtered node features into a lightweight prediction module to obtain the predicted regional cold air process index.

[0025] The present invention also provides a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute any of the above-described methods for determining regional cold air process indices based on spatiotemporal graph neural networks.

[0026] The present invention also provides an electronic device, including a processor and a memory, wherein the processor is electrically connected to the memory, the memory is used to store instructions and data, and the processor is used in the steps of the method for determining the regional cold air process index based on the spatiotemporal graph neural network described in any of the preceding claims.

[0027] The present invention provides a method, system, storage medium, and electronic device for determining regional cold air process indices based on spatiotemporal graph neural networks. The present invention has the following beneficial effects: (1) Optimization of features and graph structure based on meteorological physical mechanisms. This invention uses time-delay causal analysis to screen key meteorological factors, replacing the traditional synchronous correlation screening, thus giving the model input clear physical precursor significance. At the same time, by fusing real-time wind field data to construct a dynamic directed spatiotemporal graph, replacing the static geographic map, the graph structure can accurately represent the movement path and propagation direction of cold air. These two improvements make the model input and structure itself meteorologically interpretable, improving the reliability of forecast results.

[0028] (2) Coordinated optimization of prediction accuracy and computational efficiency. This invention designs a collaborative architecture of "gated filtering + lightweight prediction". A content-aware gating unit is introduced after the spatiotemporal attention network to effectively filter out noise and improve feature quality. Subsequently, a simplified minLSTM is used as the predictor, which significantly reduces model complexity and computational overhead while maintaining the core temporal modeling capability. This architecture significantly improves the training and inference efficiency of the model while ensuring prediction accuracy.

[0029] (3) An end-to-end solution for professional meteorological forecasting has been formed. This invention provides a complete technology chain from causal feature selection, dynamic graph construction, gated spatiotemporal learning to lightweight forecasting. The solution is customized for professional scenarios of cold air process forecasting, and the modules are optimized collaboratively to overcome the problem of the disconnect between general models and meteorological knowledge. The whole method has a clear logic, verifiable intermediate results, and is easy for business personnel to understand and trust. It has good business applicability and the potential to be extended to similar meteorological forecasting tasks such as rainstorms and typhoons. Attached Figure Description

[0030] The technical solution and other beneficial effects of the present invention will become apparent from the following detailed description of specific embodiments of the invention, in conjunction with the accompanying drawings.

[0031] Figure 1 A flowchart of a method for determining regional cold air process indices based on spatiotemporal graph neural networks, provided in an embodiment of the present invention.

[0032] Figure 2 This is a schematic diagram of the lightweight prediction module provided in an embodiment of the present invention.

[0033] Figure 3 A schematic diagram of the structure of a regional cold air process index determination system based on a spatiotemporal graph neural network provided in an embodiment of the present invention.

[0034] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] (1) In the application of spatiotemporal graph neural networks for meteorological forecasting, existing schemes mainly adopt the paradigm of "static graph structure + general spatiotemporal convolution". The typical process is as follows: meteorological stations or grid points are regarded as graph nodes, a static adjacency matrix is ​​constructed using the latitude and longitude distance between stations, spatial information is aggregated through graph convolutional networks (GCN) or graph attention networks (GAT), and finally combined with temporal convolution (TCN) or recurrent neural networks (RNN) for prediction. For example, a study used an architecture combining GAT and GRU to predict urban PM2.5 concentration, and its graph structure is based on the fixed geographical distance between stations. However, these schemes have the following problems: 1) Rigid coupling of spatiotemporal features: Most models adopt "spatial aggregation first, then temporal modeling" or simple spatiotemporal alternating convolution, failing to design a tightly coupled spatiotemporal attention mechanism to simultaneously capture the most critical "when-where-what elements". During cold air processes, the key influencing areas evolve over time, and the existing architecture is insufficient in its ability to dynamically focus on this.

[0037] 2) The graph structure is detached from meteorological dynamics: The propagation of cold air is essentially the result of three-dimensional atmospheric motion, especially near fronts, where meteorological elements are highly correlated and have a clear direction (downwind propagation). Fixed geographic adjacency maps completely ignore this dynamic and directional physical correlation determined by real-time flow fields, causing models to be unable to accurately model the movement path and intensity transport of cold air, resulting in large prediction errors in areas with rapid frontal movement or complex terrain.

[0038] (2) Regarding multivariate time-series feature selection, existing solutions mainly rely on filtering or embedded methods. For example, Pearson correlation coefficient, mutual information, or LASSO regression are used to screen features before model training, or attention weights are added to the neural network to implicitly select features. However, these solutions have the following drawbacks: 1) Ignoring time-lag causal relationships: The impact of cold air systems has significant time-lag effects (e.g., changes in upstream variables precede downstream responses). Traditional correlation analysis (such as the Pearson coefficient) can only capture synchronous linear relationships and cannot identify time-lag driving factors with forecast significance, resulting in unclear physical meaning of the selected features and limited forecast lead time.

[0039] 2) Feature selection is separate from the prediction model: The feature selection stage is independent of the prediction model training, which is a "two-step" strategy. The selected feature set is static and cannot be dynamically adjusted according to different types of cold air (such as strong cold waves and moderate cold air) or different forecast periods. It has poor flexibility and is difficult to cope with complex and changeable weather conditions.

[0040] To address the aforementioned problems, embodiments of the present invention provide a method, system, storage medium, and electronic device for determining regional cold air process indices based on spatiotemporal graph neural networks. The regional cold air process index determination system based on spatiotemporal graph neural networks provided by embodiments of the present invention can be integrated into an electronic device, which can be a terminal, server, or other device. The terminal can include a tablet computer, laptop computer, personal computer (PC), microprocessor box, or other devices.

[0041] Please see Figure 1 , Figure 1 The flowchart illustrates a method for determining a regional cold air process index based on a spatiotemporal graph neural network, provided in an embodiment of the present invention. This method, applied in electronic devices, includes the following steps: S1. Obtain the multivariate time series dataset of historical meteorological observations and the target cold air process index sequence. Select several key variables with the strongest time-delay causal driving relationship to the target cold air process index sequence from the multivariate time series dataset.

[0042] S2 constructs a dynamic directed spatiotemporal graph using meteorological stations as nodes and multiple key variables as node features.

[0043] In regional cold air process forecasting, the interaction between meteorological stations is not static but dynamically changing, dominated by atmospheric circulation and evolving with the weather system. Most existing methods employ static adjacency maps based on geographical distance, which cannot characterize key physical processes such as the movement and intensity transport of cold air fronts. Furthermore, using all meteorological variables as input introduces significant noise, and traditional feature selection based on synchronous correlation cannot identify time-lag driving factors with forecast precursor significance, resulting in short forecast lead times and poor interpretability.

[0044] As a solution, this invention proposes a physically-guided dynamic graph construction and causal feature selection strategy. It mainly consists of two steps: First, using a nonlinear time-delay causal discovery algorithm, key driving variables with significant causal correlations to the target cold air process index are automatically identified from high-dimensional time-series data, ensuring the physical meaning and predictive value of the input features (corresponding to step S1). Second, based on dynamic feature selection, static geographical constraints and real-time atmospheric flow fields are integrated to construct a dynamic directed spatiotemporal graph whose topology evolves over time. This allows the graph structure to accurately reflect the three-dimensional propagation dynamics of the cold air system, providing a physically interpretable, high-fidelity data structure for subsequent models (corresponding to step S2).

[0045] In one embodiment, step S1 includes the following steps: S11, for each candidate variable and target, calculate the causal strength under different time lags.

[0046] Specifically, given a multivariate time series dataset of historical meteorological observations (in, T For time step, N For the number of stations, D (as the original variable dimension) and the target cold air process index sequence This embodiment aims to... D Select the candidate variables from the candidate variables. Y Having the strongest time-delay causal driving relationship K Key variables ( K < D ).

[0047] Because meteorological systems are nonlinear, traditional linear correlations (such as the Pearson coefficient) cannot capture nonlinear causality. Therefore, this invention's embodiments are based on nonlinear causality discovery algorithms such as "convergent cross-mapping," combined with the dynamical system "phase space reconstruction" theory to achieve calculation. Causality strength Quantified by "the ability to predict nearest neighbors in phase space": Delay range setting: Determine the time delay Possible values ​​(e.g.) ), To determine the maximum reasonable time delay, domain experience or data length must be considered. (Time delay) Represents candidate vectors Impact on target "Time delay" - that is exist The state at any given moment, corresponding to The state at time t.

[0048] S111, Calculate a single time delay The following causal strength: S1111, by embedding dimension m and delay step size (here) That is, causal time delay), which transforms the one-dimensional time series of candidate vectors into m-dimensional phase space trajectories.

[0049] An m-dimensional phase space trajectory can be exemplified as follows: .

[0050] S1112, Select the nearest neighbor point, and find the k historical states that are closest to the current state at time t in the phase space of the candidate vector (nearest neighbor search).

[0051] S1113, using the phase space information of these k neighbors, obtain the target exist The state at any given time is measured by the prediction error or the correlation coefficient between the predicted value and the true value, which measures the ability to predict the target using the candidate vector phase space, i.e., the causal strength.

[0052] The "quality" of this predictive skill directly corresponds to the strength of causality. The better the prediction skill (the smaller the error and the higher the correlation coefficient), the better the prediction ability. right The stronger the causal drive, the greater the causal strength. The larger.

[0053] S112, Traversal Delay Repeat step S111 (i.e., S1111-S1113) above to calculate the candidate vector under all candidate time delays. causal strength Finally, the distribution of causal intensity under different time delays was obtained.

[0054] S12, for each candidate variable Take the maximum causal strength The top K candidate variables with the highest global causal contribution were selected as key variables, forming a set of key variables. .

[0055] The selection process can be dynamically adjusted based on different types of cold air processes or forecast lead times, forming a "scenario-variable" mapping and improving the model's adaptability. Mathematically, this process can be formalized as follows:

[0056] in, and A physically meaningful time delay search range (e.g., 0 to 48 hours) is defined.

[0057] In one embodiment, the K key variables selected in step S1 are used as the feature vectors of each station (graph node) at time t. The edge structure of the dynamic directed spatiotemporal graph in step S2 consists of two parts and evolves dynamically over time, including static basic edges and dynamic meteorological flow edges.

[0058] Static base edges: Based on the latitude and longitude distances between stations, a static adjacency matrix is ​​constructed using a Gaussian kernel function or the K-nearest neighbor method to maintain the spatial continuity of the base. Static base edges are represented by the following formula:

[0059] in, For static base edges, Indicates site i With the site j Spatial distance between them It is the bandwidth parameter of the Gaussian kernel function (belonging to the algorithm hyperparameters).

[0060] Dynamic meteorological flow edge: reading time t Reanalysis of wind field data (e.g., 850 hPa horizontal wind vector) If the site i The wind vector at the location points to the station. j In that direction, a path is established from... i arrive j The directed edges of the array have weights that are positively correlated with wind speed, simulating the advection transport of cold air.

[0061] Dynamic meteorological flow edges are expressed by the following formula:

[0062] in, Let be the dynamic meteorological flow edge at time t. For the site i exist t Wind vector at any moment From i point to j The position vector, Scaling factor It is the dot product.

[0063] The final dynamic spatiotemporal graph adjacency matrix A weighted fusion of the two:

[0064] in, For static base edges, For dynamic meteorological flow edges, It is an adjustable hyperparameter or a learnable scalar.

[0065] Thus, a dynamic directed spatiotemporal graph is obtained, in which node features are key variables and the edge structure evolves in real time with the circulation:

[0066] in, V This represents all stations in a dynamic directed spacetime graph sequence (each station is a node); Let the set of edges at time t be defined by the static basic edges. With dynamic meteorological flow edge It is a weighted fusion that determines the connection relationships between nodes; and The difference is, It is a matrix representation of the connection weights. It is a set representation of connection relationships; , indicating time t A matrix composed of the eigenvectors of all sites.

[0067] This invention uses time-delay causal analysis to screen key inputs from a physical mechanism perspective and constructs a dynamic graph that integrates real-time wind fields to characterize the propagation path of cold air. This overcomes the limitations of static graphs and all-variable inputs, providing a high-value, physically interpretable data foundation for subsequent deep spatiotemporal feature learning.

[0068] S3. Input the dynamic directed spatiotemporal graph into the spatiotemporal graph attention network to obtain the initial node feature matrix; input the initial node feature matrix into the content-aware gating unit to obtain the adaptively filtered node features; input the adaptively filtered node features into the lightweight prediction module to obtain the predicted regional cold air process index.

[0069] After obtaining a physically guided, dynamically directed spatiotemporal graph sequence, the core of subsequent model design lies in efficiently and robustly extracting key spatiotemporal patterns and ultimately accurately predicting the cold air process index. Existing spatiotemporal graph attention networks often produce outputs unrelated to the target when processing such high-dimensional, noisy data; directly inputting these into the predictor affects accuracy and stability. Simultaneously, to meet the timeliness requirements of meteorological operations, the prediction module must balance efficiency and accuracy. Existing solutions often simply stack lightweight predictors (such as minLSTM) with complex feature extraction networks, failing to achieve synergistic optimization in feature representation capacity and computational efficiency.

[0070] As a solution, this invention proposes a highly efficient gated spatiotemporal network and forecasting collaborative architecture for meteorological processes. It mainly consists of two steps: First, a lightweight content-aware gating unit is introduced after the spatiotemporal graph attention network to adaptively filter the attention output, suppressing noise and enhancing the focusing ability on key cold air systems (such as fronts and key regions), thereby improving the purity and task relevance of features. Second, the refined high-quality spatiotemporal features are input into a collaboratively designed, capacity-matched lightweight sequence prediction module (minLSTM) to achieve efficient and accurate end-to-end decoding from feature extraction to exponential prediction.

[0071] In one embodiment, step S3, inputting the dynamic directed spatiotemporal graph into the spatiotemporal graph attention network to obtain the initial node feature matrix, specifically includes: To input a dynamic directed spatiotemporal graph into a spatiotemporal graph attention network, the graph snapshots at each time step must first be encoded to fit the network's input format. The core of the network lies in combining spatial graph attention mechanisms with temporal sequence modeling capabilities: spatially, a multi-head attention mechanism aggregates the features of each node on its directed neighbors, considering the influence of edge direction on information transmission, thereby extracting the spatial dependency features of nodes within each time step; temporally, temporal attention or recurrent structures are used to capture the changing patterns of node features over time, fusing and updating information from different time steps. After this series of operations, the network outputs a node feature matrix that integrates spatiotemporal context information. Each row of this matrix corresponds to a node, and each column corresponds to a feature dimension, forming the initial node feature matrix.

[0072] In one embodiment, step S3, inputting the initial node feature matrix into the content-aware gating unit to obtain adaptively filtered node features, specifically includes: S31 performs a linear transformation on the initial node feature matrix and generates a channel-node fine-grained vector gate through an activation function.

[0073] Specifically, the spatiotemporal graph attention network at time... t The output node feature matrix is .in N For the number of nodes, F This is the feature dimension. The traditional approach is to directly pass it to the next layer or prediction module. However, It may contain background noise or redundant information that is unrelated to the current cold air process.

[0074] To perform adaptive filtering, this embodiment of the invention designs a lightweight content-aware gating unit. This unit first... Perform a linear transformation and generate a channel-node fine-grained vector gate using the Sigmoid activation function. Each element has a value between 0 and 1, representing the information retention strength of the corresponding node and feature dimension. This process can be expressed by the following formula:

[0075] in, and These are learnable parameters. This is the Sigmoid function.

[0076] S32, multiply the vector gate and the initial node feature matrix element by element to obtain the adaptively filtered node features.

[0077] Specifically, it is expressed by the following formula:

[0078] in, This indicates element-wise multiplication.

[0079] This operation allows the model to dynamically suppress information flows identified as noise or irrelevant, while enhancing characteristic signals highly correlated with the intensity, location, and evolution phase of cold air processes. For example, during the onset of a cold air outbreak, gating may tend to enhance channels related to temperature gradients and wind shear near stations on the frontal zone, while suppressing changes in stable meteorological features far from the main path.

[0080] In one embodiment, step S3, which involves inputting the adaptively filtered node features into the lightweight prediction module to obtain the predicted regional cold air process index, specifically includes: Node feature sequence after gating filtering It features a higher signal-to-noise ratio and better task relevance. To perform the final regional cold air process index prediction, this embodiment of the invention selects a lightweight yet capacity-matched minLSTM as the sequence prediction module. Figure 2 This is a schematic diagram of the lightweight prediction module provided in an embodiment of the present invention, as shown below. Figure 2 As shown. Compared with the standard LSTM, minLSTM significantly reduces the number of parameters and computational complexity while maintaining the core temporal modeling capabilities through structural simplification. It is particularly suitable for efficient decoding of high-quality, low-noise feature sequences refined by front-end deep networks.

[0081] For the input feature vector at a time step The core computational steps of minLSTM can be described as follows:

[0082]

[0083]

[0084]

[0085]

[0086] in, For the Gate of Oblivion These are the weight matrix and bias term of the forget gate, respectively. The input to the lightweight prediction module is the adaptively filtered node features. For input gate, These are the weight matrix and bias term of the input gate, respectively. In the candidate hidden state, , To generate the weight matrix and bias terms for candidate states, For the normalized forgetting gate, For the normalized input gate, This represents the final hidden state at the current moment. To hide the time for discussion. This indicates element-wise multiplication.

[0087] Finally, the hidden state of the last time step. Alternatively, by aggregating the hidden states of all time steps and mapping them through a fully connected layer, the predicted regional cold air process index can be obtained.

[0088] Synergy is demonstrated by the fact that minLSTM's lightweight design results in fewer parameters. Directly applying it to raw or coarsely selected features could easily lead to underfitting due to insufficient model capacity. However, in this architecture, the input to minLSTM consists of high-order features deeply refined through a series of complex modules, including dynamic causal selection, physical mapping, spatiotemporal attention learning, and content-aware gating filtering. These features already contain rich spatiotemporal evolution patterns strongly correlated with the target, and noise is significantly suppressed. At this point, minLSTM's lightweight advantage is fully realized—it can efficiently learn temporal evolution patterns from these "high-quality features" and make accurate predictions with extremely low computational cost, avoiding the overfitting risks and computational waste that complex predictors might bring, achieving a perfect balance between the "depth" of feature extraction and the "efficiency" of prediction decoding.

[0089] In summary, this invention improves the purity and task relevance of spatiotemporal features through "content-aware gating," and then efficiently decodes high-quality features using a "lightweight minLSTM," forming a collaborative optimization loop from feature construction to final prediction. This architecture not only ensures high prediction accuracy but also significantly improves the overall computational efficiency and inference speed of the model, meeting the dual requirements of meteorological operations for timeliness and accuracy.

[0090] Based on the method described in the above embodiments, this embodiment will further describe it from the perspective of a regional cold air process index determination system based on a spatiotemporal graph neural network. This regional cold air process index determination system based on a spatiotemporal graph neural network can be implemented as an independent entity or integrated into an electronic device. The electronic device can be a terminal, server, or other devices. The terminal can include a tablet computer, a laptop computer, a personal computer (PC), a microprocessor box, or other devices.

[0091] Please see Figure 3 , Figure 3 This invention specifically describes a regional cold air process index determination system based on a spatiotemporal graph neural network, which is applied to electronic devices. The system may include: The acquisition and filtering module is used to acquire a multivariate time series dataset of historical meteorological observations and a target cold air process index sequence, and to filter out several key variables with the strongest time-delay causal driving relationship to the target cold air process index sequence from the multivariate time series dataset. The dynamic directed spatiotemporal graph construction module is used to construct a dynamic directed spatiotemporal graph with meteorological stations as nodes and the aforementioned key variables as node features. The regional cold air process index prediction module is used to input the dynamic directed spatiotemporal graph into a spatiotemporal graph attention network to obtain an initial node feature matrix; input the initial node feature matrix into a content-aware gating unit to obtain adaptively filtered node features; and input the adaptively filtered node features into a lightweight prediction module to obtain the predicted regional cold air process index.

[0092] In specific implementation, the above modules and / or units can be implemented as independent entities, or they can be arbitrarily combined and implemented as the same or several entities. For the specific implementation of the above modules and / or units, please refer to the previous method embodiments. For the specific beneficial effects that can be achieved, please also refer to the beneficial effects in the previous method embodiments, which will not be repeated here.

[0093] In addition, embodiments of the present invention also provide an electronic device, which may be a computer, tablet computer, or other similar device. This electronic device can implement the steps of any embodiment of the method for determining regional cold air process indices based on spatiotemporal graph neural networks provided in the embodiments of the present invention. Therefore, it can achieve the beneficial effects achievable by any of the methods for determining regional cold air process indices based on spatiotemporal graph neural networks provided in the embodiments of the present invention, as detailed in the preceding embodiments, and will not be repeated here.

[0094] Figure 3 A specific structural block diagram of an electronic device provided in an embodiment of the present invention is shown. This electronic device can be used to implement the regional cold air process index determination method based on spatiotemporal graph neural network provided in the above embodiments. The electronic device 500 can be a terminal, server, or other device. The terminal can include a tablet computer, laptop computer, personal computer (PC), microprocessor box, or other devices.

[0095] The memory 520 can be used to store software programs and modules, such as the program instructions / modules corresponding to those in the above embodiments. The processor 580 executes various functional applications and data processing by running the software programs and modules stored in the memory 520. The memory 520 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 520 may further include memory remotely located relative to the processor 580, and these remote memories can be connected to the electronic device 500 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0096] The input unit 530 can be used to receive input numeric or character information, and to generate a keyboard and mouse related to user settings and function control. Display unit 540 can be used to display information input by the user or information provided to the user, as well as various graphical user interfaces, which can be composed of graphics, text, icons, video, and any combination thereof. Display unit 540 may include display panel 541, which may optionally be configured in the form of LCD (Liquid Crystal Display), OLED (Organic Light-Emitting Diode), or other similar forms.

[0097] Electronic device 500, through transmission module 570 (e.g., Wi-Fi module), can help users receive requests, send information, etc., providing users with wireless broadband internet access. Although transmission module 570 is shown in the figure, it is understood that it is not an essential component of electronic device 500 and can be omitted as needed without changing the essence of the invention.

[0098] The processor 580 is the control center of the electronic device 500. It connects to various parts of the phone via various interfaces and lines, and performs various functions and processes data of the electronic device 500 by running or executing software programs and / or modules stored in the memory 520, and by calling data stored in the memory 520, thereby providing overall monitoring of the electronic device. Optionally, the processor 580 may include one or more processing cores; in some embodiments, the processor 580 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into the processor 580.

[0099] Electronic device 500 also includes a power supply 590 (such as a battery) that supplies power to various components. In some embodiments, the power supply may be logically connected to processor 580 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 590 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0100] Although not shown, the electronic device 500 also includes cameras (such as front-facing cameras and rear-facing cameras), Bluetooth modules, etc., which will not be described in detail here. Specifically, in this embodiment, the display unit of the electronic device is a touch screen display, and the mobile terminal also includes a memory and one or more programs, wherein one or more programs are stored in the memory and configured to be executed by one or more processors. One or more programs contain instructions for performing the following operations: A multivariate time series dataset of historical meteorological observations and a target cold air process index sequence are obtained. Several key variables with the strongest time-delay causal driving relationship to the target cold air process index sequence are selected from the multivariate time series dataset. A dynamic directed spatiotemporal graph is constructed using meteorological stations as nodes and the aforementioned key variables as node features. The dynamic directed spatiotemporal graph is input into the spatiotemporal graph attention network to obtain an initial node feature matrix; the initial node feature matrix is ​​input into the content-aware gating unit to obtain adaptively filtered node features; the adaptively filtered node features are input into the lightweight prediction module to obtain the predicted regional cold air process index.

[0101] In practice, the above modules can be implemented as independent entities or combined in any way to be implemented as the same or several entities. For the specific implementation of the above modules, please refer to the previous method implementation examples, which will not be repeated here.

[0102] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor. Therefore, embodiments of the present invention provide a storage medium storing multiple instructions that can be loaded by a processor to execute the steps of any embodiment of the regional cold air process index determination method based on a spatiotemporal graph neural network provided by the present invention.

[0103] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0104] Since the instructions stored in the storage medium can execute the steps in any embodiment of the method for determining the regional cold air process index based on a spatiotemporal graph neural network provided in the embodiments of the present invention, the beneficial effects that any method for determining the regional cold air process index based on a spatiotemporal graph neural network provided in the embodiments of the present invention can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0105] The foregoing has provided a detailed description of a method, system, storage medium, and electronic device for determining regional cold air process indices based on spatiotemporal graph neural networks, as provided in the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for determining regional cold air process indices based on spatiotemporal graph neural networks, characterized in that, The method includes: A multivariate time series dataset of historical meteorological observations and a target cold air process index sequence are obtained. Several key variables with the strongest time-delay causal driving relationship to the target cold air process index sequence are selected from the multivariate time series dataset. A dynamic directed spatiotemporal graph is constructed using meteorological stations as nodes and the aforementioned key variables as node features. The dynamic directed spatiotemporal graph is input into the spatiotemporal graph attention network to obtain an initial node feature matrix; the initial node feature matrix is ​​input into the content-aware gating unit to obtain adaptively filtered node features; the adaptively filtered node features are input into the lightweight prediction module to obtain the predicted regional cold air process index.

2. The method for determining regional cold air process indices based on spatiotemporal graph neural networks according to claim 1, characterized in that, The multivariate time series dataset includes multiple candidate variables; Several key variables with the strongest time-delay causal driving relationship to the target cold air process index sequence were selected from the multivariate time series dataset, including: For each candidate variable and objective, calculate the causal strength under different time lags; For each candidate variable, the maximum causal strength is taken as the global causal contribution of the candidate variable. The top K candidate variables with the highest global causal contribution are selected as key variables, and a set of key variables is formed.

3. The method for determining the regional cold air process index based on a spatiotemporal graph neural network according to claim 2, characterized in that, For each candidate variable and target, the causal strength under different time lags is calculated, including: Calculate the causality strength under a single time delay: By embedding dimension m and delay step The one-dimensional time series of candidate vectors is transformed into an m-dimensional phase space trajectory; Select the nearest neighbor points, and in the phase space of the candidate vectors, find the k historical states that are closest to the current state at time t. Using the phase space information of these k neighbors, we can obtain the target's position in the phase space. The state at any given time is measured by the prediction error or the correlation coefficient between the predicted value and the true value, which measures the ability to predict the target using the candidate vector phase space, i.e., the causal strength. Traversal delay Repeat the above steps to calculate the candidate vector under all candidate time delays. The causal strength.

4. The method for determining the regional cold air process index based on a spatiotemporal graph neural network according to claim 1, characterized in that, The edges of the dynamic directed spatiotemporal graph include static basic edges and dynamic meteorological flow edges; The static basic edge is represented by the following formula: in, For static base edges, Indicates site i With the site j Spatial distance between them It is the bandwidth parameter of the Gaussian kernel function; The dynamic meteorological flow edge is represented by the following formula: in, Let be the dynamic meteorological flow edge at time t. For the site i exist t Wind vector at any moment From i point to j The position vector, Scaling factor It is the dot product.

5. The method for determining the regional cold air process index based on a spatiotemporal graph neural network according to claim 4, characterized in that, The dynamic directed spacetime graph is represented by the following formula: in, It is a dynamic directed spacetime graph sequence. This represents all stations in a dynamic directed spatiotemporal graph sequence. Let represent the set of edges at time t. It is a set representation of connection relationships. Denotes the edge at time t. It is a matrix representation of the connection weights. This represents the matrix consisting of the eigenvectors of all stations at time t; in, For static base edges, For dynamic meteorological flow edges, It is an adjustable hyperparameter or a learnable scalar.

6. The method for determining regional cold air process indices based on spatiotemporal graph neural networks according to claim 1, characterized in that, The initial node feature matrix is ​​input into the content-aware gating unit to obtain adaptively filtered node features, including: The initial node feature matrix is ​​linearly transformed, and a channel-node fine-grained vector gate is generated through an activation function; The vector gate and the initial node feature matrix are multiplied element-wise to obtain the adaptively filtered node features.

7. The method for determining regional cold air process indices based on spatiotemporal graph neural networks according to claim 1, characterized in that, The processing procedure of the lightweight prediction module is represented by the following formula: in, For the Gate of Oblivion These are the weight matrix and bias term of the forget gate, respectively. The input to the lightweight prediction module is the adaptively filtered node features. For input gate, These are the weight matrix and bias term of the input gate, respectively. In the candidate hidden state, , To generate the weight matrix and bias terms for candidate states, For the normalized forgetting gate, For the normalized input gate, This represents the final hidden state at the current moment. To hide the time for discussion. This indicates element-wise multiplication.

8. A system for determining regional cold air process indices based on spatiotemporal graph neural networks, characterized in that, include: The acquisition and filtering module is used to acquire a multivariate time series dataset of historical meteorological observations and a target cold air process index sequence, and to filter out several key variables with the strongest time-delay causal driving relationship to the target cold air process index sequence from the multivariate time series dataset. The dynamic directed spatiotemporal graph construction module is used to construct a dynamic directed spatiotemporal graph with meteorological stations as nodes and the aforementioned key variables as node features. The regional cold air process index prediction module is used to input the dynamic directed spatiotemporal graph into the spatiotemporal graph attention network to obtain the initial node feature matrix; The initial node feature matrix is ​​input into the content-aware gating unit to obtain the adaptively filtered node features; The adaptively filtered node features are input into the lightweight prediction module to obtain the predicted regional cold air process index.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted to be loaded by a processor to execute the method for determining regional cold air process indices based on a spatiotemporal graph neural network as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, The device includes a processor and a memory, the processor being electrically connected to the memory, the memory being used to store instructions and data, and the processor being used to execute the steps in the method for determining the regional cold air process index based on a spatiotemporal graph neural network as described in any one of claims 1 to 7.

Images