Precipitation prediction map convolution method based on graph reconstruction and spatial self-attention

Abstract

The invention relates to a rainfall prediction map convolution method based on graph reconstruction and spatial self-attention, and belongs to the technical field of crossing of artificial intelligence and urban meteorological early warning technologies. The method and the device are used for solving the problems of excessive smoothness and insufficient spatial dependency capture caused by dependence on a geographic adjacency relation in precipitation prediction of a traditional graph convolutional network. A sparse graph structure reflecting real meteorological association is constructed by calculating a Pearson's correlation coefficient of Beidou multi-base-station meteorological time series data, and a traditional geographic adjacency matrix is replaced; designing a dynamic space self-attention mechanism, fusing the static association graph and a dynamic attention weight, and realizing adaptive aggregation of long-range dependence between nodes; and in combination with multilevel feature propagation of graph convolution and an attention matrix, the modeling capability of a heterogeneous meteorological spatial mode is improved. According to the method, the accuracy of short temporary rainfall prediction is remarkably improved, the method is particularly suitable for association mining of base stations which are far in geographic distance but similar in meteorological feature, and a high-robustness spatial feature extraction scheme is provided for a meteorological early warning system.

Description

TECHNICAL FIELD

[0001] The application relates to a precipitation prediction graph convolution method based on graph reconstruction and spatial self-attention, and belongs to the technical field of artificial intelligence and urban meteorological warning. BACKGROUND

[0002] Meteorological prediction, as an important public service, has a profound impact on social and economic development and people's life safety. Among them, short-term precipitation prediction, due to its key role in disaster warning, has always been an important topic in meteorological research. Traditional prediction methods mainly rely on numerical weather prediction models and statistical learning techniques, but these methods have obvious limitations in terms of accuracy and real-time performance.

[0003] Numerical weather prediction models simulate atmospheric motion processes by solving complex physical equation systems. Although such models perform well in large-scale weather system prediction, they have high computational costs and are limited by grid resolution, making it difficult to capture local sudden precipitation events. In the context of urban waterlogging warning, numerical models often cannot accurately predict local severe convective precipitation.

[0004] With the development of machine learning technology, statistical methods have been gradually introduced into the field of meteorological prediction. Early models such as ARIMA and SVM can handle single-site meteorological time series data, but cannot effectively utilize spatial correlation information. The advent of deep learning has brought improvements, with time series models such as LSTM and GRU being able to better handle the temporal dependence of meteorological data, and CNN being used to process grid-based meteorological data. However, these methods still have problems with insufficient modeling capabilities when faced with non-uniformly distributed base station data.

[0005] In recent years, graph convolutional networks (GCN) have shown unique advantages in meteorological prediction due to their ability to handle non-Euclidean space data. Existing technologies usually treat meteorological base stations as graph nodes and construct graph structures through geographical adjacency relationships. Common construction methods include Delaunay triangulation or fixed distance threshold connection. Although these methods are simple and direct, they have significant drawbacks in practical applications.

[0006] Geographical adjacency graph construction methods often result in the forced connection of unrelated base stations. For example, mountain base stations and plain base stations may be geographically close, but due to topographical influences, their meteorological characteristics differ significantly. This unreasonable connection can cause ineffective mixing of feature information, leading to decreased prediction accuracy. More seriously, in deep GCN, this inappropriate aggregation can cause over-smoothing problems, where the feature representations of different nodes gradually converge, losing discriminability.

[0007] Another key issue is that existing methods are difficult to capture long-range weather dependencies. In actual weather systems, certain base stations that are geographically far apart may have high correlation due to being affected by the same weather system. For example, multiple base stations in the periphery of a typhoon may exhibit similar patterns of pressure changes, but traditional geographically contiguous graphs cannot establish effective connections at such long distances.

[0008] In addition, weather correlations themselves have dynamic changing characteristics. In the process of rapid evolution of weather systems, the correlation between base stations will change significantly over time. The static graph structure adopted by existing technologies cannot adapt to this dynamic nature, resulting in insufficient response capability of the model to sudden weather events.

[0009] These limitations seriously restrict the actual application effect of existing technologies in short- and long-range precipitation prediction. Especially in the context of frequent extreme weather events, meteorological warning systems have higher requirements for the accuracy and reliability of prediction technology. How to break through the limitations of geographical adjacency and build a more reasonable graph structure has become a key challenge to improve precipitation prediction performance.

[0010] Disadvantages of existing technologies:

[0011] 1. Over-smoothing problem: Geographical adjacency graph can cause irrelevant base station features to be mixed, and node features in deep GCN tend to converge, making it difficult to distinguish key weather patterns.

[0012] 2. Long-range dependency missing: Fixed graph structure cannot capture the association of base stations with similar weather data but far apart in geographical distance.

[0013] 3. Static graph limitations: Weather correlations change dynamically over time, but existing graph structures cannot adaptively adjust. SUMMARY

[0014] The technical problem solved by the present application is to overcome the shortcomings of existing technologies and propose a precipitation prediction graph convolution method based on graph reconstruction and spatial self-attention, which:

[0015] 1. Avoids the over-smoothing problem caused by traditional GCN due to geographical adjacency;

[0016] 2. Dynamically captures long-range weather dependency relationships between base stations;

[0017] 3. Constructs a graph structure that can reflect both stable associations and adapt to short-term changes.

[0018] The present application proposes an innovative precipitation prediction graph convolution method, which breaks through the limitations of traditional geographical adjacency graphs through meteorological data-driven graph reconstruction and dynamic spatial self-attention mechanism. The method first abandons the graph construction method based on physical distance and instead starts from the correlation of meteorological data itself to establish a graph structure that better reflects the real weather relationship.

[0019] The technical solution of the present application is:

[0020] A precipitation prediction graph convolution method based on graph reconstruction and spatial self-attention, the steps of the method comprising:

[0021] First, continuously collect real-time monitoring data streams of Beidou weather stations at 30-second intervals, process the collected monitoring data streams using a sliding window to obtain input features, and set labels for the obtained input features, the labels being the amount of precipitation after 40 minutes;

[0022] Second, standardize the input features obtained in the first step to obtain standardized input features;

[0023] Third, construct a precipitation prediction model based on graph convolution and self-attention mechanism;

[0024] Fourth, input the standardized input features obtained in the second step into the graph convolution neural network constructed in the third step for training;

[0025] Fifth, use the trained graph convolution neural network in the fourth step to predict precipitation, i.e., the inference process.

[0026] In the first step, the monitoring data stream includes temperature, humidity, and air pressure, etc. key meteorological parameters;

[0027] The size of the sliding window is 10 minutes, and the step is 2 minutes;

[0028] In the second step, the Z-score standardization method is used for standardization, specifically:

[0029] Calculate the Z-score of temperature in each window, and keep the mean and standard deviation of temperature;

[0030] Calculate the Z-score of humidity in each window, and keep the mean and standard deviation of humidity;

[0031] Calculate the Z-score of air pressure in each window, and keep the mean and standard deviation of air pressure;

[0032] In the third step, the precipitation prediction model based on graph convolution and self-attention mechanism includes a graph structure construction module, a multi-order graph convolution layer, a spatial self-attention encoder, and a precipitation prediction head;

[0033] Each module is connected by a residual connection to form a complete network for processing spatio-temporal data of meteorological station networks and outputting future precipitation prediction;

[0034] The graph structure construction module is composed of one geographical adjacency calculation unit and one dynamic similarity calculation unit, and is responsible for generating a topological relationship graph of meteorological base stations, wherein the geographical adjacency unit adopts a KNN algorithm (k=5) to establish an initial connection, and the dynamic similarity unit calculates the meteorological data correlation between base stations through a DTW distance in a sliding window, and finally outputs a mixed adjacency matrix A∈R^(N×N) (N is the number of base stations);

[0035] The multi-order graph convolution layer is composed of two graph convolution sub-layers, which are a first sub-layer and a second sub-layer.

[0036] Each graph convolution sub-layer includes one adjacency matrix normalization unit (using a symmetric normalization method), one feature transformation unit (weight matrix W∈R^(d×d')) and one gated residual connection (gating coefficient γ=0.3).

[0037] The input dimension of the first sub-layer is N×20×3 (base station×time step×feature), and the output is N×20×128.

[0038] The second sub-layer maintains the same dimension and uses a skip connection to prevent gradient disappearance.

[0039] The spatial self-attention encoder is composed of the following components:

[0040] 1 static-dynamic fusion unit: dynamically fuses static features output by the graph convolution and real-time meteorological data;

[0041] 2 layer normalization modules: before and after attention calculation, respectively;

[0042] 1 multi-head graph attention module (4 heads, each head dimension 32) and 1 spatio-temporal projection feedforward network (hidden layer dimension 512);

[0043] 3 residual connections, respectively, a skip connection from the fusion unit input to the layer normalization output, a connection from the first layer normalization to the attention output, and a connection from the second layer normalization to the feedforward network output.

[0044] The precipitation prediction uses an RNN layer.

[0045] The residual connection uses a gating mechanism, and the gating coefficient is dynamically calculated according to the current meteorological conditions.

[0046] During the inference process, the graph reconstruction is triggered every 2 minutes in the real-time running phase. After the new meteorological data is input, it is first standardized and the dynamic similarity matrix is calculated, and after updating the adjacency relationship, the graph convolution operation is performed. The fusion result of the spatial feature and the time series feature is sent to the prediction module to generate a double-scale precipitation probability distribution for the next 40 minutes and 60 minutes.

[0047] Advantages

[0048] The application provides a precipitation prediction graph convolution method based on meteorological data correlation graph reconstruction and spatial self-attention, which is used for solving the problems of over-smoothing and insufficient spatial dependence capture caused by dependence on geographical adjacency in traditional graph convolution network in precipitation prediction. The method comprises the following steps: 1) a sparse graph structure reflecting real meteorological correlation is constructed by calculating the Pearson correlation coefficient of Beidou multi-base station meteorological time series data, and the traditional geographical adjacency matrix is replaced; 2) a dynamic spatial self-attention mechanism is designed, the static correlation graph and the dynamic attention weight are fused, and adaptive aggregation of long-range dependence between nodes is realized; 3) the multi-level feature propagation of the graph convolution and the attention matrix is combined, and the modeling capability for heterogeneous meteorological spatial patterns is improved. The method significantly improves the accuracy of short-impending precipitation prediction, is especially suitable for base station correlation mining of geographical distance far but meteorological characteristics similar, and provides a high-robustness spatial feature extraction scheme for a meteorological warning system. DETAILED DESCRIPTION

[0049] The application will be further described below in combination with examples.

[0050] Real-time meteorological monitoring data from Beidou multi-base stations are received, including temperature, humidity, air pressure and other key parameters. These data are continuously collected at an interval of 30 seconds to form high-precision time series. For the data in each sliding time window, the system calculates the Pearson correlation coefficient between the meteorological sequences of each base station, and evaluates the correlation strength between the base stations by quantifying the similarity of the data curves. This evaluation method based on data similarity can discover important correlations that cannot be captured by traditional geographical distance.

[0051] An adaptive threshold strategy is adopted in the graph reconstruction process to determine the effective edge connection. The system retains the highest correlation coefficient for each base station and removes weakly correlated connections. This sparse processing not only reduces the computational complexity, but more importantly, avoids feature mixing between irrelevant nodes. In order to ensure that the node's own features are not diluted, each base station will retain a self-loop edge pointing to itself. The graph structure constructed in this way can accurately reflect the real correlation pattern in the meteorological system.

[0052] To further enhance the dynamic adaptability of the model, the application introduces a spatial self-attention mechanism. This mechanism can dynamically adjust the information propagation weight between nodes according to real-time meteorological features. Specifically, the system will consider both the static correlation coefficient graph and the dynamically calculated attention weight, and obtain the final adjacency matrix through weighted fusion. This design enables the model to both grasp stable meteorological relationships and flexibly respond to sudden weather changes.

[0053] In the feature propagation stage, the system adopts a multi-layer graph convolutional network to aggregate node information. Each layer of convolution operation is based on the fused adjacency matrix, ensuring efficient flow of information between truly relevant nodes. To prevent over-smoothing problems in deep networks, residual connections are added between layers to preserve the original features of the nodes. This design not only achieves deep feature extraction, but also maintains the discrimination between nodes.

[0054] The output of the entire graph convolution module is the spatially aggregated node feature representation, which not only contains local weather information, but also incorporates global spatial correlation patterns. Experiments show that compared to traditional geographical adjacency graphs, this graph structure based on data correlation reconstruction can significantly improve the model's ability to recognize complex weather patterns, especially when dealing with long-range correlations and sudden weather events.

[0055] In the method of the present application, the meteorological data-driven graph reconstruction: a sparse graph is constructed by Pearson correlation coefficient, replacing the geographical adjacency relationship, solving the problem of over-smoothing. Dynamic spatial self-attention: fusion of static graph and dynamic attention weight, adaptive capture of long-range meteorological dependence. Residual connection design: preserve the original features of the nodes between GCN layers to prevent feature degradation. Graph reconstruction focuses the model on strongly correlated base stations, improving prediction accuracy; dynamic attention mechanism improves the response speed of sudden weather events, enhancing the robustness of the system, and the method can be extended to other spatiotemporal prediction tasks, with strong universality.

[0056] Similarity measure replacement: Pearson correlation coefficient can be replaced by cosine similarity or mutual information.

[0057] Graph sparsification strategy: kNN can be replaced by Top-k percentage threshold.

[0058] Attention fusion method: weighted sum can be replaced by a gating mechanism (such as GRU).

[0059] Embodiment

[0060] To verify the actual effect of the present application, real-time monitoring data of 25 Beidou meteorological base stations in Beijing urban area and surrounding areas (covering Haidian, Chaoyang, Fengtai, etc. administrative areas and Yanqing, Miyun, etc. remote suburbs) in July 2023 were selected for prediction. Each base station collected temperature, humidity, and pressure data at 30-second intervals to form continuous time series.

[0061] The original data stream is processed using a 10-minute sliding window (step size 2 minutes). Taking the Zhongguancun base station in Haidian District as an example, the mean temperature in a certain window is 28.6°C, and the standard deviation is 0.8; the mean humidity is 78%, and the standard deviation is 5.2; the mean pressure is 1002.3 hPa, and the standard deviation is 0.7. After Z-score standardization, the temperature sequence in this period is converted to normalized values [-0.25, 1.02, 0.33,...], and humidity and pressure are processed synchronously. After standardization of all base station data, an input tensor with dimensions 25x20x3 (25 base stations x 20 time steps x 3 features) is formed.

[0062] The graph reconstruction module first calculates the Pearson correlation coefficient of the time series data between the base stations, and retains the top 5 strong correlation connections for each base station by adaptive thresholding, generating a sparse adjacency matrix. At the same time, the spatial self-attention module detects that the pressure of a certain base station in a certain area has abnormally decreased in the later period of the window, dynamically enhancing its connection weight (attention score from 0.3 to 0.7) with the base stations in the southwest direction, effectively capturing the signal of the thunderstorm cloud cluster moving eastward.

[0063] The graph convolution network performs feature aggregation on the reconstructed graph structure, avoiding noise interference from irrelevant nodes. The gated residual connection preserves the original features of the nodes, preventing over-smoothing problems in deep networks. The spatial self-attention encoder further identifies dynamic weather patterns and enhances information propagation on key paths.

[0064] The final output is a 40-minute precipitation probability distribution, showing a high-probability zone in the northwest-southeast direction, consistent with the actual radar echo movement path. Compared with the traditional geographical adjacency GCN model, the prediction time of short-term heavy precipitation (>20mm / h) in Haidian District is improved from 12 minutes to 23 minutes, and the false positive rate is reduced by 34%.

[0065] This embodiment demonstrates that this method can effectively adapt to local meteorological differences caused by complex underlying surfaces in Beijing, accurately express the heterogeneous relationships of mountainous areas, urban areas and suburban areas through data-driven graph reconstruction, and significantly improve the tracking ability of rapidly evolving severe convective weather through dynamic attention mechanisms.

[0066] In summary, the above is only a preferred embodiment of the present application, and is not intended to limit the scope of protection of the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A precipitation prediction graph convolution method based on graph reconstruction and spatial self-attention, characterized by The steps of the method include: First, collect the real-time monitoring data stream of the Beidou weather base at a set time interval, process the collected monitoring data stream using a sliding window to obtain input features, and set labels for the obtained input features, the labels being the precipitation after 40 minutes; Second, standardize the input features with labels obtained in the first step to obtain standardized input features; Third, construct a precipitation prediction model based on graph convolution and self-attention mechanism; Fourth, input the standardized input features obtained in the second step into the graph convolution neural network constructed in the third step for training; Fifth, use the trained graph convolution neural network in the fourth step to predict precipitation.

2. The precipitation prediction graph convolution method based on graph reconstruction and spatial self-attention according to claim 1, wherein: In the first step, the time interval is set to 30 seconds.

3. The precipitation prediction graph convolution method based on graph reconstruction and spatial self-attention according to claim 1, wherein: In the first step, the monitoring data stream includes temperature, humidity, and air pressure.

4. The precipitation prediction graph convolution method based on graph reconstruction and spatial self-attention according to claim 1, wherein: The size of the sliding window is 10 minutes, and the step is 2 minutes.

5. The precipitation prediction graph convolution method based on graph reconstruction and spatial self-attention according to claim 1, wherein: In the second step, the Z-score standardization method is used for standardization, specifically: Calculate the Z-score of temperature in each window, and keep the mean and standard deviation of temperature; Calculate the Z-score of humidity in each window, and keep the mean and standard deviation of humidity; Calculate the Z-score of air pressure in each window, and keep the mean and standard deviation of air pressure.

6. The precipitation prediction graph convolution method based on graph reconstruction and spatial self-attention according to claim 1, wherein: In the third step, the precipitation prediction model based on graph convolution and self-attention mechanism includes a graph structure construction module, a multi-order graph convolution layer, a spatial self-attention encoder, and a precipitation prediction head.

7. The precipitation prediction graph convolution method based on graph reconstruction and spatial self-attention according to claim 6, wherein: The graph structure construction module is composed of 1 geographical adjacency calculation unit and 1 dynamic similarity calculation unit, and is used to generate a topological relationship graph of the weather base; The geographical adjacency unit uses the KNN algorithm to establish an initial connection; The dynamic similarity unit calculates the correlation of meteorological data between stations through the DTW distance in the sliding window, and finally outputs a mixed adjacency matrix A∈R^(N×N), N being the number of stations.

8. The precipitation prediction graph convolution method based on graph reconstruction and spatial self-attention according to claim 6, wherein: The multi-order graph convolution layer is composed of 2 graph convolution sublayers, which are the first sublayer and the second sublayer; Each graph convolution sublayer includes: 1 adjacency matrix normalization unit, which uses a symmetric normalization method; 1 feature transformation unit, weight matrix W ∈ R^(d×d'); 1 gated residual connection, gating coefficient γ = 0.3; The first sub-layer input dimension is N × 20 × 3, and the output is N × 20 × 128; The second sub-layer keeps the dimension unchanged, and adopts a skip connection to prevent gradient disappearance.

9. The precipitation prediction graph convolution method based on graph reconstruction and spatial self-attention according to claim 6, characterized in that: The spatial self-attention encoder is composed of the following components: 1 static-dynamic fusion unit: dynamically fuses the static features output by the graph convolution and the real-time meteorological data; 2 layer normalization modules: before and after attention calculation respectively; 1 multi-head graph attention module: including 4 heads and 1 spatio-temporal projection feedforward network, each head dimension 32, spatio-temporal projection feedforward network hidden layer dimension 512; 3 residual connections: skip connection from fusion unit input to layer normalization output, connection from first layer normalization to attention output, and connection from second layer normalization to feedforward network output. The precipitation prediction adopts an RNN layer; The residual connection adopts a gating mechanism, and the gating coefficient is dynamically calculated according to the current meteorological conditions.

10. The precipitation prediction graph convolution method based on graph reconstruction and spatial self-attention according to claim 1, characterized in that: In the fifth step, during the prediction process, the graph reconstruction is triggered every 2 minutes in the real-time running stage, after the new meteorological data is input, it is first normalized and the dynamic similarity matrix is calculated, the graph convolution operation is performed after the adjacency relationship is updated, and the fusion result of the spatial features and the time series features is sent to the prediction module to generate the double-scale precipitation probability distribution of the next 40 minutes and 60 minutes.