Photovoltaic power prediction method based on multi-head graph attention mechanism of multi-dimensional meteorological elements

Abstract

The present invention proposes a photovoltaic power generation prediction method based on a multi-head graph attention mechanism of multi-dimensional meteorological elements. On the one hand, the importance of multi-dimensional meteorological elements that affect photovoltaic power generation is studied, and the influence weight graph between multi-dimensional meteorological elements and photovoltaic power generation is modeled, and the influence weight of multi-dimensional meteorological elements on photovoltaic power generation is captured with the help of a multi-head graph attention mechanism. On the other hand, based on the spatial autocorrelation described in the first law of geography, the spatial dependence between photovoltaic power stations is modeled, and a method for mining the correlation between power generation between photovoltaic power stations is used with the help of a multi-head graph attention mechanism.

Description

Technical Field

[0001] The present invention belongs to the field of photovoltaic technology, and in particular relates to a photovoltaic power generation prediction method based on a multi-dimensional meteorological element multi-head graph attention mechanism. Background Art

[0002] The efficient use of photovoltaics is an important driving force for the production and development of new energy. However, factors such as the geographical environment and extreme weather make the power generation of photovoltaic stations volatile, which has a certain impact on the stable operation of photovoltaic stations. Therefore, the accurate prediction of photovoltaic power is of great significance to the stable operation and dispatch of the power grid.

[0003] The traditional photovoltaic power prediction method is a mechanism-driven method. According to the physical principles of photovoltaic power generation, a mechanism prediction model containing meteorological data and photovoltaic plant parameters is established. A modified current-voltage relationship of a single diode model is proposed in an existing literature. The model starts with a single solar cell, expands to photovoltaic modules, and then expands to photovoltaic arrays. The model parameters are regressed with experimental data at different irradiance and temperature levels. Another existing literature establishes an empirical model between photovoltaic power and weather conditions, including irradiance, ambient temperature and humidity. Then, the parameters of the model are regressed using the actual observations of the photovoltaic modules.

[0004] The actual photovoltaic system has time-varying characteristics and time-varying parameters, so the power generation modeling is often very complicated. Regression fitting based on complex physical models cannot provide satisfactory performance, especially on cloudy and rainy days with complex weather changes.

[0005] The data-driven method achieves accurate power prediction by machine learning the mapping relationship between historical meteorological data and power generation. One existing document proposed photovoltaic power prediction based on wavelet neural network, but its defect is that it does not fully consider the time correlation of historical data. Another existing document uses support vector machine to learn the weather condition mapping between satellite images and photovoltaic power, which relatively improves the prediction accuracy. Another existing document uses convolutional neural networks (CNN) and recurrent neural networks (RNN) technology to further improve the power prediction effect.

[0006] In recent years, some studies have attempted to use graphs to model the spatial relationship between photovoltaic power stations. The existing patent established a graph convolutional neural network model based on the topological structure of regional photovoltaic power stations, aggregated the spatial features of neighboring photovoltaic fields, and output a set of spatial features containing spatial evolution patterns. It is worth noting that the existing patent constructs the adjacency matrix in the graph convolution process based on the Pearson coefficient between photovoltaic power generation powers.

[0007] Unmodeled weights of the impact of multidimensional meteorological elements on the power generation of photovoltaic power stations: Traditional photovoltaic power generation prediction models often fail to fully capture the complex impact of multidimensional meteorological elements on the power generation of photovoltaic power stations. These models may rely on simple linear relationships or fixed parameterization methods, and cannot dynamically identify and quantify the specific impact weights of different meteorological elements on power generation under conditions of spatial heterogeneity. The multi-head graph attention mechanism can automatically learn the complex relationships between meteorological elements and dynamically adjust the weights of each element to reflect its true impact on power generation. Through the graph structure, the mechanism not only considers the impact of a single meteorological element, but also the interaction between elements, providing a more comprehensive analysis method. The present invention aims to solve the limitations of the prior art in dealing with the weights of the impact of multidimensional meteorological elements by introducing a multi-head graph attention mechanism.

[0008] The prior art has not fully considered the impact of the meteorological characteristics of spatially related photovoltaic power stations on the predicted power generation of the target photovoltaic power station: The first law of geography states that "Everything is related to everything else, but near things are more related than distant things". The present invention interprets this law in the task of predicting the power generation of photovoltaic power stations as follows: when predicting the power generation of a photovoltaic power station, it is necessary not only to make predictions based on its own meteorological elements and historical power generation, but also to consider the meteorological elements and historical power generation of adjacent photovoltaic power stations. The relationship between the power generation of adjacent photovoltaic power stations stems from the fact that the meteorological elements that affect photovoltaic power generation have spatial autocorrelation. However, the prior art only constructs a graphical model from the perspective of the correlation between the power generation of photovoltaic power stations, while ignoring the important impact of the meteorological elements of adjacent power stations on the predicted power generation of the target power station, resulting in the accuracy of the prediction model being restricted. Summary of the invention

[0009] The purpose of the present invention is to overcome the above shortcomings of the prior art. By establishing a weight map that captures the impact of multidimensional meteorological factors on the power generation of photovoltaic power stations and a spatial map of spatial correlation, and combining a multi-head graph attention neural network (GAT) that focuses on deep learning technology for graph structure, the weight of the impact of multidimensional meteorological factors on the power generation of photovoltaic power stations, as well as the multidimensional meteorological factor characteristics and historical power generation characteristics of spatially correlated photovoltaic power stations, the accuracy of photovoltaic power generation prediction is improved.

[0010] On the one hand, the present invention studies the importance of multidimensional meteorological factors that affect photovoltaic power generation, studies the modeling of the influence weight graph between multidimensional meteorological factors and photovoltaic power generation, and uses the multi-head graph attention mechanism to capture the influence weight of multidimensional meteorological factors on photovoltaic power generation. On the other hand, based on the spatial autocorrelation described in the first law of geography, the spatial dependence between photovoltaic power stations is modeled, and a method for mining the correlation between power generation between photovoltaic power stations is used with the help of the multi-head graph attention mechanism.

[0011] The present invention includes the following four aspects:

[0012] (1) Construction of a weighted graph of the impact of multidimensional meteorological factors on photovoltaic power generation: The present invention uses a single photovoltaic power station as a node in the graph model, and uses multidimensional meteorological factors that affect photovoltaic power generation, including air pressure, mean sea level air pressure, ambient temperature, dew point temperature, total precipitation, surface solar radiation to the ground, latitudinal wind speed at a height of 10 meters, longitudinal wind speed at a height of 10 meters, total cloud cover, low cloud cover, medium cloud cover, high cloud cover, and clear sky irradiance, a total of 13 dimensions of meteorological factors, as nodes in the graph model. The present invention models the impact weights between multidimensional meteorological factors and photovoltaic power generation based on a graph structure. The greater the impact, the higher the weight. The present invention uses a parameterized impact weight graph as an adjacency matrix for a subsequent multi-head graph attention module, which can effectively capture the impact weights of multidimensional meteorological factors on photovoltaic power generation, and enhance the understanding of the impact of meteorological factors on photovoltaic power generation.

[0013] (2) Construction of a spatial graph of photovoltaic power stations based on Euclidean space distance: The present invention uses a single photovoltaic power station as a node in the graph model. The present invention sets a distance threshold to measure whether there is spatial correlation between photovoltaic power stations. If the Euclidean space distance between two photovoltaic power stations is greater than the set threshold, the spatial correlation between the other photovoltaic power station and it is not considered when predicting the power generation trend of one of the photovoltaic power stations. Edges are established between photovoltaic power station nodes whose Euclidean space distance is less than the set threshold, and the weights of the edges are set as learnable parameters to form a spatial graph. The present invention uses the spatial graph as an adjacency matrix for the subsequent multi-head graph attention module, and can effectively aggregate the meteorological and power generation characteristics of adjacent photovoltaic power stations according to the weights to enhance the ability to predict the power generation trend of the target photovoltaic power station.

[0014] (3) Photovoltaic power generation graph neural network modeling: Graph neural network (GNN) is a deep learning technology that focuses on graph structure. It can aggregate information of related regions and help simulate complex associations in geographic space. Multi-head graph attention neural network (GAT) is one type of graph neural network. The present invention establishes a photovoltaic power generation graph neural network model, a multi-head graph attention module that uses the influence weight graph as the adjacency matrix to capture the influence weights of multi-dimensional meteorological factors on the power generation of photovoltaic power stations, and a multi-head graph attention module that uses the spatial graph as the adjacency matrix to model the spatial dependence between photovoltaic power stations, and convolves the time dimension to achieve end-to-end photovoltaic power generation time series prediction.

[0015] (4) Construction of sample data sets: It is necessary to construct a sample data set of the geographical unit where the photovoltaic power station is located, including air pressure (Pa), mean sea level air pressure (Pa), ambient temperature (°C), dew point temperature (°C), total precipitation (M), surface solar radiation to the ground (W / m²), latitudinal wind speed at 10 meters (m / s), longitudinal wind speed at 10 meters (m / s), total cloud cover, low cloud cover, medium cloud cover, high cloud cover, hourly sine, hourly cosine, clear sky irradiance (W / m²), solar zenith angle, solar azimuth angle and photovoltaic power generation (W) for model training, verification and testing. These data sets are key resources in the research and development process.

[0016] More specifically, the present invention proposes a photovoltaic power prediction method based on a multi-dimensional meteorological element multi-head graph attention mechanism, characterized in that the method comprises the following steps:

[0017] Step 1: construct a weighted graph of the impact of multidimensional meteorological factors on the power generation of photovoltaic power stations; use the parameterized weighted graph as an adjacency matrix for the subsequent multi-head graph attention layer to capture the weights of the impact of multidimensional meteorological factors on the power generation of photovoltaic power stations, and enhance the understanding of the impact of meteorological factors on the power generation of photovoltaic power stations;

[0018] Step 2: Construct a spatial graph of PV plants based on Euclidean space distance. Take a single PV plant as a node in the graph model, establish edges between PV plant nodes whose Euclidean space distance is less than a set threshold, and set the weights of the edges as learnable parameters to form a spatial graph. Use the spatial graph as an adjacency matrix for the subsequent multi-head graph attention layer, and effectively aggregate the meteorological and power generation characteristics of adjacent PV plants according to the weights to enhance the prediction capability of the power generation trend of the target PV plant.

[0019] Step 3: PV power generation graph neural network modeling; the multi-head graph attention layer uses the impact weight graph as the adjacency matrix to capture the impact weights of multi-dimensional meteorological factors on the power generation of PV power stations, and the spatial graph is used as the adjacency matrix to model the spatial dependency between PV power stations. The time dimension is convolved to achieve end-to-end PV power generation power time series prediction;

[0020] Step 4, sample data set construction: construct a time series sample data set of the geographical unit where the photovoltaic power station is located, including air pressure, mean sea level air pressure, ambient temperature, dew point temperature, total precipitation, surface solar radiation to the ground, latitudinal wind speed at a height of 10 meters, longitudinal wind speed at a height of 10 meters, total cloud cover, low cloud cover, medium cloud cover, high cloud cover, hourly sine, hourly cosine, clear sky irradiance, solar zenith angle, solar azimuth angle and photovoltaic power generation power for model training, verification and testing.

[0021] The present invention proposes a method for constructing a weighted graph of the impact of multidimensional meteorological factors on the power generation of photovoltaic power stations. The greater the impact, the higher the weight. The present invention uses the parameterized impact weight graph as an adjacency matrix for the subsequent multi-head graph attention module, aiming to enable the model to capture the impact weights between meteorological factors and the impact weights of multidimensional meteorological factors on the power generation of photovoltaic power stations, and enhance the understanding of the impact of meteorological factors on the power generation of photovoltaic power stations.

[0022] A distance threshold is set to measure whether there is spatial correlation between photovoltaic power stations. Edges are established between photovoltaic power station nodes whose Euclidean space distance is less than the set threshold, and the weights of the edges are set as learnable parameters to form a spatial graph. The present invention uses the spatial graph as an adjacency matrix for the subsequent multi-head graph attention module, aiming to effectively aggregate the meteorological and power generation characteristics of adjacent photovoltaic power stations according to the weights.

[0023] By establishing a weight map that captures the impact of multidimensional meteorological factors on the power generation of photovoltaic power stations and a spatial map of spatial correlation, and combining the multi-head graph attention neural network (GAT) which is a deep learning technology focusing on graph structure, we can fully utilize the weights of the impact of multidimensional meteorological factors on the power generation of photovoltaic power stations, as well as the multidimensional meteorological factor characteristics and historical power generation characteristics of spatially related photovoltaic power stations, to improve the accuracy of photovoltaic power generation prediction.

[0024] The present invention has the following beneficial technical effects:

[0025] The present invention proposes a photovoltaic power generation prediction method based on multidimensional meteorological elements and graph neural networks. By establishing a weight graph that captures the impact of multidimensional meteorological elements on the power generation of photovoltaic power stations and a spatial graph of spatial correlation, and combining the multi-head graph attention neural network (GAT) which is a deep learning technology focusing on graph structure, the weight of the impact of multidimensional meteorological elements on the power generation of photovoltaic power stations, as well as the impact of multidimensional meteorological elements of spatially correlated photovoltaic power stations on the power generation of target photovoltaic power stations are comprehensively considered, forming a graph model that is different from only considering the meteorological elements of the target photovoltaic power station to be predicted and only constructing a graph model from the perspective of the correlation between power generation.

[10] A new method is developed to better understand the power generation law of photovoltaic power stations, providing a technical approach to achieve more accurate power generation prediction.

[0026] Traditional photovoltaic power prediction models often fail to fully capture the complex impact of multidimensional meteorological factors on photovoltaic power generation. These models may rely on simple linear relationships or fixed parameterization methods, and cannot dynamically identify and quantify the specific impact weights of different meteorological factors on power generation under conditions of spatial heterogeneity. The multi-head graph attention mechanism can automatically learn the complex relationships between meteorological factors and dynamically adjust the weights of each factor to reflect its true impact on power generation. Through the graph structure, the mechanism not only considers the impact of a single meteorological factor, but also the interaction between the factors, providing a more comprehensive analysis method. The present invention aims to solve the limitations of the prior art in dealing with the impact weights of multidimensional meteorological factors by introducing a multi-head graph attention mechanism.

[0027] The prior art only constructs the graph model from the perspective of the correlation between the power generation of photovoltaic power stations, while ignoring the important influence of the meteorological factors of adjacent power stations on the prediction of the power generation of the target power station. Different from the graph model used in the prior art, the present invention sets a distance threshold to measure whether there is spatial correlation between photovoltaic power stations. Edges are established between photovoltaic power station nodes whose Euclidean space distance is less than the set threshold, and the weights of the edges are set to learnable parameters to form a spatial graph. The present invention uses the spatial graph as an adjacency matrix for the subsequent multi-head graph attention module, which can effectively aggregate the meteorological and power generation characteristics of adjacent photovoltaic power stations according to the weights to enhance the ability to predict the power generation trend of the target photovoltaic power station. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Schematic diagram of the adjacency matrix that influences the weight graph;

[0029] Figure 2 It is a schematic diagram of the adjacency matrix of the spatial graph;

[0030] Figure 3 Schematic diagram of the structure of the neural network model for photovoltaic power generation power graph. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical scheme and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in each embodiment of the present invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above-mentioned purpose, the present invention adopts the following technical scheme.

[0032] The present invention proposes a photovoltaic power generation power prediction method based on a multi-dimensional meteorological element graph neural network, which specifically comprises the following steps:

[0033] Step 1: construct a weight map of the impact of multi-dimensional meteorological factors on the power generation of photovoltaic power stations;

[0034] The present invention uses a single photovoltaic power station as a node in the graph model, and uses multidimensional meteorological factors that affect photovoltaic power generation, including air pressure (Pa), average sea level air pressure (Pa), ambient temperature (°C), dew point temperature (°C), total precipitation (M), surface solar radiation to the ground (W / m²), latitudinal wind speed at a height of 10 meters (m / s), longitudinal wind speed at a height of 10 meters (m / s), total cloud cover, low cloud cover, medium cloud cover, high cloud cover, clear sky irradiance (W / m²), a total of 13 dimensional meteorological factors are also used as nodes in the graph model, such as Figure 1 shown.

[0035] is a learnable parameter, which represents the influence weights between multidimensional meteorological elements and multidimensional meteorological elements on the power generation of photovoltaic power stations, forming an adjacency matrix of the influence weight graph .

[0036] .

[0037] 4.2 Construction of spatial map of photovoltaic power station based on Euclidean space distance;

[0038] The present invention uses a single photovoltaic power station as a node in the graph model. The present invention sets a distance threshold To measure whether there is spatial correlation between photovoltaic power stations. If the Euclidean space distance between two photovoltaic power stations Greater than the set threshold , then when predicting the power generation trend of one PV power station, the spatial correlation between the other PV station and it is not considered.

[0039] Distance in Euclidean space Less than the set threshold The edges are established between the nodes of the photovoltaic power station and the weights of the edges are set as learnable parameters. , forming a spatial weight matrix .

[0040] .

[0041] Based on the spatial weight matrix Calculate the cosine similarity between photovoltaic power plants and establish the spatial graph adjacency matrix , as shown in formula (1) and Figure 2 shown.

[0042] (1)

[0043] and Indicates Meteorological elements In photovoltaic power plants and The value on . The values ​​of constitute the spatial weight matrix .

[0044] 4.3 Photovoltaic power generation power graph neural network modeling;

[0045] Based on research objectives Photovoltaic power stations in (Length is Time series) of meteorological elements and power generation, forecast Photovoltaic power stations in (Length is The power generation capacity is the time series of the power generation capacity.

[0046] The structure of the photovoltaic power generation power graph neural network model is as follows Figure 3 The photovoltaic power generation multi-head graph attention mechanism model includes: input layer, temporal convolution layer, two fully connected layers, and two multi-head graph attention layers. Among them, each layer input includes parameters , , , ,in, for , is the number of samples used to update the model weights in one training. is the feature dimension, which refers to the number of features in the input data. is the length of the time series. In time series forecasting, longer time series can provide more time-dependent information, but will increase computational complexity. The choice of time series length depends on the requirements of the specific task and the amount of available data. is the number of nodes. In this invention, the number of nodes is the number of photovoltaic power stations. , will not change due to the action of the neural network.

[0047] Input layer: The input layer is the first layer of the neural network and is responsible for receiving raw input data. In photovoltaic power generation prediction, the input data includes meteorological elements and power generation. The input layer does not perform any calculations, but simply passes the data to the next layer.

[0048] Temporal convolutional layers: Temporal convolutional layers are specifically designed to process time series data. They extract temporal correlations through convolution operations. There are usually multiple temporal convolutional layers stacked together to capture features at different time scales.

[0049] Fully connected layer: maps the output of the previous layer to a high-dimensional space for more complex nonlinear transformations. After the temporal convolutional layer extracts temporal features, the fully connected layer can further extract and combine these features to provide richer information for the final prediction task.

[0050] The inputs of the two multi-head graph attention layers are and : The multi-head graph attention layer is another form of graph neural network, which allows the model to learn the importance or weights between different nodes in the graph. Through the multi-head attention mechanism, the model can identify the most important nodes and relationships for the prediction task, thereby improving the accuracy of the prediction.

[0051] The calculation formula of self-attention in the multi-head graph attention layer is shown in formulas (2)-(6), which is used to calculate the node Its neighboring nodes (satisfy Take the attention coefficient and weighted sum as an example.

[0052] (2)

[0053] (3)

[0054] (4)

[0055] (5)

[0056] (6)

[0057] in Indicates that the node is The feature vector of the layer, It means that after convolution, the node is The feature vector of the layer. , , , represents the learnable parameters, Indicates concatenation of vectors. For the number of heads, The function is used to convert the scores into probabilities, ensuring that all attention scores add up to 1. is the scaling factor used when computing the dot product to maintain numerical stability, Indicates the multi-head attention calculation process for the vector, Indicates the attention mechanism calculation for the vector. Refers to The attention calculation results of each person.

[0058] 4.4 Sample data set construction and model optimization;

[0059] Collection research objectives Photovoltaic power stations in (Length is The data set consists of meteorological elements (time series) and historical data of power generation.

[0060] Model input: Photovoltaic power stations in (Length is The meteorological elements and power generation power of the photovoltaic power station in a certain time series The similarity is established ; Based on the Euclidean space distance of photovoltaic power stations ; Model tag: Photovoltaic power stations in (Length is The power generation of the time series (time series);

[0061] Build the sample as above The samples are divided into training set, validation set and test set in a 6:2:2 ratio. The training set is used to train the model; the validation set is used to adjust the model hyperparameters; and the test set is used to test the final accuracy of the model.

[0062] When applying the model, collect data and build Photovoltaic power stations in (Length is The feature matrix, attribute graph and spatial graph of the time series are input into the trained model for forward propagation to obtain Photovoltaic power stations in (Length is The power generation prediction results of the time series (time series).

Claims

1. A photovoltaic power generation prediction method based on multi-dimensional meteorological elements multi-head graph attention mechanism, characterized in that: The method comprises the following steps, Step 1: construct a weighted graph of the impact of multidimensional meteorological factors on the power generation of photovoltaic power stations; use the parameterized weighted graph as an adjacency matrix for the subsequent multi-head graph attention layer to capture the weights of the impact of multidimensional meteorological factors on the power generation of photovoltaic power stations, and enhance the understanding of the impact of meteorological factors on the power generation of photovoltaic power stations; Step 2: Construct a spatial graph of PV plants based on Euclidean space distance. Take a single PV plant as a node in the graph model, establish edges between PV plant nodes whose Euclidean space distance is less than a set threshold, and set the weights of the edges as learnable parameters to form a spatial graph. Use the spatial graph as an adjacency matrix for the subsequent multi-head graph attention layer, and effectively aggregate the meteorological and power generation characteristics of adjacent PV plants according to the weights to enhance the prediction capability of the power generation trend of the target PV plant. Step 3: PV power generation graph neural network modeling; the multi-head graph attention layer uses the impact weight graph as the adjacency matrix to capture the impact weights of multi-dimensional meteorological factors on the power generation of PV power stations, and the spatial graph is used as the adjacency matrix to model the spatial dependency between PV power stations. The time dimension is convolved to achieve end-to-end PV power generation power time series prediction; Step 4, sample data set construction: construct a sample data set of the geographical unit where the photovoltaic power station is located, including air pressure, mean sea level air pressure, ambient temperature, dew point temperature, total precipitation, surface solar radiation to the ground, latitudinal wind speed at a height of 10 meters, longitudinal wind speed at a height of 10 meters, total cloud cover, low cloud cover, medium cloud cover, high cloud cover, hourly sine, hourly cosine, clear sky irradiance, solar zenith angle, solar azimuth angle and photovoltaic power generation time series sample data set for model training, verification and testing; In step 2, a single PV plant is used as a node in the graph model and a distance threshold is set. To measure whether there is spatial correlation between photovoltaic power stations, if the Euclidean space distance between two photovoltaic power stations is Greater than the set threshold , then when predicting the power generation trend of one PV power station, the spatial correlation between the other PV station and it is not considered. Less than the set threshold The edges are established between the nodes of the photovoltaic power station and the weights of the edges are set as learnable parameters. , forming a spatial weight matrix , ； Based on the spatial weight matrix Calculate the cosine similarity between photovoltaic power plants and establish the spatial graph adjacency matrix , as shown in formula (1), (1) and Indicates Meteorological elements In photovoltaic power plants and The value on The values ​​of constitute the spatial weight matrix .

2. The method according to claim 1, characterized in that In step 1, when establishing the influence weight graph, a single photovoltaic power station is used as a node in the graph model, and the multidimensional meteorological factors that affect photovoltaic power generation, including air pressure, mean sea level air pressure, ambient temperature, dew point temperature, total precipitation, surface solar radiation to the ground, latitudinal wind speed at a height of 10 meters, longitudinal wind speed at a height of 10 meters, total cloud cover, low cloud cover, medium cloud cover, high cloud cover, and clear sky irradiance, a total of 13 dimensional meteorological factors, are also used as nodes in the graph model; is a learnable parameter, which represents the influence weights between multidimensional meteorological elements and multidimensional meteorological elements on the power generation of photovoltaic power stations, forming an adjacency matrix of the influence weight graph , 。 3. The method according to claim 1, characterized in that In step 3, the photovoltaic power generation multi-head graph attention mechanism model includes: input layer, temporal convolution layer, two fully connected layers, and two multi-head graph attention layers; The inputs of the two multi-head graph attention layers are matrices and The calculation formula of the multi-head graph attention in the multi-head graph attention layer is shown in formulas (2)-(6), to calculate the node Taking the weighted sum of its attention coefficient as an example, the adjacent nodes satisfy , (2) (3) (4) (5) (6) in Indicates that the node is The feature vector of the layer, It means that after convolution, the node is The feature vector of the layer, , , , represents the learnable parameters, Indicates concatenation of vectors. For the number of heads, The function is used to convert the scores into probabilities, ensuring that all attention scores add up to 1. is the scaling factor used when computing the dot product to maintain numerical stability, Indicates the multi-head attention calculation process for the vector, Indicates the attention mechanism calculation performed on the vector, Refers to The attention calculation results of each head.

4. A computer device comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, characterized in that: When the processor executes the program, the steps in the method according to any one of claims 1 to 3 are implemented.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps in the method according to any one of claims 1 to 3 are implemented.

Images