A method for constructing a time series prediction model taking into account temporal and spatial proximity

Abstract

The present invention discloses a method for constructing a time series prediction model that takes into account both spatiotemporal proximity. A spatiotemporal weight matrix is ​​established by measuring the spatiotemporal proximity between geographical units, and combined with a graph attention neural network, a deep learning technology that focuses on graph structure, the characteristics of meteorological elements and solar radiation in past time series of a target geographical unit and its neighboring units are effectively captured. This forms a new method for predicting solar radiation that is different from only considering the characteristics of the target geographical unit itself in a time series in the past, thereby better understanding the laws of solar radiation and providing a technical approach to achieving more accurate solar radiation prediction and prediction of extreme low light events.

Description

Technical Field

[0001] The invention belongs to the field of computers, and in particular relates to a method for constructing a time series prediction model taking into account both temporal and spatial proximity. Background Art

[0002] Solar energy is a renewable clean energy that does not directly emit carbon dioxide or air pollutants during operation. It has great potential to improve climate resilience, energy security and social justice. Photovoltaic power generation is a technology that directly converts solar energy into electrical energy by using the photovoltaic effect at the interface of semiconductors. The efficient use of photovoltaics is an important driving force for the production and development of new energy.

[0003] Due to the influence of factors such as precipitation and cloud cover, solar energy resources are intermittent, volatile and random, resulting in high uncertainty in power generation in different regions and at different times. Among the many factors that interfere with photovoltaic power generation, the unknown risks brought about by random extreme weather events are one of the important reasons for the uncertainty in the current use of solar energy resources. At present, China's mainstream photovoltaic facilities are under the condition that the solar radiation is less than 100W / m 2 When the solar shortwave radiation incident on the ground is less than 100W / m 2 The situation is regarded as low light, and an extreme low light event is defined. An extreme low light event refers to an event in which the solar radiation deviates seriously from its average state and has a great impact on photovoltaic power generation due to long-term, large-scale extreme weather such as continuous rain, sandstorms, haze, hail, etc., which is a low-probability event that is not easy to occur. An extreme low light event is defined according to the following judgment conditions: (1) When the number of consecutive extreme low light days is equal to 3 days, it is regarded as the beginning of an extreme low light event; (2) When there are no extreme low light days for 2 consecutive days, it is regarded as the end of an extreme low light event; (3) Non-continuous normal radiation days are allowed during an extreme low light event.

[0004] The prediction of extreme low light events depends on the prediction of solar radiation. Various mathematical methods have been used to predict solar radiation, such as numerical weather prediction (NWP) models. NWP models apply to fields of atmospheric variables. To predict the future state of the field, the model is initialized with observational data and a set of mathematical equations of atmospheric physics and dynamics are solved. Another method is model output statistics (MOS). MOS is a statistical method applied to the output of NWP models to predict other variables or improve the forecast.

[0005] Recently, a considerable number of studies have shown that the accuracy of forecasts can be improved by using artificial intelligence models. The literature carefully reviews the methods of solar radiation forecasting using artificial neural network techniques. Other methods, such as least squares support vector machine (LS-SVM)) , fuzzy and wavelet-based models have also been used as solar irradiation prediction tools.

[0006] From another perspective, the solar radiation series can be considered as a time series. This means that we can use time series analysis methods to study the underlying dynamics and employ mathematical methods to predict the next value. In the literature, there are some attempts to characterize the frequency content of solar irradiance time series. One existing literature uses data mining algorithms to develop four data-driven models for solar radiation forecasting. In another existing literature, researchers tried to find the most relevant input parameters for solar radiation forecasting. In another existing literature, time series models have been used for large-scale global solar radiation forecasts based on geostationary satellite data.

[0007] Time series forecasting is a regression forecasting method and belongs to quantitative forecasting. Its basic principle is: on the one hand, it recognizes the continuity of the development of things, uses past time series data for statistical analysis, and infers the development trend of things; on the other hand, it fully considers the randomness caused by accidental factors. In order to eliminate the impact of random fluctuations, it uses historical data for statistical analysis, and appropriately processes the data to make trend forecasts.

[0008] Some studies believe that the solar radiation series can be regarded as a time series. This means that we can use time series analysis methods to study the underlying dynamics and use mathematical methods to predict the next value. Therefore, the application of time series prediction methods in solar radiation prediction is one of the current research hotspots.

[0009] Some scholars argue that due to the spatial dependence of geographical space, when predicting solar radiation for a geographical unit, not only its own meteorological elements should be considered, but also the meteorological elements of neighboring geographical units.

[0010] The existing technology only considers the characteristics of the target geographical unit itself in the past time series, and does not consider the impact of the meteorological elements and solar radiation of the target geographical unit and its neighboring units in the past time series on the future solar radiation of the target geographical unit.

[0011] The first law of geography states that "Everything is related to everything else, but nearthings are more related than distant things". This law is interpreted in the task of predicting solar radiation as follows: when predicting solar radiation for a geographic unit, it is necessary not only to predict based on its own meteorological elements and historical solar radiation, but also to consider the meteorological elements and historical solar radiation of adjacent geographic units. The relationship between solar radiations of adjacent geographic units stems from the spatial autocorrelation of meteorological elements that affect solar radiation. However, the existing technology only considers the relationship between the meteorological elements of the geographic unit itself and solar radiation to construct the prediction model, while ignoring the important influence of the meteorological elements of the adjacent geographic units on the predicted solar radiation of the target geographic unit, resulting in the accuracy of the prediction model being restricted. Summary of the invention

[0012] The purpose of the present invention is to overcome the shortcomings of the prior art. The present invention measures the spatiotemporal proximity between geographic units to establish a spatiotemporal weight matrix, and combines the graph attention neural network (GAT), a deep learning technology that focuses on graph structure, to effectively capture the characteristics of meteorological elements and solar radiation of the target geographic unit and its neighboring units in the past time series, forming a new method for solar radiation prediction that is different from only considering the characteristics of the target geographic unit itself in the past time series, thereby better understanding the law of solar radiation and providing a technical approach to achieve more accurate solar radiation prediction and prediction of extreme low light events.

[0013] The present invention aims to propose a spatiotemporal weight matrix for modeling temporal and spatial influencing factors at the same time. The matrix will help the prediction model capture the characteristics of meteorological elements and solar radiation of the target geographical unit and its neighboring units in the past time series, thereby achieving more accurate solar radiation prediction and prediction of extreme low light events. Study the method of measuring the spatiotemporal proximity between geographical units. The present invention establishes a spatiotemporal weight matrix based on spatiotemporal proximity to model the impact of meteorological elements and solar radiation of the target geographical unit and its neighboring units in the past time series on the future solar radiation of the target geographical unit. Study how to use the spatiotemporal weight matrix, with the help of graph attention neural network technology, to establish a solar radiation prediction method that can take into account the influence of neighboring geographical units in time and space.

[0014] On this basis, the present invention proposes a method for constructing a time series prediction model taking into account the temporal and spatial proximity, which specifically includes:

[0015] (1) Spatiotemporal data feature representation and graph construction method: It is necessary to adopt a feature representation method suitable for geospatial data. This method can extract air pressure (Pa), mean sea level pressure (Pa), ambient temperature ( )、Dew point temperature( ), total precipitation (M), 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 and high cloud cover, so as to comprehensively and quantitatively describe the meteorological factors affecting solar radiation. The present invention divides the prediction target area into A square grid with a resolution of (m 2 ). The present invention uses a graph structure in which each unit in the geographic space is represented by a node and the spatial relationship between the geographic units is represented by an edge. The purpose of graph construction is to represent the spatial connection between adjacent geographic units and effectively capture the characteristics of adjacent geographic units for subsequent analysis and modeling.

[0016] (2) Method for measuring spatiotemporal proximity and constructing a spatiotemporal weight matrix: The present invention proposes a method for measuring the spatiotemporal proximity between geographical units, and establishes a spatiotemporal weight matrix based on the spatiotemporal proximity to model the impact of the meteorological elements and solar radiation of the target geographical unit and its neighboring units in the past time series on the future solar radiation of the target geographical unit.

[0017] (3) Extreme low light event time series prediction model considering temporal and spatial proximity: Graph Attention Neural Network (GAT) is a deep learning technology that focuses on graph structure. It can aggregate information of neighboring geographic units and help simulate complex associations in geographic space. Combining feature representation methods and Graph Attention Neural Network (GAT), a comprehensive model can be constructed to effectively capture the characteristics of meteorological elements and solar radiation of the target geographic unit and its neighboring units in the past time series, and realize end-to-end solar radiation prediction and prediction of extreme low light events.

[0018] (4) Sample dataset construction: It is necessary to construct a dataset containing historical solar radiation (W / m²), air pressure (Pa), average sea level pressure (Pa), ambient temperature ( )、Dew point temperature( ), total precipitation (M), zonal wind speed at 10 meters (m / s), meridional wind speed at 10 meters (m / s), total cloud cover, low cloud cover, medium cloud cover and high cloud cover, etc., for model training, verification and testing. These datasets are key resources in the research and development process.

[0019] More specifically, the present invention proposes a method for constructing a time series prediction model taking into account both temporal and spatial proximity, which specifically comprises the following steps:

[0020] Step 1: Spatiotemporal data feature representation and graph construction: Divide the prediction target area into geographic units, represent the geographic units with nodes, and represent the spatial relationship between the geographic units with edges to form a graph structure;

[0021] Step 2: Spatiotemporal proximity measurement and spatiotemporal weight matrix construction: Based on spatiotemporal proximity, a spatiotemporal weight matrix is ​​established to model the impact of meteorological elements and solar radiation of the target geographical unit and its neighboring geographical units in the past time series on the future solar radiation of the target geographical unit;

[0022] Step 3: A time series prediction model for extreme low light events that takes into account both temporal and spatial proximity is used to effectively capture the characteristics of meteorological elements and solar radiation in the target geographic unit and its neighboring geographic units in the past time series, and to achieve end-to-end solar radiation prediction and prediction of extreme low light events.

[0023] Step 4: Sample data set construction and model optimization: Construct a sample data set of multi-source spatiotemporal data for model training, verification and testing.

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

[0025] The present invention proposes a method for measuring the spatiotemporal proximity between geographical units, and establishes a spatiotemporal weight matrix based on the spatiotemporal proximity to model the impact of meteorological elements and solar radiation of a target geographical unit and its neighboring units in the past time series on the future solar radiation of the target geographical unit.

[0026] By introducing the spatiotemporal weight matrix and using the graph attention neural network (GAT), a deep learning technology that focuses on graph structure, the spatiotemporal correlation between the target geographic unit and its spatially neighboring units is fully utilized to more effectively capture the characteristics of meteorological elements and solar radiation of the target geographic unit and its neighboring units in the past time series, thereby improving the accuracy of solar radiation prediction.

[0027] This paper proposes a method for constructing a time series prediction model that takes into account the temporal and spatial proximity - taking the prediction of photovoltaic extreme low light events as an example. By measuring the temporal and spatial proximity between geographical units to establish a spatiotemporal weight matrix, and combining the graph attention neural network (GAT), a deep learning technology that focuses on graph structure, it effectively captures the characteristics of meteorological elements and solar radiation in the past time series of the target geographical unit and its neighboring units, forming a solar radiation prediction method that is different from only considering the characteristics of the target geographical unit itself in the past time series. [9] A new method is developed to better understand the laws of solar radiation, providing a technical approach for more accurate solar radiation forecasts and prediction of extreme low light events.

[0028] The present invention innovatively proposes a method for measuring the spatiotemporal proximity between geographical units in a solar radiation prediction scenario and establishing a spatiotemporal weight matrix. The spatiotemporal weight matrix is ​​used to model the impact of meteorological elements and solar radiation of a target geographical unit and its neighboring units in the past time series on the future solar radiation of the target geographical unit. However, the prior art only considers the characteristics of the target geographical unit itself in a past time series, and does not fully consider the impact of the characteristics of meteorological elements and solar radiation of the target geographical unit and its neighboring units in the past time series on the future solar radiation of the target geographical unit.

[0029] Different from the deep learning model used in the prior art, the present invention introduces a spatiotemporal weight matrix and uses the graph attention neural network (GAT), a deep learning technology that focuses on graph structure, to fully utilize the spatiotemporal correlation between the target geographic unit itself and its spatially neighboring units, so as to more effectively capture the characteristics of meteorological elements and solar radiation of the target geographic unit and its neighboring units in the past time series, thereby improving the accuracy of solar radiation prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Predicting target areas for solar radiation Schematic diagram;

[0031] Figure 2 It is a characteristic diagram of multi-dimensional meteorological elements;

[0032] Figure 3 This is a schematic diagram of the method for obtaining values ​​of multidimensional meteorological elements in geographical units;

[0033] Figure 4 To divide the geographical unit The dimensional feature map is a node Dimensional feature diagram;

[0034] Figure 5 Schematic diagram of spatiotemporal proximity and spatiotemporal weight matrix;

[0035] Figure 6 is the spatial graph adjacency matrix Schematic diagram of the construction method;

[0036] Figure 7 It is a schematic diagram of the influence of the target unit itself and its spatial neighboring units on the target unit;

[0037] Figure 8 This is a schematic diagram of the structure of the extreme low-light event time series prediction model;

[0038] Fig. 9 This is the input diagram of GATBlock1;

[0039] Fig.10This is the input diagram of GATBlock2. DETAILED DESCRIPTION

[0040] 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.

[0041] The present invention proposes a method for constructing a time series prediction model taking into account both temporal and spatial proximity, which specifically includes:

[0042] Step 1: spatiotemporal data feature representation and graph construction;

[0043] The solar radiation is predicted for the target area Divide in space into square geographic units, the resolution of the geographic unit is (m 2 ).like Figure 1 shown.

[0044] Each geographic unit contains air pressure (Pa), mean sea level pressure (Pa), ambient temperature ( )、Dew point temperature( ), total precipitation (M), 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 and high cloud cover meteorological elements Dimension. Figure 2 shown.

[0045] The geographical unit determines the value of the meteorological element according to the size of the area that intersects with the spatial range of the meteorological element.

[0046] By plot composition; set The areas are , the temperature values ​​are ,like Figure 3 shown.

[0047] Geographical unit The area is , the temperature value is , the calculation method is shown in formula (1):

[0048] (1)

[0049] Map geographic units to nodes. The dimensional feature map is a node dimensional features. Figure 4 shown.

[0050] Step 2, spatiotemporal proximity measurement and spatiotemporal weight matrix construction;

[0051] Spatiotemporal proximity is the spatiotemporal proximity between any two homogeneous areas (including points) in geographic space. , for a given "flow", is proportional to the total flow between the two , which is inversely proportional to the average time it takes to get from one end to the other , as shown in formula (2):

[0052] (2)

[0053] The present invention assumes that the total flow of meteorological elements affecting solar radiation is the same between geographical units, focusing on the average propagation time of meteorological elements between geographical units. The present invention measures the average time from three aspects: Euclidean spatial distance, geographical unit pattern and time lag degree. , to represent the spatiotemporal proximity between two geographic units. represents the spatiotemporal proximity between all geographic units, such as Figure 5 shown.

[0054] Geographical unit exist The solar radiation at a given moment is affected by itself and the neighboring units in space. , and exist In with The influence of meteorological elements and solar radiation at all times.

[0055] (1) Euclidean space distance: the present invention sets a distance threshold To measure whether there is spatial influence between geographic units. If the Euclidean distance between two geographic units Greater than the set threshold , when predicting the wind speed trend of one geographical unit, the spatial influence between the other geographical unit and it is not considered.

[0056] (2) Geographical unit pattern: In Euclidean space Less than the set threshold Establish edges between geographic unit nodes and set the weights of the edges as learnable parameters , forming a spatial graph , which aims to capture the effect of geographic unit patterns on average time The impact of Figure 6 shown.

[0057] (3) Time lag: Setting parameters express Time factors (situations) The degree of influence of momentary factors (situations).

[0058] In summary, the target geographic unit itself ( ) and its spatial neighboring units ( , and ) in the past sequence ( , ) has been presented in the meteorological elements and solar radiation prediction target geographic unit in The impact of Figure 7 And as shown in formula (3). Indicates that the geographical unit is The feature vector at time, and The product of is the spatiotemporal weight matrix .

[0059] (3)

[0060] Step 3: A time series prediction model for extreme low light events taking into account the temporal and spatial proximity;

[0061] based on Geographical units in Meteorological elements and solar radiation, forecast Geographical units in of solar radiation.

[0062] The structure of the extreme low-light event time series prediction model taking into account the temporal and spatial proximity is as follows: Figure 8 shown. for , is the feature dimension, is the length of the time series ( :Considering the current time and the length of the previous sequence The impact of time series characteristics on current moment characteristics; : Based on the length The time series prediction length is time series), is the number of nodes.

[0063] The input layer is the first layer of the neural network and is responsible for receiving input data. The temporal convolution layer is used to extract features from time series data. It performs a convolution operation on the input data with a local receptive field through a sliding window to capture temporal correlations. The graph attention layer (GATBlock1 and GATBlock2) is an attention mechanism based on graph convolution, which can capture the complex relationships between nodes and assign different importance weights to different neighbors through the attention mechanism.

[0064] GATBlock1 and GATBlock2 are two consecutive graph attention blocks, each containing graph convolution operations and attention mechanisms. Input spatiotemporal weight matrix Depend on (spatial graph adjacency matrix) and (indicating the degree of time lag). The graph attention block is based on parameterized and The weights between nodes are calculated and the features of neighboring nodes are weighted summed to aggregate local neighborhood information.

[0065] The inputs to GATBlock1 are Geographical units in The meteorological elements and spatiotemporal weight matrix are as follows: Fig. 9 And as shown in formula (4). Represents a convolution operation with an attention mechanism.

[0066] (4)

[0067] The calculation method of graph attention in GATBlock1 is shown in formulas (5)-(7), where Its neighboring nodes Take the attention coefficient and weighted sum as an example, the node satisfy ;

[0068] (5)

[0069] (6)

[0070] (7)

[0071] in, express Time Node In the The feature vector of the layer, Indicates that the node is at the first The feature vector of the layer, represents the activation function, Indicates The parameters of the layer convolution come from (spatial graph adjacency matrix), Indicates the use of shared parameters right Perform a linear transformation, Represents the node and ( ) The transformed results are spliced, It means mapping the concatenated high-dimensional features to a real number. is the activation function, express The degree of influence of the time on the predicted target time.

[0072] The input to GATBlock2 is the research goal Geographical units in Solar radiation, research objectives Geographical units in The historical solar radiation and spatiotemporal weight matrix is ​​as follows: Fig.10 As shown in formula (8), Represents the intermediate result feature vector of solar radiation prediction output by the fully connected layer, represents the historical solar radiation eigenvector, Express and To splice, Represents the concatenated high-dimensional features. The calculation method of graph attention in GATBlock2 is shown in formulas (3)-(5).

[0073] (8)

[0074] The present invention uses a fully connected layer at the end of a neural network to map the learned high-level features to the final output.

[0075] Step 4: Sample data set construction and model optimization;

[0076] Collection research objectives Geographical units in (Length is The meteorological elements and historical solar radiation of the time series constitute the data set. The model input is: Geographical units in (Length is time series) of meteorological elements and solar radiation; spatiotemporal weight matrix ; Model labels are: Geographical units in of solar radiation.

[0077] 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.

[0078] When applying the model, collect data and build Geographical units in (Length is The feature matrix and spatiotemporal weight matrix of the time series , input the trained model for forward propagation, and you can get Geographical units in Solar radiation prediction results.

Claims

1. A method for constructing a time series prediction model taking into account both temporal and spatial proximity, characterized in that: The specific steps include: Step 1: Spatiotemporal data feature representation and graph construction: Divide the prediction target area into geographic units, represent the geographic units with nodes, and represent the spatial relationship between the geographic units with edges to form a graph structure; Step 2: Spatiotemporal proximity measurement and spatiotemporal weight matrix construction: Based on spatiotemporal proximity, a spatiotemporal weight matrix is ​​established to model the impact of meteorological elements and solar radiation of the target geographical unit and its neighboring geographical units in the past time series on the future solar radiation of the target geographical unit; Step 3: A time series prediction model for extreme low light events that takes into account both temporal and spatial proximity is used to effectively capture the characteristics of meteorological elements and solar radiation in the target geographic unit and its neighboring geographic units in the past time series, and to achieve end-to-end solar radiation prediction and prediction of extreme low light events. Step 4: Sample data set construction and model optimization: Construct a sample data set of multi-source spatiotemporal data for model training, verification and testing; In step 2, the spatiotemporal proximity is the spatiotemporal proximity between any two homogeneous regions in geographic space , which is proportional to the total flow between the two , which is inversely proportional to the average time it takes to get from one end to the other , as shown in formula (2): (2) Using meteorological elements to propagate the average time between geographic units The difference between the two measures the temporal proximity; the average time is measured from three aspects: Euclidean spatial distance, geographical unit pattern and time lag. , to represent the spatiotemporal proximity between two geographic units, and the spatiotemporal weight matrix represents the spatiotemporal proximity between all geographic units; Euclidean distance: Setting distance threshold To measure whether there is spatial influence between geographical units, if the Euclidean distance between two geographical units Greater than the set threshold , then when predicting the wind speed trend of one geographical unit, the spatial influence between the other geographical unit and it is not considered; Geographical unit pattern: distance in Euclidean space Less than the set threshold Establish edges between geographic unit nodes and set the edge weights as learnable parameters , forming a spatial graph ; Time lag: Setting parameters express The time factor The degree of influence of time factors; Prediction target geographic unit Its spatial neighboring units , and In the past sequence ( , ) has been presented in the meteorological elements and solar radiation prediction target geographic unit in The influence of is shown in formula (3), where Indicates that the geographical unit is The feature vector at time, and The product of is the spatiotemporal weight matrix ; (3) In step 3, based on Geographical units in Meteorological elements and solar radiation, forecast Geographical units in of solar radiation, of which :Considering the current time and the length of the previous sequence The impact of time series characteristics on current moment characteristics; : Based on the length The time series prediction length is time series; The parameters of the extreme low-light event time series prediction model taking into account the temporal and spatial proximity include for , is the feature dimension, is the length of the time series; The input layer is the first layer of the neural network, responsible for receiving input data. The temporal convolution layer is used to extract features from time series data. It performs convolution operations on the local receptive field of the input data through a sliding window to capture temporal correlations. The graph attention block GATBlock1 and the graph attention block GATBlock2 are an attention mechanism based on graph convolution, which can capture the complex relationship between nodes and assign different importance weights to different neighbors through the attention mechanism. GATBlock1 and GATBlock2 are two consecutive graph attention blocks, each of which contains graph convolution operations and attention mechanisms. The input spatiotemporal weight matrix Depend on and Composition, the graph attention block is based on parameterized and The weights between the calculated nodes are weighted summed up on the features of neighboring nodes to aggregate local neighborhood information; The inputs to GATBlock1 are Geographical units in The meteorological elements and spatiotemporal weight matrix are as shown in formula (4). represents a convolution operation with an attention mechanism, and the output is : (4) The calculation method of graph attention in GATBlock1 is shown in formulas (5)-(7), where Its neighboring nodes Take the attention coefficient and weighted sum as an example, the node satisfy ; ( ... ) (5) (6) (7) in, express Time Node In the The feature vector of the layer, Indicates that the node is at the first The feature vector of the layer, represents the activation function, Indicates Parameters of layer convolution, from the spatial graph adjacency matrix , Indicates the use of shared parameters right Perform a linear transformation, Represents the node and , , the transformed results are spliced, It means mapping the concatenated high-dimensional features to a real number. is the activation function, express The degree of influence of the moment on the predicted target moment, express Time Node In the The feature vector of the layer; The input to GATBlock2 is the research goal Geographical units in Solar radiation, research objectives Geographical units in The historical solar radiation and spatiotemporal weight matrix is ​​as shown in formula (8): Represents the intermediate result feature vector of solar radiation prediction output by the fully connected layer, represents the historical solar radiation eigenvector, Express and To splice, Represents the concatenated high-dimensional features. The calculation method of graph attention in GATBlock2 is shown in formulas (3)-(5); (8) A fully connected layer is used at the end of the neural network to map the learned high-level features to the final output.

2. The method according to claim 1, characterized in that In step 1, the solar radiation prediction target area Divide in space into square geographic units, the resolution of the geographic unit is m 2 Each geographic unit contains the characteristics of atmospheric pressure, mean sea level pressure, ambient temperature, dew point temperature, total precipitation, 10-meter latitudinal wind speed, 10-meter longitudinal wind speed, total cloud cover, low cloud cover, medium cloud cover and high cloud cover. dimension; geographic unit By plot composition; set The areas are , the temperature values ​​are , geographic unit The area is , the temperature value is , The calculation method is shown in formula (1): (1) Map geographic units to nodes and map the The dimensional feature map is a node Dimensional features.

3. 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 2 are implemented.

4. 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 2 are implemented.

Images