Meteorological resource interpolation method based on adaptive spatial neural network

Abstract

The application discloses a meteorological resource interpolation method based on an adaptive spatial neural network and belongs to the technical field of spatial interpolation. The method takes a gating recurrent unit as a core and constructs an adaptive spatial neural network based on the gating recurrent unit. The method solves the limitation of multi-dimensional spatial distance calculation by defining a generalized spatial distance, fits the nonlinear relationship between the generalized spatial distance and a spatial weight coefficient by using the adaptive spatial neural network, trains the adaptive spatial neural network in combination with an interactive verification method, and obtains meteorological data of a distributed photovoltaic power station by using the trained adaptive spatial neural network based on the spatial coordinates and the meteorological feature sequence of a plurality of integrated power stations in a wide range corresponding to the distributed photovoltaic power station to be generated. The application provides global meteorological data support for distributed photovoltaic power stations and the like.

Description

Technical Field

[0001] This invention belongs to the field of meteorological data processing and spatial interpolation technology, specifically relating to a meteorological resource interpolation method based on an adaptive spatial neural network. Background Technology

[0002] Spatial interpolation technology infers data from unknown regions using known discrete data samples. It is widely used in fields such as geography and meteorology, primarily for data complementation, contour line construction, and data gridding. Traditional spatial interpolation algorithms are mainly divided into deterministic interpolation methods and geographic interpolation methods. Deterministic interpolation methods include inverse distance weighted (IDW) interpolation and local polynomial interpolation, while geographic interpolation is represented by Kriging interpolation.

[0003] IDW interpolation is based on the first law of geography and assigns weights by calculating the Euclidean distance between unknown points and reference points. However, it is difficult to handle nonlinear relationships in the data, and the weight assignment is greatly affected by power parameters. Kriging interpolation introduces a variogram function to describe spatial dependence and is the best linear unbiased extrapolation method. However, it has a large prediction error in extreme regions and exhibits obvious particle delimitation and sawtooth phenomena.

[0004] Distributed photovoltaic (PV) power plants generally lack specialized meteorological monitoring equipment and rely on meteorological data from surrounding integrated power plants for power generation forecasting. Traditional interpolation methods fail to fully capture the spatial nonlinear characteristics of meteorological data, resulting in insufficient accuracy and significant errors in PV power generation forecasting, impacting power plant operational efficiency and grid dispatch efficiency. Therefore, a high-precision, adaptive spatial interpolation method is urgently needed to improve the accuracy of global extrapolation of meteorological resource data. Summary of the Invention

[0005] To address the problems in the existing technology, this invention provides a meteorological resource interpolation method based on an adaptive spatial neural network.

[0006] The technical solution of the present invention is as follows: This invention discloses a meteorological resource interpolation method based on an adaptive spatial neural network for generating meteorological data for distributed photovoltaic power stations; the method includes the following steps: 1) Identify the distributed photovoltaic power stations for which meteorological data is to be generated, obtain the spatial coordinates and historical meteorological feature sequences of multiple integrated power stations within a wide area corresponding to the distributed photovoltaic power stations, and then normalize all historical meteorological feature sequences and spatial coordinates; and based on the normalized spatial coordinates, obtain the generalized spatial distance between each pair of integrated power stations through the standard orthogonal basis spatial coordinate system. 2) Construct an adaptive spatial recurrent neural network based on gated recurrent units. Based on the generalized spatial distance and normalized historical meteorological feature sequences from step 1), train the adaptive spatial recurrent neural network using cross-validation. 3) Obtain the spatial coordinates of the distributed photovoltaic power station and perform normalization processing, then obtain the generalized spatial distance between the distributed photovoltaic power station and each integrated power station; obtain the meteorological feature sequence of each integrated power station and perform normalization processing, and obtain the spatial weight coefficient between the distributed photovoltaic power station and each integrated power station through a trained adaptive spatial recurrent neural network based on the generalized spatial distance of step 3) and the normalized meteorological feature sequence; obtain the meteorological data of the distributed photovoltaic power station based on the spatial weight coefficient and the meteorological feature sequence of each integrated power station.

[0007] Further, in step 2), training the adaptive spatial recurrent neural network includes: 21) Divide the normalized historical meteorological feature sequences of all integrated power stations into N equal parts, select N-1 parts as the training set of the adaptive spatial neural network, and use the rest as the validation set. 22) Input the normalized historical meteorological feature sequences and corresponding generalized spatial distances of the integrated power stations in the training set into the adaptive spatial neural network, and train the adaptive spatial neural network by gradient descent. 23) Input the normalized historical meteorological feature sequence and corresponding generalized spatial distance of the integrated power stations in the verification set into the adaptive spatial neural network obtained in step 22), and calculate multiple RMSE values. If the RMSE value continues to decrease, continue to step 25); otherwise, continue to step 24). 24) Determine whether the training rounds of the adaptive spatial neural network have reached the preset number of rounds. If they have, continue to step 25); otherwise, continue to step 22). At the same time, change the order in which the normalized historical meteorological feature sequences and the corresponding generalized spatial distances of the integrated power stations in the training set are input into the adaptive spatial neural network. 25) Select N-1 sets as the training set for the adaptive spatial neural network, and the rest as the validation set, and execute steps 22)-25) until no new training set can be selected; obtain the trained adaptive spatial neural network based on the adaptive spatial neural network trained on each training set.

[0008] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention addresses the problem of the lack of professional meteorological resource monitoring devices for distributed photovoltaic power stations by proposing an adaptive spatial recurrent neural network based on GRU for interpolating meteorological resource data of distributed photovoltaic power stations. By integrating the meteorological data from point-to-point monitoring of the integrated power station, the network interpolation method is used to obtain meteorological data for the locations of distributed photovoltaic power stations throughout the entire region. The adaptive spatial recurrent neural network of this invention achieves the effect of extrapolating from data monitoring points to surface data across the entire region.

[0009] (2) In the wide-area meteorological data interpolation experiment, the adaptive spatial recurrent neural network can better solve the problems of severe particle boundary and jagged phenomenon in the traditional geographic interpolation model. The interpolation transition surface of the adaptive spatial recurrent neural network is relatively smooth.

[0010] (3) The adaptive spatial recurrent neural network utilizes meteorological data from a wide-area integrated power station and analyzes the sequence relationships and error data characteristics between data samples to achieve rapid interpolation to the target global area. The adaptive spatial recurrent neural network effectively solves the problem of insufficient data volume in distributed photovoltaic power stations and also significantly improves prediction accuracy. Attached Figure Description

[0011] Figure 1 This is a flowchart of the meteorological resource interpolation method based on adaptive spatial neural networks of the present invention; Figure 2 This is a structural diagram of a gated loop unit; Figure 3 This is a flowchart of the training process for the adaptive spatial neural network of the present invention; Figure 4 This is a diagram illustrating the calculation process of the spatial weighting coefficients in this invention; Figure 5 This is a comparison diagram of the interpolation effects of IDW interpolation, Kriging interpolation, and adaptive spatial neural network interpolation in the embodiments of the present invention; Figure 6 This is a comparison chart of the Kriging interpolation results and the actual values ​​in an embodiment of the present invention; Figure 7 This is a comparison chart of the IDW interpolation results and the actual values ​​in an embodiment of the present invention; Figure 8 This is a comparison diagram between the interpolation results of the adaptive spatial neural network in this embodiment of the invention and the actual values. Detailed Implementation

[0012] The present invention will be further described and illustrated below with reference to specific embodiments. The embodiments described are merely examples of the content of this disclosure and do not limit the scope of the invention. The technical features of each embodiment in the present invention can be combined accordingly, provided that there is no mutual conflict.

[0013] Spatial interpolation algorithms construct spatial correlation models by utilizing the distances between known and unknown points within a region and spatial weighting factors. Spatial interpolation typically relies on mathematical formulas and basic computational methods to generate spatial grid data, but this approach may not fully capture the nonlinear variations and complex relationships within the spatial data. High-precision calculations when considering spatial distance weights also have limitations. Therefore, this invention proposes using a neural network algorithm to optimize the spatial weight calculation process during interpolation. Through the autonomous learning and training of the neural network, the fitting degree between spatial distance and weights is enhanced.

[0014] The core objective of this invention is to address the problem of missing meteorological data caused by the lack of professional meteorological monitoring devices in distributed photovoltaic power plants. By constructing an adaptive spatial interpolation model (G-SARNN) based on gated cyclic unit (GRU) networks, and utilizing the meteorological feature sequences of integrated power plants within a wide area corresponding to the distributed photovoltaic power plants to be generated, high-precision interpolation of the meteorological feature sequences of distributed photovoltaic power plants can be achieved, providing reliable data support for the prediction of power generation of distributed photovoltaic power plants.

[0015] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0016] like Figure 1 The diagram shows a flowchart of the meteorological resource interpolation method based on an adaptive spatial neural network according to the present invention. The method of the present invention includes the following steps: Step 1: Identify the distributed photovoltaic power stations from which meteorological data is to be generated. Obtain the spatial coordinates and historical meteorological feature sequences of multiple integrated power stations within a wide area corresponding to the distributed photovoltaic power stations. Then, normalize all historical meteorological feature sequences and spatial coordinates to unify the measurement scale of meteorological data in all spatial coordinates and historical meteorological feature sequences. Based on the normalized spatial coordinates, obtain the generalized spatial distance between every two integrated power stations through a standard orthogonal basis spatial coordinate system.

[0017] Specifically, the historical meteorological characteristic sequence includes the total radiation, direct radiation, diffuse radiation, ambient temperature, air pressure, and relative humidity of the integrated power station location within a historical time period. The meteorological data are total radiation, direct radiation, diffuse radiation, ambient temperature, air pressure, or relative humidity.

[0018] Normalizing all historical meteorological feature sequences is crucial for scaling. Different meteorological elements have vastly different dimensions and numerical ranges. Without normalization, features with large values ​​(such as irradiance) would dominate model training, hindering the model's learning of smaller but important features (such as humidity). Furthermore, during optimization (e.g., gradient descent), models using neural networks find the optimal solution more quickly and smoothly when all meteorological elements are on the same scale, improving training efficiency and stability. Normalization can be achieved using min-max normalization and Z-score standardization.

[0019] Normalizing spatial coordinates is necessary because the original spatial coordinates (such as latitude and longitude, distance, etc.) may have different dimensions or value ranges. If they are directly used in distance calculations or neural network training, certain dimensions may dominate the model's judgment. After normalization, all spatial coordinates will be on a uniform scale, which helps to avoid numerical explosion or gradient vanishing problems in multi-layer neural networks.

[0020] The key to spatial interpolation algorithms lies in the calculation of spatial weight coefficients. The accuracy of these coefficients directly affects the quality of the interpolation results, and their magnitude is closely related to the spatial layout. Spatial interpolation essentially reveals the correlation between the distances between known sampling points and the unknown regions to be inferred. Two-dimensional spatial distance calculation, based on Euclidean distance, is a typical method in classical interpolation algorithms. However, classical interpolation algorithms are only applicable to calculations in different directions within two-dimensional space. This is because in three-dimensional and multi-dimensional spaces, directions and spatial elements exhibit variability and fluctuating trends, leading to errors in distance calculations and interpolation results compared to actual values. Therefore, a new expression can be derived for any vector in multi-dimensional space: first, the normalized spatial coordinates of the integrated power stations are converted to coordinates in a standard orthogonal basis spatial coordinate system; then, the generalized spatial distance between any two integrated power stations is obtained based on their coordinates. The formula for calculating the generalized spatial distance is as follows: in, This represents the generalized spatial distance between two integrated power stations; A metric function representing the difference between vectors in a multidimensional space; , Let be the coordinates of an integrated power station in a standard orthogonal basis space coordinate system. The coordinates of the integrated power station in the standard orthogonal basis space coordinate system The nth dimension sub-coordinate; , Let the coordinates of another integrated power station be in the standard orthogonal basis space coordinate system. The coordinates of the integrated power station in the standard orthogonal basis space coordinate system The nth dimension sub-coordinate.

[0021] Step 2: Construct an adaptive spatial recurrent neural network based on gated recurrent units. The adaptive spatial recurrent neural network is trained using cross-validation based on the generalized spatial distance and normalized historical meteorological feature sequences obtained in Step 1.

[0022] Specifically, the adaptive spatial recurrent neural network consists of an input layer, a first GRU layer, a second GRU layer, a fully connected layer, and an input layer connected in sequence.

[0023] The Gated Recurrent Unit (GRU) is a model abstracted from the structure of biological neural networks. GRU simplifies LSTM by having only two control gate functions: the update gate and the reset gate. While their working mechanisms are largely similar, GRU has fewer parameters, making it easier to train and improving training efficiency to some extent. The structure of GRU is shown in the diagram below. Figure 2 As shown.

[0024] The output state from the previous time step is passed to the current time step through the update gate. The update gate determines whether to retain the output state from the previous time step. The current input and the output state from the previous time step are written into the update gate as the current candidate state. The combination of candidate states is determined by the reset gate. Decide, The larger the value, the higher the output state of the previous time step. The more content is written, the more the current output status changes. Output state from the previous time step and candidate states The decision was made at the same time.

[0025] Common activation functions include the Sigmoid function, Tanh function, and ReLU function. Among them, the ReLU function is widely used due to its high computational accuracy and good output sparsity. The main function of the activation function is to convert the input information in the GRU into an output signal, i.e., a nonlinear mapping function. : It possesses the characteristics of nonlinearity, monotonicity, continuous differentiability, and unsaturation. To avoid the problem that the derivative of the ReLU function being zero when the input value is less than or equal to zero might cause neuron inhibition, leading to premature model termination, this invention proposes a Parametric Rectified Linear Unit (PRELU) activation function, which is an improvement on the Leaky Rectified Linear Unit (LReLU) activation function. The mathematical form of the PReLU activation function is: in, This represents the parameterized modified linear unit activation function; This represents the input value of the parameterized modified linear unit activation function; Represents the maximum value function; Describes the minimum value function; This represents the hyperparameter, specifically the slope of the parameterized modified linear unit activation function when the input value is less than or equal to 0. It will adaptively adjust and optimize during the training of the adaptive spatial recurrent neural network.

[0026] Using the PReLU activation function to adjust hyperparameters By adaptively adjusting, the network function can continuously learn, train, and improve on the variable input data. By adding a very small number of parameters, the training effect and accuracy of the neural network are continuously improved, which has high practical value.

[0027] Compared to traditional nonlinear activation functions, the PReLU activation function is less sensitive to parameter initialization than the sigmoid function. However, an efficient initialization strategy is still needed during training to effectively eliminate the instability caused by the exponential changes in deep learning network parameters. Assuming independent distribution of network function variables, this paper analyzes many properties of the PReLU activation function and calculates its variance using the following formula. in, Represents the variance of the neural network input (current layer); This represents the input data for the k-th layer; Indicates the total number of layers in the neural network; This represents the input dimension of the PReLU activation function; This represents the variance of the weight parameters in the k-th layer; This represents the weighting parameter. When... When the condition of equal to 1 is met, the convergence failure caused by the exponential scaling of parameters can be avoided, thus improving the accuracy of network training.

[0028] Regarding the training of adaptive spatial neural networks, this invention proposes a novel network training framework based on interactive verification, addressing the problems of gradient vanishing and insufficient accuracy in traditional transfer learning algorithms. This method demonstrates good performance in parameter configuration and network optimization algorithm training. By constructing an adaptive spatial neural network based on gated recurrent units (GRUs), the aim is to improve the accuracy of training results and the precision of interpolation results. GRUs have a strong ability to extract and analyze sequence relationships between data, and their accuracy continuously improves on the training set during network training, but errors increase on the validation set. To address the overfitting problem during computation, an error magnitude assessment is performed at the end of each cycle, using the root mean square error (RMSE) as the error evaluation metric. Abnormal error changes trigger the network's natural protection measures, prematurely stopping the computation process. The overall process is as follows: Figure 3 As shown.

[0029] Interactive validation, as a core method for network training and evaluation, is similar in concept to neural network training strategies. It splits the overall data into a training subset for network learning and a validation subset for performance evaluation. The training set is used for the main training process, while the validation set is used to evaluate the performance of the predictive network. Interactive validation effectively mines the sequence features between data samples while effectively avoiding overfitting. The main computational steps are to divide the data samples into N equal parts, selecting n-1 parts as the network's training set and the remainder as the validation set. The training set is used to train the network to capture spatial feature relationships in the data samples. After network training, the validation set is used to validate the network's performance and evaluate its accuracy and generalization ability on unknown data. By averaging the results of each group in this way, the bias of regional network training is reduced. This invention adopts the interactive validation method, and the main steps for training an adaptive spatial recurrent neural network are as follows: 21) Divide the normalized historical meteorological feature sequences of all integrated power stations into N equal parts, select N-1 parts as the training set of the adaptive spatial neural network, and use the rest as the validation set. 22) Input the normalized historical meteorological feature sequences and corresponding generalized spatial distances of the integrated power stations in the training set into the adaptive spatial neural network, and train the adaptive spatial neural network by gradient descent. 23) Input the normalized historical meteorological feature sequence and corresponding generalized spatial distance of the integrated power stations in the verification set into the adaptive spatial neural network obtained in step 22), and calculate multiple RMSE values. If the RMSE value continues to decrease, continue to step 25); otherwise, continue to step 24). 24) Determine whether the training rounds of the adaptive spatial neural network have reached the preset number of rounds. If they have, continue to step 25); otherwise, continue to step 22). At the same time, change the order in which the normalized historical meteorological feature sequences and the corresponding generalized spatial distances of the integrated power stations in the training set are input into the adaptive spatial neural network. 25) Select N-1 sets as the training set for the adaptive spatial neural network, and the rest as the validation set, and execute steps 22)-25) until no new training set can be selected; obtain the trained adaptive spatial neural network based on the adaptive spatial neural network trained on each training set.

[0030] Specifically, step 22) includes: using the normalized historical meteorological feature sequence of any integrated power station in the training set as a label, and based on the normalized historical meteorological feature sequences of the remaining integrated power stations in the training set and the corresponding generalized spatial distance, obtaining the spatial weight coefficient between the integrated power station and each of the other integrated power stations through an adaptive spatial neural network; then, based on the spatial weight coefficient and the normalized historical meteorological feature sequence of the integrated power station, obtaining the predicted meteorological feature sequence of the integrated power station corresponding to the label; training the adaptive spatial neural network through gradient descent; and updating the parameters in the adaptive spatial neural network.

[0031] In a specific embodiment of the present invention, step 23) includes: 231) Using the normalized historical meteorological feature sequence of any integrated power station in the validation set as the label, and based on the normalized historical meteorological feature sequences of the other integrated power stations in the validation set and the corresponding generalized spatial distance, the spatial weight coefficient between the integrated power station and each of the other integrated power stations is obtained through the adaptive spatial neural network obtained in step 22). Then, based on the spatial weight coefficient and the normalized historical meteorological feature sequence of the integrated power station, the predicted meteorological feature sequence of the integrated power station corresponding to the label is obtained. The RMSE value is calculated based on the predicted meteorological feature sequence of the integrated power station corresponding to the label and the normalized historical meteorological feature sequence. 232) Using the normalized historical meteorological feature sequence of another integrated power station in the validation set as a label, and calculating the RMSE value; repeat step 232) to obtain multiple RMSE values. If the RMSE value continues to decrease, continue to step 25); otherwise, continue to step 24).

[0032] Generalized spatial distance calculation overcomes the limitations of traditional interpolation by using a functional relationship between spatial weighting coefficients and generalized spatial distance to interpolate unknown regions. The nonlinear function between the spatial weighting coefficients and the generalized spatial distance is defined as follows: in, Indicates the first A set of spatial weighting coefficients for integrated power plants; Indicates the first The first integrated power station and the first Spatial weighting coefficients among integrated power stations; This represents the mapping function between generalized spatial distance and spatial weight coefficients; Indicates the first The first integrated power station and the first The generalized spatial distance between integrated power stations.

[0033] To avoid overfitting, this invention sets the spatial weighting coefficient between the integrated power station itself to 0, and only calculates the spatial weighting coefficient between one integrated power station and another. That is: in, Indicates the first The first integrated power station and the first Weight distribution coefficients among integrated power plants This represents the standard weight component, ensuring that the above requirements and related weights are 0, and that the value is 1 in the spatial weighting process between other regions.

[0034] This invention improves upon GRU for spatiotemporal sequence prediction by constructing a GRU-based adaptive spatial neural network, abbreviated as G-SARNN. G-SARNN fully utilizes the nonlinear relationships between data sequences learned by the neural network to fit and optimize the spatial weight coefficients, namely: in, Indicates the first A set of weight distribution coefficients for an integrated power station; Represents the spatially weighted distribution generating function; Indicates the first The generalized spatial distance between an integrated power station and other integrated power stations Indicates the first The first integrated power station and the first Weight distribution coefficients among integrated power plants. Spatial weight coefficients. It can be obtained by multiplying the weight component coefficient and the weight standard component, and the calculation process is shown in the following formula: in, This represents the spatial weighting coefficient matrix of the integrated power station; This represents the weight distribution coefficient matrix of the integrated power plant; This represents the weighted standard component matrix of the integrated power plant.

[0035] In calculating the multidimensional generalized spatial distance and spatial weight coefficients, two different spatial weight component calculation methods are used as network inputs. The network is trained using data, and different data are needed to obtain the topology weights for different distributed photovoltaic power station interpolation processes, thus obtaining accurate weight coefficients. The spatial weight calculation process is as follows: Figure 4 As shown. Figure 4 This paper demonstrates how a neural network structure can transform the spatial location of sample points into weight coefficients for a target region in a multidimensional generalized space. The model input is spatial data samples from an integrated power plant. After spatial distance calculation, hidden layer nonlinear fitting, and spatial weight quantization, the final output is a comprehensive weight for interpolation prediction. This structure effectively improves the adaptability and prediction accuracy of the interpolation process.

[0036] The calculation of spatial weight coefficients is based on the generalized spatial distance calculation method, which differs slightly from the calculation of spatial distances in different directions within the same space using two-dimensional Euclidean spatial distance. Therefore, this invention addresses the problem of spatial direction fragmentation by proposing a generalized spatial distance calculation method. Based on a gated recurrent unit (GRU) neural network, it learns the correlation of data sample space, further fitting the nonlinear coefficient relationship between spatial distance and spatial weights, and thus constructing a G-SARNN. G-SARNN combines the advantages of deep learning and grid spatial interpolation, utilizing GRU to optimize the spatial weight coefficients, thereby better uncovering the correlation between spatial distance and spatial weights and improving the accuracy and precision of spatial interpolation.

[0037] Step 3: Obtain the spatial coordinates of the distributed photovoltaic power station and perform normalization processing, then obtain the generalized spatial distance between the distributed photovoltaic power station and each integrated power station; obtain the meteorological feature sequence of each integrated power station and perform normalization processing; based on the generalized spatial distance of step 3) and the normalized meteorological feature sequence, obtain the spatial weight coefficient between the distributed photovoltaic power station and each integrated power station through a trained adaptive spatial recurrent neural network; obtain the meteorological data of the distributed photovoltaic power station based on the spatial weight coefficient and the meteorological feature sequence of each integrated power station.

[0038] Specifically, meteorological data of distributed photovoltaic power stations are obtained based on spatial weighting coefficients and meteorological characteristic sequences of each integrated power station; this includes multiplying the meteorological characteristic sequences of each integrated power station with the corresponding spatial weighting coefficients, and then adding all the multiplication results to obtain the meteorological characteristic sequences of the distributed photovoltaic power stations, that is, obtaining the meteorological data of the distributed photovoltaic power stations.

[0039] In one embodiment of the present invention, meteorological data from 77 widely distributed integrated power stations in and around a certain location were selected, and G-SARNN was used to perform spatial interpolation of the meteorological data network for the distributed photovoltaic power stations within this area. This embodiment uses solar radiance as an example for experimental analysis. The adaptive spatial neural network, utilizing its superior analytical and construction capabilities, performed nonlinear fitting on the spatial distances and spatial weight coefficients of the integrated power stations used in the experiment to verify the feasibility of the network.

[0040] Regarding the interpolation experiments, this embodiment uses a method of comparing the interpolation results of traditional interpolation methods such as weighted average interpolation and Kriging interpolation with those of the adaptive spatial neural network to verify the effectiveness of the proposed adaptive spatial neural network. The G-SARNN function is set as a... The mapping process involves normalizing the spatial coordinates of the integrated power station with solar irradiance data samples to unify the data measurement scale. By balancing network accuracy and complexity, the sequential relationships between data samples are accurately extracted, thereby effectively simulating the nonlinear relationship between generalized spatial distance and spatial weight coefficients.

[0041] In an adaptive spatial recurrent neural network (RNN), the first GRU layer, the second GRU layer, and the fully connected layer serve as hidden layers. The selection of the number of hidden layers is crucial. Too few hidden layers prevent the network from accurately fitting the generalized spatial distance and spatial weight coefficients, leading to poor prediction accuracy. Too many hidden layers increase network complexity, making it prone to local minima, reducing computational efficiency, and wasting time. Therefore, determining the appropriate number of hidden layers requires fine-tuning using various methods, such as growth pruning, grid search, and transfer learning.

[0042] Based on the characteristics of G-SARNN, ​​in order to more effectively extract the relationships between data samples and fit the interaction between spatial distance and weight coefficients, the experiment selected hidden layer neurons as... The number of layers and the number of input samples in the training set are... The steps for calculating the number of hidden layers are as follows.

[0043] Step 1: Training begins, input the number of data samples. Initialize index variables d=0 Set the maximum power ,satisfy .

[0044] Step 2: Judgment d Check if it equals 0. If yes, proceed to Step 3; otherwise, update. d That is, for the current dAdd 1, then jump to Step 3.

[0045] Step 3: Judgment Is it greater than If yes, then output If not, then update. d That is, for the current d Add 1, then jump to Step 2.

[0046] Step 4: Output the hidden layer neuron structure to end training.

[0047] The number of hidden layers is obtained through the above steps. m The parameters of the adaptive spatial recurrent neural network are set according to the number of neurons in the hidden layer, as shown in Tables 1 and 2.

[0048] Table 1. Number of neurons in a neural network Table 2 Network Hyperparameter Set After each training cycle, the Reliability and Validation Performance Index (RMSE) of the G-SARNN training and validation is evaluated. If the RMSE remains relatively constant and shows an upward trend, the training is stopped immediately, and the current model parameters are recorded. This situation is considered overfitting.

[0049] Table 3 shows the analysis of interpolation results for the data sample space data interaction validation set using weighted average interpolation, Kriging interpolation, and GRU-based adaptive spatial neural network. The analysis results include model-related parameters ( The evaluation indexes are: root mean square error (RMSE), mean square error (MAE), and mean relative error (MAPE).

[0050] Table 3. Data Sample Interaction Validation Results Table 3 shows that, based on the prediction evaluation metrics of cross-validation, the IDW model, Kriging model, and G-SARNN can all achieve the purpose of interpolation prediction. The relevant parameters of the three models are as follows: All reached above 0.7. According to the model evaluation criteria... Higher numerical values ​​indicate better model performance. The G-SARNN parameter is 0.7982, representing a 7.4% and 10.34% improvement over IDW interpolation and the Kriging model, respectively. RMSE, a commonly used metric to measure the difference between predicted and actual data, indicates higher interpolation model accuracy. The RMSE values ​​for G-SARNN, ​​Kriging interpolation, and IDW are 0.0069, 0.1040, and 0.0082, respectively, indicating that G-SARNN performs better in interpolation tasks. The root mean square error (MAE) and mean relative error (MAPE) metrics both validate that G-SARNN outperforms other models. Therefore, the results show that the GRU-based adaptive spatial neural network exhibits high accuracy and low prediction error in interpolation, demonstrating its superior interpolation performance. To analyze and compare the interpolation effects of the three models, the interpolation results of these models on the solar power generation dataset are graphically displayed, such as... Figure 5 As shown.

[0051] Three different models were used to perform data interpolation analysis on the lower right part of the climate data sample to fit the spatial and temporal variation of meteorological data. Both the traditional geographic interpolation model and the GRU-based adaptive spatial neural network achieved a certain degree of interpolation effect. Specifically, the Kriging model showed the worst prediction performance, with the most obvious particle boundary separation and a noticeable jagged effect; the IDW interpolation model had a much smoother boundary than Kriging, but some local protrusions still existed; G-SARNN performed better than IDW and Kriging in interpolation prediction at higher values, with a smoother interpolation trend surface, and achieved the best results.

[0052] The line graph comparison of the true and predicted values ​​of the interpolation prediction results of the three models is as follows: Figure 6 , Figure 7 and Figure 8 As shown in the figure, the analysis results indicate that Kriging interpolation is the worst at fitting actual high and low values, with its interpolation results generally lower than the actual values ​​and having the largest error. IDW interpolation and the G-SARNN model provide relatively better prediction results, both reducing the overestimation of high values ​​and underestimation of low values ​​that occur with Kriging interpolation to some extent. Among them, the proposed G-SARNN model is relatively accurate in predicting extreme points, with the smallest error between its prediction and the actual values, and exhibits high interpolation prediction accuracy.

[0053] By comparing and analyzing the interpolation fitting of various models for meteorological data, the interpolation prediction results show that Kriging interpolation has a significant upward bias in high-value regions and a significant downward interpolation error in low-value regions. Although G-SARNN has some prediction bias near the maximum value, its interpolation prediction accuracy is significantly better than IDW interpolation and Kriging interpolation, and it performs stably and well in low-value regions. Therefore, compared with traditional geographic interpolation models, the G-SARNN proposed in this invention is relatively more sensitive to the prediction of extreme values, the interpolation results are closer to the actual values, and there are fewer instances of predicting lower values ​​near the maximum value or predicting higher values ​​near the minimum value; the prediction results are basically without deviation from the actual values.

[0054] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. Those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A meteorological resource interpolation method based on an adaptive spatial neural network for generating meteorological data for distributed photovoltaic power stations; characterized in that, The method includes the following steps: 1) Identify the distributed photovoltaic power stations for which meteorological data is to be generated, obtain the spatial coordinates and historical meteorological feature sequences of multiple integrated power stations within a wide area corresponding to the distributed photovoltaic power stations, and then normalize all historical meteorological feature sequences and spatial coordinates; and based on the normalized spatial coordinates, obtain the generalized spatial distance between each pair of integrated power stations through the standard orthogonal basis spatial coordinate system. 2) Construct an adaptive spatial recurrent neural network based on gated recurrent units. Based on the generalized spatial distance and normalized historical meteorological feature sequences from step 1), train the adaptive spatial recurrent neural network using cross-validation. 3) Obtain the spatial coordinates of the distributed photovoltaic power station and perform normalization processing, then obtain the generalized spatial distance between the distributed photovoltaic power station and each integrated power station; obtain the meteorological feature sequence of each integrated power station and perform normalization processing, and obtain the spatial weight coefficient between the distributed photovoltaic power station and each integrated power station through a trained adaptive spatial recurrent neural network based on the generalized spatial distance of step 3) and the normalized meteorological feature sequence; obtain the meteorological data of the distributed photovoltaic power station based on the spatial weight coefficient and the meteorological feature sequence of each integrated power station.

2. The meteorological resource interpolation method based on adaptive spatial neural networks according to claim 1, characterized in that, In step 1), the generalized spatial distance between every two integrated power stations is obtained based on the normalized spatial coordinates through a standard orthogonal basis spatial coordinate system; this includes: The normalized spatial coordinates corresponding to the integrated power stations are converted into coordinates in the standard orthogonal basis spatial coordinate system. Then, the generalized spatial distance between each pair of integrated power stations is obtained based on the coordinates of each pair of integrated power stations. The formula for calculating the generalized spatial distance is as follows: ； in, This represents the generalized spatial distance between two integrated power stations; A metric function representing the difference between vectors in a multidimensional space; , Let be the coordinates of an integrated power station in a standard orthogonal basis space coordinate system. The coordinates of the integrated power station in the standard orthogonal basis space coordinate system The nth dimension sub-coordinate; , Let the coordinates of another integrated power station be in the standard orthogonal basis space coordinate system. The coordinates of the integrated power station in the standard orthogonal basis space coordinate system The nth dimension sub-coordinate.

3. The meteorological resource interpolation method based on adaptive spatial neural networks according to claim 1, characterized in that, In step 2), the adaptive spatial recurrent neural network consists of an input layer, a first GRU layer, a second GRU layer, a fully connected layer, and the input layer in sequence; wherein, the activation function in the GRU layer is a parameterized modified linear unit activation function, and the expression of the parameterized modified linear unit activation function is: ； in, This represents the parameterized modified linear unit activation function; This represents the input value of the parameterized modified linear unit activation function; Represents the maximum value function; Describes the minimum value function; This represents the hyperparameters, which are adaptively adjusted and optimized during the training of the adaptive spatial recurrent neural network.

4. The meteorological resource interpolation method based on adaptive spatial neural networks according to claim 1, characterized in that, Step 2) involves training the adaptive spatial recurrent neural network, including: 21) Divide the normalized historical meteorological feature sequences of all integrated power stations into N equal parts, select N-1 parts as the training set of the adaptive spatial neural network, and use the rest as the validation set. 22) Input the normalized historical meteorological feature sequences and corresponding generalized spatial distances of the integrated power stations in the training set into the adaptive spatial neural network, and train the adaptive spatial neural network by gradient descent. 23) Input the normalized historical meteorological feature sequence and corresponding generalized spatial distance of the integrated power stations in the verification set into the adaptive spatial neural network obtained in step 22), and calculate multiple RMSE values. If the RMSE value continues to decrease, continue to step 25); otherwise, continue to step 24). 24) Determine whether the training rounds of the adaptive spatial neural network have reached the preset number of rounds. If they have, continue to step 25); otherwise, continue to step 22). At the same time, change the order in which the normalized historical meteorological feature sequences and the corresponding generalized spatial distances of the integrated power stations in the training set are input into the adaptive spatial neural network. 25) Select N-1 sets as the training set for the adaptive spatial neural network, and the rest as the validation set, and execute steps 22)-25) until no new training set can be selected; obtain the trained adaptive spatial neural network based on the adaptive spatial neural network trained on each training set.

5. The meteorological resource interpolation method based on adaptive spatial neural networks according to claim 4, characterized in that, Step 22) includes: Using the normalized historical meteorological feature sequence of any integrated power station in the training set as a label, and based on the normalized historical meteorological feature sequences of the other integrated power stations in the training set and their corresponding generalized spatial distances, an adaptive spatial neural network is used to obtain the spatial weight coefficients between the integrated power station and each of the other integrated power stations. Then, based on the spatial weight coefficients and the normalized historical meteorological feature sequences of the integrated power stations, the predicted meteorological feature sequence of the integrated power station corresponding to the label is obtained. The adaptive spatial neural network is trained using the gradient descent method, and the parameters in the adaptive spatial neural network are updated.

6. The meteorological resource interpolation method based on adaptive spatial neural networks according to claim 4, characterized in that, Step 23) includes: 231) Using the normalized historical meteorological feature sequence of any integrated power station in the validation set as the label, and based on the normalized historical meteorological feature sequences of the other integrated power stations in the validation set and the corresponding generalized spatial distance, the spatial weight coefficient between the integrated power station and each of the other integrated power stations is obtained through the adaptive spatial neural network obtained in step 22). Then, based on the spatial weight coefficient and the normalized historical meteorological feature sequence of the integrated power station, the predicted meteorological feature sequence of the integrated power station corresponding to the label is obtained. The RMSE value is calculated based on the predicted meteorological feature sequence of the integrated power station corresponding to the label and the normalized historical meteorological feature sequence. 232) Using the normalized historical meteorological feature sequence of another integrated power station in the validation set as a label, and calculating the RMSE value; repeat step 232) to obtain multiple RMSE values. If the RMSE value continues to decrease, continue to step 25); otherwise, continue to step 24).

7. The meteorological resource interpolation method based on adaptive spatial neural networks according to claim 4, characterized in that, In step 25), the trained adaptive spatial neural network is obtained, including: The average of the same parameters of the adaptive spatial neural networks trained on each training set is used as the corresponding new parameters, thus obtaining the trained adaptive spatial neural network.

8. The meteorological resource interpolation method based on adaptive spatial neural networks according to claim 1, characterized in that, In step 3), the meteorological data of the distributed photovoltaic power station is obtained based on the spatial weighting coefficient and the meteorological characteristic sequence of each integrated power station; including: The meteorological characteristic sequence of each integrated power station is multiplied by the corresponding spatial weight coefficient, and then all the multiplication results are added together to obtain the meteorological characteristic sequence of the distributed photovoltaic power station, that is, the meteorological data of the distributed photovoltaic power station.

Images