Intelligent integration method of multiple numerical weather forecast models

Abstract

The present invention provides a multi-value weather forecast model intelligent integration method, which matches the model forecast data of the same target forecast point within a preset period with the observation data of the target observation point corresponding to the same target forecast point to obtain time series data with the same time resolution, extracts the time series of multiple meteorological elements from the time series data and performs meteorological element correlation analysis based on the time series, determines the target meteorological element based on the meteorological element correlation analysis result and constructs a wind and light sample set based on the time series of the target meteorological element, trains a pre-established neural network model based on the wind and light sample set and the respective position information of each target observation point and each target forecast point, and adjusts the parameters of the neural network model during the training process to obtain an intelligent multi-mode integration model. The present invention can improve the efficiency and accuracy of multi-value weather forecast model integration.

Description

Technical Field

[0001] The invention relates to the technical field of weather forecasting, and in particular to an intelligent integration method of multiple numerical weather forecast models. Background Art

[0002] As numerical weather forecast products continue to be enriched and gradually mature, different numerical weather forecast models have different advantages and disadvantages. The method of integrating multiple numerical weather forecast models for weather forecasting has higher forecast accuracy than the method of using a single numerical weather forecast model for weather forecasting.

[0003] Traditional multi-numerical weather forecast model integration methods mainly include Monte Carlo forecasting method and multiple regression method. The Monte Carlo forecasting method estimates the error distribution range of the initial values, and then obtains a set of initial values, and then obtains the ensemble forecast value; the multiple regression method includes the ensemble average method and machine learning algorithm. The ensemble average method is a weighted average of the forecast results of all numerical weather forecast models participating in the multi-numerical weather forecast model integration. The machine learning method mainly uses the decision tree model to perform nonlinear regression on the forecast results of multiple numerical weather forecast models.

[0004] However, the traditional Monte Carlo forecasting method has high computational complexity, requires a lot of computing resources and takes a long time. The error obtained from the simulation is a probabilistic error, and the forecast results are still uncertain. The ensemble average method is too simple and cannot effectively exert the advantages of different numerical weather forecast models, and has certain limitations. The machine learning method is very sensitive to training data, is easily affected by noise and leads to overfitting, and is difficult to capture the evolution characteristics of numerical weather forecast model data over time, and has poor interpretability. Summary of the invention

[0005] In view of this, an object of the present invention is to provide a method for intelligent integration of multiple numerical weather forecast models to alleviate the above-mentioned problems existing in the related art.

[0006] The embodiment of the present invention provides a method for intelligent integration of multiple numerical weather forecast models, comprising: obtaining observation data of multiple target observation points within a preset time period and model forecast data of multiple numerical weather forecast models corresponding to the target forecast points within the preset time period; wherein the target forecast points are the central forecast points closest to the corresponding target observation points and the neighborhood points of each central forecast point; matching the model forecast data of the same target forecast point with the observation data of the target observation point corresponding to the same target forecast point to obtain time series data with the same time resolution; extracting time series of multiple meteorological elements from the time series data, and performing meteorological element correlation analysis based on the time series; wherein the The time series includes a first time series of corresponding meteorological elements in a corresponding numerical weather forecast mode and a second time series of corresponding meteorological elements corresponding to the observation data; the target meteorological elements are determined based on the results of meteorological element correlation analysis, and a wind and light sample set is constructed based on the time series of the target meteorological elements; based on the wind and light sample set and the respective position information of each target observation point and each target forecast point, a pre-established neural network model is trained and the parameters of the neural network model are adjusted during the training process to obtain an intelligent multi-mode integrated model; wherein the neural network model includes a backbone network, a branch network and a decoder, and the output ends of the backbone network and the branch network are respectively connected to the input end of the decoder.

[0007] An embodiment of the present invention provides an intelligent integration method for multiple numerical weather forecast models. The method first obtains observation data of multiple target observation points within a preset time period and model forecast data of multiple numerical weather forecast models corresponding to the target forecast points within the preset time period, and then matches the model forecast data of the same target forecast point with the observation data of the target observation point corresponding to the same target forecast point to obtain time series data with the same time resolution. Then, the time series of multiple meteorological elements are extracted from the time series data and meteorological element correlation analysis is performed based on the time series. Then, the target meteorological element is determined based on the result of the meteorological element correlation analysis and a wind and light sample set is constructed based on the time series of the target meteorological element. Then, a pre-established neural network model is trained based on the wind and light sample set and the respective position information of each target observation point and each target forecast point, and the parameters of the neural network model are adjusted during the training process to obtain an intelligent multi-mode integration model. By adopting the above technology, the observation data and location information of multiple target observation points and the multi-numerical weather forecast model data and location information of the target forecast point can be used to train the neural network model, so that the real-time multi-numerical weather forecast model data can be integrated through the trained neural network model, thereby realizing the intelligent integration of the multi-numerical weather forecast model. Compared with the traditional multi-numerical weather forecast model integration method, the efficiency and accuracy of the multi-numerical weather forecast model integration are improved.

[0008] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0009] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0011] Figure 1 A schematic diagram of a flow chart of a method for intelligent integration of multiple numerical weather forecast models in an embodiment of the present invention;

[0012] Figure 2 This is a schematic diagram of the arrangement of characteristic channels in an embodiment of the present invention;

[0013] Figure 3 This is an example diagram of the workflow of the intelligent multi-mode integration model in an embodiment of the present invention;

[0014] Figure 4 This is a diagram showing an example structure of a backbone network in an embodiment of the present invention;

[0015] Figure 5 This is a diagram showing an example of a branch network structure in an embodiment of the present invention;

[0016] Figure 6 This is a structural example diagram of a decoder in an embodiment of the present invention;

[0017] Figure 7 It is an overall flow chart of intelligent integration of multiple numerical weather forecast modes in an embodiment of the present invention;

[0018] Figure 8 It is a time series diagram of model forecast values ​​of different numerical weather forecast models and multi-model integration results and observation values ​​of some samples of the first test set in an embodiment of the present invention;

[0019] Fig. 9 It is a time series diagram of model forecast values ​​of different numerical weather forecast models for some samples of the second test set in an embodiment of the present invention, as well as multi-model integration results and observation values. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described in combination with the embodiments below. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] At present, the traditional multi-numerical weather forecast model integration method mainly includes Monte Carlo forecasting method and multiple regression method, and the multiple regression method includes ensemble average method and machine learning algorithm. However, the traditional Monte Carlo forecasting method has high computational complexity, requires a large amount of computing resources and takes a long time. The error obtained by simulation is a probabilistic error, and the forecast result is still uncertain; the ensemble average method is too simple and cannot effectively play the advantages of different numerical weather forecast models, and has certain limitations; the machine learning method is very sensitive to training data, and is easily affected by noise, resulting in overfitting, and it is difficult to capture the evolution characteristics of numerical weather forecast model data over time, and the interpretability is poor. Based on this, the implementation of the present invention provides a multi-numerical weather forecast model intelligent integration method, which can alleviate the above-mentioned problems existing in the relevant technology.

[0022] See also Figure 1 As shown, a method for intelligent integration of multiple numerical weather forecast models provided by an embodiment of the present invention may include the following steps:

[0023] Step S101, obtaining observation data of a plurality of target observation points within a preset time period and model forecast data of a plurality of numerical weather forecast models corresponding to target forecast points within the preset time period.

[0024] Among them, the above-mentioned target forecast points can be the central forecast points closest to the corresponding target observation points and the neighborhood points of each central forecast point.

[0025] The above-mentioned multiple numerical weather forecast models may include CMA-WSP (China Meteorological Administration Wind and Solar Weather Forecast System), CMA-MESO (China Meteorological Administration Mesoscale Weather Numerical Forecast System), ECMWF (European Centre for Medium-Range Weather Forecasts), NCEP-GFS (U.S. National Centers for Environmental Prediction Global Forecast System), etc., without limitation.

[0026] Step S102, matching the model forecast data of the same target forecast point with the observation data of the target observation point corresponding to the same target forecast point to obtain time series data with the same time resolution.

[0027] The time resolution of the above-mentioned time series data can adopt the time resolution of the model forecast data of the numerical weather forecast mode with the highest time resolution among multiple numerical weather forecast modes. For example, the above-mentioned multiple numerical weather forecast modes include CMA-WSP, CMA-MESO, ECMWF and NCEP-GFS, among which the time resolution of the model forecast data of CMA-WSP is the highest, which is 15 minutes. Then the time resolution of the above-mentioned time series data can be 15 minutes.

[0028] Step S103, extracting time series of multiple meteorological elements from the time series data, and performing meteorological element correlation analysis based on the time series.

[0029] The above-mentioned time series may include a first time series of corresponding meteorological elements in a corresponding numerical weather forecast mode and a second time series of corresponding meteorological elements corresponding to the above-mentioned observation data.

[0030] The above-mentioned multiple meteorological elements may include meteorological elements related to radiation (such as surface temperature, downward shortwave radiation from the surface, normal shortwave radiation, surface scattered radiation, direct downward radiation from the surface, direct downward radiation from the surface in clear sky, etc.) and meteorological elements related to wind speed (such as 10-meter wind speed, 50-meter wind speed, 70-meter wind speed, 80-meter wind speed, 100-meter wind speed, 120-meter wind speed, 1000hPa wind speed, 925hPa wind speed, etc.), and there is no limitation on this.

[0031] Step S104, determining the target meteorological element based on the meteorological element correlation analysis result, and constructing a scenery sample set based on the time series of the target meteorological element.

[0032] Step S105, based on the scenery sample set and the respective position information of each target observation point and each target forecast point, a pre-established neural network model is trained and the parameters of the neural network model are adjusted during the training process to obtain an intelligent multi-mode integrated model.

[0033] The neural network model may include a trunk network, a branch network and a decoder, and the output ends of the trunk network and the branch network are respectively connected to the input end of the decoder.

[0034] An embodiment of the present invention provides an intelligent integration method for multiple numerical weather forecast models. The method first obtains observation data of multiple target observation points within a preset time period and model forecast data of multiple numerical weather forecast models corresponding to the target forecast points within the preset time period, and then matches the model forecast data of the same target forecast point with the observation data of the target observation point corresponding to the same target forecast point to obtain time series data with the same time resolution. Then, the time series of multiple meteorological elements are extracted from the time series data and meteorological element correlation analysis is performed based on the time series. Then, the target meteorological element is determined based on the result of the meteorological element correlation analysis and a wind and light sample set is constructed based on the time series of the target meteorological element. Then, a pre-established neural network model is trained based on the wind and light sample set and the respective position information of each target observation point and each target forecast point, and the parameters of the neural network model are adjusted during the training process to obtain an intelligent multi-mode integration model. By adopting the above technology, the observation data and location information of multiple target observation points and the multi-numerical weather forecast model data and location information of the target forecast point can be used to train the neural network model, so that the real-time multi-numerical weather forecast model data can be integrated through the trained neural network model, thereby realizing the intelligent integration of the multi-numerical weather forecast model. Compared with the traditional multi-numerical weather forecast model integration method, the efficiency and accuracy of the multi-numerical weather forecast model integration are improved.

[0035] As a possible implementation manner, the above-mentioned wind-light sample set may include a radiation sample set and a wind speed sample set, and the above-mentioned target meteorological element may include a first target meteorological element related to radiation and a second target meteorological element related to wind speed; based on this, constructing the wind-light sample set based on the time series of the target meteorological element in the above-mentioned step S104 may include:

[0036] Step 1: construct the radiation sample set based on the first time series and the second time series of the first target meteorological element.

[0037] Exemplarily, the first time series of each first target meteorological element can be used as a first channel, and the second time series of the above-mentioned first target meteorological element corresponding to the above-mentioned observation data can be used as the first label, so that the first channel corresponding to the above-mentioned first target meteorological element is composed into an initial radiation sample and the first label corresponding to the above-mentioned first target meteorological element is marked for the initial radiation sample; each initial radiation sample with the first label is sampled by the first sliding window sampling method, and the first radiation sample obtained after sampling is normalized, and then the normalized second radiation sample is composed of the above-mentioned radiation sample set.

[0038] Step 2: construct the wind speed sample set based on the first time series and the second time series of the second target meteorological element.

[0039] Exemplarily, the first time series of each second target meteorological element can be used as a second channel, and the second time series of the above-mentioned second target meteorological element corresponding to the above-mentioned observation data can be used as a second label, so that the second channel corresponding to the above-mentioned second target meteorological element is composed of an initial wind speed sample and the second label corresponding to the above-mentioned second target meteorological element is marked for the initial wind speed sample; each initial wind speed sample with the second label is sampled by a second sliding window sampling method, and the first wind speed sample obtained after sampling is normalized, and then the normalized second wind speed samples are combined into the above-mentioned wind speed sample set.

[0040] In actual application, solar radiation has significant daily variation patterns, while the daily variation of wind speed is characterized by multiple fluctuations and nonlinearity. In order to make the above neural network model better capture the evolution characteristics of radiation and wind speed over time and improve the accuracy of the integration of multiple numerical weather forecast models, the required wind and solar sample set needs to contain the time series of each feature (each feature can represent a meteorological element). Figure 2 As shown in the figure, the time series of each feature (radiation-related meteorological element or wind speed-related meteorological element) can be used as a channel, and the time series of observation data shared by multiple channels can be used as a label. Multiple channels together form an initial sample (initial radiation sample or initial wind speed sample), and each initial sample is labeled with a corresponding label. Figure 2 The specific arrangement of the characteristic channels is shown.

[0041] Due to the limitations of the input data size and structure of the neural network model, Figure 2 After obtaining the initial samples by arranging the characteristic channels shown in the figure, the time series length of each sample (radiation sample or wind speed sample) can be set to a fixed length D (take ), set the window size of the sliding window to D, and set the sliding step size of the sliding window, so that each initial sample is slidingly sampled through the sliding window according to its window size and sliding step size, and then the first sample obtained after sliding sampling constitutes an initial scenery sample set.

[0042] In order to make different features have the same scale while maintaining their original distribution, the features and labels contained in each first sample can be normalized to [0, 1] according to the maximum and minimum values, and then the normalized second samples are used to form a landscape sample set. This normalization operation can improve the performance and stability of the neural network. The calculation formula for the maximum and minimum value normalization is as follows:

[0043]

[0044] in, is the original value (the original value of the feature or the original value of the label), and are the maximum value (feature maximum value or label maximum value) and minimum value (feature minimum value or label minimum value) of the feature respectively. is the normalized value (feature normalized value or label normalized value).

[0045] For the convenience of description, the obtained radiation sample set and wind speed sample set can be denoted as DS and DW respectively. The dimension of DS is M1×C1×D, M1 is the total radiation sample size, C1 is the number of radiation features, and the dimension of DW is M2×C2×D, M2 is the total wind speed sample size, and C2 is the number of wind speed features.

[0046] As a possible implementation, see Figure 3 As shown in the figure, the intelligent multi-mode integration model can be composed of three parts: a backbone network, a branch network and a decoder. The main workflow of the intelligent multi-mode integration model can be as follows: the time series with different characteristics under each numerical weather forecast mode are input into the backbone network to extract key information to obtain f1, the spatiotemporal information including longitude, latitude and month is input into the branch network for calculation to obtain f2, and then f1 and f2 are spliced ​​and input into the decoder for calculation, and finally the multi-mode integration result is obtained.

[0047] Based on the workflow of the intelligent multi-mode integration model, the structures of the backbone network, the branch network and the decoder are described as follows:

[0048] (1) The backbone network may include a first residual block, three second residual blocks and a first processing block connected in sequence; the first residual block includes a first convolutional layer, a first batch normalization layer, a first activation layer, a first random dropout layer, a second convolutional layer and a first pooling layer connected in sequence, the input end of the first convolutional layer is connected to the input end of the first pooling layer, and the output end of the second convolutional layer is connected to the output end of the first pooling layer; the second residual block includes a second batch normalization layer, a second activation layer, a second random dropout layer, a third convolutional layer, a third batch normalization layer, a third activation layer, a third random dropout layer and a fourth convolutional layer and a second pooling layer connected in sequence, the input end of the second batch normalization layer is connected to the input end of the second pooling layer, and the output end of the fourth convolutional layer is connected to the output end of the second pooling layer; the first processing block includes a fourth batch normalization layer and a fourth activation layer connected in sequence.

[0049] In practical applications, in order to fully extract the time evolution information of different numerical weather forecast model data, give full play to the advantages of each numerical weather forecast model, and fuse an integrated weather forecast result that is closest to the observed data, see Figure 4 As shown, the ResNet model can be used as a reference, with multiple residual blocks connected as the main part of the backbone network to extract the characteristic time series information of multiple numerical weather forecast modes; Figure 4In , the backbone network is composed of a Block A, three Block Bs and a Block C connected in sequence. Block A and Block B are two residual blocks with different network layer depths. Block A and Block B introduce skip connections to allow part of the feature information to pass through the intermediate layer and enter the deeper network layer, thereby improving the problems of gradient diffusion / explosion and neural network model degradation in deep neural networks. Block A and Block B both include stacked Conv1d layers, BN layers, APReLU layers and Dropout layers (with a discard probability of 0.5) and a skip-connected MaxPool layer; BlockA is composed of stacked Conv1d layers, BN layers, APReLU layers, Dropout layers and Conv1d layers in sequence and a skip-connected MaxPool layer. The input and output of the MaxPool layer in Block A correspond to the input of the first Conv1d layer and the output of the second Conv1d layer, respectively; Block Block B consists of a BN layer, an APReLU layer, a Dropout layer, a Conv1d layer, a BN layer, an APReLU layer, a Dropout layer, a Conv1d layer, and a jump-connected MaxPool layer. The input and output of the MaxPool layer in Block B correspond to the input of the first BN layer and the output of the second Conv1d layer, respectively. Block C consists of a BN layer and an APReLU layer stacked in sequence. All Conv1d layers in the backbone network use a convolution kernel of size 3×1 to extract features. Figure 4 The backbone network shown uses the Adaptively ParametricRectifier Linear Unit (APReLU) instead of the Rectified Linear Unit (ReLU) in the ResNet model, avoiding the problem of ReLU activation failure leading to model training failure when all features are less than 0; when the original feature is negative, APReLU can use a small fully connected network to generate a set of weight coefficients for each sample and act on the original feature, providing a more flexible nonlinear transformation method. The adaptability of APReLU enables the model to have better expressiveness and generalization capabilities.

[0050] Continuing from the previous example, see Figure 3As shown in the figure, the input data of the backbone network is the feature time series, and the dimension of the feature time series is N×C×D, where N is the number of samples (equal to M1 or M2), C is the number of channels (equal to C1 or C2), and D is the length of the time series; the output data of the backbone network is the feature vector f1, and the dimension of f1 is N×B, where B is the number of output features of the backbone network and is consistent with the number of filters in the initial convolutional layer of the backbone network. B can be set according to actual needs.

[0051] (2) The branch network includes two second processing blocks and one third processing block connected in sequence; the second processing block includes a first fully connected layer, a fifth batch normalization layer, and a fifth activation layer connected in sequence; the third processing block includes a second fully connected layer, a sixth batch normalization layer, and a sixth activation layer connected in sequence.

[0052] In actual applications, the spatiotemporal information such as latitude, longitude and month are all one-dimensional, which is not suitable to be used as a channel input neural network model together with the characteristic time series. Therefore, a small branch network is constructed based on MLP to learn the influence of spatiotemporal information in the multi-numerical weather forecast model integration task with high computational efficiency, avoiding the problem of one-dimensional information loss due to the backbone network being too deep. Figure 5 As shown, the branch network mainly consists of two L D and a L E Connected in sequence, L D and L E There are two different BN-MLP modules. Each BN-MLP module is composed of a Linear (fully connected) layer, a BN layer, and an activation layer stacked in sequence; Figure 5 middle, L D The activation function is ELU, which performs exponential smoothing on negative values ​​and can also transfer gradients in negative areas, which can help the model converge faster. L E The activation function is LeakyReLU. Like ReLU, LeakyReLU introduces nonlinear characteristics, enabling the model to learn complex patterns. However, LeakyReLU does not completely become zero in the negative region, retaining a certain sparsity, which helps to improve the efficiency and performance of the model. L D and a L E The number of neurons in each first layer can be set according to actual needs.

[0053] Continuing from the previous example, see Figure 3As shown in the figure, the input data of the branch network is one-dimensional spatiotemporal information with a dimension of N×F, where N is the number of samples and F is the number of one-dimensional information (the default is 3, and the one-dimensional information includes longitude, latitude and month). The output data of the branch network is the feature vector f2, and the dimension of f2 is N×E, where E represents the number of output features of the branch network, and E can be set according to actual needs.

[0054] (3) The decoder includes two second processing blocks and a fourth processing block connected in sequence; the fourth processing block includes a third fully connected layer.

[0055] In actual applications, the decoder can also be built based on MLP. The decoder is used to decode and fuse f1 and f2, and use its powerful nonlinear fitting ability to simulate the time evolution characteristics of the observed data, and finally obtain the multi-mode integrated forecast result. Figure 6 As shown, the decoder mainly consists of two L D and a L F Connected in sequence, L D is the BN-MLP module described above, L F is a single fully connected layer. L D and a L F The number of neurons in each first layer can be set according to actual needs.

[0056] Continuing from the previous example, see Figure 3 As shown in the figure, the input data of the decoder is a vector concatenated by f1 and f2, and the dimension of the vector is N×(B+E), where B and E are the number of features of f1 and f2 respectively; the output data of the decoder is the multi-mode integration result, and the dimension of the multi-mode integration result is N×G, where G represents the time series length of the forecast result, and G can be set to be consistent with the time series length of the sample (radiation sample or wind speed sample).

[0057] As a possible implementation, the meteorological element correlation analysis result may include a first correlation coefficient between each first time series of each meteorological element and its corresponding second time series in each numerical weather forecast mode; based on this, determining the target meteorological element based on the meteorological element correlation analysis result in the above step S104 may include:

[0058] Step A1, sort the obtained first correlation coefficients in descending order, and determine the meteorological elements whose sorting positions are before the preset sorting positions as candidate meteorological elements; or, determine the meteorological elements whose first correlation coefficients are not less than the preset correlation coefficient threshold as candidate meteorological elements.

[0059] Step A2, determining the maximum value of the first correlation coefficient corresponding to the same candidate meteorological element under the same numerical weather forecast mode, and taking the candidate meteorological element at the target forecast point corresponding to each maximum value as the corresponding target meteorological element.

[0060] In actual application, in order to reduce the number of features, reduce model complexity and reduce the risk of model overfitting, and avoid low-correlation features interfering with model training, the calculated first correlation coefficients of each meteorological element can be sorted and the meteorological elements with the top K largest first correlation coefficients can be selected as candidate meteorological elements according to the Top-K algorithm. Meteorological elements with first correlation coefficients not less than a preset correlation coefficient threshold can also be selected as candidate meteorological elements based on the calculated first correlation coefficients of each meteorological element. For the same numerical weather forecast mode and the same candidate meteorological element, only the candidate meteorological element at the target forecast point with the highest correlation (i.e., the largest first correlation coefficient) is selected as the key influencing factor (i.e., the target meteorological element).

[0061] As a possible implementation, the first time series may include the model forecast values ​​of the corresponding meteorological elements at each target moment within a preset time period under the corresponding numerical weather forecast mode, and the second time series may include the observation values ​​of the corresponding meteorological elements at each target moment within the preset time period; based on this, the meteorological element correlation analysis based on the time series in the step S103 may include: determining the first correlation coefficient corresponding to each meteorological element under each numerical weather forecast mode based on the model forecast values ​​of each meteorological element corresponding to each target moment under each numerical weather forecast mode and the observation values ​​of each meteorological element corresponding to each target moment.

[0062] In actual application, after obtaining the model forecast value of a meteorological element at n target times within a preset period under a certain numerical weather forecast mode and the observed value of the meteorological element at n target times, the first correlation coefficient r1 between the model forecast value and the observed value of the meteorological element under the numerical weather forecast mode can be calculated based on the obtained model forecast value and observed value using the following formula:

[0063]

[0064] in, is the model forecast value of the meteorological element at the i-th target time under the numerical weather forecast model, It is the average value of the model forecast of the meteorological element in the preset period under the numerical weather forecast model. is the time series of the observed value (observed radiation value / observed wind speed value) of the meteorological element at the i-th target time. It is the observed average value of the meteorological element in the preset period. According to this calculation method, the first correlation coefficient corresponding to each meteorological element in each numerical weather forecast mode can be calculated.

[0065] As a possible implementation, the time resolution of the observation data may be a preset time length; based on this, the step S102 (i.e., matching the model forecast data of the same target forecast point with the observation data of the target observation point corresponding to the same target forecast point to obtain time series data with the same time resolution) may include:

[0066] Step a1, converting the model forecast data of the same target forecast point in each numerical weather forecast model into first time series data with a time resolution of a preset time length by interpolation.

[0067] Step a2, correlating the first time series data and the observation data having the same time series corresponding to the same target forecast point to obtain the above time series data.

[0068] As a possible implementation, the intelligent multi-mode integrated model may include a radiation model and a wind speed model; based on this, the above step S105 (i.e., based on the wind and light sample set and the respective position information of each target observation point and each target forecast point, training a pre-established neural network model and adjusting the parameters of the neural network model during the training process to obtain the intelligent multi-mode integrated model) may include:

[0069] Step B1, obtaining the timing information and position information corresponding to the radiation samples contained in the radiation sample set, and constructing a first training set, a first verification set and a first test set based on the radiation sample set, the timing information and the position information, and then using the first training set to train the neural network model, and using the first verification set to verify the performance of the neural network model and adjust the parameters of the neural network model during the training process, until the first test set is used to test the performance of the neural network model after the training is completed, to obtain the radiation model.

[0070] Among them, the target observation points corresponding to the samples contained in the first training set, the first validation set and the first test set are different, and each sample contained in the first training set, the first validation set and the first test set is bound with its corresponding time series information.

[0071] The loss function of the radiation model can be the root mean square error (RMSE) between the output of the radiation model and its corresponding observation value.

[0072] Step B2, obtaining the position information corresponding to the wind speed samples contained in the wind speed sample set, and constructing a second training set, a second verification set and a second test set based on the wind speed sample set and the position information, using the second training set to train the neural network model, and using the second verification set during the training process to verify the performance of the neural network model and adjust the parameters of the neural network model, until the second test set is used to test the performance of the neural network model after the training is completed, to obtain the wind speed model.

[0073] The second training set, the second validation set, and the second test set each contain wind speed samples corresponding to different combinations of target observation points and time series information.

[0074] The loss function of the wind speed model may be determined by determining the RMSE and the second correlation coefficient between the output result of the wind speed model and its corresponding observation value, and determining the loss function of the wind speed model based on the RMSE and the second correlation coefficient.

[0075] In practical applications, the loss function of the wind speed model is The calculation formula can be:

[0076]

[0077] in, is the RMSE between the wind speed model forecast value (i.e., the output data of the wind speed model) and the wind speed observation value (i.e., the label of the input sample of the wind speed model), The calculation formula is:

[0078]

[0079]

[0080] r2 represents the second correlation coefficient of the corresponding data set (the second training set, the second validation set or the second test set) of the wind speed model, m represents the total number of samples of the corresponding data set, represents the model forecast value of the jth sample in the corresponding data set (i.e., the output result after the wind speed model predicts the jth sample in the corresponding data set). represents the average value of the model forecast of the corresponding data set (i.e., the average value of the output results after the wind speed model predicts m samples of the corresponding data set). represents the observed value of the j-th sample in the corresponding dataset (i.e., the label of the j-th sample in the corresponding dataset), represents the observed mean of the corresponding dataset (i.e., the average of the labels of the m samples in the corresponding dataset).

[0081] For ease of understanding, the implementation of the above-mentioned multi-value weather forecast model intelligent integration method is exemplarily described as follows by taking a specific application as an example.

[0082] See also Figure 7 As shown in the figure, the overall process of intelligent integration of multiple numerical weather forecast models can mainly include five steps: data analysis and feature engineering; preparation of wind and solar sample sets; construction of intelligent multi-mode integration model; model training and optimization; and verification of wind and solar integrated forecast results.

[0083] Below Figures 2 to 7 Take the following as an example to describe the specific implementation methods of the above five steps.

[0084] 1. Data Analysis and Feature Engineering

[0085] Data analysis and feature engineering mainly include the following steps 1.1 to 1.3:

[0086] Step 1.1, site timing matching.

[0087] The model forecast data used in this scheme include model forecast data under four numerical weather forecast modes, namely CMA-WSP, CMA-MESO, ECMWF and NCEP-GFS, in the same period. The spatiotemporal resolution and forecast time of the model forecast data are shown in Table 1. The observation data OBS is the measured data from discrete observation sites (i.e., radiation observation sites and wind speed observation sites) (the time resolution is generally 15 minutes). In order to facilitate subsequent analysis and modeling, the model forecast data and observation data of the same site and the same time series are first matched one by one, and the time resolution is unified to 15 minutes. The model forecast data with an original time resolution greater than 15 minutes needs to be interpolated by data interpolation to obtain time series data with a time resolution of 15 minutes (i.e., the first time series data at this time).

[0088] Table 1 Information table of model forecast data used in this scheme

[0089]

[0090] Step 1.2, correlation analysis of meteorological elements.

[0091] The meteorological elements contained in the model forecast data are rich in variety and quantity. This scheme pre-selects the model forecast data of the central point corresponding to the location of the observation site (that is, the forecast grid point closest to the observation site, which can be recorded as g1) and its adjacent points (that is, the forecast grid points adjacent to the center point in the four directions above, below, left and right, which can be recorded as g2~g5) under the above-mentioned four numerical weather forecast modes, and extracts the meteorological elements related to temperature, air pressure, relative humidity, wind field, radiation, etc. in each numerical weather forecast mode from the selected model forecast data. For all radiation observation stations, the sub-stations calculate the correlation coefficient between each meteorological element in each numerical weather forecast mode and its corresponding observed radiation value, and obtain meteorological elements related to the observed radiation, including surface temperature, downward shortwave radiation from the surface, normal shortwave radiation, surface scattered radiation, downward direct radiation from the surface, and downward direct radiation from the surface in clear sky, etc. For all wind speed stations, the sub-stations calculate the correlation coefficient between each meteorological model element in each numerical weather forecast mode and its corresponding observed wind speed value (i.e., the first correlation coefficient r1 at this time), and obtain meteorological elements related to the observed wind speed, including 10-meter wind speed, 50-meter wind speed, 70-meter wind speed, 80-meter wind speed, 100-meter wind speed, 120-meter wind speed, 1000hPa wind speed, and 925hPa wind speed, etc. The calculation formula of r1 is as follows:

[0092]

[0093] Among them, X is the time series of a meteorological element in a certain numerical weather forecast mode, and Y is the time series of the observed radiation value / observed wind speed value of the meteorological element.

[0094] Step 1.3, screening of key influencing factors.

[0095] In order to reduce the number of features, reduce the complexity of the model, reduce the risk of overfitting, and avoid the interference of low-correlation features in model training, the meteorological elements with r1≥0.5 are selected as key influencing factors according to the correlation coefficient of each meteorological element calculated in the previous step, and only the point with the largest r1 is selected for the same mode and the same meteorological element (that is, one is selected from g1~g5). After screening, 11 key influencing factors of radiation are obtained, namely: RMAPS_swdown (RMAPS surface downward shortwave radiation), RMAPS_swddif (RMAPS surface scattered radiation), RMAPS_swddirc (RMAPS clear sky surface downward direct radiation), RMAPS_swnorm (RMAPS normal shortwave radiation), RMAPS_swddir (RMAPS surface downward direct radiation) and RMAPS_swddirc (RMAPS clear sky surface downward direct radiation), RMAPS_swnorm (RMAPS normal shortwave radiation), RMAPS_swddir (RMAPS surface downward direct radiation) of RMAPS (Rapid Update Multiscale Analysis and Forecast System). APS_tsk (surface temperature from RMAPS), EC_ws10 (10-m wind speed from ECMWF) from ECMWF, CMA_MESO_tsk (surface temperature from CMA-MESO) and CMA_MESO_swdown (surface downward shortwave radiation from CMA-MESO) from CMA-MESO, NCEP_GFS_SWDOWN (surface downward shortwave radiation from NCEP-GFS) and NCEP_GFS_TSK (surface temperature from NCEP-GFS) from NCEP-GFS. After screening, 13 key influencing factors of wind speed were obtained, namely: RMAPS_ws70 (70-meter wind speed of RMAPS), RMAPS_ws80 (80-meter wind speed of RMAPS), RMAPS_ws100 (100-meter wind speed of RMAPS), RMAPS_ws10 (10-meter wind speed of RMAPS), RMAPS_ws120 (120-meter wind speed of RMAPS), RMAPS_ws_1000hPa (1000hPa wind speed of RMAPS) and RMAPS_ws_925hPa (925hPa wind speed of RMAPS). hpa wind speed), CMA_MESO_ws_925hpa (925hpa wind speed of CMA-MESO), CMA_MESO_ws_1000hpa (1000hpa wind speed of CMA-MESO), CMA_MESO_ws100 (100-meter wind speed of CMA-MESO), CMA_MESO_ws50 (50-meter wind speed of CMA-MESO) and CMA_MESO_ws10 (10-meter wind speed of CMA-MESO) of CMA-MESO and EC_ws100 (100-meter wind speed of ECMWF) of ECMWF.

[0096] 2. Production of Scenery Sample Set

[0097] The production of the scenery sample set mainly includes the following steps 2.1 to 2.3:

[0098] Step 2.1, feature channel arrangement.

[0099] According to Figure 2 In the arrangement shown, the time series of each feature is used as a channel, the time series of observation data shared by multiple channels is used as a label, and multiple channels together constitute an initial sample, and each initial sample is marked with a corresponding label.

[0100] Step 2.2, slide and crop the time series.

[0101] After all the data are arranged according to the previous step, a fixed-length time series is sampled by sliding. Due to the limitations of the input data size and structure of the neural network model, the time series of the sample is set to 12h (length is 48), and the sliding step is set to 24. Each initial sample can be sampled by sliding according to the set time series length and sliding step. The radiation samples obtained after sampling are composed of the radiation sample set DS, and the wind speed samples obtained after sampling are composed of the wind speed sample set DW. The DS dimension is M1×11×48, where M1 is the total radiation sample size, 11 is the number of radiation features, and 48 is the time series length; the DW dimension is M2×13×48, where M2 is the total wind speed sample size, 13 is the number of wind speed features, and 48 is the time series length. In the actual model reasoning process, the model will complete the 240h multi-mode integrated forecast through rolling iteration. Since the essence of multi-mode integration in this scheme is to fuse the information between different numerical weather forecast modes and different meteorological elements to obtain a result that is closest to the observed value, this is different from the process of time series prediction from scratch (that is, in time series prediction, if a rolling iterative forecast is performed, it is necessary to first predict the result of the next 1 hour based on historical data, and then predict the result of the next 2 hours based on the result, and so on. This will cause the error to gradually accumulate and the accuracy to gradually decrease). Therefore, the rolling iteration of multi-mode integration in this scheme will not cause error accumulation.

[0102] Step 2.3: Normalize the maximum and minimum values.

[0103] Each feature and each label is normalized and scaled to [0, 1] according to the maximum and minimum values, so that different features have the same scale while maintaining the original distribution, which can improve the performance and stability of the neural network. The calculation formula for maximum and minimum normalization is as follows:

[0104]

[0105] in, is the original value of the feature, and are the maximum and minimum values ​​of the feature, respectively. is the normalized eigenvalue.

[0106] 3. Construction of intelligent multi-mode integration model

[0107] The intelligent multi-modal integration model built by Figure 3 The time series with different characteristics under each numerical weather forecast mode are input into the backbone network to extract key information to obtain f1, and the spatiotemporal information such as longitude, latitude and month are input into the branch network to obtain f2, and then f1 and f2 are spliced ​​and input into the decoder to finally obtain the multi-mode integration result.

[0108] The construction of intelligent multi-mode integration model mainly includes the following steps 3.1 to 3.3:

[0109] Step 3.1, build the backbone network.

[0110] The structure of the backbone network is as follows Figure 4 As shown. You can first build Block A, Block B, and Block C, and then connect one Block A, three Block Bs, and one Block C in sequence to build the backbone network. The input data of the backbone network is a feature time series with a dimension of N×C×D, where N is the number of samples, C is the number of channels (equal to the number of key influencing factors in step 1.3), and D is the length of the time series (consistent with the sample time series length set in step 2.2); the output data of the backbone network is a feature vector f1 with a dimension of N×B, where B is consistent with the number of filters in the initial convolutional layer, and B can be set to 88 based on experience.

[0111] Step 3.2, branch network construction.

[0112] The branch network structure is as follows Figure 5 As shown. You can build L D and L E , and then put the two L D and a L E Connect in sequence to build a branch network. The two L D and a L EThe number of neurons in the first layer is 3, 16 and 8 respectively. The input data of the branch network is one-dimensional spatiotemporal information, with a dimension of N×F, where N is the number of samples and F is the number of one-dimensional information (the default is 3, and the one-dimensional information includes longitude, latitude and month); the output data of the branch network is a feature vector f2, with a dimension of N×E, where E represents the number of output features of the branch network, and here E can be set to 4 based on experience.

[0113] Step 3.3, decoder construction.

[0114] The decoder is used to decode and fuse f1 and f2 to obtain the multi-mode integrated forecast result. The decoder structure is as follows: Figure 6 As shown. You can build L D and L F , and then put the two L D and a L F Connect them in sequence to build the decoder. L D and a L F The number of neurons in the first layer is 92, 128 and 64 respectively. The input data of the decoder is a vector concatenated from f1 and f2, with a dimension of N×(B+E), where B and E are the number of features of f1 and f2, respectively. At this time, the values ​​of B and E are 88 and 4 respectively; the output data of the decoder is the multi-mode integrated forecast result, with a dimension of N×G, where G represents the time series length of the forecast result, and G is set to 48 at this time.

[0115] 4. Model training and optimization

[0116] Model training and optimization mainly include the following steps 4.1 to 4.4:

[0117] Step 4.1, radiation model training.

[0118] All data (including observation data and model forecast data) from 32 radiation stations in 2022 are used to produce the radiation sample set DS. In order to improve the generalization ability of the radiation multi-mode integrated model at different stations, the training set, validation set and test set of the radiation multi-mode integrated model (i.e., the radiation model) can be divided according to the selection of radiation stations: first, the data of 4 radiation stations are randomly or manually selected to construct the test set of the radiation model, and then the data of 3 radiation stations are randomly or manually selected to construct the validation set of the radiation model, and the data of the remaining 25 radiation stations are used to construct the training set of the radiation model. When manually selecting radiation stations, radiation stations can be selected in different geographical regions to divide the training set, validation set and test set of the radiation model. For example, the data of one radiation station in East China, North China, Southwest China and South China are selected to construct the test set, the data of one radiation station in Northeast China, Central China and Northwest China are selected to construct the validation set, and the data of the remaining 25 radiation stations are used to construct the training set.

[0119] All experiments used the adaptive motion estimation (Adam) algorithm to optimize the radiation model training and the Adam optimizer used the same parameter settings. The learning rate scheduler ReduceLROnPlateau was used to achieve adaptive adjustment of the learning rate, thereby improving the training effect of the radiation model. In all experiments, the loss function of the radiation model was RMSE, batchsize=6144, and the number of training iterations epoch=200.

[0120] Step 4.2, radiation model optimization.

[0121] The time resolution of the multiple numerical weather forecast models used for integration is not consistent. Some model forecast data are 1 hour or even 6 hours. When corresponding to the 15-minute observation data, the missing values ​​of the model forecast data will seriously affect the learning effect of the model. The characteristic time series coarser than 15 minutes can be linearly interpolated to obtain a time series with a time resolution of 15 minutes, thereby improving the integration effect of the multiple numerical weather forecast models.

[0122] During the experiment, it was found that the model's prediction of the maximum radiation value was too low. This may be due to the relatively small number of samples with large radiation values. Therefore, it is necessary to perform data enhancement on the training set of the radiation model to obtain more samples with large radiation values, thereby improving the prediction accuracy of the radiation model for the maximum radiation value. The data enhancement method can be: during the sliding sampling of each initial sample, if the maximum observed radiation value of the sample collected by the current sliding exceeds 500W / m 2 , then the next sliding step size is reduced to 2 and the sliding sampling is continued until the maximum value of the observed radiation collected by sliding does not exceed 500W / m 2After sampling, the sliding step size is restored from 2 to 24. Compared with the sliding sampling method with a fixed sliding step size of 24, this data enhancement method can obtain more radiation large value samples.

[0123] In addition, since solar radiation has significant daily variations, the time information (i.e., the specific moment of the characteristic time series) can also be arranged as a feature channel when constructing the test set, validation set, training set and radiation sample set to be predicted of the radiation model, so that the time information can be included in the sample and input into the radiation model together with different characteristic time series, thereby further improving the integration effect of multiple numerical weather forecast models.

[0124] Step 4.3, wind speed model training.

[0125] The wind speed sample set DW is produced using all data from the four wind speed stations in 2022 (including observation data and model forecast data). Due to the small number of stations, the training set, validation set, and test set are divided by selecting wind speed stations and months: first, data from 8 wind speed stations + months are randomly or manually selected to construct the test set of the wind speed model, and then data from 4 wind speed stations + months are randomly selected to construct the validation set of the wind speed model. The data from the remaining 36 wind speed stations + months are used to construct the training set of the wind speed model.

[0126] All experiments used the adaptive motion estimation (Adam) algorithm to optimize the wind speed model training and the Adam optimizer used the same parameter settings. The learning rate scheduler ReduceLROnPlateau was used to achieve adaptive adjustment of the learning rate, thereby improving the training effect of the wind speed model. In all experiments, the loss function of the wind speed model was RMSE, batchsize=6144, and the number of training iterations epoch=200.

[0127] Step 4.4, wind speed model optimization.

[0128] Since a small number of wind speed samples will lead to poor integration effect of the multi-numerical weather forecast model of the wind speed model, the number of wind speed samples can be increased by reducing the sliding step size in step 2.2 to improve the integration effect of the multi-numerical weather forecast model of the wind speed model. Specifically, if the number of wind speed samples obtained after sampling in step 2.2 is less than 100,000, the sliding step size can be reduced to 2 and each initial wind speed sample can be re-sampled to re-produce the wind speed sample set.

[0129] During the experiment, it was found that the wind speed curve of the multi-mode integrated wind speed model was generally flat and had a small change range. This may be due to the limitations of the single loss function. Therefore, the correlation coefficient of the corresponding data set (training set, validation set or test set) of the wind speed model was added to the loss function of the wind speed model (that is, the second correlation coefficient r2 at this time) to further improve the effect of multi-numerical weather forecast model integration. The modified loss function of the wind speed model The calculation formula is as follows:

[0130]

[0131] in, is the RMSE between the wind speed model forecast and the wind speed observation, The calculation formula is:

[0132]

[0133]

[0134] in, represents the wind speed model forecast value, Represents the observed wind speed value.

[0135] (V) Verification of wind and solar integrated forecast results

[0136] The verification of wind and solar integrated forecast results mainly includes the following steps 5.1 and 5.2:

[0137] Step 5.1, radiation model verification.

[0138] The RMSE between the model predicted radiation value and the observed radiation value under the original numerical weather forecast model is calculated on the test set of the reserved radiation model, and the numerical weather forecast model with the smallest RMSE is taken as the optimal original numerical weather forecast model (which can be referred to as the optimal original model). Then, the RMSE between the multi-model integration result of the radiation model and the observed radiation value is calculated, and the decrease in the RMSE corresponding to the multi-model integration result relative to the RMSE corresponding to the optimal original model is compared (the greater the decrease, the better the radiation model effect), so as to test the effect of the radiation model.

[0139] Specifically, the constructed radiation model was used to perform 240 hours of multi-mode integration and verification on the test set. The test set of the radiation model contains a total of 1408 samples, of which 63.57% of the samples have a smaller error (RMSE) than the optimal original mode CMA-WSP, and 41.48% of the samples have a RMSE corresponding to the multi-mode integration results that is more than 20% lower than the RMSE corresponding to CMA-WSP. The average RMSE corresponding to the multi-mode integration results is 157.7135W / m 2The average RMSE corresponding to CMA-WSP is 128.6862 W / m 2 The average RMSE of the multi-mode integration results when aligned with the CMA-WSP data (i.e., the first 84 hours) is 110.9445 W / m 2 The average decrease of RMSE corresponding to the multi-model integration results compared with the RMSE corresponding to CMA-WSP is 13.79%. Randomly select samples from the test set of the radiation model to visualize the model forecast values ​​of different numerical weather forecast models and the multi-model integration results and observations. The visualization comparison results are shown in Figure 8 shown.

[0140] Step 5.2, wind speed model verification.

[0141] The RMSE between the model forecast wind speed value and the observed wind speed value under the original numerical weather forecast mode is calculated on the test set of the reserved wind speed model, and the numerical weather forecast mode with the smallest RMSE is taken as the optimal original mode. Then, the RMSE between the multi-mode integration result of the wind speed model and the observed wind speed value is calculated, and the decrease in the RMSE corresponding to the multi-mode integration result relative to the RMSE corresponding to the optimal original mode is compared (the greater the decrease, the better the wind speed model effect), so as to verify the effect of the wind speed model.

[0142] Specifically, the constructed wind speed model was used to perform multi-mode integration and verification on the test set for 240 hours. The test set of the wind speed model contains a total of 186 samples, of which 72.04% of the samples have a smaller error (i.e., RMSE) in the multi-mode integration results compared with the optimal original model CMA-WSP, and 39.25% of the samples have a RMSE corresponding to the multi-mode integration results that is more than 20% lower than the RMSE corresponding to CMA-WSP. The average RMSE corresponding to the multi-mode integration results is 2.4047m / s, and the average RMSE corresponding to CMA-WSP is 2.0913m / s; when the multi-mode integration results are aligned with the CMA-WSP data (i.e., the first 84 hours), the RMSE corresponding to the multi-mode integration results is reduced by an average of 20.63% compared with the RMSE corresponding to CMA-WSP. Samples were randomly selected from the test set of the wind speed model to visualize the model forecast values ​​of different numerical weather forecast models, as well as the multi-mode integration results and observations. The visualization comparison results are shown in the figure. Fig. 9 shown.

[0143] The above-mentioned intelligent integration method of multiple numerical weather forecast models has the following advantages:

[0144] 1) Different from the regression mapping of single-time numerical weather forecast model products and observation data, this technology inputs the time series of key influencing factors in each numerical weather forecast model into the model as different channels, so that the model can better capture the evolution characteristics of radiation and wind speed over time.

[0145] 2) The structures of the backbone network and branch networks can process heterogeneous data such as numerical weather forecast model products and spatiotemporal information in a targeted manner, while realizing the integration of multiple numerical weather forecast models and retaining the information of the original model forecast data to the maximum extent, which helps to improve the accuracy of the integration of multiple numerical weather forecast models.

[0146] 3) By enhancing the data of the radiation model training set, more radiation large value samples are added, which helps to improve the prediction accuracy of the radiation model for radiation large values. In addition, time information is also input into the radiation model as a feature channel, so that the radiation model can better learn the daily variation of solar radiation.

[0147] 4) By introducing the correlation coefficient as part of the loss function during the wind speed model training process, the problem of the overall flatness and small variation of the multi-model integrated wind speed curve was effectively improved, and the integration effect of the multi-numerical weather forecast model was further improved.

[0148] Unless otherwise specifically stated, the relative steps, numerical expressions and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present invention.

[0149] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0150] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for intelligent integration of multiple numerical weather forecast models, characterized in that: include: Obtaining observation data of multiple target observation points within a preset time period and model forecast data of multiple numerical weather forecast models corresponding to target forecast points within the preset time period; wherein the target forecast points are the central forecast points closest to the corresponding target observation points and the neighboring points of each central forecast point; Matching the model forecast data of the same target forecast point with the observation data of the target observation point corresponding to the same target forecast point to obtain time series data with the same time resolution; Extracting a plurality of time series of meteorological elements from the time series data, and performing meteorological element correlation analysis based on the time series; wherein the time series includes a first time series of the corresponding meteorological elements in the corresponding numerical weather forecast mode and a second time series of the corresponding meteorological elements corresponding to the observation data; Determine the target meteorological elements based on the results of meteorological element correlation analysis, and construct a landscape sample set based on the time series of the target meteorological elements; Based on the scenery sample set and the respective position information of each target observation point and each target forecast point, a pre-established neural network model is trained and the parameters of the neural network model are adjusted during the training process to obtain an intelligent multi-mode integrated model; wherein the neural network model includes a backbone network, a branch network and a decoder, and the output ends of the backbone network and the branch network are respectively connected to the input end of the decoder; The meteorological element correlation analysis result includes a first correlation coefficient between each first time series of each meteorological element and its corresponding second time series in each numerical weather forecast mode; determining the target meteorological element based on the meteorological element correlation analysis result includes: Sorting the obtained first correlation coefficients in descending order of the first correlation coefficients, and determining the meteorological elements whose sorting positions are before the preset sorting positions as candidate meteorological elements; or determining the meteorological elements whose first correlation coefficients are not less than the preset correlation coefficient threshold as candidate meteorological elements; The maximum value of the first correlation coefficient corresponding to the same candidate meteorological element under the same numerical weather forecast mode is determined, and the candidate meteorological element at the target forecast point corresponding to each maximum value is used as the corresponding target meteorological element.

2. The intelligent integration method of multiple numerical weather forecast models according to claim 1, characterized in that: The backbone network includes a first residual block, three second residual blocks and a first processing block connected in sequence; the first residual block includes a first convolutional layer, a first batch normalization layer, a first activation layer, a first random dropout layer, a second convolutional layer and a first pooling layer connected in sequence, the input end of the first convolutional layer is connected to the input end of the first pooling layer, and the output end of the second convolutional layer is connected to the output end of the first pooling layer; the second residual block includes a second batch normalization layer, a second activation layer, a second random dropout layer, a third convolutional layer, a third batch normalization layer, a third activation layer, a third random dropout layer and a fourth convolutional layer and a second pooling layer connected in sequence, the input end of the second batch normalization layer is connected to the input end of the second pooling layer, and the output end of the fourth convolutional layer is connected to the output end of the second pooling layer; the first processing block includes a fourth batch normalization layer and a fourth activation layer connected in sequence; The branch network includes two second processing blocks and one third processing block connected in sequence; the second processing block includes a first fully connected layer, a fifth batch normalization layer, and a fifth activation layer connected in sequence; the third processing block includes a second fully connected layer, a sixth batch normalization layer, and a sixth activation layer connected in sequence; The decoder includes two second processing blocks and a fourth processing block connected in sequence; the fourth processing block includes a third fully connected layer.

3. The intelligent integration method of multiple numerical weather forecast models according to claim 2, characterized in that: The wind and light sample set includes a radiation sample set and a wind speed sample set, and the target meteorological element includes a first target meteorological element related to radiation and a second target meteorological element related to wind speed; The scenery sample set is constructed based on the time series of the target meteorological elements, including: Constructing the radiation sample set based on the first time series and the second time series of the first target meteorological element; The wind speed sample set is constructed based on the first time series and the second time series of the second target meteorological element.

4. The intelligent integration method of multiple numerical weather forecast models according to claim 3 is characterized in that: Constructing the radiation sample set based on the first time series and the second time series of the first target meteorological element includes: The first time series of each first target meteorological element is used as a first channel, and the second time series of the first target meteorological element corresponding to the observation data is used as a first label, so that the first channel corresponding to the first target meteorological element is formed into an initial radiation sample and the first label corresponding to the first target meteorological element is marked for the initial radiation sample; Each initial radiation sample with the first label is sampled by a first sliding window sampling method, and the first radiation sample obtained after sampling is normalized, and then the normalized second radiation samples are used to form the radiation sample set.

5. The intelligent integration method of multiple numerical weather forecast models according to claim 4 is characterized in that: Constructing the wind speed sample set based on the first time series and the second time series of the second target meteorological element includes: The first time series of each second target meteorological element is used as a second channel, and the second time series of the second target meteorological element corresponding to the observed data is used as a second label, so that the second channel corresponding to the second target meteorological element is formed into an initial wind speed sample and the second label corresponding to the second target meteorological element is marked for the initial wind speed sample; Each initial wind speed sample with the second label is sampled by a second sliding window sampling method, and the first wind speed samples obtained after sampling are normalized, and then the normalized second wind speed samples are used to form the wind speed sample set.

6. The intelligent integration method of multiple numerical weather forecast models according to claim 5, characterized in that: The first time series includes the model forecast values ​​of the corresponding meteorological elements at each target time within a preset period in the corresponding numerical weather forecast mode, and the second time series includes the observation values ​​of the corresponding meteorological elements at each target time within the preset period; The meteorological element correlation analysis is performed based on the time series, including: Based on the model forecast value of each meteorological element corresponding to each target time in each numerical weather forecast mode and the observation value of each meteorological element corresponding to each target time, the first correlation coefficient corresponding to each meteorological element in each numerical weather forecast mode is determined.

7. The intelligent integration method of multiple numerical weather forecast models according to claim 1, characterized in that: The time resolution of the observation data is a preset time length; matching the model forecast data of the same target forecast point with the observation data of the target observation point corresponding to the same target forecast point to obtain time series data with the same time resolution, including: Converting the model forecast data of the same target forecast point in each numerical weather forecast model into first time series data with a time resolution of the preset time length by interpolation; The first time series data and the observation data having the same time series corresponding to the same target forecast point are correspondingly associated to obtain the time series data.

8. The intelligent integration method of multiple numerical weather forecast models according to claim 6, characterized in that: The intelligent multi-mode integrated model includes a radiation model and a wind speed model; based on the wind and light sample set and the respective position information of each target observation point and each target forecast point, a pre-established neural network model is trained and the parameters of the neural network model are adjusted during the training process to obtain an intelligent multi-mode integrated model, including: Obtaining timing information and position information corresponding to the radiation samples contained in the radiation sample set, and constructing a first training set, a first validation set and a first test set based on the radiation sample set, the timing information and the position information, then using the first training set to train the neural network model, and using the first validation set to verify the performance of the neural network model and adjust the parameters of the neural network model during the training process, until the first test set is used to test the performance of the neural network model after the training is completed, to obtain the radiation model; wherein the target observation points corresponding to the samples contained in the first training set, the first validation set and the first test set are different, and each sample contained in the first training set, the first validation set and the first test set is bound to its corresponding timing information; The position information corresponding to the wind speed samples contained in the wind speed sample set is obtained, and a second training set, a second verification set and a second test set are constructed based on the wind speed sample set and the position information, the second training set is used to train the neural network model, and the second verification set is used to verify the performance of the neural network model and adjust the parameters of the neural network model during the training process, until the performance of the neural network model is tested with the second test set after the training is completed, so as to obtain the wind speed model; wherein the second training set, the second verification set and the second test set each contain different combinations of target observation points and timing information corresponding to the wind speed samples.

9. The intelligent integration method of multiple numerical weather forecast models according to claim 8, characterized in that: The loss function of the radiation model is the root mean square error between the output result of the radiation model and its corresponding observation value; Determining the loss function of the wind speed model includes: determining a root mean square error and a second correlation coefficient between an output result of the wind speed model and its corresponding observation value, and determining the loss function of the wind speed model based on the root mean square error and the second correlation coefficient.

Images