A short-term photovoltaic power prediction method based on LSGAN-DCN network

Abstract

This paper discloses a short-term photovoltaic power forecasting method based on the LSGAN-DCN network, comprising the following steps: 1. Collecting raw sample data, dividing the data by climate type, screening out sudden weather sample data, and performing normalization preprocessing on the sample data; 2. Using a multi-generator-based least squares generative adversarial network (MAD-LSGANs) to learn the sample distribution patterns of the original real photovoltaic data with sudden weather, generating new samples that are similar to the original sample data and have a high degree of diversity; 3. Establishing a deep cross network (DCN) model consisting of a cross network and a deep network; 4. Using the whale optimization algorithm (WOA) to determine the model's hyperparameters, and using the trained model to perform short-term photovoltaic power forecasting. This method effectively solves the problem of insufficient sudden weather sample data when predicting photovoltaic power generation in photovoltaic power stations, and enhances the acquisition of high-order feature cross information, thereby effectively improving the accuracy of photovoltaic power forecasting.

Description

Technical Field

[0001] The present invention relates to the technical field of short-term photovoltaic power prediction, and in particular to a short-term photovoltaic power prediction method based on an LSGAN-DCN network. Background Art

[0002] As the installed capacity of large-scale, centralized, grid-connected photovoltaic power plants continues to grow, the challenges they pose to power systems are becoming increasingly severe. Efficient and accurate short-term photovoltaic power forecasting can help power companies formulate rational operation and dispatch plans, and is a crucial foundation for ensuring the safe and economical operation of new power systems.

[0003] At present, photovoltaic power prediction mostly uses feature engineering to start from the feature set of original sample data. Through clustering, extraction and algorithm selection of features, the feature subset with the greatest correlation with photovoltaic power generation is obtained, and the impact of important information on the power generation prediction of photovoltaic power stations is strengthened, thereby improving its prediction accuracy while simplifying the model.

[0004] Currently, the classification of weather types in photovoltaic power forecasting is mostly limited to non-mutated weather, with little research addressing sudden weather changes such as sunny to cloudy or cloudy to rainy. To address this issue, some studies have used ensemble empirical mode decomposition (EMD) to decompose hourly photovoltaic power station output data into relatively stable component sequences. This analysis extracts different meteorological factors for each type of data, and then uses support vector machines with different kernel functions and parameters to construct prediction models for each component sequence, thereby obtaining the optimal prediction value for each component sequence. However, most existing studies assume sufficient raw sample data sets, ignoring the inherent sample shortage of raw sample data for sudden weather changes or newly built photovoltaic power stations, which can affect photovoltaic power forecast results. Summary of the Invention

[0005] In order to overcome the deficiencies in the above-mentioned prior art, the present invention provides a short-term photovoltaic power prediction method based on the LSGAN-DCN network, in order to enhance the matching degree and sample capacity of sample data of different categories, strengthen the acquisition of high-order feature cross-information, thereby effectively solving the problem of insufficient sample data of sudden weather changes when predicting photovoltaic power generation of photovoltaic power stations, and improving the prediction accuracy of photovoltaic power generation under sudden weather changes.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] The short-term photovoltaic power prediction method based on the LSGAN-DCN network of the present invention is characterized in that it is carried out according to the following steps:

[0008] Step 1: Collection and classification of sudden weather sample data;

[0009] Step 1.1: Obtain the original photovoltaic data sequence from the photovoltaic power station and divide the original photovoltaic data sequence according to the set climate type split point to obtain the photovoltaic data sequence under each climate type; and filter out the l-type sudden weather sample data from the photovoltaic data sequence under each climate type according to the weather type;

[0010] Step 1.2: Select n samples from each type of sudden weather sample data to form a sample data set, and each sample in the sample data set consists of m types of meteorological features. After normalizing the sample data set, a meteorological feature matrix with a dimension of n×m is obtained. Among them, x ij represents the jth meteorological feature of the i-th sample after normalization, i∈1,2,…,n, j∈1,2,…,m;

[0011] Step 1.3: Collect the photovoltaic power data corresponding to the m meteorological characteristics in the n samples in the sample data set, and perform normalization processing to obtain the power sequence P = [p1, p2, ... p i …p n ], p i represents the power of the i-th sample after normalization;

[0012] Step 1.4: The meteorological feature matrix k n×m Splice with the power sequence P to obtain the sample matrix Among them, p i,m+1 Represents the sample matrix K n×(m+1) The element in the i-th row and m+1-th column, and p i,m+1 =p i , i∈1,2,…,n, j∈1,2,…,m+1;

[0013] Step 1.5: Set the sample matrix K n×(m+1) Divide into training samples E a×(m+1) and test sample R e×(m+1) , a+e=n;

[0014] Step 2: Construct a least squares generative adversarial network based on multiple generators to generate new samples;

[0015] The least squares generative adversarial network includes: d generative models G and a discriminative model D; let the loss function of the least squares generative adversarial network be the least squares loss function;

[0016] Step 2.1: Use a fully connected neural network to construct d generative models G = [G1, G2, ... G g ,…,G d ], G g represents the g-th generative model, g∈1,2,…,d;

[0017] Step 2.2, constructing the discriminant model D using a fully connected neural network;

[0018] Step 2.3: Input d one-dimensional random noises of length c into the generative model G and obtain pseudo samples Q. G,(c×d)×(m+1) =[Q 1,(c)×(m+1) ,Q 2,(c)×(m+1) ,…,Q g,(c)×(m+1) …,Q d,(c)×(m+1) ], Q g,(c)×(m+1) Indicates that the dimension of the g-th generative model generated is c×(m+1) pseudo samples;

[0019] Step 2.4, fix the parameters of the generative model G, train the discriminative model D; a×(m+1) and pseudo samples Q G,(c×d)×(m+1) The discriminant model D is input for binary classification processing, and d one-dimensional thermal tensors of length c are output; samples marked as "1" in the one-dimensional thermal tensor represent true samples; samples marked as "0" are false samples; based on the one-dimensional thermal tensor, the least squares generative adversarial network model is trained using the stochastic gradient descent method to update the parameters of the discriminant model D, thereby obtaining a trained discriminant model D';

[0020] Step 2.5: Fix the parameters of the trained discriminant model D' and use it to train the generative model G:

[0021] Input one-dimensional random noise of length c into the generative model G and generate new pseudo samples Q' G,(c×d)×(m+1) ; The training sample E a×(m+1) and new pseudo sample Q' G,(c×d)×(m+1) Input into the trained discriminant model D' to determine whether the sample is true or false; according to the heat tensor output by the discriminant model D', the parameters of each generative model are updated using the stochastic gradient descent method to obtain the trained generative model G';

[0022] Step 2.6: Fix the parameters of the trained generative model G' to optimize the trained discriminative model D':

[0023] Input one-dimensional random noise of length c into the trained generative model G' to obtain new pseudo samples;

[0024] The new pseudo sample and the training sample E a×(m+1) Input it into the trained discriminant model D' for optimization training to obtain the optimized discriminant model D";

[0025] Step 2.7: Repeat steps 2.3 to 2.6 until all pseudo samples generated by the discriminant model are marked as "1". The training is completed, and the latest sample Q" is generated by the generative model after the final optimization training. G,(c×d)×(m+1) ;

[0026] Step 2.8: The latest sample set Q″ G,(c×d)×(m+1) With training sample E a×(m+1) After merging, we get the new training sample set X s×(m+1) , s=a+(c×d);

[0027] Step 3: Build the DCN network, which includes: the embedding and stacking layers at the top, the parallel cross neural network layer and deep neural network layer in the middle, and the combined output layer at the bottom;

[0028] Step 3.1: New training sample set X s×(m+1) Input into the DCN network and determine the new training sample set X s×(m+1) Each meteorological feature is divided into sparse feature R and dense feature Y dense ;

[0029] In the embedding and stacking layers, the sparse feature R is converted into a low-dimensional dense vector R using formula (1). emb Otherwise,

[0030] R emb,r =W emd,r R r (1)

[0031] In formula (1): R r represents the rth sparse feature, Indicates the total number of sparse features; W emd,r Represents the weight parameter when the rth sparse feature is converted into a low-dimensional dense vector; R emb,r Represents the low-dimensional dense vector transformed from the rth sparse feature;

[0032] Step 3.2: The embedding and stacking layers use formula (2) to transform the low-dimensional dense vector R emb With the dense feature Y dense After splicing, the common input vector λ0 of the cross neural network and the deep neural network is obtained using formula (2):

[0033]

[0034] In formula (2): represents the transpose of the r-th low-dimensional dense vector; represents the transpose of a dense vector;

[0035] Step 3.3, construct the cross neural network layer and the deep neural network layer;

[0036] Step 3.3.1, the cross neural network layer is composed of U cross layers;

[0037] The input vector λ0 is input into the cross neural network for explicit feature cross learning, and the explicit feature cross information λ output by the u-th cross layer is obtained using formula (3): u , so that the cross neural network outputs explicit high-order feature cross information λ U ;

[0038]

[0039] In formula (3), ω u ,b u is the weight and bias of the u-th layer;

[0040] Step 3.3.2, the deep neural network layer is composed of H fully connected layers;

[0041] The input vector λ0 is input into the deep neural network layer, and the implicit high-order feature cross information φ output by the hth fully connected layer is obtained using formula (4) h , so that the deep neural network outputs implicit high-order feature cross information φ H ;

[0042] φ h =f(ω h φ h-1 +b h ), h∈1,2,…,h,…,H (4)

[0043] In formula (4): h , b h are the weight and bias of the hth fully connected layer respectively; f(·) is the ReLU function;

[0044] Step 3.4: The combined output layer is composed of a fully connected layer, and the explicit feature cross information λ U and implicit feature cross information φ H After merging, the combined output layer is input, and the predicted value v is obtained using formula (5):

[0045]

[0046] In formula (5), Represents the explicit feature cross information λ U The transposed and implicit feature cross information φ H The transpose of ω denseis the weight vector of the combined output layer; σ(·) represents the ReLU function;

[0047] Step 3.5: The predicted value v output by the DCN network is denormalized to obtain the actual photovoltaic power prediction value v o ;

[0048] Step 4: Whale algorithm optimizes DCN network;

[0049] Step 4.1: Use the whale optimization algorithm to determine the hyperparameters of the DCN network, including: the number of cross network layers U, the number of deep network layers H, the number of neurons θ, and the training period τ;

[0050] Step 4.2: New training sample set X s×(m+1) The data is input into a deep cross-network model with determined hyperparameters for training, thereby obtaining a trained photovoltaic power generation prediction model under sudden weather changes.

[0051] Step 4.3, using the test set R e×(m+1) The trained photovoltaic power generation power prediction model under sudden weather changes is used to predict the power generation of photovoltaic power stations, thereby obtaining the short-term photovoltaic power generation power prediction results under sudden weather changes.

[0052] Compared with the prior art, the beneficial effects of the present invention are embodied in:

[0053] 1 Aiming at the impact of insufficient original sample data of sudden weather on photovoltaic power generation prediction, the present invention adopts the MAD-LSGANs network model based on the sample distribution law of real photovoltaic data of original samples of sudden weather, which can generate new samples that are similar to the original sample data and have high diversity, thereby suppressing the impact of insufficient original sample data of sudden weather on photovoltaic power generation prediction results.

[0054] 2 Aiming at the problem of extracting different information features, the present invention adopts a deep cross network to efficiently obtain explicit feature cross information and implicit feature cross information, reducing the process of using feature engineering to obtain feature cross information, and improving the photovoltaic power prediction accuracy while obtaining high-order feature cross information.

[0055] 3 Aiming at the difficulty in selecting hyperparameters in the DCN network model, the present invention proposes to use the global search capability of the whale algorithm to find the optimal number of network layers of the deep cross neural network, the number of neurons corresponding to each layer, and the training cycle to obtain the optimal solution model, thereby avoiding falling into the local optimal solution during the prediction process and greatly improving the convergence rate and accuracy of photovoltaic prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a schematic diagram of the process of short-term photovoltaic power prediction of the present invention;

[0057] Figure 2 Schematic diagram of the DCN model structure. DETAILED DESCRIPTION

[0058] In this embodiment, a short-term photovoltaic power prediction method based on the LSGAN-DCN network takes into account the impact of insufficient sample data of sudden weather on the prediction accuracy of photovoltaic power generation. The MAD-LSGANs network model is used to learn the sample distribution law of real photovoltaic data of sudden weather, and generate new samples that are similar to the original sample data and have high diversity, so as to achieve the purpose of expanding the sample data training set. Secondly, a deep cross network is used to obtain high-order feature cross information of the sample data, and the deep cross network model optimized by the whale algorithm is used to predict the short-term photovoltaic power of the photovoltaic power station under sudden weather, thereby obtaining high-order feature cross information while expanding the sample training set to improve the accuracy of short-term photovoltaic power prediction. Specifically, if Figure 1 As shown, the steps are as follows:

[0059] Step 1: Collection and classification of sudden weather sample data;

[0060] Step 1.1: Obtain the original photovoltaic data sequence from the photovoltaic power station and divide the original photovoltaic data sequence according to the set climate type split point to obtain the photovoltaic data sequence under each climate type; and filter out the l-type sudden weather sample data from the photovoltaic data sequence under each climate type according to the weather type;

[0061] Step 1.2: Select n samples from each type of sudden weather sample data to form a sample data set, and each sample in the sample data set consists of m types of meteorological features. After normalizing the sample data set, a meteorological feature matrix with a dimension of n×m is obtained. Among them, x ij represents the jth meteorological feature of the i-th sample after normalization, i∈1,2,…,n, j∈1,2,…,m;

[0062] Step 1.3: Collect the photovoltaic power data corresponding to the m meteorological characteristics in the n samples in the sample data set, and perform normalization processing to obtain the power sequence P = [p1, p2, ... p i …p n ], p i represents the power of the i-th sample after normalization;

[0063] Step 1.4: The meteorological feature matrix k n×m Splice with the power sequence P to obtain the sample matrix Among them, p i,m+1 Represents the sample matrix K n×(m+1)The element in the i-th row and m+1-th column, and p i,m+1 =p i , i∈1,2,…,n, j∈1,2,…,m+1;

[0064] Step 1.5: Set the sample matrix K n×(m+1) Divide into training samples E a×(m+1) and test sample R e×(m+1) , a+e=n;

[0065] Step 2: Construct a least squares generative adversarial network based on multiple generators to generate new samples;

[0066] The least squares generative adversarial network includes: d generative models G and a discriminative model D; the loss function of the least squares generative adversarial network is set as the least squares loss function; the stochastic gradient descent method is set as the optimizer of the least squares generative adversarial network model to find the optimal solution of the model and adjust the model parameters.

[0067] Step 2.1: Use a fully connected neural network to construct d generative models G = [G1, G2, ... G g ,…,G d ], G g represents the g-th generative model, g∈1,2,…,d;

[0068] Step 2.2, constructing the discriminant model D using a fully connected neural network;

[0069] Step 2.3: Input d one-dimensional random noises of length c into the generative model G and obtain pseudo samples Q. G,(c×d)×(m+1) =[Q 1,(c)×(m+1) ,Q 2,(c)×(m+1) ,…,Q g,(c)×(m+1) …,Q d,(c)×(m+1) ], Q g,(c)×(m+1) Indicates that the dimension of the g-th generative model generated is c×(m+1) pseudo samples;

[0070] Step 2.4, fix the parameters of the generative model G, train the discriminative model D; a×(m+1) and pseudo samples Q G,(c×d)×(m+1) The discriminant model D is input for binary classification processing, and d one-dimensional thermal tensors of length c are output; samples marked as "1" in the one-dimensional thermal tensor represent true samples; samples marked as "0" are false samples; based on the one-dimensional thermal tensor, the least squares generative adversarial network model is trained using the stochastic gradient descent method to update the parameters of the discriminant model D, thereby obtaining a trained discriminant model D';

[0071] Step 2.5: Fix the parameters of the trained discriminant model D' and use it to train the generative model G:

[0072] Input one-dimensional random noise of length c into the generative model G and generate new pseudo samples Q' G,(c×d)×(m+1) ; The training sample E a×(m+1) and new pseudo sample Q' G,(c×d)×(m+1) Input into the trained discriminant model D' to determine whether the sample is true or false; according to the heat tensor output by the discriminant model D', the parameters of each generative model are updated using the stochastic gradient descent method to obtain the trained generative model G';

[0073] Step 2.6: Fix the parameters of the trained generative model G' to optimize the trained discriminative model D':

[0074] Input one-dimensional random noise of length c into the trained generative model G' to obtain new pseudo samples;

[0075] The new pseudo sample and the training sample E a×(m+1) Input it into the trained discriminant model D' for optimization training to obtain the optimized discriminant model D";

[0076] Step 2.7: Repeat steps 2.3 to 2.6 until all pseudo samples generated by the discriminant model are marked as "1". The training is completed, and the latest sample Q" is generated by the generative model after the final optimization training. G,(c×d)×(m+1) The generative model G improves the similarity between the generated samples and the real training samples, deceiving the discriminative model D as much as possible. The discriminative model learns to avoid being deceived by fake samples as much as possible. By alternating the training of the generative model and the discriminative model, the parameters of the two models are continuously optimized, thus improving the capabilities of each model.

[0077] Step 2.8: The new sample set Q″ G,(c×d)×(m+1) With training sample E a×(m+1) After merging, we get the new training sample set X s×(m+1) , s = a + (c × d), in order to achieve the purpose of expanding the original sample training set.

[0078] Step 3: Build the DCN network, such as Figure 2 As shown, it includes: embedding and stacking layers at the top, parallel cross neural network layers and deep neural network layers in the middle, and combined output layers at the bottom;

[0079] Step 3.1: New training sample set X s×(m+1) Input into the DCN network and judge the training sample set X s×(m+1) Each meteorological feature is divided into sparse feature R and dense feature Y dense ;

[0080] In the embedding and stacking layers, the sparse feature R is converted into a low-dimensional dense vector R using formula (1). emb Otherwise,

[0081] R emb,r =W emd,r R r (1)

[0082] In formula (1): R r represents the rth sparse feature, Indicates the total number of sparse features; W emd,r Represents the weight parameter when the rth sparse feature is converted into a low-dimensional dense vector; R emb,r Represents the low-dimensional dense vector transformed from the rth sparse feature;

[0083] Step 3.2, embedding and stacking layer uses formula (2) to transform the low-dimensional dense vector R emb With dense features Y dense After splicing, the common input vector λ0 of the cross neural network and the deep neural network is obtained using formula (2):

[0084]

[0085] In formula (2): represents the transpose of the r-th low-dimensional dense vector; represents the transpose of a dense vector;

[0086] Step 3.3, construct the cross neural network layer and the deep neural network layer;

[0087] Step 3.3.1, the cross neural network layer is composed of U cross layers;

[0088] The input vector λ0 is input into the cross neural network for explicit feature cross learning, and the explicit feature cross information λ output by the u-th cross layer is obtained using formula (3): u , so that the cross neural network outputs explicit high-order feature cross information λ U ;

[0089]

[0090] In formula (3), ω u ,b u is the weight and bias of the u-th layer;

[0091] Step 3.3.2, the deep neural network layer consists of H fully connected layers;

[0092] The input vector λ0 is input into the deep neural network layer, and the implicit high-order feature cross information φ output by the hth fully connected layer is obtained using formula (4):h , so that the deep neural network outputs implicit high-order feature cross information φ H ;

[0093] φ h =f(ω h φ h-1 +b h ), h∈1,2,…,h,…,H (4)

[0094] In formula (4): h , b h are the weight and bias of the hth fully connected layer respectively; f(·) is the ReLU function;

[0095] Step 3.4, the combined output layer is composed of a fully connected layer, and the explicit feature cross information λ U and implicit feature cross information φ H After merging, the combined output layer is input, and the predicted value v is obtained using formula (5):

[0096]

[0097] In formula (5), Represents the explicit feature cross information λ U The transposed and implicit feature cross information φ H The transpose of ω dense is the weight vector of the combined output layer; σ(·) represents the ReLU function;

[0098] Step 3.5: The predicted value v output by the DCN network is denormalized to obtain the actual photovoltaic power prediction value v o ;

[0099] Step 4: Whale algorithm optimizes DCN network;

[0100] Step 4.1: Use the whale optimization algorithm to determine the hyperparameters of the DCN network, including: the number of cross network layers U, the number of deep network layers H, the number of neurons θ, and the training period τ;

[0101] Step 4.2: New training sample set X s×(m+1) The data is input into a deep cross-network model with determined hyperparameters for training, thereby obtaining a trained photovoltaic power generation prediction model under sudden weather changes.

[0102] Step 4.3, using the test set R e×(m+1) The trained photovoltaic power generation power prediction model under sudden weather changes is used to predict the power generation of photovoltaic power stations, thereby obtaining the short-term photovoltaic power generation power prediction results under sudden weather changes.

[0103] Traditional methods for predicting photovoltaic power generation under sudden weather conditions often involve finely segmenting sample datasets and selecting different meteorological factors for different weather types to build different network models. However, sudden weather sample data inherently suffers from insufficient sample capacity. Once the data is segmented into different weather types, the sample capacity is significantly reduced. To address this issue, a least-squares generative adversarial network is used to generate new samples and increase the sample capacity.

[0104] In summary, the present invention learns the sample distribution law of the original real photovoltaic data of sudden weather through the least squares generative adversarial network of multiple generators, generates new samples that are similar to the original sample data and have high diversity, and improves the capacity of the sudden weather training set samples; in order to obtain high-order feature cross-information, a deep cross network is used to learn the explicit feature cross-information and the implicit feature cross-information; at the same time, the whale algorithm is used to optimize the deep cross network to obtain the optimal network model, thereby improving the prediction accuracy of the photovoltaic power generation power under sudden weather.

Claims

1. A short-term photovoltaic power prediction method based on LSGAN-DCN network, characterized in that: The steps are as follows: Step 1: Collection and classification of sudden weather sample data; Step 1.1, obtain the original photovoltaic data sequence from the photovoltaic power station, and divide the original photovoltaic data sequence according to the set climate type segmentation point to obtain the photovoltaic data sequence under each climate type; filter the photovoltaic data sequence under each climate type according to the weather type. Mutation-like weather sample data; Step 1.2: Select from each type of sudden weather sample data samples constitute a sample data set, and each sample in the sample data set consists of After normalizing the sample data set, the dimension is obtained. Meteorological characteristic matrix ;in, Represents the normalized The first sample Meteorological characteristics, , ; Step 1.3: Collect and centralize the sample data In the samples The photovoltaic power generation data corresponding to the various meteorological characteristics are normalized to obtain the power series , Represents the normalized The power of the samples; Step 1.4: Transform the meteorological feature matrix With power sequence Perform splicing to obtain the sample matrix ;in, Represents the sample matrix Middle Rank elements of the column, and , , ; Step 1.5: Sample matrix Divide into training samples and test samples , ; Step 2: Construct a least squares generative adversarial network based on multiple generators to generate new samples; The least squares generative adversarial network includes: Generative Model and a discriminant model ; Let the loss function of the least squares generative adversarial network be the least squares loss function; Step 2.1: Use fully connected neural network to construct Generative Model , Indicates the A generative model, ; Step 2.2: Use a fully connected neural network to build the discriminant model ; Step 2.3, The length is One-dimensional random noise is input to the generative model and obtain pseudo samples , Indicates the The dimension generated by the generative model is pseudo samples; Step 2.4: Fix the generative model Parameters of training discriminant model ; The training samples and pseudo samples Input the discriminant model Perform binary classification processing and output The length is One-dimensional thermal tensor; samples marked as "1" in the one-dimensional thermal tensor represent true samples; samples marked as "0" are false samples; based on the one-dimensional thermal tensor, the least squares generative adversarial network model is trained using the stochastic gradient descent method to update the discriminant model Parameters of the trained discriminant model ; Step 2.5: Fix the trained discriminant model Parameters for training the generative model : The length is One-dimensional random noise is input to the generative model and generate new pseudo samples ; The training samples and new pseudo samples Input to the trained discriminant model To judge the authenticity of the sample; according to the discriminant model The output heat tensor uses the stochastic gradient descent method to update the parameters of each generative model to obtain the trained generative model ; Step 2.6: Fix the trained generative model Parameters for the trained discriminant model Perform optimization training: The length is One-dimensional random noise is input to the trained generative model , get new pseudo samples; The new pseudo samples and training samples Input to the trained discriminant model Optimize the training and get the discriminant model of the optimized training ; Step 2.7: Repeat steps 2.3 to 2.6 until all pseudo samples generated by the discriminant model are marked as "1". The training is completed, and the latest samples are generated by the generative model with the final optimized training. ; Step 2.8: The latest sample set With training samples After merging, we get a new training sample set , ; Step 3: Build the DCN network, which includes: the embedding and stacking layers at the top, the parallel cross neural network layer and deep neural network layer in the middle, and the combined output layer at the bottom; Step 3.1: New training sample set Input into the DCN network and judge the new training sample set Each meteorological feature is divided into sparse features and dense features ; In the embedding and stacking layers, the sparse features are transformed into Convert to a low-dimensional dense vector Otherwise, (1) In formula (1): Indicates the sparse features, ; Indicates the total number of sparse features; Indicates the The weight parameter when converting a sparse feature into a low-dimensional dense vector; Indicates the A low-dimensional dense vector transformed from sparse features; Step 3.2: The embedding and stacking layers use formula (2) to transform the low-dimensional dense vector With the dense features After splicing, use formula (2) to obtain the common input vector of the cross neural network and the deep neural network : (2) In formula (2): Indicates the The transpose of a low-dimensional dense vector; represents the transpose of a dense vector; Step 3.3, construct the cross neural network layer and the deep neural network layer; Step 3.3.1, the cross neural network layer is composed of Cross-layer composition; The input vector Input into the cross neural network for explicit feature cross learning, and use formula (3) to obtain the first The explicit feature cross information output by the cross layer , so that the cross neural network outputs explicit high-order feature cross information ; ， (3) In formula (3), For the Layer weights and biases; Step 3.3.2, the deep neural network layer is composed of Fully connected layers; The input vector Input into the deep neural network layer, and use formula (4) to obtain the first The implicit feature cross information output by the fully connected layer , so that the deep neural network outputs implicit high-order feature cross information ; ， (4) In formula (4): , They are The weights and biases of the fully connected layers; is the ReLU function; Step 3.4: The combined output layer is composed of a fully connected layer, and the explicit high-order feature cross information and implicit high-level feature cross-information After merging, the combined output layer is input to obtain the predicted value using formula (5): : (5) In formula (5), , Represent explicit high-order feature cross information The transposition and implicit high-level feature cross information The transpose of is the weight vector of the combined output layer; represents the ReLU function; Step 3.5: The predicted value output by the DCN network After denormalization, the actual photovoltaic power generation power prediction value is obtained ; Step 4: Whale algorithm optimizes DCN network; Step 4.1: Use the whale optimization algorithm to determine the hyperparameters of the DCN network, including: the number of cross-network layers , number of deep network layers , the number of neurons , training cycle ; Step 4.2: New training sample set The data is input into a deep cross-network model with determined hyperparameters for training, thereby obtaining a trained photovoltaic power generation prediction model under sudden weather changes. Step 4.3: Using the test set The trained photovoltaic power generation power prediction model under sudden weather changes is used to predict the power generation of photovoltaic power stations, thereby obtaining the short-term photovoltaic power generation power prediction results under sudden weather changes.

Images