A method for constructing a photovoltaic power prediction model

Abstract

The present invention belongs to the field of power prediction of photovoltaic clusters in microgrids, and specifically relates to a method for constructing a photovoltaic power prediction model, comprising: constructing a training sample set; constructing RBLS using the training sample set with power labels; constructing a residual prediction network: respectively searching for the most similar sample pair X to each labeled training sample X from the unlabeled training sample set and the first q labeled training samples adjacent to it in time, inputting them into the RBLS to obtain a first predicted power sequence, respectively inputting the pair X and the first q unlabeled data samples adjacent to it in time into the RBLS to obtain a second predicted power sequence, respectively inputting the two predicted power sequences into the residual prediction network, and outputting a residual a. o , with a o is equal to the true residual and a o The residual prediction network parameters are optimized with the goal of being equal to , and the resulting RBLS and residual prediction network are constructed as a photovoltaic power prediction model. The present invention can improve power prediction accuracy while ensuring the lightweight model.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of micro-grid photovoltaic cluster power generation power prediction, and more particularly to a photovoltaic power prediction model construction method. BACKGROUND

[0002] With the increasing demand for sustainable energy worldwide, photovoltaic cluster power generation as a clean and renewable energy form has gradually increased its penetration rate in the power grid. However, the accuracy of photovoltaic cluster power prediction is affected by various factors such as weather conditions, equipment performance, etc., making photovoltaic data exhibit complex and variable characteristics. These time series data not only have multi-dimensional characteristics, but also have significant randomness and volatility, posing a considerable challenge to accurate photovoltaic power prediction. On the other hand, in reality, there is only a small amount of data with power value labels, and few studies use massive unlabeled data to optimize the effect of the prediction model.

[0003] Typical photovoltaic power prediction methods include physical models, mathematical statistics, and artificial intelligence. Physical models construct a mapping relationship with power through information such as the geographical location of the power station, photovoltaic panel characteristics parameters, and weather forecast data, but it is difficult to model in photovoltaic cluster scenarios and complex systems. Mathematical statistics mainly includes linear regression and time series, which utilizes a large amount of historical data containing a lot of redundant parts, and the model is simple but suitable for situations where precision is not required. In the artificial intelligence method, vector machine-based prediction usually needs to classify the original data by weather, and then train model parameters under different categories to improve accuracy. The entire process is relatively complex, especially when applied to photovoltaic cluster scenarios, and there are certain difficulties in achieving high-precision prediction. Deep learning-based methods require high computing resources for training devices, however, mobile computing devices in the photovoltaic cluster power prediction scenario are usually difficult to meet the requirements of deep large models. SUMMARY

[0004] In view of the defects and improvement needs of the prior art, the present application provides a photovoltaic power prediction model construction method, which aims to improve the photovoltaic power prediction accuracy while ensuring the lightweight of the model.

[0005] To achieve the above-mentioned purpose, according to one aspect of the present application, a photovoltaic power prediction model construction method is provided, comprising:

[0006] Constructing a training sample set, the dimension of each training sample is N*w, N is the feature dimension of photovoltaic time series data, and w is the time width;

[0007] Using the training sample set with power labels to construct RBLS as a power preliminary prediction module;

[0008] Residual prediction network is constructed: find the most similar training sample to each active power labeled training sample X from the training sample set without power label Input the X and the training samples with the previous q active power labels into RBLS respectively to obtain a first predicted power sequence, input the X and the data samples with the previous q active power labels into RBLS respectively to obtain a second predicted power sequence, input the first predicted power sequence into the residual prediction network to output the residual a of the predicted power of the X , and input the second predicted power sequence into the residual prediction network to output the residual of the predicted power of the X o . Take a o equal to the real residual and a o equal to as the target to optimize the residual prediction network parameters.

[0009] The constructed RBLS and residual prediction network are used as the photovoltaic power prediction model to complete the construction, wherein the output of the residual prediction network is used to compensate for the power prediction error of the RBLS.

[0010] Further, the residual prediction network is a fully connected neural network.

[0011] Further, the mean square error L MSE is used to represent the deviation between a o and the real residual, and the square error L SQR is used to represent the deviation between a o and .

[0012] Further, the most similar training sample is found by the Euclidean distance between samples

[0013] Further, the implementation of the RBLS is as follows:

[0014] For each active power labeled training sample X, learn the N-dimensional feature at the current time and the time sequence feature at the previous time according to the time sequence from front to back to obtain the data time sequence feature in the data sample before the current time as the time sequence feature at the current time, and the initial time sequence feature is preset as 0; meanwhile, learn the word importance of the N-dimensional feature at each time in the data sample; take the time sequence feature Z w ​and the word importance of each moment N-dimensional feature is spliced to form a feature vector; through ridge regression, the output weight of the RBLS power output layer corresponding to the time sequence feature of the last moment in the data sample with a dimension of N*w is calculated according to the power label value corresponding to each X and the feature vector, and the output weight of the RBLS power output layer corresponding to the word importance of each moment is obtained, and the RBLS is obtained.

[0015] Further, the cumulative learning is performed in the following manner:

[0016] Z p =f(X p α z +Z p-1 U+β z )

[0017] In the formula, Z p is the data time sequence feature corresponding to the current moment p, f(·) is an activation function, X p is the N-dimensional feature of the current moment p, Z p-1 is the data time sequence feature corresponding to the last moment p-1, alpha z and U are respectively randomly generated weight matrices, beta z is a bias.

[0018] Further, the word importance learning is performed in the following manner:

[0019] H p =g(X p alpha h +beta h )

[0020] In the formula, H p is the word importance of the N-dimensional feature of the moment p, X p is the N-dimensional feature of the moment p, g(·) is an activation function, alpha h and beta h are respectively randomly generated weight matrices and a bias.

[0021] Further, the training sample set is constructed in the following manner: the normalized photovoltaic time sequence data containing N-dimensional features are cut and divided along the time axis direction with a preset step length t with a window shape of N*w, to obtain the training sample set; wherein w is the time width of each window.

[0022] The application also provides a photovoltaic power prediction method, which uses the photovoltaic power prediction model constructed by the construction method to perform photovoltaic power prediction.

[0023] The application also provides a computer readable storage medium comprising a stored computer program, wherein the computer program, when executed by a processor, controls a device in which the storage medium is located to perform the steps of the method described above.

[0024] Overall, the above technical solutions conceived by the present application can achieve the following beneficial effects:

[0025] (1) The method for constructing a photovoltaic power prediction model proposed by the present application first constructs an RBLS as a preliminary power prediction module, and then constructs a residual prediction network to compensate for the power prediction error of the RBLS, which can significantly improve the accuracy of power prediction, and the model composed of the RBLS and the residual prediction network has low complexity, so the method of the present application can improve the photovoltaic power prediction accuracy while ensuring the lightweight of the model. Among them, the RBLS can effectively extract time series features, and the residual prediction network is introduced to predict the output residual of the RBLS, further improving the prediction accuracy. When training the residual prediction network, a o The goal is to make the overall prediction result closer to the true value. In addition, when training the residual prediction network, the present application adopts a semi-supervised learning method, that is, from the training sample set without power label, find the training sample When optimizing the residual network parameters, a o The goal is to make the overall prediction result closer to the true value. In addition, when training the residual prediction network, the present application adopts a semi-supervised learning method, that is, from the training sample set without power label, find the training sample The goal is to make the overall prediction result closer to the true value. In addition, when training the residual prediction network, the present application adopts a semi-supervised learning method, that is, from the training sample set without power label, find the training sample BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 A flow chart of a method for constructing a photovoltaic power prediction model is provided for the embodiments of the present application.

[0027] Figure 2 A flow chart of a method for constructing a photovoltaic power prediction model is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solutions and advantages of the present application clearer and more apparent, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0029] Embodiment one

[0030] A method for constructing a photovoltaic power prediction model, as shown in Figure 1 , comprising:

[0031] Construct a training sample set, each training sample has a dimension of N*w, N is the feature dimension of photovoltaic time series data, and w is the time width;

[0032] Construct RBLS using the training sample set with active power labels as a preliminary power prediction module;

[0033] Construct a residual prediction network: find the most similar training sample to each active power labeled training sample X from the training sample set without active power labels , respectively, and input RBLS to obtain a first predicted power sequence, and input and the previous q training samples without active power labels to RBLS to obtain a second predicted power sequence, input the first predicted power sequence into the residual prediction network, and output the residual a o of the predicted power of the X , and input the second predicted power sequence into the residual prediction network to output the residual o of the predicted power of the X o . , and optimize the residual prediction network parameters with the goal of a o equal to the real residual and a o equal to

[0034] Use the constructed RBLS and residual prediction network as a photovoltaic power prediction model to complete the construction, wherein the output of the residual prediction network is used to compensate for the power prediction error of RBLS.

[0035] The method for constructing a photovoltaic power prediction model proposed in this embodiment first constructs RBLS as a preliminary power prediction module, and then constructs a residual prediction network to compensate for the power prediction error of RBLS, which can significantly improve the accuracy of power prediction. The model composed of RBLS and the residual prediction network has low complexity, so the method of this embodiment can improve the photovoltaic power prediction accuracy while ensuring the lightweight of the model. Among them, RBLS (width learning model) has high learning efficiency because it only needs to learn and calculate the output weight matrix, and RBLS can effectively extract time series features. Further, the residual prediction network is introduced to predict the output residual of RBLS, which further improves the prediction accuracy. When training the residual prediction network, the goal is to make a o equal to the real residual, so that the overall prediction result is closer to the true value.

[0036] In addition, the application adopts a semi-supervised learning method when training the residual prediction network, that is, finding the training sample most similar to each active power-labeled training sample X from the training sample set without power label In the optimization of the residual network parameters, a o With As the goal, so that the model can have high prediction accuracy when encountering data samples that are not completely consistent with the training samples. Therefore, the training method improves the generalization ability of the entire photovoltaic power prediction model, that is, has strong generalization performance. Specifically, due to the introduction of unlabeled data, the model can access more data distribution and feature space, thereby enhancing its adaptability to new data. This means that in actual application, the model can more accurately predict photovoltaic power under various conditions, providing strong support for monitoring and analysis of photovoltaic power stations. In summary, in the construction of the photovoltaic power prediction model, the active perception method is further introduced on the basis of the introduction of the residual prediction network, and samples are selected from the unlabeled data as random enhancement samples of the labeled data, so that the most informative data is actively selected for learning, thereby improving the prediction performance.

[0037] In terms of data utilization efficiency, the embodiment fully utilizes the advantages of semi-supervised learning. Traditional supervised learning requires a large amount of labeled data for training, while the application combines labeled data and unlabeled data, and uses unlabeled data as random enhancement samples of the labeled data to be selected, thereby effectively reducing the demand for training data. This innovation not only reduces the difficulty and cost of data collection, but also greatly shortens the training time and improves the training efficiency of the model.

[0038] Therefore, the embodiment is a semi-supervised learning active perception power prediction method based on time series width learning and conventional prediction network stacking applied to a microgrid photovoltaic cluster power station monitoring and analysis system. The RBLS width learning model is used to extract time series features, the residual prediction network is used to predict the residual of the RBLS, and the random enhancement samples of the labeled data are selected from the unlabeled data and input into the semi-supervised model to improve the prediction accuracy and algorithm generalization.

[0039] Preferably, the training sample set is constructed by: sliding and dividing the normalized photovoltaic time series data containing N-dimensional features with a window shape of N*w along the time axis direction with a preset step size t, to obtain the training sample set; wherein w is the time width of each window.

[0040] Specifically, for photovoltaic data with N-dimensional features (temperature, humidity, irradiance, etc.), in order to improve the efficiency of the subsequent training model, the maximum value D max and the minimum value D min, the one-dimensional feature data is standardized to the interval [0, 1] by the following formula:

[0041]

[0042] After the standardized processing of all dimension feature data, the window width w and the sliding step t are set, and the window with the shape of N*w is obtained, which is slid along the time axis direction on the data set with the step t, and the data sample set cut out is represented as X∈R N·w .

[0043] In one embodiment, the implementation of the above RBLS construction is as follows:

[0044] For each training sample X with a power label, the N-dimensional features at the current time and the time sequence features at the previous time are learned in the forward time order, the data time sequence features before the current time located in the data sample are obtained as the time sequence features at the current time, and the initial time sequence features are preset to 0; at the same time, the word importance of the N-dimensional features at each time in the data sample is learned; the time sequence feature Z w at the wth time corresponding to each X and the word importance of the N-dimensional features at each time are spliced to form a feature vector; through the way of ridge regression, the output weight of the time sequence feature at the last time in the data sample with the dimension of N*w mapped to the RBLS power output layer and the output weight of the RBLS power output layer obtained by mapping the word importance corresponding to each time are calculated according to the power label value and the feature vector corresponding to each X, and the RBLS is obtained.

[0045] Preferably, the cumulative learning is performed in the following way:

[0046] Z p =f(X p α z +Z p-1 U+β z )

[0047] In the formula, Z p is the data time sequence feature corresponding to the current time p, f(·) is an activation function, X p is the N-dimensional feature at the current time p, Z p-1 is the data time sequence feature corresponding to the previous time p-1, α z and U are respectively randomly generated weight matrices, and β z is a bias.

[0048] Preferably, the word importance learning is performed in the following way:

[0049] H p =g(X p α h +β h )

[0050] Where H p is the word importance of the N-dimensional feature at time p, X p is the N-dimensional feature at time p, g(·) is the activation function, α h and β h are randomly generated weight matrices and biases.

[0051] Specifically, RBLS (recurrent broad learning system) needs to learn two features, "temporal features" and "word importance", separately, and then combine the two features through ridge regression to obtain the weight matrix mapped to the output layer.

[0052] Step 1: Learn time series features. The vector composed of all features at the pth moment can be expressed as X p ∈R 1×N , let X p Z with the same dimensions p Represents the currently accumulated learned time series information. Z0 is a zero vector, so calculate Z p Not only the current input X is required p , we also need the time series feature Z of the previous moment p-1 , the specific calculation formula is:

[0053] Z p =f(X p α z +Z p-1 U+β z )

[0054] Among them, f represents the activation function such as sigmoid, α z and U is a randomly generated weight matrix, β z Represents the bias. After w iterations, Z w Contains all the temporal features learned from w moments, Z w The calculation formula is:

[0055] Z w =f(X w α z +Z w-1 U+β z )

[0056] Z w Through the weight matrix W z Connect to output Y, W z It reflects the impact of timing characteristics on the output.

[0057] Step 2: Learn the importance of words. Take time p as an example, andp As input, word importance H p The calculation method of p = 1, 2,..., w is as follows:

[0058] H p = g(X p α h + β h )

[0059] Wherein, g is an activation function, α h and β h are randomly generated weights and bias, and the obtained H p The dimension is N x m. H p , p = 1, 2,..., w is connected to the output Y through the weight matrix W h Reflects the proportion of the influence of each word on the output.

[0060] Step 3: Calculate the weight matrix W.

[0061] The weight matrix W = [W z |W h ] is the weight matrix connecting the features and the output layer. By solving this matrix through ridge regression, the influence of the time sequence feature Z w And the word importance [H1, H2,..., H w ] on the output can be learned at the same time, and the calculation formula is as follows:

[0062] W = A + Y

[0063] A + = (A T A) -1 A T

[0064] A = [Z k |H1,...,H w ]

[0065] After training the RBLS, the residual error of the RBLS is predicted through a fully connected neural network (which can be preferred) to further reduce the error. The number of input layer neurons of the fully connected layer is equal to the number of inputs q, and the number of output layer neurons is 1 to realize the regression task.

[0066] Taking a labeled sample X ∈ R N·w As an example, first find the most similar sample X ~ from the unlabeled sample, and the standard of similarity is the Euclidean distance:

[0067]

[0068] X and the previous q-1 samples, and the previous q-1 samples are input into the RBLS to obtain a predicted power value sequence [y T-q ,y T-q+1 ,...,y T ] and [y T-q ,y T-q+1 ,...,y T ] are input into the full connection neural network, and the expected output of the full connection neural network is X, and the expected output of the latter is [y T-q ,y T-q+1 ,...,y T ]. Therefore, as preferred, the loss value L of the full connection neural network can be divided into the calculation of the mean square error L MSE and the square error L SQR .

[0069] First step: calculate the mean square error. The input a l vector of the previous layer is passed to the calculation process of the next layer as follows:

[0070] a l =f(a l-1 w l +b l )

[0071] where w l and b l are the weight matrix and bias between the l-1 and l layers, and f is the activation function. Since the last layer has only one neuron, the mean square error loss

[0072] L MSE =(y-a o ) 2

[0073] where y is the actual residual error of X after being predicted by the RBLS, and a o is the output of the full connection neural network.

[0074] Second step: calculate the square error. The power value sequence [y T-q ,y T-q+1 ,...,y T ] is output by the full connection neural network as a o , and the corresponding output is Since is a random augmented sample of X and the full connection neural network has dropout, a o and are not equal with high probability. The closer the two values are, the stronger the model generalization is, and the square error is calculated as

[0075]

[0076] Third step: fusion loss value. The mean square error L MSE and square difference L SQR According to a certain proportion, the loss L of the full connection neural network is fused, and the calculation method is

[0077] L=λL MSE +μL SQR

[0078] Then, the parameters are updated by the back propagation algorithm, and the first step to the third step are repeated until the error is less than the threshold or the set number of cycles is reached.

[0079] The flow chart can be seen Figure 2 In general, the embodiment proposes to apply the semi-supervised learning active perception algorithm of stacking time series width learning and residual prediction network to photovoltaic cluster power generation power prediction. First, the time series features of photovoltaic data are extracted by RBLS, then the full connection neural network is trained by using the semi-supervised learning method to predict the RBLS residual, and the random enhanced sample of labeled data is found in the unlabeled sample by adding active perception, which improves the prediction accuracy and algorithm generalization ability.

[0080] Embodiment two

[0081] A photovoltaic power prediction method, using the photovoltaic power prediction model constructed by the construction method in embodiment one to predict photovoltaic power.

[0082] Power prediction is to predict the power at a certain time in the future by using historical photovoltaic data. The related technical solution is the same as that in embodiment one, and will not be repeated here.

[0083] Embodiment three

[0084] The application also relates to an electronic device, which comprises a memory and a processor, the memory stores a computer program, and the processor realizes the steps of the above method when executing the computer program.

[0085] The electronic device can be a desktop computer, a notebook computer, a palm computer, a cloud server, and the like. The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic, discrete hardware components, and the like. The memory can be used to store computer programs and / or modules, and the processor can implement various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by calling data stored in the memory.

[0086] The related technical solutions are the same as those of Embodiment One, and will not be described herein.

[0087] Embodiment Four

[0088] The present application also relates to a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the above method.

[0089] Specifically, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state memory devices.

[0090] The related technical solutions are the same as those of Embodiment One, and will not be described herein.

[0091] Those skilled in the art can easily understand that the above description is only a preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, and improvement within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for constructing a photovoltaic power prediction model, characterized in that: include: Construct a training sample set, the dimension of each training sample is , N is the characteristic dimension of photovoltaic time series data, is the time width; The training sample set with power labels is used to construct RBLS as the preliminary power prediction module; Construct a residual prediction network: Find the training samples with power labels from the training sample set without power labels respectively. The most similar training sample , for this And the first q training samples with power labels adjacent to them in time are input into RBLS respectively to obtain the first predicted power sequence. The first q power-labeled data samples adjacent to it in time are input into RBLS to obtain the second predicted power sequence, and the first predicted power sequence is input into the residual prediction network to output the The residual of the predicted power , input the second prediction power sequence into the residual prediction network, and output the The residual of the predicted power ,by is equal to the true residual and and Optimizing the residual prediction network parameters with equality as the goal; The constructed RBLS and residual prediction network are used as the photovoltaic power prediction model to complete the construction, wherein the output of the residual prediction network is used to compensate for the RBLS power prediction error; The implementation of building RBLS is as follows: For each training sample with power label , according to the time sequence from the front to the back, learn the N-dimensional features of the current moment and the time series features of the previous moment, and obtain the time series features of the data in the data sample before the current moment as the time series features of the current moment. The initial time series features are preset to 0; at the same time, learn the word importance of the N-dimensional features at each moment in the data sample; The corresponding Temporal characteristics of moments And the word importance of N-dimensional features at each moment is concatenated to form a feature vector; through ridge regression, according to each The corresponding power label value and feature vector are calculated in the dimension of The time series feature corresponding to the last moment in the data sample is mapped to the output weight of the RBLS power output layer and the word importance mapping corresponding to each moment is used to obtain the output weight of the RBLS power output layer, and the RBLS is obtained.

2. The construction method according to claim 1, characterized in that The residual prediction network is a fully connected neural network.

3. The construction method according to claim 1, characterized in that Mean square error express The deviation from the true residual, using squared error express and The deviation between.

4. The construction method according to claim 1, wherein Find the most similar training samples by using the Euclidean distance between samples .

5. The construction method according to claim 1, characterized in that The following method is used to learn the N-dimensional features of the current moment and the time series features of the previous moment: Where, For the current moment p The corresponding data timing characteristics, is the activation function, For the current moment p N-dimensional features, is the data time series feature corresponding to the previous moment p-1, and are randomly generated weight matrices, For bias.

6. The construction method according to claim 1, characterized in that The following method is used to learn word importance: Where, For the moment p The word importance of N-dimensional features, For the moment p N-dimensional features, is the activation function, and are randomly generated weight matrices and biases.

7. The construction method according to any one of claims 1 to 5, characterized in that: The training sample set is constructed by: normalizing the photovoltaic time series data containing N-dimensional features with the shape The window is arranged along the time axis with a preset step size t Perform sliding segmentation to obtain a training sample set; among them, is the time width of each window.

8. A photovoltaic power prediction method, characterized in that: Photovoltaic power prediction is performed using the photovoltaic power prediction model constructed by the construction method according to any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed by a processor, the device where the storage medium is located is controlled to perform the steps of the method according to any one of claims 1 to 7.

Images