Photovoltaic power prediction method, device, feeder terminal, system, medium and chip

Abstract

The present disclosure relates to the field of intelligent power distribution, and specifically discloses a photovoltaic power prediction method, device, feeder terminal, system, medium and chip, the method comprising: obtaining the true value of photovoltaic power generation in a sample time period and multiple types of feature data corresponding to the sample time period that affect photovoltaic power generation as a sample data set; using a random forest algorithm to determine the important types that affect photovoltaic power generation from multiple types; inputting the feature data of important types corresponding to multiple historical time periods before the prediction time period into a pre-trained convolutional neural network CNN model to obtain the feature vector output by the CNN model; inputting the feature vector output by the CNN model into a pre-trained RVM model to obtain the photovoltaic power generation in the prediction time period output by the RVM model. This technical solution can accurately predict photovoltaic power generation, and is mainly used for photovoltaic equipment management, photovoltaic power generation grid-connected regulation, and efficient linkage of source, grid, load and storage.

Description

Technical Field

[0001] The present disclosure relates to the field of intelligent power distribution technology, and in particular to a photovoltaic power prediction method, device, feeder terminal, system, medium and chip. Background Art

[0002] Solar energy is one of the most important clean energy sources in human society. Photovoltaic power generation converts solar energy into electricity, which can be consumed locally or connected to the power grid and sent to distant places. However, due to the great randomness and volatility of photovoltaic power generation, it poses great challenges to grid connection and local consumption. In order to harness such energy, there is an urgent need for a solution that can accurately and efficiently predict photovoltaic power generation. Summary of the invention

[0003] In order to solve the problems in the related art, the embodiments of the present disclosure provide a photovoltaic power prediction method, device, feeder terminal, system, medium and chip.

[0004] In a first aspect, an embodiment of the present disclosure provides a photovoltaic power prediction method, the method being applied to a feeder terminal, the method comprising:

[0005] Acquire a sample data set, the sample data set including a true value of photovoltaic power generation in a sample time period and multiple types of characteristic data affecting the photovoltaic power generation corresponding to the sample time period;

[0006] Based on the sample data set, a random forest algorithm is used to determine the important types that affect the photovoltaic power generation power from the multiple types;

[0007] Obtain important types of feature data corresponding to multiple historical time periods before the forecast time period;

[0008] Inputting important types of feature data corresponding to multiple historical time periods before the forecast time period into a pre-trained convolutional neural network (CNN) model to obtain a feature vector output by the CNN model;

[0009] The feature vector output by the CNN model is input into a pre-trained RVM model to obtain the photovoltaic power generation power in the predicted time period output by the RVM model.

[0010] In a possible implementation manner, the multiple types of characteristic data affecting the photovoltaic power generation power corresponding to the sample time period include at least two types of characteristic data as follows:

[0011] PV power generation, temperature, air pressure, humidity, PV module temperature, total radiation, direct radiation, and diffuse radiation for multiple time periods before the sample time period.

[0012] In a possible implementation, the acquiring the sample data set includes:

[0013] Acquire an original data set, the original data set including an original photovoltaic power generation power in a sample time period and multiple types of original feature data affecting the photovoltaic power generation power corresponding to the sample time period;

[0014] The data in the original data set is preprocessed to obtain the sample data set, wherein the preprocessing includes at least one of outlier processing, missing value processing, and normalization processing.

[0015] In a possible implementation, the method further includes:

[0016] Acquire a training sample from the sample data set, the training sample including a true value of photovoltaic power generation in a sample time period and important type of feature data affecting the photovoltaic power generation corresponding to the sample time period;

[0017] The training samples are used to train a CNN-RVM model, wherein the CNN-RVM model includes the CNN model and the RVM model.

[0018] In a possible implementation, the using the training samples to train a CNN-RVM model includes:

[0019] The grid search method is used to tune the three parameters of the number of neurons, convolution kernel size and learning rate in the pre-trained CNN-RVM model;

[0020] The mean square error is used as the optimization target of the CNN-RVM model, and the formula of the mean square error is as follows:

[0021]

[0022] Where T represents the number of samples in the training sample, y i represents the photovoltaic power prediction value corresponding to the i-th sample output by the CNN-RVM model, represents the true value of the photovoltaic power generation power in the i-th sample.

[0023] In a possible implementation, the initial settings of the model parameters in the CNN model include: the first convolution layer is initially set to 64 neurons, the second convolution layer is initially set to 32 neurons, the third convolution layer is initially set to 16 neurons, the window size of each convolution layer is initially set to 1×3, the pooling layer connected to each convolution layer is a maximum pooling layer, the pooling layer window size is initially set to 2×2, the step size is initially set to 2, and the activation function is a linear rectification function;

[0024] The penalty coefficient γ of the RVM model is initially set to 0.1, and the kernel function of the RVM model is a Gaussian function;

[0025] The batch size of the CNN-RVM model is initially set to 120, the initial learning rate is initially set to 0.01, and the initial number of training times is initially set to 500.

[0026] In a possible implementation, the method further includes:

[0027] Acquire samples other than the training samples from the sample data set as test samples;

[0028] Using the test sample, testing the CNN-RVM model trained using the training sample;

[0029] If the test passes, the trained CNN-RVM model is the trained CNN-RVM model;

[0030] If the test fails, continue to use the CNN-RVM model trained with the training samples.

[0031] In a second aspect, an embodiment of the present disclosure provides a method for training a photovoltaic power prediction model, the method comprising:

[0032] Acquire a sample data set, the sample data set including a true value of photovoltaic power generation in a sample time period and multiple types of characteristic data affecting the photovoltaic power generation corresponding to the sample time period;

[0033] Based on the sample data set, a random forest algorithm is used to determine the important types that affect the photovoltaic power generation power from the multiple types;

[0034] Acquire a training sample from the sample data set, the training sample including a true value of photovoltaic power generation in a sample time period and important type of feature data affecting the photovoltaic power generation corresponding to the sample time period;

[0035] The training samples are used to train a CNN-RVM model, wherein the CNN-RVM model includes a CNN model and an RVM model, wherein the CNN model is used to extract feature vectors from important types of feature data corresponding to a prediction time period, and the RVM model is used to predict the photovoltaic power generation power of the prediction time period based on the feature vectors extracted by the CNN model.

[0036] In a possible implementation, the using the training samples to train a CNN-RVM model includes:

[0037] The grid search method is used to optimize the three parameters of the number of neurons, convolution kernel size and learning rate in the pre-training CNN-RVM model, and the mean square error is used as the optimization target of the CNN-RVM model. The formula of the mean square error is as follows:

[0038]

[0039] Where T represents the number of samples in the training sample, y i represents the photovoltaic power prediction value corresponding to the i-th sample output by the CNN-RVM model, represents the true value of the photovoltaic power generation power in the i-th sample.

[0040] In a third aspect, an embodiment of the present disclosure provides a photovoltaic power prediction device, which is applied to a feeder terminal and includes:

[0041] A sample acquisition module is configured to acquire a sample data set, wherein the sample data set includes a true value of photovoltaic power generation in a sample time period and multiple types of characteristic data affecting the photovoltaic power generation corresponding to the sample time period;

[0042] A data type determination module is configured to determine, based on the sample data set, an important type that affects the photovoltaic power generation power from the multiple types using a random forest algorithm;

[0043] A data acquisition module is configured to acquire important types of feature data corresponding to a plurality of historical time periods before the prediction time period;

[0044] A feature extraction module is configured to input feature data of important types corresponding to multiple historical time periods before the predicted time period into a pre-trained convolutional neural network (CNN) model to obtain a feature vector output by the CNN model;

[0045] The prediction module is configured to input the feature vector output by the CNN model into a pre-trained RVM model to obtain the photovoltaic power generation power in the predicted time period output by the RVM model.

[0046] In a possible implementation manner, the multiple types of characteristic data affecting the photovoltaic power generation power corresponding to the sample time period include at least two types of characteristic data as follows:

[0047] PV power generation, temperature, air pressure, humidity, PV module temperature, total radiation, direct radiation, and diffuse radiation for multiple time periods before the sample time period.

[0048] In a possible implementation, the sample acquisition module is configured as follows:

[0049] Acquire an original data set, wherein the original data set includes the original photovoltaic power generation power of a sample time period and various types of original feature data corresponding to the sample time period that affect the photovoltaic power generation power; preprocess the data in the original data set to obtain the sample data set, wherein the preprocessing includes at least one of outlier processing, missing value processing, and normalization processing.

[0050] In a possible implementation, the device further includes:

[0051] The model training module is configured to obtain training samples from the sample data set, wherein the training samples include the true value of the photovoltaic power generation power in the sample time period and the important type of feature data that affects the photovoltaic power generation power corresponding to the sample time period; and use the training samples to train a CNN-RVM model, wherein the CNN-RVM model includes the CNN model and the RVM model.

[0052] In a possible implementation manner, the model training module uses the training sample to train a portion of the CNN-RVM model, which is configured as follows:

[0053] The grid search method is used to tune the three parameters of the number of neurons, convolution kernel size and learning rate in the pre-trained CNN-RVM model;

[0054] The mean square error is used as the optimization target of the CNN-RVM model, and the formula of the mean square error is as follows:

[0055]

[0056] Where T represents the number of samples in the training sample, y i represents the photovoltaic power prediction value corresponding to the i-th sample output by the CNN-RVM model, represents the true value of the photovoltaic power generation power in the i-th sample.

[0057] The initial settings of the model parameters in the CNN model include: the first convolution layer is initially set to 64 neurons, the second convolution layer is initially set to 32 neurons, the third convolution layer is initially set to 16 neurons, the window size of each convolution layer is initially set to 1×3, the pooling layer connected to each convolution layer is a maximum pooling layer, the window size of the pooling layer is initially set to 2×2, the step size is initially set to 2, and the activation function is a linear rectification function;

[0058] The penalty coefficient γ of the RVM model is initially set to 0.1, and the kernel function of the RVM model is a Gaussian function;

[0059] The batch size of the CNN-RVM model is initially set to 120, the initial learning rate is initially set to 0.01, and the initial number of training times is initially set to 500.

[0060] In a possible implementation, the device further includes:

[0061] The testing module is configured to obtain samples other than the training samples from the sample data set as test samples; use the test samples to test the CNN-RVM model trained using the training samples; if the test passes, the trained CNN-RVM model is the trained CNN-RVM model; if the test fails, continue to use the CNN-RVM model trained using the training samples.

[0062] In a fourth aspect, an embodiment of the present disclosure provides a training device for a photovoltaic power prediction model, the device comprising:

[0063] A sample data acquisition module is configured to acquire a sample data set, wherein the sample data set includes a true value of photovoltaic power generation in a sample time period and multiple types of characteristic data affecting the photovoltaic power generation corresponding to the sample time period;

[0064] An important type determination module is configured to determine the important type that affects the photovoltaic power generation power from the multiple types using a random forest algorithm based on the sample data set;

[0065] A training sample acquisition module is configured to acquire training samples from the sample data set, wherein the training samples include a true value of photovoltaic power generation in a sample time period and important type of feature data that affects the photovoltaic power generation corresponding to the sample time period;

[0066] A training module is configured to use the training samples to train a CNN-RVM model, wherein the CNN-RVM model includes a CNN model and an RVM model, wherein the CNN model is used to extract feature vectors from important types of feature data corresponding to a prediction time period, and the RVM model is used to predict the photovoltaic power generation power of the prediction time period based on the feature vectors extracted by the CNN model.

[0067] In a possible implementation, the training module is configured as follows:

[0068] The grid search method is used to optimize the three parameters of the number of neurons, convolution kernel size and learning rate in the pre-training CNN-RVM model, and the mean square error is used as the optimization target of the CNN-RVM model. The formula of the mean square error is as follows:

[0069]

[0070] Where T represents the number of samples in the training sample, y i represents the photovoltaic power prediction value corresponding to the i-th sample output by the CNN-RVM model, represents the true value of the photovoltaic power generation power in the i-th sample.

[0071] In a fifth aspect, an embodiment of the present disclosure provides an electronic device, comprising a memory and a processor, wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the method described in any one of the first aspect or the second aspect.

[0072] In a sixth aspect, an embodiment of the present disclosure provides a system, comprising the above-mentioned electronic device.

[0073] In a seventh aspect, an embodiment of the present disclosure provides a readable storage medium on which computer instructions are stored, and the computer instructions implement the above-mentioned method steps when executed by a processor.

[0074] In an eighth aspect, a chip is provided in an embodiment of the present disclosure, and the chip includes the device as described in any one of the third aspect or the fourth aspect above.

[0075] According to the technical solution provided by the embodiment of the present disclosure, after obtaining a sample data set, a random forest algorithm can be used based on the sample data set to determine the important types that affect the photovoltaic power generation power from the multiple types involved in the sample data set, thereby screening out several important types of feature data that will have a greater impact on the photovoltaic power generation power. Then, the feature data of the important types corresponding to multiple historical time periods before the prediction time period are input into a pre-trained convolutional neural network CNN model to obtain a feature vector output by the CNN model; the feature vector output by the CNN model is input into a pre-trained RVM model to obtain the photovoltaic power generation power of the prediction time period output by the RVM model. In this way, the combination of the random forest method and the CNN-RVM prediction model can predict the photovoltaic power generation power more accurately and in real time. In addition, the photovoltaic power prediction function is implemented at the feeder terminal instead of the master station, which can avoid resource congestion at the master station and waste of resources on the terminal side.

[0076] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Other features, objectives and advantages of the present disclosure will become more apparent through the following detailed description of non-limiting embodiments in conjunction with the accompanying drawings. In the accompanying drawings:

[0078] Figure 1 A flow chart of a photovoltaic power prediction method according to an embodiment of the present disclosure is shown.

[0079] Figure 2A A schematic diagram of a CNN-RVM model according to an embodiment of the present disclosure is shown.

[0080] Figure 2B A flow chart showing a method for training a photovoltaic power prediction model according to an embodiment of the present disclosure is shown.

[0081] Figure 3 A structural block diagram of a photovoltaic power prediction device according to an embodiment of the present disclosure is shown.

[0082] Figure 4 A structural block diagram of a photovoltaic power prediction model training device according to an embodiment of the present disclosure is shown.

[0083] Figure 5 A structural block diagram of a feeder terminal according to an embodiment of the present disclosure is shown.

[0084] Figure 6 A schematic diagram showing the structure of a computer system suitable for implementing the method of the embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0085] Hereinafter, exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily implement them. In addition, for the sake of clarity, parts not related to the description of the exemplary embodiments are omitted in the accompanying drawings.

[0086] In the present disclosure, it should be understood that terms such as "include" or "have" are intended to indicate the presence of features, numbers, steps, behaviors, components, parts, or a combination thereof disclosed in the present specification, and are not intended to exclude the possibility that one or more other features, numbers, steps, behaviors, components, parts, or a combination thereof exist or are added.

[0087] It should also be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present disclosure may be combined with each other. The present disclosure will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0088] Figure 1 FIG. 1 is a flow chart showing a photovoltaic power prediction method according to an embodiment of the present disclosure. Figure 1 As shown, the photovoltaic power prediction method includes the following steps S101-S105:

[0089] In step S101, a sample data set is obtained, wherein the sample data set includes a true value of photovoltaic power generation in a sample time period and multiple types of characteristic data affecting the photovoltaic power generation corresponding to the sample time period;

[0090] In step S102, based on the sample data set, a random forest algorithm is used to determine important types that affect the photovoltaic power generation power from the multiple types;

[0091] In step S103, important types of feature data corresponding to the historical time period before the prediction time period are obtained;

[0092] In step S104, the feature data of important types corresponding to the historical time period before the predicted time period is input into a pre-trained convolutional neural network (CNN) model to obtain a feature vector output by the CNN model;

[0093] In step S105, the feature vector output by the CNN model is input into a pre-trained RVM model to obtain the photovoltaic power generation power of the predicted time period output by the RVM model.

[0094] In a possible implementation, the photovoltaic power prediction method is applicable to a feeder terminal that can perform photovoltaic power prediction, and the feeder terminal may be a FTU (Feeder Terminal Unit, distribution switch monitoring terminal), a DTU (Distribution Terminal Unit, switch terminal equipment), an intelligent fusion terminal, etc.

[0095] In a possible implementation, the photovoltaic power generation power of the sample time period may be the photovoltaic power generation power at each moment within the sample time period. For example, if the sample time period is one day, the photovoltaic power generation power of the sample time period may be the photovoltaic power generation power at each moment of the 24 hours of the day.

[0096] In a possible implementation, there are various types of characteristic data that affect the photovoltaic power generation of a sample time period. For example, the photovoltaic power generation at each moment in multiple time periods before the sample time period (such as two days before a certain day) can, to a certain extent, reflect the photovoltaic power generation in the sample time period. Therefore, the characteristic data can include the photovoltaic power generation in multiple time periods before the sample time period; the meteorological data in multiple time periods before the sample time period will also affect the photovoltaic power generation. Therefore, the characteristic data can also include the meteorological data at multiple moments before the sample time period. For example, the meteorological data can include at least one of temperature, air pressure, humidity, total radiation, direct radiation, and diffuse radiation; when the photovoltaic module generates electricity, the self-condition of the photovoltaic module will also affect the photovoltaic power generation. Therefore, the characteristic data can also include the component characteristic data of the photovoltaic module in multiple time periods before the sample time period. For example, the component characteristic data of the photovoltaic module can be the photovoltaic module temperature. It should be noted here that the characteristic data that affects the photovoltaic power generation of the sample time period can be the characteristic data at each moment in multiple time periods before the sample time period.

[0097] In a possible implementation, the photovoltaic power prediction problem is a non-linear time period prediction problem. The time series data will continuously change with the change of the time period, and the factors affecting the photovoltaic power output are very complex, and there are many non-time series data. If all the characteristic data is directly input into the neural network model, the time series relationship and non-time series relationship in the characteristic data cannot be effectively extracted, resulting in inaccurate photovoltaic power prediction. Therefore, in this implementation, the random forest algorithm can be used to screen various types of characteristic data first.

[0098] In a possible implementation, assume that there are N samples in the sample dataset, and each sample includes the true value of the photovoltaic power generation in the sample time period and various types of characteristic data corresponding to the sample time period that affect the photovoltaic power generation. The random forest algorithm can randomly select N samples (because sampling is with replacement, so although it is N, it is impossible to traverse all samples) each time with replacement, and use the selected N samples to train a decision tree. Assume that the number of types of characteristic data in the sample dataset is d, and only k (k < d) types of characteristic data are selected each time a decision tree is trained. In this way, multiple decision trees can be trained, and these multiple decision trees can form a random forest. Each decision tree will have a voting result, and the type of characteristic data used by the decision tree with the most votes in the final voting result is the important type. In this way, the random forest algorithm can be used to determine the important types of characteristic data that affect the photovoltaic power generation from the various types of characteristic data, and then use the characteristic data of the important types to predict the photovoltaic power generation in the prediction time period.

[0099] In a possible implementation, a CNN-RVM prediction model can be used to predict the photovoltaic power generation in the prediction time period, and the CNN-RVM prediction model includes a pre-trained CNN (Convolutional Neural Networks) model and a pre-trained RVM (Relevance Vector Machine) model. The feature data of important types corresponding to multiple historical time periods before the prediction time period can be obtained. For example, assuming that the prediction time period is one day in the future, the multiple historical time periods can be the previous two days, and the feature data of important types corresponding to the historical time periods can be input into the CNN-RVM prediction model to predict the photovoltaic power generation in the prediction time period.

[0100] In a possible implementation, the important types of feature data corresponding to the historical time period before the predicted time period can be input into a pre-trained CNN model. The CNN model can mine the feature information and detailed information of the changes of these important types of feature data in the time dimension, fully learn the correlation between the important types of feature data, and extract a set of feature vectors.

[0101] In one possible implementation, the RVM (Relevance Vector Machine) model is a Bayesian sparse kernel method for regression and classification problems. It avoids the main limitations of the support vector machine (SVM), such as requiring the kernel function to be a positive definite kernel, etc. Compared with this SVM, the RVM will usually be a sparser model, making it faster on the test set. The use of the RVM model can not only enhance the fitting ability of the entire prediction model, but also greatly reduce the complexity of the model and improve the real-time performance of photovoltaic power prediction. In this implementation, the feature vector extracted by the CNN model can be input into the RVM model. The RVM model can further extract the features of the feature data and perform nonlinear fitting to output the photovoltaic power generation power of the prediction time period.

[0102] After obtaining the sample data set, the present embodiment can use a random forest algorithm based on the sample data set to determine the important types that affect the photovoltaic power generation from the multiple types involved in the sample data set, thereby screening out several important types of feature data that will have a greater impact on the photovoltaic power generation. Then, the feature data of the important types corresponding to the historical time period before the prediction time period are input into a pre-trained convolutional neural network CNN model to obtain the feature vector output by the CNN model; the feature vector output by the CNN model is input into a pre-trained RVM model to obtain the photovoltaic power generation in the prediction time period output by the RVM model. In this way, the combination of the random forest method and the CNN-RVM prediction model can predict the photovoltaic power generation more accurately and in real time. In addition, the photovoltaic power prediction function is implemented at the feeder terminal instead of the master station, which can avoid resource congestion at the master station and waste of resources on the terminal side.

[0103] In a possible implementation manner, the multiple types of characteristic data affecting the photovoltaic power generation power corresponding to the sample time period include at least two types of characteristic data as follows:

[0104] PV power generation, temperature, air pressure, humidity, PV module temperature, total radiation, direct radiation, and diffuse radiation for multiple time periods before the sample time period.

[0105] In this embodiment, the sample time period may be one day, the multiple time periods before the sample time period may be two days before the sample time period, and the photovoltaic power generation, temperature, air pressure, humidity, photovoltaic module temperature, total radiation, direct radiation, and scattered radiation of the multiple time periods before the sample time period may be the photovoltaic power generation, temperature, air pressure, humidity, photovoltaic module temperature, total radiation, direct radiation, and scattered radiation at each moment of the two days before the sample time period. Of course, the characteristic data may also include various other data that can affect the photovoltaic power generation, which are not listed here one by one.

[0106] In a possible implementation, the acquiring the sample data set includes:

[0107] Acquire an original data set, the original data set including an original photovoltaic power generation power in a sample time period and multiple types of original feature data affecting the photovoltaic power generation power corresponding to the sample time period;

[0108] The data in the original data set is preprocessed to obtain the sample data set, wherein the preprocessing includes at least one of outlier processing, missing value processing, and normalization processing.

[0109] In this embodiment, raw characteristic data for each hour of the 24 hours a day can be collected, such as hourly photovoltaic power values, temperature, air pressure, humidity, component temperature, total radiation, direct radiation and scattered radiation. A total of no less than 30 days of raw characteristic data is collected to form an original data set.

[0110] In this embodiment, when collecting data, abnormal data may occur due to various reasons. For example, the collected temperature is 80 degrees, which is obviously beyond the normal temperature range, or data is missing, for example, the humidity at a certain time is not collected. At this time, the abnormal value in the data can be identified according to the predetermined normal data range, and the abnormal value can be revised. For example, the average value of the data values ​​at two times adjacent to the abnormal value is used as the revised value corresponding to the abnormal value. For missing values, a filling value can be filled in. For example, the average value of the data values ​​at two times adjacent to the missing value is used as the filling value corresponding to the missing value.

[0111] In this implementation, since multiple types of original feature data are collected, the amplitudes of different types of original feature data are not in the same range. In order to reduce the interference caused by the difference in the value ranges of each type of feature data, the data in the original data set can be normalized to normalize the amplitudes of these data to the same range. Usually, each type of feature data in the original data set can be normalized to the range of 0 to 1.

[0112] This implementation can perform outlier processing, missing value processing or normalization processing on the data in the original data set that is originally collected, so as to reduce the interference caused by the original collected data not meeting the standards.

[0113] In a possible implementation, the method further includes:

[0114] Acquire a training sample from the sample data set, the training sample including a true value of photovoltaic power generation in a sample time period and important type of feature data affecting the photovoltaic power generation corresponding to the sample time period;

[0115] The training samples are used to train a CNN-RVM model, wherein the CNN-RVM model includes the CNN model and the RVM model.

[0116] In this embodiment, the true value of the photovoltaic power in the sample time period (such as 24 hours on a certain day) can be used as the label value, and the characteristic data of important types in the two days before the sample time period (48 hours in total) can be used as input variables. In this way, one sample is constructed, and finally m samples are constructed; 70% of the samples can be selected from the m samples as training samples.

[0117] In this embodiment, the input variables in the training sample, i.e., the feature data of important types that affect the photovoltaic power generation corresponding to the sample time period, i.e., the feature data of important types two days (48 hours in total) before the sample time period, can be input into the CNN-RVM model to be trained to obtain the photovoltaic power generation power prediction value corresponding to the sample time period predicted by the CNN-RVM model; by comparing the photovoltaic power generation power prediction value corresponding to the sample time period with the corresponding label value, i.e., the true value of the photovoltaic power in the sample time period, the prediction accuracy of the CNN-RVM model can be calculated, and the model parameters of the CNN-RVM model are continuously adjusted until the accuracy of the photovoltaic power generation power prediction value predicted by the CNN-RVM model reaches a predetermined value, and a trained CNN-RVM model can be obtained, and the trained CNN-RVM model includes a trained CNN model and a trained RVM model.

[0118] In a possible implementation, the using the training samples to train a CNN-RVM model includes:

[0119] The grid search method is used to optimize the three parameters of the number of neurons, convolution kernel size and learning rate in the pre-training CNN-RVM model, and the mean square error is used as the optimization target of the CNN-RVM model. The formula of the mean square error is as follows:

[0120]

[0121] Where T represents the number of samples in the training sample, y i represents the photovoltaic power prediction value corresponding to the i-th sample output by the CNN-RVM model, represents the true value of the photovoltaic power generation power in the i-th sample.

[0122] In this embodiment, hyperparameters are some configuration options of the model, such as the number of neurons, convolution kernel size, and learning rate described here. Usually, multiple different sets of hyperparameter combinations are tried, and then the best performing set is selected as the final parameter setting of the model. This process is called hyperparameter tuning or parameter adjustment. In order to adjust the parameters more effectively, a grid search method can be used to automatically find the optimal hyperparameters. It will generate all possible parameter combinations, and then train the model for each set of parameters and evaluate it, and finally select the best performing parameter combination. In this embodiment, the grid search method can be used to tune the three parameters of the number of neurons, convolution kernel size, and learning rate in the CNN-RVM model before training to obtain the best performing parameter combination of the number of neurons, convolution kernel size, and learning rate.

[0123] In this embodiment, when training the CNN-RVM model, the mean square error can be used as the optimization target of the CNN-RVM model, that is, the mean square error between the predicted value and the true value of T samples is compared according to the above formula, and the parameters are continuously adjusted to obtain the CNN-RVM model corresponding to the minimum mean square value, which is the trained CNN-RVM model.

[0124] In a possible implementation, the initial setting of the model parameters in the CNN model includes: the first convolution layer is initially 64 neurons, the second convolution layer is initially 32 neurons, the third convolution layer is initially 16 neurons, the window size of each convolution layer is initially 1×3, the pooling layer connected to each convolution layer is a maximum pooling layer, the pooling layer window size is 2×2, the step size is 2, and the activation function is a Relu function;

[0125] The penalty coefficient γ of the RVM model is initially set to 0.1;

[0126] The initial learning rate of the CNN-RVM model is 0.01.

[0127] In this embodiment, Figure 2A A schematic diagram of the model structure of a CNN-RVM model according to an embodiment of the present disclosure is shown, as Figure 2A As shown in the figure, the CNN-RVM model includes a CNN model and an RVM model. The CNN model includes three convolutional layers, namely the first convolutional layer, the second convolutional layer and the third convolutional layer. Each convolutional layer is connected to a maximum pooling layer, and the last maximum pooling layer is connected to a fully connected layer. After the input variable is input into the first convolution layer, the convolution kernel is convolved with the input variable to obtain a convolution vector, and the convolution vector is sent to the maximum pooling layer. The maximum value in the pooling window is selected as the pooling result, and the pooling result is output; three convolution and pooling steps are performed to deeply mine the long-term dependence and time series correlation of photovoltaic data; the output of the last maximum pooling layer is flattened into a one-dimensional vector, and the flattened vector is connected to the fully connected layer. A feature space transformation is performed through the fully connected layer to extract and integrate the previously extracted features, and finally a set of feature vectors is obtained. The feature vector extracted by the CNN model is used as the input of the RVM model to obtain the predicted value output by the RVM model, and the predicted value is compared with the corresponding label value to calculate the prediction accuracy of the CNN-RVM model, and the model parameters of the CNN-RVM model are continuously adjusted until the accuracy of the photovoltaic power prediction value predicted by the CNN-RVM model reaches a predetermined value.

[0128] In this embodiment, the initial parameters of the CNN model and the RVM model before training may be as shown in Table 1 below:

[0129]

[0130]

[0131] Table 1

[0132] In this implementation, batch_size refers to the number of samples used in each iteration during the deep learning model training process; the learning rate can determine the speed and direction of the model updating the weight parameters during the training process; the number of model training times refers to the number of times the model iteratively updates the parameters during the training process. The more training times, the stronger the model's learning ability, but its complexity and computational cost will also increase accordingly. In each training iteration, the model will update its internal parameters based on the training data and optimization algorithm to gradually improve the model's prediction accuracy and generalization ability.

[0133] In this implementation, the initial parameters of the model in Table 1 above can be used as a reference to continuously adjust these parameter values ​​until a CNN-RVM model with accurate prediction is obtained through training.

[0134] The trained CNN-RVM model in this embodiment is trained based on the above initial parameters, and a better model can be quickly trained.

[0135] In a possible implementation, the method further includes:

[0136] Acquire samples other than the training samples from the sample data set as test samples;

[0137] Using the test sample, testing the CNN-RVM model trained using the training sample;

[0138] If the test passes, the trained CNN-RVM model is the trained CNN-RVM model;

[0139] If the test fails, continue to use the CNN-RVM model trained with the training samples.

[0140] In this embodiment, test samples can also be obtained from the sample data set. The test samples can be other samples in the sample data set except the training samples. For example, after obtaining the sample data set, 70% can be selected as test samples and the remaining 30% can be selected as test samples.

[0141] In this embodiment, after the CNN-RVM model is trained using the training sample, the test sample can be used to test the accuracy of the CNN-RVM model, and the input variables in the test sample, i.e., the important types of feature data that affect the photovoltaic power generation corresponding to the sample time period, are input into the CNN-RVM model to obtain the photovoltaic power generation power prediction value of the sample time period output by the CNN-RVM model, and the photovoltaic power generation power prediction value and the photovoltaic power generation power true value of each sample time period in the test sample are compared to calculate the accuracy of the CNN-RVM model. If the accuracy meets the predetermined conditions, for example, the accuracy reaches 90%, it means that the test has passed, and the trained CNN-RVM model can be determined as the trained CNN-RVM model; if the test fails, it is necessary to continue to use the CNN-RVM model trained by the training sample until the test passes.

[0142] The present disclosure also provides a method for training a photovoltaic power prediction model. Figure 2B A flow chart of a method for training a photovoltaic power prediction model according to an embodiment of the present disclosure is shown. Figure 2B As shown, the training method of the photovoltaic power prediction model includes the following steps S201-S204:

[0143] In step S201, a sample data set is obtained, wherein the sample data set includes a true value of photovoltaic power generation in a sample time period and multiple types of characteristic data affecting the photovoltaic power generation corresponding to the sample time period;

[0144] In step S202, based on the sample data set, a random forest algorithm is used to determine important types that affect the photovoltaic power generation power from the multiple types;

[0145] In step S203, a training sample is obtained from the sample data set, wherein the training sample includes a true value of photovoltaic power generation in a sample time period and important type of feature data affecting the photovoltaic power generation corresponding to the sample time period;

[0146] In step S204, the training samples are used to train a CNN-RVM model, wherein the CNN-RVM model includes a CNN model and an RVM model, wherein the CNN model is used to extract feature vectors from important types of feature data corresponding to a prediction time period, and the RVM model is used to predict the photovoltaic power generation power of the prediction time period based on the feature vectors extracted by the CNN model.

[0147] In a possible implementation manner, the training of the photovoltaic power prediction model is applicable to a device capable of executing the training of the photovoltaic power prediction model.

[0148] In a possible implementation, the photovoltaic power generation power of the sample time period may be the photovoltaic power generation power at each moment within the sample time period. For example, if the sample time period is one day, the photovoltaic power generation power of the sample time period may be the photovoltaic power generation power at each moment of the 24 hours of the day.

[0149] In a possible implementation, there are multiple types of characteristic data that affect the photovoltaic power generation power in the sample time period. For example, the photovoltaic power generation power at each moment in multiple time periods before the sample time period (such as two days before a certain day) can reflect the photovoltaic power generation power in the sample time period to a certain extent, so the characteristic data may include the photovoltaic power generation power in multiple time periods before the sample time period; the meteorological data in multiple time periods before the sample time period will also affect the photovoltaic power generation power, so the characteristic data may also include the meteorological data at multiple moments before the sample time period. For example, the meteorological data may include at least one of temperature, air pressure, humidity, total radiation, direct radiation and scattered radiation; when the photovoltaic component generates electricity, the photovoltaic component's own conditions will also affect the photovoltaic power generation power, so the characteristic data may also include the component characteristic data of the photovoltaic component in multiple time periods before the sample time period. For example, the component characteristic data of the photovoltaic component may be the temperature of the photovoltaic component. It should be noted here that the characteristic data that affects the photovoltaic power generation power in the sample time period may be the characteristic data at each moment in multiple time periods before the sample time period.

[0150] In one possible implementation, the photovoltaic power prediction problem is a nonlinear time period prediction problem. The time series data will continue to change with the change of time period, and the factors affecting the photovoltaic power output are very complex, and there are many non-time series data. If all feature data are directly input into the neural network model, the time series relationship and non-time series relationship in the feature data cannot be effectively extracted, which will lead to inaccurate photovoltaic power prediction. Therefore, this implementation can first use the random forest algorithm to screen multiple types of feature data.

[0151] In a possible implementation, assume that there are N samples in the sample dataset, and each sample includes the true value of the photovoltaic power generation during the sample time period and various types of feature data corresponding to the sample time period that affect the photovoltaic power generation. The random forest algorithm can randomly select N samples (with replacement, so although it is N, it is impossible to traverse all samples) each time, and use the selected N samples to train a decision tree. Assume that the number of types of feature data in the sample dataset is d, and only k (k < d) types of feature data are selected each time a decision tree is trained. In this way, multiple decision trees can be trained, and these multiple decision trees can form a random forest. Each decision tree will have a voting result, and the type of feature data used by the decision tree with the most votes in the final voting result is the important type. In this way, the random forest algorithm can be used to determine the important types that affect the photovoltaic power generation from the various types of feature data, and then the feature data of the important types can be used to predict the photovoltaic power generation during the prediction time period.

[0152] In this implementation, the true value of the photovoltaic power during the sample time period (such as 24 hours of a certain day) can be used as the label value, and the feature data of the important types in the two days before the sample time period (a total of 48 hours) can be used as the input variables. In this way, 1 sample is constructed, and finally m samples are constructed; 70% of the samples can be selected from the m samples as the training samples.

[0153] In this implementation, the input variables in the training samples, that is, the feature data of the important types that affect the photovoltaic power generation corresponding to the sample time period, that is, the feature data of the important types in the two days before the sample time period (a total of 48 hours), can be input into the CNN-RVM model to be trained, and the predicted value of the photovoltaic power generation corresponding to the sample time period predicted by the CNN-RVM model can be obtained; by comparing the predicted value of the photovoltaic power generation corresponding to the sample time period with the corresponding label value, that is, the true value of the photovoltaic power during the sample time period, the prediction accuracy of the CNN-RVM model can be calculated, and the model parameters of the CNN-RVM model can be continuously adjusted until the accuracy of the predicted value of the photovoltaic power generation predicted by the CNN-RVM model reaches a predetermined value, and then the trained CNN-RVM model can be obtained. The trained CNN-RVM model includes a trained CNN model and a trained RVM model.

[0154] In a possible implementation, training the CNN-RVM model using the training samples includes:

[0155] Using the grid search method, optimize the three parameters of the number of neurons, convolution kernel size, and learning rate in the CNN-RVM model before training, and use the mean square error as the optimization target of the CNN-RVM model. The formula for the mean square error is as follows:

[0156]

[0157] Where T represents the number of samples in the training sample, y i represents the photovoltaic power prediction value corresponding to the i-th sample output by the CNN-RVM model, represents the true value of the photovoltaic power generation power in the i-th sample.

[0158] In this embodiment, hyperparameters are some configuration options of the model, such as the number of neurons, convolution kernel size, and learning rate described here. Usually, multiple different sets of hyperparameter combinations are tried, and then the best performing set is selected as the final parameter setting of the model. This process is called hyperparameter tuning or parameter adjustment. In order to adjust the parameters more effectively, a grid search method can be used to automatically find the optimal hyperparameters. It will generate all possible parameter combinations, and then train the model for each set of parameters and evaluate it, and finally select the best performing parameter combination. In this embodiment, the grid search method can be used to tune the three parameters of the number of neurons, convolution kernel size, and learning rate in the CNN-RVM model before training to obtain the best performing parameter combination of the number of neurons, convolution kernel size, and learning rate.

[0159] In this embodiment, when training the CNN-RVM model, the mean square error can be used as the optimization target of the CNN-RVM model, that is, the mean square error between the predicted value and the true value of T samples is compared according to the above formula, and the parameters are continuously adjusted to obtain the CNN-RVM model corresponding to the minimum mean square value, which is the trained CNN-RVM model.

[0160] The present disclosure also provides a photovoltaic power prediction device, Figure 3 The structure block diagram of the photovoltaic power prediction device according to the embodiment of the present disclosure is shown, and the device can be implemented as part or all of the electronic device through software, hardware or a combination of both. Figure 3 As shown, the photovoltaic power prediction device comprises:

[0161] The sample acquisition module 301 is configured to acquire a sample data set, wherein the sample data set includes a true value of photovoltaic power generation in a sample time period and multiple types of characteristic data affecting the photovoltaic power generation corresponding to the sample time period;

[0162] A data type determination module 302 is configured to determine, based on the sample data set, an important type that affects the photovoltaic power generation from the multiple types using a random forest algorithm;

[0163] The data acquisition module 303 is configured to acquire important types of feature data corresponding to multiple historical time periods before the prediction time period;

[0164] The feature extraction module 304 is configured to input the feature data of important types corresponding to multiple historical time periods before the predicted time period into a pre-trained convolutional neural network (CNN) model to obtain a feature vector output by the CNN model;

[0165] The prediction module 305 is configured to input the feature vector output by the CNN model into a pre-trained RVM model to obtain the photovoltaic power generation power in the predicted time period output by the RVM model.

[0166] In a possible implementation, the photovoltaic power prediction device is applicable to a feeder terminal capable of performing photovoltaic power prediction, and the feeder terminal may be a FTU, a DTU, an intelligent fusion terminal, or the like.

[0167] In a possible implementation, the photovoltaic power generation power of the sample time period may be the photovoltaic power generation power at each moment within the sample time period. For example, if the sample time period is one day, the photovoltaic power generation power of the sample time period may be the photovoltaic power generation power at each moment of the 24 hours of the day.

[0168] In a possible implementation, there are multiple types of characteristic data that affect the photovoltaic power generation power in the sample time period. For example, the photovoltaic power generation power at each moment in multiple time periods before the sample time period (such as two days before a certain day) can reflect the photovoltaic power generation power in the sample time period to a certain extent, so the characteristic data may include the photovoltaic power generation power in multiple time periods before the sample time period; the meteorological data in multiple time periods before the sample time period will also affect the photovoltaic power generation power, so the characteristic data may also include the meteorological data at multiple moments before the sample time period. For example, the meteorological data may include at least one of temperature, air pressure, humidity, total radiation, direct radiation and scattered radiation; when the photovoltaic component generates electricity, the photovoltaic component's own conditions will also affect the photovoltaic power generation power, so the characteristic data may also include the component characteristic data of the photovoltaic component in multiple time periods before the sample time period. For example, the component characteristic data of the photovoltaic component may be the temperature of the photovoltaic component. It should be noted here that the characteristic data that affects the photovoltaic power generation power in the sample time period may be the characteristic data at each moment in multiple time periods before the sample time period.

[0169] In a possible implementation, the photovoltaic power prediction problem is a non-linear time period prediction problem. The time series data changes continuously with the change of the time period, and the factors affecting the photovoltaic power output are very complex, and there are many non-time series data. If all the feature data is directly input into the neural network model, the time series relationship and non-time series relationship in the feature data cannot be effectively extracted, resulting in inaccurate photovoltaic power prediction. Therefore, in this implementation, the random forest algorithm can be used to screen various types of feature data first.

[0170] In a possible implementation, assume that there are N samples in the sample dataset, and each sample includes the true value of the photovoltaic power generation in the sample time period and various types of feature data corresponding to the sample time period that affect the photovoltaic power generation. The random forest algorithm can randomly select N samples (because sampling is with replacement, so although it is N, it is impossible to traverse all samples) each time, and use the selected N samples to train a decision tree. Assume that the number of types of feature data in the sample dataset is d, and only k (k < d) types of feature data are selected each time when training a decision tree. In this way, multiple decision trees can be trained, and these multiple decision trees can form a random forest. Each decision tree will have a voting result, and the type of feature data used by the decision tree with the most votes in the end is the important type. In this way, the random forest algorithm can be used to determine the important types of feature data that affect the photovoltaic power generation from the various types of feature data, and then use the feature data of the important types to predict the photovoltaic power generation in the prediction time period.

[0171] In a possible implementation, the CNN-RVM prediction model can be used to predict the photovoltaic power generation in the prediction time period. The CNN-RVM prediction model includes a pre-trained CNN (Convolutional Neural Networks) model and a pre-trained RVM (Relevance Vector Machine) model. The feature data of the important types corresponding to multiple historical time periods before the prediction time period can be obtained. For example, assume that the prediction time period is the next day, then these multiple historical time periods can be the previous two days. The feature data of the important types corresponding to the historical time period can be input into the CNN-RVM prediction model to predict the photovoltaic power generation in the prediction time period.

[0172] In a possible implementation, the feature data of the important types corresponding to the historical time period before the prediction time period can be input into the pre-trained CNN model. The CNN model can mine the feature information and the detailed information of the changes in the time dimension of these important types of feature data, fully learn the correlation between the important types of feature data, and extract a set of feature vectors.

[0173] In one possible implementation, the RVM (Relevance Vector Machine) model is a Bayesian sparse kernel method for regression and classification problems. It avoids the main limitations of the support vector machine (SVM), such as requiring the kernel function to be a positive definite kernel, etc. Compared with this SVM, the RVM will usually be a sparser model, making it faster on the test set. The use of the RVM model can not only enhance the fitting ability of the entire prediction model, but also greatly reduce the complexity of the model and improve the real-time performance of photovoltaic power prediction. In this implementation, the feature vector extracted by the CNN model can be input into the RVM model. The RVM model can further extract the features of the feature data and perform nonlinear fitting to output the photovoltaic power generation power of the prediction time period.

[0174] After obtaining the sample data set, the present embodiment can use a random forest algorithm based on the sample data set to determine the important types that affect the photovoltaic power generation from the multiple types involved in the sample data set, thereby screening out several important types of feature data that will have a greater impact on the photovoltaic power generation. Then, the feature data of the important types corresponding to the historical time period before the prediction time period are input into a pre-trained convolutional neural network CNN model to obtain the feature vector output by the CNN model; the feature vector output by the CNN model is input into a pre-trained RVM model to obtain the photovoltaic power generation in the prediction time period output by the RVM model. In this way, the combination of the random forest method and the CNN-RVM prediction model can predict the photovoltaic power generation more accurately and in real time. In addition, the photovoltaic power prediction function is implemented at the feeder terminal instead of the master station, which can avoid resource congestion at the master station and waste of resources on the terminal side.

[0175] In a possible implementation manner, the multiple types of characteristic data affecting the photovoltaic power generation power corresponding to the sample time period include at least two types of characteristic data as follows:

[0176] PV power generation, temperature, air pressure, humidity, PV module temperature, total radiation, direct radiation, and diffuse radiation for multiple time periods before the sample time period.

[0177] In this embodiment, the sample time period may be one day, the multiple time periods before the sample time period may be two days before the sample time period, and the photovoltaic power generation, temperature, air pressure, humidity, photovoltaic module temperature, total radiation, direct radiation, and scattered radiation of the multiple time periods before the sample time period may be the photovoltaic power generation, temperature, air pressure, humidity, photovoltaic module temperature, total radiation, direct radiation, and scattered radiation at each moment of the two days before the sample time period. Of course, the characteristic data may also include various other data that can affect the photovoltaic power generation, which are not listed here one by one.

[0178] In a possible implementation, the sample acquisition module is configured as follows:

[0179] Acquire an original data set, wherein the original data set includes the original photovoltaic power generation power of a sample time period and various types of original feature data corresponding to the sample time period that affect the photovoltaic power generation power; preprocess the data in the original data set to obtain the sample data set, wherein the preprocessing includes at least one of outlier processing, missing value processing, and normalization processing.

[0180] In this embodiment, the sample acquisition module can collect original characteristic data for each hour of the 24 hours a day, such as the hourly photovoltaic power value, temperature, air pressure, humidity, component temperature, total radiation, direct radiation and scattered radiation, and collect a total of no less than 30 days of original characteristic data to form an original data set.

[0181] In this embodiment, when collecting data, abnormal data may occur due to various reasons. For example, the collected temperature is 80 degrees, which is obviously beyond the normal temperature range, or data is missing, for example, the humidity at a certain time is not collected. At this time, the abnormal value in the data can be identified according to the predetermined normal data range, and the abnormal value can be revised. For example, the average value of the data values ​​at two times adjacent to the abnormal value is used as the revised value corresponding to the abnormal value. For missing values, a filling value can be filled in. For example, the average value of the data values ​​at two times adjacent to the missing value is used as the filling value corresponding to the missing value.

[0182] In this implementation, since multiple types of original feature data are collected, the amplitudes of different types of original feature data are not in the same range. In order to reduce the interference caused by the difference in the value ranges of each type of feature data, the data in the original data set can be normalized to normalize the amplitudes of these data to the same range. Usually, each type of feature data in the original data set can be normalized to the range of 0 to 1.

[0183] This implementation can perform outlier processing, missing value processing or normalization processing on the data in the original data set that is originally collected, so as to reduce the interference caused by the original collected data not meeting the standards.

[0184] In a possible implementation, the device further includes:

[0185] The model training module is configured to obtain training samples from the sample data set, wherein the training samples include the true value of the photovoltaic power generation power in the sample time period and the important type of feature data that affects the photovoltaic power generation power corresponding to the sample time period; and use the training samples to train a CNN-RVM model, wherein the CNN-RVM model includes the CNN model and the RVM model.

[0186] In this implementation, the true value of the photovoltaic power in a sample time period (such as 24 hours on a certain day) can be used as the label value, and the important types of characteristic data two days before the sample time period (48 hours in total) can be used as input variables, so as to construct one sample and finally construct m samples; the model training module can select 70% of the samples from the m samples as training samples.

[0187] In this embodiment, the model training module can input the input variables in the training sample, i.e., the feature data of important types that affect the photovoltaic power generation corresponding to the sample time period, i.e., the feature data of important types two days (48 hours in total) before the sample time period, into the CNN-RVM model to be trained, and obtain the photovoltaic power generation power prediction value corresponding to the sample time period predicted by the CNN-RVM model; compare the photovoltaic power generation power prediction value corresponding to the sample time period with the corresponding label value, i.e., the true value of the photovoltaic power in the sample time period, and calculate the prediction accuracy of the CNN-RVM model, and continuously adjust the model parameters of the CNN-RVM model until the accuracy of the photovoltaic power generation power prediction value predicted by the CNN-RVM model reaches a predetermined value, and a trained CNN-RVM model can be obtained, and the trained CNN-RVM model includes a trained CNN model and a trained RVM model.

[0188] In a possible implementation manner, the model training module uses the training sample to train a portion of the CNN-RVM model, which is configured as follows:

[0189] The grid search method is used to tune the three parameters of the number of neurons, convolution kernel size and learning rate in the pre-trained CNN-RVM model;

[0190] The mean square error is used as the optimization target of the CNN-RVM model, and the formula of the mean square error is as follows:

[0191]

[0192] Where T represents the number of samples in the training sample, y i represents the photovoltaic power prediction value corresponding to the i-th sample output by the CNN-RVM model, represents the true value of the photovoltaic power generation power in the i-th sample.

[0193] In this embodiment, hyperparameters are some configuration options of the model, such as the number of neurons, convolution kernel size and learning rate described here, etc., and usually multiple groups of different hyperparameter combinations are tried, and then the group with the best performance is selected as the final parameter setting of the model. This process is called hyperparameter tuning or parameter adjustment. In order to adjust the parameters more effectively, a grid search method can be used to automatically find the optimal hyperparameters. It will generate all possible parameter combinations, and then train the model for each set of parameters and evaluate it, and finally select the best performing parameter combination. In this embodiment, the model training module can use a grid search method to tune the three parameters of the number of neurons, convolution kernel size and learning rate in the CNN-RVM model before training, and obtain the best performing parameter combination of the number of neurons, convolution kernel size and learning rate.

[0194] In this embodiment, when training the CNN-RVM model, the model training module can use the mean square error as the optimization target of the CNN-RVM model, that is, compare the mean square error between the predicted value and the true value of T samples according to the above formula, and continuously adjust the parameters to obtain the CNN-RVM model that minimizes the mean square value, which is the trained CNN-RVM model.

[0195] In a possible implementation, the initial settings of the model parameters in the CNN model include: the first convolution layer is initially set to 64 neurons, the second convolution layer is initially set to 32 neurons, the third convolution layer is initially set to 16 neurons, the window size of each convolution layer is initially set to 1×3, the pooling layer connected to each convolution layer is a maximum pooling layer, the pooling layer window size is initially set to 2×2, the step size is initially set to 2, and the activation function is a linear rectification function;

[0196] The penalty coefficient γ of the RVM model is initially set to 0.1, and the kernel function of the RVM model is a Gaussian function;

[0197] The batch size of the CNN-RVM model is initially set to 120, the initial learning rate is initially set to 0.01, and the initial number of training times is initially set to 500.

[0198] In this embodiment, if Figure 2AAs shown in the figure, the CNN-RVM model includes a CNN model and an RVM model. The CNN model includes three convolutional layers, namely the first convolutional layer, the second convolutional layer and the third convolutional layer. Each convolutional layer is connected to a maximum pooling layer, and the last maximum pooling layer is connected to a fully connected layer. After the input variable is input into the first convolution layer, the convolution kernel is convolved with the input variable to obtain a convolution vector, and the convolution vector is sent to the maximum pooling layer. The maximum value in the pooling window is selected as the pooling result, and the pooling result is output; three convolution and pooling steps are performed to deeply mine the long-term dependence and time series correlation of photovoltaic data; the output of the last maximum pooling layer is flattened into a one-dimensional vector, and the flattened vector is connected to the fully connected layer. A feature space transformation is performed through the fully connected layer to extract and integrate the previously extracted features, and finally a set of feature vectors is obtained. The feature vector extracted by the CNN model is used as the input of the RVM model to obtain the predicted value output by the RVM model, and the predicted value is compared with the corresponding label value to calculate the prediction accuracy of the CNN-RVM model, and the model parameters of the CNN-RVM model are continuously adjusted until the accuracy of the photovoltaic power prediction value predicted by the CNN-RVM model reaches a predetermined value.

[0199] In this embodiment, the initial parameters of the CNN model and the RVM model before training may be as shown in Table 1 above.

[0200] In this implementation, batch_size refers to the number of samples used in each iteration during the deep learning model training process; the learning rate can determine the speed and direction of the model updating the weight parameters during the training process; the number of model training times refers to the number of times the model iteratively updates the parameters during the training process. The more training times, the stronger the model's learning ability, but its complexity and computational cost will also increase accordingly. In each training iteration, the model will update its internal parameters based on the training data and optimization algorithm to gradually improve the model's prediction accuracy and generalization ability.

[0201] In this implementation, the initial parameters of the model in Table 1 above can be used as a reference to continuously adjust these parameter values ​​until a CNN-RVM model with accurate prediction is obtained through training.

[0202] In this embodiment, the CNN-RVM model is trained based on the above initial parameters, and a better model can be quickly trained.

[0203] In a possible implementation, the device further includes:

[0204] A testing module is configured to obtain a test sample from the sample data set, wherein the test sample includes the true value of the photovoltaic power generation power in the sample time period and the important type of feature data corresponding to the sample time period that affects the photovoltaic power generation power, and the test sample is different from the training sample; using the test sample, testing the CNN-RVM model trained using the training sample; if the test passes, the trained CNN-RVM model is the trained CNN-RVM model; if the test fails, continue to use the CNN-RVM model trained using the training sample.

[0205] In this embodiment, test samples can also be obtained from the sample data set. The test samples can be other samples in the sample data set except the training samples. For example, after obtaining the sample data set, the test module can select 70% of them as test samples and the remaining 30% as test samples.

[0206] In this embodiment, after the CNN-RVM model is trained using the training sample, the test module can use the test sample to test the accuracy of the CNN-RVM model, input the input variables in the test sample, i.e., the important types of feature data affecting the photovoltaic power generation corresponding to the sample time period, into the CNN-RVM model, obtain the photovoltaic power generation power prediction value of the sample time period output by the CNN-RVM model, compare the photovoltaic power generation power prediction value and the photovoltaic power true value of each sample time period in the test sample, calculate the accuracy of the CNN-RVM model, and if the accuracy meets the predetermined conditions, for example, the accuracy reaches 90%, it means that the test is passed, and the trained CNN-RVM model can be determined as the trained CNN-RVM model; if the test fails, it is necessary to continue to use the CNN-RVM model trained by the training sample until the test passes.

[0207] The present disclosure also provides a photovoltaic power prediction model training device, Figure 4 The structure block diagram of the photovoltaic power prediction model training device according to the embodiment of the present disclosure is shown, and the device can be implemented as part or all of the electronic device through software, hardware or a combination of both. Figure 4 As shown, the training device of the photovoltaic power prediction model includes:

[0208] The sample data acquisition module 401 is configured to acquire a sample data set, wherein the sample data set includes a true value of photovoltaic power generation in a sample time period and multiple types of characteristic data affecting the photovoltaic power generation corresponding to the sample time period;

[0209] An important type determination module 402 is configured to determine the important type that affects the photovoltaic power generation power from the multiple types using a random forest algorithm based on the sample data set;

[0210] The training sample acquisition module 403 is configured to acquire training samples from the sample data set, wherein the training samples include the true value of the photovoltaic power generation in the sample time period and the important type of feature data affecting the photovoltaic power generation corresponding to the sample time period;

[0211] The training module 404 is configured to use the training samples to train a CNN-RVM model, wherein the CNN-RVM model includes a CNN model and an RVM model, wherein the CNN model is used to extract feature vectors from important types of feature data corresponding to a prediction time period, and the RVM model is used to predict the photovoltaic power generation power of the prediction time period based on the feature vectors extracted by the CNN model.

[0212] In a possible implementation manner, the training of the photovoltaic power prediction model is applicable to a device capable of executing the training of the photovoltaic power prediction model.

[0213] In a possible implementation, the photovoltaic power generation power of the sample time period may be the photovoltaic power generation power at each moment within the sample time period. For example, if the sample time period is one day, the photovoltaic power generation power of the sample time period may be the photovoltaic power generation power at each moment of the 24 hours of the day.

[0214] In a possible implementation, there are multiple types of characteristic data that affect the photovoltaic power generation power in the sample time period. For example, the photovoltaic power generation power at each moment in multiple time periods before the sample time period (such as two days before a certain day) can reflect the photovoltaic power generation power in the sample time period to a certain extent, so the characteristic data may include the photovoltaic power generation power in multiple time periods before the sample time period; the meteorological data in multiple time periods before the sample time period will also affect the photovoltaic power generation power, so the characteristic data may also include the meteorological data at multiple moments before the sample time period. For example, the meteorological data may include at least one of temperature, air pressure, humidity, total radiation, direct radiation and scattered radiation; when the photovoltaic component generates electricity, the photovoltaic component's own conditions will also affect the photovoltaic power generation power, so the characteristic data may also include the component characteristic data of the photovoltaic component in multiple time periods before the sample time period. For example, the component characteristic data of the photovoltaic component may be the temperature of the photovoltaic component. It should be noted here that the characteristic data that affects the photovoltaic power generation power in the sample time period may be the characteristic data at each moment in multiple time periods before the sample time period.

[0215] In a possible implementation manner, the photovoltaic power prediction problem is a non-linear time period prediction problem. The time series data will continuously change with the change of the time period, and the factors affecting the photovoltaic power output are very complex, and there are many non-time series data. If all the feature data is directly input into the neural network model, the time series relationship and non-time series relationship in the feature data cannot be effectively extracted, resulting in inaccurate photovoltaic power prediction. Therefore, in this implementation manner, the random forest algorithm can be first used to screen various types of feature data.

[0216] In a possible implementation manner, it is assumed that there are N samples in the sample data set, and each sample includes the true value of the photovoltaic power generation in the sample time period and various types of feature data corresponding to the sample time period that affect the photovoltaic power generation. The random forest algorithm can randomly select N samples (because of sampling with replacement, although it is N, it is impossible to traverse all samples) each time with replacement, and use the selected N samples to train a decision tree. It is assumed that the number of types of feature data in the sample data set is d, and only k (k < d) types of feature data are selected each time a decision tree is trained. In this way, multiple decision trees can be trained, and these multiple decision trees can form a random forest. Each decision tree will have a voting result, and the type of feature data used by the decision tree with the most votes in the final voting result is the important type. In this way, the important type of feature data that affects the photovoltaic power generation can be determined from the various types of feature data by using the random forest algorithm, and then the feature data of the important type can be used to predict the photovoltaic power generation in the prediction time period.

[0217] In this implementation manner, the true value of the photovoltaic power in the sample time period (such as 24 hours of a certain day) can be used as the label value, and the feature data of the important type in the two days before the sample time period (a total of 48 hours) can be used as the input variable. In this way, 1 sample is constructed, and finally m samples are constructed; 70% of the samples can be selected from the m samples as the training samples.

[0218] In this embodiment, the input variables in the training sample, i.e., the feature data of important types that affect the photovoltaic power generation corresponding to the sample time period, i.e., the feature data of important types two days (48 hours in total) before the sample time period, can be input into the CNN-RVM model to be trained to obtain the photovoltaic power generation power prediction value corresponding to the sample time period predicted by the CNN-RVM model; by comparing the photovoltaic power generation power prediction value corresponding to the sample time period with the corresponding label value, i.e., the true value of the photovoltaic power in the sample time period, the prediction accuracy of the CNN-RVM model can be calculated, and the model parameters of the CNN-RVM model are continuously adjusted until the accuracy of the photovoltaic power generation power prediction value predicted by the CNN-RVM model reaches a predetermined value, and a trained CNN-RVM model can be obtained, and the trained CNN-RVM model includes a trained CNN model and a trained RVM model.

[0219] In a possible implementation, the training module is configured as follows:

[0220] The grid search method is used to optimize the three parameters of the number of neurons, convolution kernel size and learning rate in the pre-training CNN-RVM model, and the mean square error is used as the optimization target of the CNN-RVM model. The formula of the mean square error is as follows:

[0221]

[0222] Where T represents the number of samples in the training sample, y i represents the photovoltaic power prediction value corresponding to the i-th sample output by the CNN-RVM model, represents the true value of the photovoltaic power generation power in the i-th sample.

[0223] In this embodiment, hyperparameters are some configuration options of the model, such as the number of neurons, convolution kernel size, and learning rate described here. Usually, multiple different hyperparameter combinations are tried, and then the best performing group is selected as the final parameter setting of the model. This process is called hyperparameter tuning or parameter adjustment. In order to adjust the parameters more effectively, a grid search method can be used to automatically find the optimal hyperparameters. It will generate all possible parameter combinations, and then train the model for each set of parameters and evaluate it, and finally select the best performing parameter combination. In this embodiment, the training module can use a grid search method to tune the three parameters of the number of neurons, convolution kernel size, and learning rate in the CNN-RVM model before training to obtain the best performing parameter combination of the number of neurons, convolution kernel size, and learning rate.

[0224] In this embodiment, when training the CNN-RVM model, the mean square error can be used as the optimization target of the CNN-RVM model, that is, the mean square error between the predicted value and the true value of T samples is compared according to the above formula, and the parameters are continuously adjusted to obtain the CNN-RVM model corresponding to the minimum mean square value, which is the trained CNN-RVM model.

[0225] The present disclosure also discloses a feeder terminal. Figure 5 A structural block diagram of a feeder terminal according to an embodiment of the present disclosure is shown.

[0226] like Figure 5 As shown, the feeder terminal includes a memory 501 and a processor 502, wherein the memory 501 is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor 502 to implement the method according to the embodiment of the present disclosure.

[0227] In one possible implementation, a photovoltaic power prediction method can be built using Python language. Relying on the Python library, the photovoltaic power prediction method can be packaged into an APP. The packaged APP is deployed on the software platform of the feeder terminal to realize the photovoltaic power prediction function on the feeder terminal side.

[0228] In one possible implementation, the one or more computer instructions are executed by the processor 502 to implement a photovoltaic power prediction method according to an embodiment of the present disclosure, the photovoltaic power prediction method comprising:

[0229] Acquire a sample data set, the sample data set including a true value of photovoltaic power generation in a sample time period and multiple types of characteristic data affecting the photovoltaic power generation corresponding to the sample time period;

[0230] Based on the sample data set, a random forest algorithm is used to determine the important types that affect the photovoltaic power generation power from the multiple types;

[0231] Obtain important types of feature data corresponding to multiple historical time periods before the forecast time period;

[0232] Inputting important types of feature data corresponding to multiple historical time periods before the forecast time period into a pre-trained convolutional neural network (CNN) model to obtain a feature vector output by the CNN model;

[0233] The feature vector output by the CNN model is input into a pre-trained RVM model to obtain the photovoltaic power generation power in the predicted time period output by the RVM model.

[0234] In a possible implementation manner, the multiple types of characteristic data affecting the photovoltaic power generation power corresponding to the sample time period include at least two types of characteristic data as follows:

[0235] PV power generation, temperature, air pressure, humidity, PV module temperature, total radiation, direct radiation, and diffuse radiation for multiple time periods before the sample time period.

[0236] In a possible implementation, the acquiring the sample data set includes:

[0237] Acquire an original data set, the original data set including an original photovoltaic power generation power in a sample time period and multiple types of original feature data affecting the photovoltaic power generation power corresponding to the sample time period;

[0238] The data in the original data set is preprocessed to obtain the sample data set, wherein the preprocessing includes at least one of outlier processing, missing value processing, and normalization processing.

[0239] In a possible implementation, the method further includes:

[0240] Acquire a training sample from the sample data set, the training sample including a true value of photovoltaic power generation in a sample time period and important type of feature data affecting the photovoltaic power generation corresponding to the sample time period;

[0241] The training samples are used to train a CNN-RVM model, wherein the CNN-RVM model includes the CNN model and the RVM model.

[0242] In a possible implementation, the using the training samples to train a CNN-RVM model includes:

[0243] The grid search method is used to tune the three parameters of the number of neurons, convolution kernel size and learning rate in the pre-trained CNN-RVM model;

[0244] The mean square error is used as the optimization target of the CNN-RVM model, and the formula of the mean square error is as follows:

[0245]

[0246] Where T represents the number of samples in the training sample, y i represents the photovoltaic power prediction value corresponding to the i-th sample output by the CNN-RVM model, represents the true value of the photovoltaic power generation power in the i-th sample.

[0247] In a possible implementation, the initial setting of the model parameters in the CNN model includes: the first convolution layer is initially 64 neurons, the second convolution layer is initially 32 neurons, the third convolution layer is initially 16 neurons, the window size of each convolution layer is initially 1×3, the pooling layer connected to each convolution layer is a maximum pooling layer, the pooling layer window size is 2×2, the step size is 2, and the activation function is a linear rectification function;

[0248] The penalty coefficient γ of the RVM model is initially set to 0.1, and the kernel function of the RVM model is a Gaussian function;

[0249] The batch size of the CNN-RVM model is initially set to 120, the initial learning rate is initially set to 0.01, and the initial number of training times is initially set to 500.

[0250] In a possible implementation, the method further includes:

[0251] Acquire samples other than the training samples from the sample data set as test samples;

[0252] Using the test sample, testing the CNN-RVM model trained using the training sample;

[0253] If the test passes, the trained CNN-RVM model is the trained CNN-RVM model;

[0254] If the test fails, continue to use the CNN-RVM model trained with the training samples.

[0255] The present disclosure also discloses a system, which includes the above Figure 5 Feeder terminal shown.

[0256] In a possible implementation, when the above object data is photovoltaic power generation, the system can be a distribution network system, which includes a distribution master station, a distribution substation (usually located in a substation, optional), the above feeder terminals and a communication network. The distribution master station is located in the city dispatching center, the distribution substation is deployed in the substation, the substation is responsible for communicating with the feeder terminals and other power terminal equipment in the area under its jurisdiction, and the master station is responsible for communicating with each substation. Figure 5 The feeder terminal shown can accurately and efficiently predict the photovoltaic power generation. Based on the prediction capability of the electronic equipment, the distribution network system can achieve effective photovoltaic equipment management, photovoltaic power generation grid-connected regulation, and efficient linkage between source, grid, load and storage, which greatly improves the efficiency and benefits of distribution network management.

[0257] The disclosed embodiments also provide a chip, which includes the photovoltaic power prediction device mentioned above. The chip can be any chip that can realize the photovoltaic power prediction device, or the chip can be a chip of a photovoltaic power prediction model training device. The device can be realized as part or all of the chip through software, hardware or a combination of both.

[0258] Figure 6 A schematic diagram showing the structure of a computer system suitable for implementing the method of the embodiment of the present disclosure is shown.

[0259] like Figure 6 As shown, the computer system 600 includes a processing unit 601, which can perform various processes in the above-mentioned embodiments according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage part 608 into a random access memory (RAM) 603. In the RAM 603, various programs and data required for the operation of the computer system 600 are also stored. The processing unit 601, the ROM 602, and the RAM 603 are connected to each other via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0260] The following components are connected to the I / O interface 605: an input part 606 including keys, etc.; an output part 607 including a display, etc.; a storage part 608 including a hard disk, etc.; and a communication part 609 including a network interface card such as a LAN card, a modem, etc. The communication part 609 performs communication processing via a network such as the Internet. The drive 610 is also connected to the I / O interface 605 as needed. Removable media 611, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., are installed on the drive 610 as needed, so that the computer program read therefrom is installed into the storage part 608 as needed. Among them, the processing unit 601 can be implemented as a processing unit such as a CPU, a GPU, a TPU, an FPGA, an NPU, etc.

[0261] In particular, according to an embodiment of the present disclosure, the method described above can be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product, which includes computer instructions, and the computer instructions are executed by a processor to implement the method steps described above. In such an embodiment, the computer program product can be downloaded and installed from a network through the communication part 609, and / or installed from a removable medium 611.

[0262] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, a program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0263] The units or modules involved in the embodiments described in the present disclosure may be implemented by software or programmable hardware. The units or modules described may also be set in a processor, and the names of these units or modules do not constitute limitations on the units or modules themselves in some cases.

[0264] As another aspect, the present disclosure further provides a computer-readable storage medium, which may be a computer-readable storage medium included in the electronic device or computer system in the above embodiment; or a computer-readable storage medium that exists independently and is not assembled into a device. The computer-readable storage medium stores one or more programs, and the programs are used by one or more processors to execute the method described in the present disclosure.

[0265] The above description is only a preferred embodiment of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the present disclosure is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the inventive concept. For example, the above features are replaced with the technical features with similar functions disclosed in the present disclosure (but not limited to) by each other.

Claims

1. A photovoltaic power prediction method, characterized in that: The method is applied to a feeder terminal, and the method comprises: Acquire a sample data set, the sample data set including a true value of photovoltaic power generation in a sample time period and multiple types of characteristic data affecting the photovoltaic power generation corresponding to the sample time period; Based on the sample data set, a random forest algorithm is used to determine the important types that affect the photovoltaic power generation power from the multiple types; Obtain important types of feature data corresponding to multiple historical time periods before the forecast time period; Inputting important types of feature data corresponding to multiple historical time periods before the forecast time period into a pre-trained convolutional neural network (CNN) model to obtain a feature vector output by the CNN model; Inputting the feature vector output by the CNN model into a pre-trained relevance vector machine (RVM) model to obtain the photovoltaic power generation in the predicted time period output by the RVM model; Wherein, the CNN model and the RVM model are jointly trained.

2. The method according to claim 1, characterized in that The multiple types of characteristic data that affect the photovoltaic power generation power corresponding to the sample time period include at least two types of characteristic data as follows: The photovoltaic power generation power, temperature, air pressure, humidity, photovoltaic module temperature, total radiation, direct radiation, and scattered radiation in multiple time periods before the sample time period.

3. The method according to claim 1, characterized in that The obtaining of the sample data set comprises: Acquire an original data set, the original data set including an original photovoltaic power generation power in a sample time period and multiple types of original feature data affecting the photovoltaic power generation power corresponding to the sample time period; The data in the original data set is preprocessed to obtain the sample data set, wherein the preprocessing includes at least one of outlier processing, missing value processing, and normalization processing.

4. The method according to claim 1, characterized in that: The method further comprises: Acquire a training sample from the sample data set, the training sample including a true value of photovoltaic power generation in a sample time period and important type of feature data affecting the photovoltaic power generation corresponding to the sample time period; The training samples are used to train a CNN-RVM model, wherein the CNN-RVM model includes the CNN model and the RVM model.

5. The method according to claim 4, characterized in that The method of using the training samples to train a CNN-RVM model includes: The grid search method is used to tune the three parameters of the number of neurons, convolution kernel size and learning rate in the pre-trained CNN-RVM model; The mean square error is used as the optimization target of the CNN-RVM model, and the formula of the mean square error is as follows: Where T represents the number of samples in the training sample, y i represents the photovoltaic power prediction value corresponding to the i-th sample output by the CNN-RVM model, represents the true value of the photovoltaic power generation power in the i-th sample.

6. The method according to claim 4, characterized in that The initial settings of the model parameters in the CNN model include: the first convolution layer is initially set to 64 neurons, the second convolution layer is initially set to 32 neurons, the third convolution layer is initially set to 16 neurons, the window size of each convolution layer is initially set to 1×3, the pooling layer connected to each convolution layer is a maximum pooling layer, the window size of the pooling layer is initially set to 2×2, the step size is initially set to 2, and the activation function is a linear rectification function; The penalty coefficient γ of the RVM model is initially set to 0.1, and the kernel function of the RVM model is a Gaussian function; The batch size of the CNN-RVM model is initially set to 120, the initial learning rate is initially set to 0.01, and the initial number of training times is initially set to 500.

7. The method according to claim 4, characterized in that The method further comprises: Acquire samples other than the training samples from the sample data set as test samples; Using the test sample, testing the CNN-RVM model trained using the training sample; If the test passes, the trained CNN-RVM model is the trained CNN-RVM model; If the test fails, continue to use the CNN-RVM model trained with the training samples.

8. A method for training a photovoltaic power prediction model, characterized in that: The method comprises: Acquire a sample data set, the sample data set including a true value of photovoltaic power generation in a sample time period and multiple types of characteristic data affecting the photovoltaic power generation corresponding to the sample time period; Based on the sample data set, a random forest algorithm is used to determine the important types that affect the photovoltaic power generation power from the multiple types; Acquire a training sample from the sample data set, the training sample including a true value of photovoltaic power generation in a sample time period and important type of feature data affecting the photovoltaic power generation corresponding to the sample time period; The training samples are used to train a CNN-RVM model, wherein the CNN-RVM model includes a convolutional neural network (CNN) model and a relevance vector machine (RVM) model, wherein the CNN model is used to extract feature vectors from feature data of important types corresponding to a prediction time period, and the RVM model is used to predict the photovoltaic power generation power of the prediction time period based on the feature vectors extracted by the CNN model.

9. The method according to claim 8, characterized in that The method of using the training samples to train a CNN-RVM model includes: The grid search method is used to optimize the three parameters of the number of neurons, convolution kernel size and learning rate in the pre-training CNN-RVM model, and the mean square error is used as the optimization target of the CNN-RVM model. The formula of the mean square error is as follows: Where T represents the number of samples in the training sample, y i represents the photovoltaic power prediction value corresponding to the i-th sample output by the CNN-RVM model, represents the true value of the photovoltaic power generation power in the i-th sample.

10. A photovoltaic power prediction device, characterized in that: The device is applied to a feeder terminal, and comprises: A sample acquisition module is configured to acquire a sample data set, wherein the sample data set includes a true value of photovoltaic power generation in a sample time period and multiple types of characteristic data affecting the photovoltaic power generation corresponding to the sample time period; A data type determination module is configured to determine, based on the sample data set, an important type that affects the photovoltaic power generation power from the multiple types using a random forest algorithm; A data acquisition module is configured to acquire important types of feature data corresponding to a plurality of historical time periods before the prediction time period; A feature extraction module is configured to input feature data of important types corresponding to multiple historical time periods before the predicted time period into a pre-trained convolutional neural network (CNN) model to obtain a feature vector output by the CNN model; The prediction module is configured to input the feature vector output by the CNN model into a pre-trained relevance vector machine (RVM) model to obtain the photovoltaic power generation power of the predicted time period output by the RVM model; wherein the CNN model and the RVM model are jointly trained.

11. The device according to claim 10, characterized in that The multiple types of characteristic data that affect the photovoltaic power generation power corresponding to the sample time period include at least two types of characteristic data as follows: The photovoltaic power generation power, temperature, air pressure, humidity, photovoltaic module temperature, total radiation, direct radiation, and scattered radiation in multiple time periods before the sample time period.

12. The device according to claim 10, characterized in that The sample acquisition module is configured as follows: Acquire an original data set, wherein the original data set includes the original photovoltaic power generation power of a sample time period and various types of original feature data corresponding to the sample time period that affect the photovoltaic power generation power; preprocess the data in the original data set to obtain the sample data set, wherein the preprocessing includes at least one of outlier processing, missing value processing, and normalization processing.

13. The device according to claim 10, characterized in that The device also includes: The model training module is configured to obtain training samples from the sample data set, wherein the training samples include the true value of the photovoltaic power generation power in the sample time period and the important type of feature data that affects the photovoltaic power generation power corresponding to the sample time period; and use the training samples to train a CNN-RVM model, wherein the CNN-RVM model includes the CNN model and the RVM model.

14. The device according to claim 13, characterized in that The model training module uses the training samples to train the CNN-RVM model, and the part is configured as follows: The grid search method is used to tune the three parameters of the number of neurons, convolution kernel size and learning rate in the pre-trained CNN-RVM model; The mean square error is used as the optimization target of the CNN-RVM model, and the formula of the mean square error is as follows: Where T represents the number of samples in the training sample, y i represents the photovoltaic power prediction value corresponding to the i-th sample output by the CNN-RVM model, represents the true value of the photovoltaic power generation power in the i-th sample.

15. The device according to claim 13, characterized in that The initial settings of the model parameters in the CNN model include: the first convolution layer is initially set to 64 neurons, the second convolution layer is initially set to 32 neurons, the third convolution layer is initially set to 16 neurons, the window size of each convolution layer is initially set to 1×3, the pooling layer connected to each convolution layer is a maximum pooling layer, the window size of the pooling layer is initially set to 2×2, the step size is initially set to 2, and the activation function is a linear rectification function; The penalty coefficient γ of the RVM model is initially set to 0.1, and the kernel function of the RVM model is a Gaussian function; The batch size of the CNN-RVM model is initially set to 120, the initial learning rate is initially set to 0.01, and the initial number of training times is initially set to 500.

16. The device according to claim 13, characterized in that The device also includes: The testing module is configured to obtain samples other than the training samples from the sample data set as test samples; use the test samples to test the CNN-RVM model trained using the training samples; if the test passes, the trained CNN-RVM model is the trained CNN-RVM model; if the test fails, continue to use the CNN-RVM model trained using the training samples.

17. A photovoltaic power prediction model training device, characterized in that: The device comprises: A sample data acquisition module is configured to acquire a sample data set, wherein the sample data set includes a true value of photovoltaic power generation in a sample time period and multiple types of characteristic data affecting the photovoltaic power generation corresponding to the sample time period; An important type determination module is configured to determine the important type that affects the photovoltaic power generation power from the multiple types using a random forest algorithm based on the sample data set; A training sample acquisition module is configured to acquire training samples from the sample data set, wherein the training samples include a true value of photovoltaic power generation in a sample time period and important type of feature data that affects the photovoltaic power generation corresponding to the sample time period; A training module is configured to use the training samples to train a CNN-RVM model, wherein the CNN-RVM model includes a convolutional neural network (CNN) model and a relevance vector machine (RVM) model, wherein the CNN model is used to extract feature vectors from feature data of important types corresponding to a prediction time period, and the RVM model is used to predict the photovoltaic power generation power of the prediction time period based on the feature vectors extracted by the CNN model.

18. The device according to claim 17, characterized in that The training module is configured as follows: The grid search method is used to optimize the three parameters of the number of neurons, convolution kernel size and learning rate in the pre-training CNN-RVM model, and the mean square error is used as the optimization target of the CNN-RVM model. The formula of the mean square error is as follows: Where T represents the number of samples in the training sample, y i represents the photovoltaic power prediction value corresponding to the i-th sample output by the CNN-RVM model, represents the true value of the photovoltaic power generation power in the i-th sample.

19. A feeder terminal, characterized in that: The method comprises a memory and a processor, wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the method according to any one of claims 1 to 9.

20. A power distribution system, characterized in that: Comprising the feeder terminal as claimed in claim 19.

21. A readable storage medium, characterized in that: Computer instructions are stored thereon, and when the computer instructions are executed by a processor, the method steps described in any one of claims 1 to 9 are implemented.

22. A chip, characterized in that: The chip comprises the device according to any one of claims 10-18.

Images