CNN-BIGRU photovoltaic power generation prediction method and device based on clustering-decomposition

Abstract

The invention discloses a CNN-BIGRU photovoltaic power generation prediction method and device based on clustering-decomposition. The method comprises the following steps: collecting original data; preprocessing the original data; inputting a density peak clustering algorithm, and automatically identifying a clustering center by calculating a local density rho and a relative distance delta to obtain a plurality of representative weather scene clusters; for each scene cluster, extracting a corresponding power sequence; generating a high-frequency sub-sequence and a low-frequency sub-sequence; a shared CNN-BiGRU network architecture is constructed; training is carried out by adopting a small-batch training strategy of different scenes and different scales; rolling prediction is completed, and prediction results of the high-frequency subsequences and the low-frequency subsequences are reconstructed and synthesized into a final prediction value; and an output result is subjected to reverse normalization processing and then is used as a photovoltaic power prediction value. The technical problems that an existing photovoltaic power prediction method is low in prediction precision and poor in model robustness under the changeable weather condition are solved, and particularly the problem that errors are suddenly increased under the weather sudden change scenes such as cloudy and sunny-to-rainy weather is solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of photovoltaic power generation prediction, and particularly relates to a CNN-BIGRU photovoltaic power generation prediction method and device based on clustering-decomposition. BACKGROUND

[0002] Photovoltaic power prediction is an important supporting technology for new energy power system dispatching and operation, and its accuracy directly affects the safety and economic benefits of the power grid. Existing methods mainly include physical mechanism-based models, statistical regression models and machine learning models, among which deep learning is widely used due to its advantages in nonlinear feature extraction and time series modeling. In recent years, scholars have proposed combining weather scene clustering with deep neural networks to improve the prediction accuracy of photovoltaic output under different weather conditions.

[0003] Support Vector Machine (SVM) is a classification and regression method based on the principle of maximum interval, suitable for small sample and high-dimensional feature prediction; Long Short-Term Memory Network (LSTM) is an improved recurrent neural network that effectively captures long-term dependencies in time series by introducing forget gates, input gates and output gates, and is widely used in wind power, photovoltaic and other new energy power prediction Physical mechanism model: based on the physical laws of solar radiation transmission, component temperature characteristics and photovoltaic component electrical characteristics, combined with geographical location, weather conditions and equipment parameters, the photovoltaic power output is calculated through mathematical equations, and the advantage is strong interpretability, but it is highly dependent on input parameters and model accuracy.

[0004] The disadvantages of the prior art are as follows: (1) Insufficient weather scene differentiation: existing methods often directly feed all historical samples into a single model, which cannot effectively identify the difference features of different weather scenes such as sunny, cloudy, rainy, etc., resulting in poor model generalization ability.

[0005] (2) Inappropriate processing of multi-time scale features: photovoltaic power contains both high-frequency fluctuations (minute level) and long-term trends (hour level), and traditional methods are difficult to simultaneously consider features of different time scales, especially in sudden change scenarios such as cloudy, sunny to rainy, etc., with a sharp increase in error.

[0006] (3) Long sequence gradient dissipation problem: traditional CNN-RNN structure is sensitive to gradient dissipation when processing long time series, and is unstable in training, making it difficult to capture long-term time series dependencies.

[0007] (4) Poor model adaptability: lack of adaptive modeling mechanism for different weather conditions, resulting in insufficient prediction accuracy and robustness under complex weather change conditions. SUMMARY

[0008] The application provides a clustering-decomposition-based CNN-BIGRU photovoltaic power generation prediction method and device, aiming to solve the technical problems of low prediction accuracy and poor model robustness of existing photovoltaic power prediction methods under variable weather conditions, especially the problem of sudden increase in error under weather mutation scenes such as cloudy, sunny to rainy, and the problems of gradient dissipation and unstable training of traditional deep learning models when processing long sequences.

[0009] In order to achieve the above-mentioned purpose, the application adopts the following technical solutions: The clustering-decomposition-based CNN-BIGRU photovoltaic power generation prediction method comprises the following steps: 1) Collect and synchronize photovoltaic output, hourly irradiance, air temperature and humidity original data; 2) Linear interpolation and denoising preprocessing are performed on missing and abnormal values in the original data; 3) Based on the preprocessed data, a weather vector is constructed, the irradiance curve and meteorological statistics in the sliding window are combined to form a multi-dimensional feature vector; a density peak clustering algorithm is input, the local density p and the relative distance d are calculated, the clustering center is automatically identified, and a plurality of representative weather scene clusters are obtained; 4) Multi-scale signal decomposition, for each scene cluster, the corresponding power sequence is extracted; adaptive decomposition is performed using complete ensemble empirical mode decomposition to obtain a plurality of intrinsic mode functions IMF1, IMF2,..., IMF n and a residual term R; according to the energy contribution degree, the high-frequency mode (IMF1-IMF3) and the low-frequency mode (IMF4-IMF n + R) are combined respectively to generate high-frequency sub-sequences and low-frequency sub-sequences; 5) A shared CNN-BiGRU network architecture is constructed for each high-frequency sub-sequence and low-frequency sub-sequence; 6) Model training strategy, for the constructed CNN-BiGRU network architecture, a small batch training strategy is adopted for training; 7) Rolling prediction reasoning, according to real-time meteorological observation data, the distance from each weather cluster center is calculated to quickly determine the belonging weather cluster; the corresponding CEEMDAN decomposer and CNN-BiGRU network architecture model weight are called; rolling prediction is completed, the prediction results of each high-frequency sub-sequence and low-frequency sub-sequence are reconstructed to form the final prediction value; and the output result is subjected to inverse normalization processing to serve as the photovoltaic power prediction value.

[0010] The application further improves that in step 1), the sampling frequency of the original data is 1 hour.

[0011] The application further improves that in step 2), 3σ criterion is used to identify abnormal values: when |x i- when μ+3σ, it is determined as an abnormal value, wherein xi represents the numerical value of the i th data point, μ represents the mean value of the data sequence, and σ represents the standard deviation of the data sequence.

[0012] The further improvement of the present application is that in step 3), the window size of the sliding window is 24 hours; and the meteorological statistics are the mean value, variance and gradient of the irradiance data in the sliding window.

[0013] The further improvement of the present application is that in step 3), the core formula of the density peak clustering is: Local density: Relative distance:

[0014] The further improvement of the present application is that in step 4), the energy contribution degree calculation formula is:

[0015] The further improvement of the present application is that in step 4), a shared CNN-BiGRU network architecture is constructed for each high-frequency sub-sequence and low-frequency sub-sequence, including: Convolution layer: one-dimensional convolution kernel is used to extract local shape features and reduce dimension, and the kernel size of the one-dimensional convolution kernel is 3 and the step is 1; Pooling layer: maximum pooling is adopted to reduce the parameter amount; BiGRU layer: the convolution feature sequence is sent into the bidirectional GRU, the number of hidden units is 64, and the parallel gating mechanism is used to capture the forward and reverse long-term dependencies; Fully connected layer: a time-distributed fully connected layer is connected, and the power prediction of future multiple time steps is output.

[0016] The further improvement of the present application is that in step 6), for the constructed CNN-BiGRU network architecture, a small-batch training strategy is adopted, including: The batch size is 32; the Adam optimizer is used, the initial learning rate is 0.001, and the cosine annealing scheduling is adopted; the loss function adopts a weighted combination of root mean square error RMSE and mean absolute error MAE: The early stopping mechanism is set, and the training is stopped when the validation set loss does not decrease for 10 consecutive epochs.

[0017] The CNN-BIGRU photovoltaic power generation prediction device based on clustering-decomposition comprises: A data collection unit collects and synchronizes photovoltaic output, hourly irradiance, temperature and humidity original data; A data processing unit performs linear interpolation and denoising preprocessing on missing and abnormal values in the original data; ​​​​The data calculation unit constructs a weather vector based on the preprocessed data, combines the irradiance curve and meteorological statistics in the sliding window to form a multi-dimensional feature vector, inputs a density peak clustering algorithm, automatically identifies a clustering center by calculating a local density p and a relative distance d, and obtains a plurality of representative weather scene clusters; The data decomposition unit decomposes a multi-scale signal, extracts a corresponding power sequence for each scene cluster, performs adaptive decomposition using a complete set empirical mode decomposition to obtain a plurality of intrinsic mode functions IMF1, IMF2,..., IMF n and a residual term R, respectively merges high-frequency modes (IMF1-IMF3) and low-frequency modes (IMF4-IMF n + R) according to energy contribution degrees, and generates high-frequency sub-sequences and low-frequency sub-sequences; The CNN-BiGRU network architecture construction unit constructs a shared CNN-BiGRU network architecture for each high-frequency sub-sequence and low-frequency sub-sequence; The model training strategy unit trains the constructed CNN-BiGRU network architecture by adopting a small-batch training strategy in a scene-by-scene and scale-by-scale manner; The rolling prediction reasoning unit calculates the distance from real-time meteorological observation data to each weather cluster center, quickly determines the belonging weather cluster, calls the corresponding CEEMDAN decomposer and CNN-BiGRU network architecture model weight, completes rolling prediction, reconstructs the prediction results of each high-frequency sub-sequence and low-frequency sub-sequence into a final prediction value, and outputs the result after inverse normalization as a photovoltaic power prediction value.

[0018] A computer-readable storage medium stores a computer program, and the computer program, when executed by a processor, implements the steps of the cluster-decomposition-based CNN-BIGRU photovoltaic power generation prediction method.

[0019] Compared with the prior art, the present application has at least the following beneficial technical effects: Significantly improve the prediction accuracy: through weather scene classification modeling, compared with the single CNN-GRU method with the closest performance, the MAE is reduced by 34.4%, the RMSE is reduced by 37.2%, the MAPE is reduced by 33.0%, and the R 2 is improved by 2.6%. Compared with the traditional LSTM method, the average improvement of each index is about 66%, which fully verifies the effectiveness of the weather scene clustering and signal decomposition strategy.

[0020] Enhance the robustness of the model: multi-scale signal decomposition effectively separates high-frequency noise and low-frequency trends, and improves the adaptability of the model to different weather conditions.

[0021] Good adaptability: through automatic weather scene recognition, no manual intervention is needed to adapt to the climate characteristics of different regions and different seasons. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the specific embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0023] Figure 1 The flow chart of the present application based on clustering-decomposition CNN-BIGRU photovoltaic power generation prediction method; Figure 2 The timing comparison chart of the embodiment of the present application; Figure 3 The prediction value vs. actual value schematic diagram of the embodiment of the present application; Figure 4 The error timing change schematic diagram of the embodiment of the present application; Figure 5 The structural block diagram of the present application based on clustering-decomposition CNN-BIGRU photovoltaic power generation prediction device. DETAILED DESCRIPTION

[0024] In the following, only some exemplary embodiments are simply described. As those skilled in the art can recognize, the described embodiments can be modified in various different ways without departing from the spirit or scope of the present application. Therefore, the drawings and the description are considered to be exemplary in nature rather than limiting.

[0025] In the description of the present application, it should be understood that when used in the specification and the appended claims, the terms "comprise" and "include" indicate the presence of described features, whole, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, whole, steps, operations, elements, components and / or sets thereof.

[0026] It should also be understood that the terms used in the present application specification are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the present application specification and the appended claims, the singular forms "a", "an" and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0027] It should be further understood that the term "and / or" as used herein in the specification and in the claims, if any, means any one of the associated listed items or a combination of them. It should be further understood that the term "comprising" as used herein in the specification and in the claims, if any, means "consisting of" or "consisting essentially of" the listed items.

[0028] Various structural diagrams according to the disclosed embodiments of the present application are shown in the accompanying drawings. These diagrams are not drawn to scale, in which certain details are exaggerated for clarity and others are omitted. The shapes and relative sizes of the various regions, layers, and their relative positions shown in the diagrams are merely exemplary, and in actuality can be deviated due to manufacturing tolerances or technical limitations, and regions / layers with different shapes, sizes, and relative positions can be additionally designed by those skilled in the art according to actual needs.

[0029] The embodiments of the present application are described in detail below in conjunction with the accompanying drawings.

[0030] Embodiment 1 As shown in 1, the present application provides a CNN-BIGRU photovoltaic power generation prediction method based on clustering-decomposition, which comprises: 1) Collect and synchronize the original data of photovoltaic output, hourly irradiance, air temperature, humidity, etc., and the data sampling frequency is 1 hour.

[0031] 2) Linear interpolation and denoising processing are performed on missing and abnormal values, and the 3σ criterion is used to identify abnormal values: when |x i - μ|>3σ, it is determined as an abnormal value 3) Construct a weather vector, based on the preprocessed data, construct a weather vector, combine the irradiance curve in the sliding window (window size is 24 hours) and the meteorological statistics (mean, variance, gradient) of the irradiance data in the sliding window to form a multi-dimensional feature vector; input the density peak clustering algorithm, automatically identify the clustering center by calculating the local density p and the relative distance d; obtain several representative weather scenario clusters, such as sunny weather, cloudy weather, and rainy weather.

[0032] The core formula of density peak clustering is: Local density: ; Relative distance: .

[0033] 4) Multi-scale signal decomposition.

[0034] For each scenario cluster, extract the corresponding power sequence; use complete ensemble empirical mode decomposition (CEEMDAN) for adaptive decomposition to obtain several intrinsic mode functions IMF1, IMF2,..., IMF nand a residual term R; high frequency modes (IMF1-IMF3) and low frequency modes (IMF4-IMF n + R) are merged according to the energy contribution degree, to generate high frequency sub-sequences and low frequency sub-sequences. The energy contribution degree calculation formula is: .

[0035] 5) A shared CNN-BiGRU network architecture is constructed for each high frequency sub-sequence and low frequency sub-sequence: Convolutional layer: one-dimensional convolution kernel (kernel size 3, step 1) is used to extract local shape features and reduce dimension; Pooling layer: maximum pooling is used to reduce the number of parameters; BiGRU layer: the convolutional feature sequence is fed into the bidirectional GRU with 64 hidden units to capture the forward and backward long-term dependencies with a parallel gating mechanism; Fully connected layer: a time-distributed fully connected layer is connected to output the power prediction for future multiple time steps (prediction length 4 hours).

[0036] 6) Model training strategy The CNN-BiGRU network architecture constructed in step 5) is optimized and trained, which includes the complete processing flow of convolutional layer feature extraction, pooling layer dimension reduction, BiGRU layer time series modeling, and fully connected layer output prediction A small batch training strategy is adopted according to the scene and scale, with a batch size of 32; the Adam optimizer is used with an initial learning rate of 0.001 and cosine annealing scheduling; the loss function uses a weighted combination of root mean square error (RMSE) and mean absolute error (MAE):

[0037] An early stopping mechanism is set to stop training when the validation set loss does not decrease for 10 consecutive epochs to avoid overfitting.

[0038] 7) Rolling prediction inference According to the real-time meteorological observation data, the distance to each weather cluster center is calculated to quickly determine the belonging weather cluster; the corresponding CEEMDAN decomposer and CNN-BiGRU network architecture (through the processing flow of convolutional layer → pooling layer → BiGRU layer → fully connected layer) model weight is called; the rolling prediction is completed, and the prediction results of each high frequency sub-sequence and low frequency sub-sequence are reconstructed to form the final prediction value; the output result is processed by inverse normalization to obtain the photovoltaic power prediction value.

[0039] 8) To verify the effectiveness and accuracy of the above clustering-decomposition CNN-BIGRU photovoltaic power prediction method, the relevant evaluation indicators are as follows

[0040]

[0041]

[0042]

[0043] where n is the number of samples, yi is the ith actual value, and yi is the ith predicted value.

[0044] Example 2: This experiment takes the photovoltaic power prediction of the Australian Desert Knowledge Solar Centre as an example, and uses the data of station No. 7 in Albury, Australia (latitude: -23.76, longitude: 133.87). The data time range is from March 1, 2019 to March 1, 2023, a total of 4 years of actual operation data, with a sampling frequency of 1 hour, a total of 35040 sample points.

[0045] After data preprocessing, the training set, validation set and test set are divided according to the ratio of 7:2:1. Through box plot analysis, it is determined that the power exceeding 110% of the installed capacity or less than 0 is abnormal data, and if it exceeds the range continuously for more than 2 hours, interpolation processing is performed. The data set features are shown in Table 1 Table 1. Measuring point data

[0046] According to the above data, the photovoltaic power is predicted by the method of the present application, and the test results of Table 2 and Figure 2 , Figure 3 , Figure 4 are obtained.

[0047] Table 2

[0048] Through the analysis of Figure 2 , Figure 3 , Figure 4 , it can be found that the predicted value and the actual value of the present application have high consistency, and more than 90% of the prediction error is controlled within ±2kW. Figure 2 The prediction results of 500 time points are shown in the time sequence comparison, the blue line is the actual value, and the yellow line is the predicted value. It can be seen that the two curves are highly coincident, especially in the power peak and trough periods, which can be accurately tracked. Figure 3 The scatter plot of predicted value vs. actual value is shown, and the data points are basically distributed near the y=x straight line, and the correlation coefficient R 2 reaches 0.981, indicating that the prediction accuracy is very high. Figure 4 The prediction error time sequence chart is shown, and it can be seen that the error is mainly concentrated in the range of ±1.5kW, and there is no obvious systematic deviation, indicating that the model has good unbiasedness.

[0049] Example 3 like Figure 5 As shown, the CNN-BIGRU photovoltaic power generation prediction device based on clustering-decomposition provided by the present invention includes: The data collection unit collects and synchronizes raw data on photovoltaic output, hourly irradiance, temperature, and humidity. The data processing unit performs linear interpolation and noise reduction preprocessing on missing and outlier values ​​in the raw data. The data calculation unit constructs weather vectors based on preprocessed data, combining irradiance curves and meteorological statistics within a sliding window to form a multidimensional feature vector; it inputs density peak clustering algorithm, which automatically identifies cluster centers by calculating local density ρ and relative distance δ, and obtains several representative weather scene clusters; The data decomposition unit performs multi-scale signal decomposition, extracting the corresponding power sequence for each scene cluster; adaptive decomposition is then performed using complete ensemble empirical mode decomposition to obtain several intrinsic mode functions (IMF1, IMF2, ..., IMF). n And a residual term R; based on energy contribution, high-frequency modes (IMF1-IMF3) and low-frequency modes (IMF4-IMF3) are separated. n + R) are merged separately to generate high-frequency subsequences and low-frequency subsequences; The CNN-BiGRU network architecture building unit constructs a shared CNN-BiGRU network architecture for each high-frequency subsequence and low-frequency subsequence; The model training strategy unit trains the constructed CNN-BiGRU network architecture using a scenario- and scale-specific mini-batch training strategy. The rolling prediction inference unit calculates the distance to the center of each weather cluster based on real-time meteorological observation data, quickly determines the weather cluster to which it belongs, calls the corresponding CEEMDAN decomposer and CNN-BiGRU network architecture model weights, completes the rolling prediction, reconstructs and synthesizes the prediction results of each high-frequency subsequence and low-frequency subsequence into the final prediction value, and outputs the result as the photovoltaic power prediction value after inverse normalization.

[0050] Example 4 The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the CNN-BIGRU photovoltaic power generation prediction method based on clustering-decomposition.

[0051] Those skilled in the art will appreciate that embodiments of the application can be readily used as software, hardware, or a combination of software and hardware. In one embodiment, the application can be implemented in software and / or firmware. In this embodiment, the software implementation can include computer readable code stored in a computer readable storage medium (alternatively referred to as a computer readable medium, a processor readable storage medium, or a processor readable code) that, when taken in whole or in part, can program one or more computer processors (alternatively referred to as computer processors, or processor) to perform one or more embodiments of the application. The computer readable storage medium can be tangible and include any mechanism for storing information in a form readable by one or more computer processors. For example, a computer readable storage medium can include read only memory (ROM); random access memory (RAM); magnetic disk storage media; optical storage media; flash memory devices; electrical, optical, acoustical or other form of propagated signals (e.g., carrier waves, infrared signals, digital signals, etc.); etc.

[0052] The present application is described in reference to the drawings, which are as follows: Figure 1 Figure 1

[0053] The computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions / acts specified in the flowchart block or blocks. Figure 1 Figure 1

[0054] The computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions / acts specified in the flowchart block or blocks. Figure 1 Figure 1

[0055] The core technical points of the application are to construct a "weather scene clustering-signal multi-scale decomposition-CNN-BiGRU" three-level processing framework, to automatically identify weather scenes through density peak clustering, to process multi-time scale features through CEEMDAN decomposition, to extract space-time dependence through CNN-BiGRU network, and to realize high-precision and strong-robust photovoltaic power prediction.

[0056] ​​​​​​The scope of protection of the present application: the protection scope of the present application includes the above-mentioned complete technical scheme, the specific implementation method of each processing module, and the related data preprocessing and post-processing technology.

[0057] The basic principles and main features of the present application are shown and described above, and the advantages of the present application are also shown and described above. It is obvious for those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-limiting, the scope of the present application is defined by the appended claims rather than the above description, and therefore all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present application. Any reference signs in the claims should not be regarded as limiting the claims involved.

[0058] In addition, it should be understood that although the present application is described in terms of embodiments, not every embodiment contains only one independent technical solution, and the description of the specification is only for the sake of clarity, those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be combined appropriately to form other embodiments that those skilled in the art can understand. The above is only to illustrate the technical idea of the present application, and cannot limit the protection scope of the present application. Any modification made on the basis of the technical solutions according to the technical idea of the present application falls within the protection scope of the claims of the present application.

Claims

1. A photovoltaic power generation prediction method based on clustering-decomposition-CNN-BIGRU, characterized in that, Comprising: 1) Collect and synchronize photovoltaic output, hourly irradiance, air temperature and humidity raw data; 2) Linear interpolation and denoising preprocessing for missing and abnormal values in raw data; 3) Based on the preprocessed data, construct a weather vector, combine the irradiance curve and meteorological statistics in the sliding window to form a multi-dimensional feature vector; input the density peak clustering algorithm, automatically identify the cluster center by calculating the local density p and the relative distance d, and obtain several representative weather scenario clusters; 4) Multi-scale signal decomposition, for each scene cluster, extract the corresponding power sequence; use complete set of empirical mode decomposition for adaptive decomposition, obtain several intrinsic mode functions IMF1, IMF2,..., IMF n and a residual term R; according to the energy contribution, the high frequency mode (IMF1-IMF3) and the low frequency mode (IMF4-IMF n + R) are respectively combined to generate high frequency subsequence and low frequency subsequence; 5) Build a shared CNN-BiGRU network architecture for each high-frequency sub-sequence and low-frequency sub-sequence; 6) Model training strategy, for the constructed CNN-BiGRU network architecture, adopt the small batch training strategy for scene and scale; 7) Rolling prediction inference, according to the real-time meteorological observation data, calculate the distance from each weather cluster center, quickly determine the belonging weather cluster; call the corresponding CEEMDAN decomposer and CNN-BiGRU network architecture model weight; Complete the rolling prediction, reconstruct the prediction results of each high-frequency sub-sequence and low-frequency sub-sequence into the final prediction value; The output result after inverse normalization processing is used as the photovoltaic power prediction value.

2. The cluster-decomposition-based CNN-BIGRU photovoltaic power generation prediction method according to claim 1, characterized in that, In step 1), the sampling frequency of the raw data is 1 hour. 3.The cluster-decomposition based CNN-BIGRU photovoltaic power generation prediction method according to claim 1, characterized in that, In step 2), outliers are identified using the 3σ criterion: when |x i - μ| > 3σ, where xi represents the value of the ith data point, μ represents the mean of the data sequence, and σ represents the standard deviation of the data sequence. 4.The cluster-decomposition based CNN-BIGRU photovoltaic power generation prediction method according to claim 1, characterized in that, In step 3), the window size of the sliding window is 24 hours; the meteorological statistics are the mean, variance and gradient of the irradiance data in the sliding window. 5.The cluster-decomposition based CNN-BIGRU photovoltaic power generation prediction method according to claim 1, characterized in that, In step 3), the core formula of density peak clustering is: Local density: ; Relative distance: . 6.The cluster-decomposition based CNN-BIGRU photovoltaic power generation prediction method according to claim 1, characterized in that, In step 4), the energy contribution degree calculation formula is: . 7.The cluster-decomposition based CNN-BIGRU photovoltaic power generation prediction method according to claim 1, characterized in that, In step 4), a shared CNN-BiGRU network architecture is built for each high-frequency sub-sequence and low-frequency sub-sequence, including: Convolution layer: use one-dimensional convolution kernel to extract local shape features and reduce dimension, the kernel size of one-dimensional convolution kernel is 3, and the step is 1; Pooling layer: use maximum pooling to reduce parameter quantity; BiGRU layer: input the convolution feature sequence into bidirectional GRU, the number of hidden units is 64, and the parallel gating mechanism is used to capture forward and backward long-term dependencies; Fully connected layer: connect the time-distributed fully connected layer, output the power prediction of future multiple time steps. 8.The cluster-decomposition based CNN-BIGRU photovoltaic power generation prediction method according to claim 1, characterized in that, In step 6), for the constructed CNN-BiGRU network architecture, adopt the small batch training strategy for scene and scale, including: The batch size is 32; the Adam optimizer is used, the initial learning rate is 0.001, the cosine annealing schedule is used; the loss function adopts a weighted combination of the root mean square error RMSE and the mean absolute error MAE: ; Set early stopping mechanism, stop training when the validation set loss does not decrease for 10 consecutive epochs.

9. The cluster-decomposition based CNN-BIGRU photovoltaic power generation prediction device, characterized in that, Comprising: Data collection unit, collecting and synchronizing photovoltaic output, hourly irradiance, air temperature and humidity raw data; Data processing unit, linear interpolation and denoising preprocessing for missing and abnormal values in raw data; Data calculation unit, based on the preprocessed data, construct a weather vector, combine the irradiance curve and meteorological statistics in the sliding window to form a multi-dimensional feature vector; Input the density peak clustering algorithm, automatically identify the cluster center by calculating the local density p and the relative distance d, and obtain several representative weather scenario clusters; A data decomposition unit, multi-scale signal decomposition, for each scene cluster, extract the corresponding power sequence; using complete set empirical mode decomposition for adaptive decomposition, obtain a number of intrinsic mode function IMF1, IMF2,..., IMF n And a residual term R; according to the energy contribution degree, the high frequency mode (IMF1-IMF3) and the low frequency mode (IMF4-IMF n + R) are respectively combined to generate high frequency subsequence and low frequency subsequence; CNN-BiGRU network architecture building unit, build a shared CNN-BiGRU network architecture for each high-frequency sub-sequence and low-frequency sub-sequence; Model training strategy unit, for the constructed CNN-BiGRU network architecture, adopt the small batch training strategy for scene and scale; The rolling prediction inference unit calculates the distance from each weather cluster center according to real-time meteorological observation data, quickly determines the belonging weather cluster, and calls the corresponding CEEMDAN decomposer and CNN-BiGRU network architecture model weight; The rolling prediction is completed, and the final prediction value is reconstructed by combining the prediction results of each high-frequency subsequence and low-frequency subsequence; The output result is subjected to inverse normalization processing and used as a photovoltaic power prediction value.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program realizes the steps of the clustering-decomposition-based CNN-BIGRU photovoltaic power generation prediction method in any one of claims 1-8 when executed by the processor.

Images