Photovoltaic power generation power prediction method and device, equipment and storage medium

Abstract

The invention discloses a photovoltaic power generation power prediction method, and the method comprises the steps: obtaining photovoltaic historical operation data, and carrying out the preprocessing of the data, and obtaining the enhanced meteorological data and numerical prediction data of a photovoltaic station; deep feature extraction is carried out on the enhanced meteorological data and numerical forecasting data, and preliminary feature information is captured; performing adaptive weighted correction on the captured preliminary feature information to obtain corrected depth feature information; carrying out decoding operation on the depth feature information, and carrying out real-time dynamic noise reduction on data output by decoding to obtain meteorological correction data; and taking the meteorological correction data obtained after noise reduction as input data of the time sequence model to obtain a photovoltaic power prediction result. According to the method, the precision of photovoltaic power prediction can be remarkably improved, the operation and maintenance management of the power grid can be optimized, and the integration and utilization efficiency of renewable energy sources is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of photovoltaic power technology, and in particular to a photovoltaic power prediction method, device, electronic device and computer storage medium based on deep learning. Background Art

[0002] Electricity is one of the most important resources in my country. Photovoltaic power generation has become more prominent under the initiative of clean energy in my country. However, due to the randomness and volatility of people's demand for electricity, power system equipment may cause certain economic losses.

[0003] The traditional photovoltaic power generation prediction technologies mainly include: ① Mathematical statistical method. This method can quickly obtain the corresponding calculation results, but its prediction accuracy is low due to the complexity of the data. ② Deep learning method. In recent years, the relatively popular deep learning algorithm has played a considerable role in the field of time series prediction.

[0004] Since the output power of photovoltaic power is affected by many factors such as photovoltaic radiation, temperature, rainfall, etc., there are certain challenges in the process of photovoltaic power prediction. To address the above problems, the second method mentioned above is used to help find the timing between data.

[0005] However, the weather data (NWP) predicted by the second method is affected by factors such as different regions, different seasons, and different weather conditions, and there are certain data anomalies, which leads to problems such as low accuracy in photovoltaic power prediction. Therefore, directly using NWP data as a training data set for a deep learning model has the problem of low prediction accuracy. Summary of the invention

[0006] One of the purposes of the present invention is to overcome the deficiencies in the prior art and provide a photovoltaic power prediction method, device, equipment and storage medium based on deep learning to solve the technical problems of inaccurate numerical forecast data, low precision of power plants and long prediction time due to the influence of southern microclimate. By effectively correcting meteorological data such as radiation to cope with its deviation problem in complex environments such as southern microclimate, that is, by improving the quality and accuracy of data and optimizing data input conditions, more accurate basic data can be provided for photovoltaic power prediction. Through data correction, not only can the accuracy of photovoltaic power prediction be significantly improved, but also the operation and maintenance of power grids can be optimized, and the integration and utilization efficiency of renewable energy can be enhanced.

[0007] In order to solve the above technical problems, in a first aspect, an embodiment of the present invention provides a photovoltaic power prediction method based on deep learning, the method comprising:

[0008] Obtain historical photovoltaic operation data and preprocess it to obtain enhanced meteorological data and numerical forecast data of photovoltaic sites;

[0009] Perform deep feature extraction on enhanced meteorological data and numerical forecast data to capture preliminary feature information;

[0010] Adaptively weight the captured preliminary feature information to obtain the corrected depth feature information;

[0011] Decoding the depth feature information, and performing real-time dynamic noise reduction on the decoded output data to obtain meteorological correction data;

[0012] The meteorological correction data obtained after noise reduction is used as the input data of the time series model to obtain the photovoltaic power prediction results.

[0013] Preferably, the deep feature extraction of the enhanced meteorological data to capture preliminary feature information specifically includes:

[0014] The DC-CNN convolution module is used to process the input meteorological data and numerical forecast data to obtain the time series-based feature matrix of the convolution output:

[0015] X Out =ReLU(BN(DC(X In ,d)));

[0016] Among them, X In represents the input numerical forecast data and meteorological data of the photovoltaic site, ReLU represents the activation function, BN represents the normalized batch operation, DC is the dilated convolution operation, and d is the dilation rate.

[0017] Preferably, the captured preliminary feature information is adaptively weighted and modified to obtain the modified depth feature information, specifically including:

[0018] ASB is used to combine the time series-based feature matrix obtained by the convolution operation to extract and correct the frequency domain features and obtain deep feature information:

[0019]

[0020] in, It represents the result of Fourier transform of the time series feature data output by convolution. represents the inverse Fourier transform, W1 and W2 represent complex weights, and M adaptive represents the adaptive spectrum mask, Represents the deep feature information output by ASB processing.

[0021] Preferably, the decoding operation of the depth feature information and performing real-time dynamic noise reduction on the decoded output data to obtain the meteorological correction data specifically includes:

[0022] The Decoder decoding structure in the Informer model is used to decode the meteorological correction data obtained by deep information extraction and information adaptive information analysis using ASB in the Encoder layer, and the data after the decoding operation is denormalized to obtain the target prediction result;

[0023] Use Kalman filtering to dynamically adjust target prediction results.

[0024] Furthermore, the method further comprises:

[0025] By changing the loss function, the original historical operation data is corrected, and the changed loss function is formulated as:

[0026]

[0027] Among them, P Mi It is expressed as the available power at power-limited time i or the actual power at unrestricted time i, P Pi is the predicted power at 4 hours, and Cap is the total installed capacity of the photovoltaic power station in that month.

[0028] In addition, to achieve the above-mentioned purpose, the present invention also proposes a photovoltaic power forecasting device, the photovoltaic power device comprising:

[0029] A data preprocessing module is used to obtain and preprocess the historical operation data of photovoltaic power plants to obtain enhanced meteorological data and numerical forecast data of photovoltaic sites;

[0030] Feature extraction module, used to perform deep feature extraction on enhanced meteorological data and numerical forecast data to capture preliminary feature information;

[0031] A feature correction module is used to perform adaptive weighted correction on the captured preliminary feature information to obtain corrected depth feature information;

[0032] A decoding and denoising module, used to decode the depth feature information and perform real-time dynamic denoising on the decoded output data to obtain meteorological correction data;

[0033] The prediction module is used to use the meteorological correction data obtained after noise reduction as the input data of the time series model to obtain the photovoltaic power prediction result.

[0034] In addition, to achieve the above-mentioned purpose, the present invention proposes a storage medium, on which a photovoltaic power prediction program based on deep learning is stored. When the photovoltaic power prediction program based on deep learning is executed by a processor, the steps described in the photovoltaic power prediction method described above are implemented.

[0035] In the third aspect, an embodiment of the present invention also provides a deep learning photovoltaic power prediction device, which includes: a solar panel, an inverter, an electric energy collection device, a data recording device, and also includes a memory, a processor, and a prediction program stored in the memory and executable on the processor, wherein the prediction program is a step for implementing the photovoltaic power prediction method as described above.

[0036] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium for reading, writing and storing by a computer, wherein the computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor executes the photovoltaic power prediction method based on deep learning as described above.

[0037] Compared with the prior art, the photovoltaic power prediction method, device, equipment and storage medium based on deep learning provided by the embodiments of the present invention have at least the following beneficial effects:

[0038] The technical solution of the embodiment of the present invention aims at the problem that when the prior art directly uses NWP data as the training data of the deep learning model, it is affected by the complex and changeable microclimate in the south, and there is a problem of low model prediction accuracy. Therefore, a photovoltaic radiation data correction algorithm model based on the Informer and RF algorithms is proposed. By performing feature enhancement and data cleaning on the data, the time dependency of the original data set is enhanced, and the NWP numerical forecast data in the southern microclimate is corrected. The time complexity is low and the correction accuracy is high. The algorithm constructs the correlation between the radiation value and the power value through the algorithm, and corrects part of the radiation data in the NWP in combination with the Informer model to cope with the complex changes in the southern microclimate, thereby improving the accuracy of short-term and ultra-short-term photovoltaic power prediction and improving the operation efficiency of the model.

[0039] This example obtains the observation data of photovoltaic stations and NWP (numerical weather forecast data), and processes the preprocessed time series data; uses deep information extraction on the photovoltaic time series data to obtain sufficient feature information; on the basis of fully extracting the time series information, obtains more efficient sequence data by using the coefficient self-attention mechanism; removes redundant sequence data by using the data distillation method, reduces the computational complexity, and retains key information. Through the decoding module, the obtained data is converted into prediction results. By improving the prediction accuracy, it can adapt to data in different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly describe the real-time examples of the present invention and highlight the different technical solutions in the prior art, the following will briefly introduce the drawings required for use. The following drawings only show some embodiments of the present invention. Relevant technicians can easily deduce other drawings based on them.

[0041] Figure 1 It is a schematic diagram of the steps of the photovoltaic power prediction method based on deep learning of the present invention;

[0042] Figure 2 It is a schematic diagram of the sub-step flow of the photovoltaic power prediction method based on deep learning of the present invention;

[0043] Figure 3 A schematic flow chart of the correction steps of the photovoltaic power prediction method based on deep learning of the present invention;

[0044] Figure 4 This is a structural block diagram of a first embodiment of a photovoltaic power generation device based on deep learning according to the present invention;

[0045] Figure 5 This is a schematic diagram of the hardware operating environment and equipment structure involved in photovoltaic power prediction based on deep learning in the present invention. DETAILED DESCRIPTION

[0046] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. The following is an example of the present invention described in detail, and those skilled in the art can easily understand the features and advantages of the present invention from the contents disclosed in this specification.

[0047] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments, but it is not intended to limit the present invention.

[0048] The following mainly takes some specific embodiments as examples to describe in detail the implementation of the technical solution of the present invention.

[0049] like Figure 1 As shown, a photovoltaic power prediction method based on deep learning in an embodiment of the present invention comprises:

[0050] S1. Obtain historical photovoltaic operation data and preprocess it to obtain enhanced meteorological data and numerical forecast data of photovoltaic sites;

[0051] S2, perform deep feature extraction on enhanced meteorological data and numerical forecast data to capture preliminary feature information;

[0052] S3, performing adaptive weighted correction on the captured preliminary feature information to obtain corrected depth feature information;

[0053] S4, decoding the depth feature information, and performing real-time dynamic noise reduction on the decoded output data to obtain meteorological correction data;

[0054] S5. The meteorological correction data obtained after noise reduction is used as input data of the time series model to obtain the photovoltaic power prediction result.

[0055] like Figure 1 As shown, it is a flow chart of an embodiment of a photovoltaic power prediction method based on deep learning according to the present invention.

[0056] Among them, in S1, the historical operation data of photovoltaic is obtained and preprocessed to obtain the enhanced meteorological data and numerical forecast data of the photovoltaic site.

[0057] First, the historical photovoltaic power generation data of the power grid department to be extracted is obtained, and the data is enhanced by data preprocessing and other methods, that is, the photovoltaic historical data and NWP numerical forecast data are preprocessed, including time feature extraction and standardization operations to ensure that the data have the same time scale, and data enhancement processing is performed based on the preset data enhancement strategy, as follows:

[0058] Perform data enhancement processing on the acquired original historical data, extract time features based on timestamps from the data in the training set, and use trigonometric functions to convert time into periodic features;

[0059] After standardizing the original data, a sliding window with a length of 16 is defined, and the data in the window is used as the input of the model.

[0060] In step S1, the photovoltaic power-related historical data of the embodiment of the present invention is obtained through ground meteorological station detection, satellite remote sensing data, etc. Specifically, through the monitoring of ground meteorological stations, the meteorological department deploys ground meteorological observation stations near photovoltaic power stations or in power grid-related areas. These observation stations collect meteorological data related to photovoltaic power generation in real time by installing radiometers, anemometers, hygrometers, thermometers and other equipment, such as solar radiation, temperature, wind speed, humidity, etc. Photovoltaic power stations may also have their own meteorological monitoring equipment to share data with meteorological departments or power grid companies.

[0061] In step S1, NWP numerical forecast data is generated by a weather forecast model, which simulates atmospheric movement and combines observation data to predict future weather conditions. NWP data mainly includes meteorological elements related to photovoltaic power generation, such as radiation, temperature, humidity, wind speed, cloud cover, etc. However, for the southern microclimate, due to complex terrain such as mountainous areas, the accuracy of NWP (numerical forecast data) data is usually low, and the prediction quality may be affected by many factors.

[0062] Specifically, the Python script can be used to preprocess the acquired time series data. For example, the first step is to remove the time node data where the radiation value exists in the two parts of the data, but the photovoltaic power does not exist. Then the proposed data and the retrograde interpolation method are used to fill in the data. Finally, the data dimensions are unified using the data normalization method, and finally a data set of 96 data items is formed every day, with 24 hours a day and 15 minutes of data. Embedding is used to encode events and convert timestamps into time codes so that the model can capture time series features.

[0063] In S2, deep feature extraction is performed on the enhanced meteorological data and numerical forecast data to capture preliminary feature information:

[0064] refer to Figure 2 , Figure 2 A schematic flow chart of sub-steps for extracting deep feature information in a photovoltaic power prediction method based on deep learning provided by an embodiment of the present invention. Step S2 specifically includes:

[0065] Use the convolution module (DC-CNN) to process the data features of the Embedding layer after data preprocessing;

[0066] Continue to refer Figure 2 , the convolution module consists of four main components: dilated convolution (DC), batch normalization (BN) layer, activation function (LeakyELU) and pooling layer (MaxPool). The corresponding formula of the convolution module is:

[0067] X Out =ReLU(BN(DC(X In ,d)))

[0068] Among them, X In is the input time series data, DC represents the dilated causal convolution operation, d is the dilation rate, BN is the batch normalization operation, ELU is the exponential linear unit activation function, and X Out is the output time series feature.

[0069] Specifically, the main function of dilated convolution (DC) is to expand the receptive field, thereby capturing a wider range of contextual information without increasing the number of convolution kernel parameters. Compared with traditional convolution, which can only handle local short-term dependencies, dilated convolution can cover farther time steps through smaller convolution kernels, thereby capturing long-term trends and patterns.

[0070] The batch normalization layer (BN) is designed to improve the stability and efficiency of model training, ensure the stable distribution of data in each layer of the network, and avoid problems such as gradient disappearance or gradient explosion.

[0071] The activation function (ELU) is designed to overcome the shortcomings of the ReLU function, especially its performance when dealing with negative values. ELU introduces nonlinearity by performing an exponential transformation on negative inputs while maintaining linear output for positive inputs, thereby improving the performance of deep neural networks.

[0072] The maximum pooling layer (MaxPool) is a downsampling operation in the convolutional network, which mainly reduces the size of the feature map and reduces the computational complexity while retaining significant features.

[0073] The function of the convolution module of the embodiment of the present invention is to perform further deep feature information extraction and nonlinear transformation on the input preliminary feature information. Specifically, the convolution module performs weighted summation of multiple consecutive time steps through the convolution kernel, thereby capturing the correlation between adjacent time steps. By capturing short-term dependencies through convolution, the model can understand the dynamic changes between adjacent time steps and extract local trends. The batch normalization layer can normalize the feature data of each channel, which helps to accelerate the convergence of the network and improve the stability of the model. The activation function introduces nonlinearity, and by setting negative values ​​to zero, the module can learn more complex and rich feature mappings.

[0074] In S3, the captured preliminary feature information is adaptively weighted and modified to obtain the corrected depth feature information:

[0075] refer to Figure 3 . Figure 3 A schematic diagram of the sub-step flow chart of step S3 of a photovoltaic power prediction method based on deep learning provided by an example of the present invention.

[0076] The main function of the adaptive spectrum block (ASB) is to improve the prediction accuracy of time series data through adaptive correction of frequency domain features. It is a key component of the key module of data correction in the embodiment of the present invention, so as to improve the data quality of time series data, enhance the robustness and generalization ability of the model, ensure that the photovoltaic power prediction model can have better data support under the meteorological conditions of the load, effectively optimize the overall performance of the model, and is a good solution to the data quality problem under the complex climatic conditions in the southern region.

[0077] S31. Perform Fourier transform on the input time series data (B, N, C) through the ASB module to convert the time series data from the time domain to the frequency domain, where B represents the batch size, N is the sequence length, and C is the number of feature channels.

[0078] That is, after the DC-CNN convolution module extracts the deep features of the data, the generated data X In Through Fourier transform (FFT), the time series data is converted from the time domain to the frequency domain, and each time data point is decomposed into sine waves of different frequencies.

[0079] Fourier transform realizes the representation of input data in the frequency dimension, that is, the amplitude and phase information of the signal at each frequency component.

[0080] In the frequency domain, the energy (sum of squares of amplitude) of each frequency component is calculated, and an adaptive spectrum mask is generated according to a preset adaptive threshold. The mask determines which frequency components need to be retained and which need to be attenuated or filtered out. Usually, frequency components with larger energy are retained, while frequency components with smaller energy are masked.

[0081] S32, performing preliminary weighted processing on the data in the frequency domain, that is, using a learnable complex weight W to adjust the input frequency components to adjust the weights of different frequency components;

[0082] Enhance the focus on specific frequencies through preliminary weighting processing and prepare for subsequent adaptive adjustment;

[0083] S33, dynamically generating an adaptive spectrum mask according to the input data frequency;

[0084] Among them, the mask is used to selectively retain important frequency components, filter irrelevant or noise frequencies, and reduce noise interference.

[0085] S34, superimposing the preliminary frequency domain weighted result and the high frequency weighted result by frequency weighted result superposition to obtain final frequency domain corrected data, thereby generating corrected complete feature data in the frequency domain;

[0086] S35. Restore the corrected result in the frequency domain back to the time domain form through inverse Fourier transform.

[0087] This facilitates the subsequent processing and prediction of the time series prediction model. After the above steps, the corrected feature data is obtained. The above process can be understood as the following formula:

[0088]

[0089] In S4, the depth feature information is decoded, and real-time dynamic noise reduction is performed on the decoded output data to obtain meteorological correction data:

[0090] The corrected depth feature information is transmitted to the decoder, which receives the output data from the encoder and generates new final output features in combination with the known information of the target data of the corrected data.

[0091] The decoder in the embodiment of the present invention uses the decoding module in the Informer to perform decoding operations:

[0092] Specifically, it can be understood that the output result X out , combined with the time features in the time series model for decoding, the multi-head sub-attention mechanism is used to generate decoding information by focusing on the important features in the feature sequence. The formula is:

[0093]

[0094] Among them, the Q, K, and V vectors represent the query, key, and value in the self-attention mechanism, which dynamically determines the most important input elements; It represents the average value selected from the top Q, sqrt(d) represents the square root of the matrix dimension, and Softmax is an activation function whose purpose is to convert the input vector into the corresponding probability. The formula is as follows:

[0095]

[0096] where σ(z) i Represents the value after the i-th element of the input vector z is converted into probability.

[0097] Furthermore, Kalman filtering is used for real-time dynamic noise reduction, the state variables and covariance matrix are initialized, the next time step and covariance matrix are predicted according to the state transfer equation, and the Kalman gain is calculated.

[0098] The corrected target value is updated through Kalman, and the state prediction and update steps are repeated to obtain the denoised estimate in real time.

[0099] The Kalman filter can be used to dynamically estimate the time series data at each time step according to the state of the model, thereby removing noise interference and ensuring smoothness and accuracy after correction.

[0100] The main function of Kalman filtering is to correct the data with large noise through real-time dynamic noise reduction in time series prediction tasks, use actual measurement values ​​for data with small noise, smooth data, and reduce the interference of noise and uncertainty on the system state. It recursively updates state estimates, combines process noise and measurement noise modeling, automatically adjusts the processing of noise of different frequencies, and enhances the prediction accuracy and robustness of the model. Kalman filtering can efficiently handle complex dynamic systems.

[0101] In S5, the meteorological correction data obtained after noise reduction is used as the input data of the time series model to obtain the photovoltaic power prediction results, which specifically include:

[0102] A photovoltaic power prediction model is established, and the corrected meteorological correction data is used as data input to predict photovoltaic power. The photovoltaic power prediction is highly correlated with factors such as weather. Through the previous steps, the historical data is cleaned up and NWP (numerical weather forecast) is extracted. Since the climate and weather in a certain area in the south are the major factors causing inaccurate photovoltaic power generation, for the photovoltaic power prediction algorithm model, the short-term predicted power is more closely related to the radiation value in the actual weather.

[0103] Therefore, we first use the cleaned data and RF algorithm in the previous step to build a regression relationship between the actual radiation data and the actual power data.

[0104] The specific steps are as follows:

[0105] Create an input sample data set, use actual direct radiation, scattered radiation and other radiation data as input data, and perform data cleaning and outlier processing.

[0106] Create new features such as the average power value, seasonal characteristics, etc., and standardize and normalize the feature data.

[0107] Data division: 70% of the data is used as training data, 20% of the data is used as a test set, and 10% of the data is used as a validation set.

[0108] Build the RF random forest model and perform model training.

[0109] By building an algorithm model through the RF model, the model can be used for short-term photovoltaic forecasting. More feature information can be extracted through the deep learning algorithm Informer, and a correction module can be added to it to compensate for the low prediction accuracy of the predicted radiation data in the NWP (numerical weather forecast) data caused by the microclimate of a southern mountainous area.

[0110] The probability coefficient attention mechanism (ProbSparse) uses probability sparsity to select query-key pairs with high correlation to reduce the algorithm's computational complexity, thereby compensating for the long algorithm training time complexity of deep learning algorithms. The encoding and decoding structure in the algorithm can pay good attention to the key features in the input data, and the extracted data features are time-series. This algorithm model has good prediction effects on nonlinear correlation data and time series data.

[0111] The loss function in the above photovoltaic power prediction algorithm is modified to the power prediction of the Southern Power Grid, which is more in line with the actual situation of the microclimate in the southern mountainous areas. By changing the loss function of the model, the loss function modified for the evaluation criteria of the Southern Power Grid can be formulated as:

[0112]

[0113] Among them, P Mi It is expressed as the available power at time i, and the unlimited power time P Mi is the actual power at time i, P Pi is the predicted power at 4 hours, and Cap is the total installed capacity of the photovoltaic power station in that month.

[0114] refer to Figure 4 The photovoltaic power prediction device proposed in the embodiment of the present invention includes:

[0115] A data preprocessing module is used to obtain and preprocess the historical operation data of photovoltaic power plants to obtain enhanced meteorological data and numerical forecast data of photovoltaic sites;

[0116] Feature extraction module, used to perform deep feature extraction on enhanced meteorological data and numerical forecast data to capture preliminary feature information;

[0117] A feature correction module is used to perform adaptive weighted correction on the captured preliminary feature information to obtain corrected depth feature information;

[0118] A decoding and denoising module, used to decode the depth feature information and perform real-time dynamic denoising on the decoded output data to obtain meteorological correction data;

[0119] The prediction module is used to use the meteorological correction data obtained after noise reduction as the input data of the time series model to obtain the photovoltaic power prediction result.

[0120] Data preprocessing module 001, used to obtain photovoltaic historical data to be processed;

[0121] It should be noted that, in a specific implementation, the photovoltaic historical data includes NWP time series data, which is obtained through data recorders at photovoltaic sites.

[0122] It should be noted that in this example, during the data collection process, the data obtained by multiple data recorders is averaged as the input module data. For input data that does not conform to the corresponding station geographical location, the input module issues a warning, requiring the user to recheck and change the corresponding data.

[0123] The preprocessing module 001 is also used to preprocess the data of the input module to obtain enhanced data;

[0124] Furthermore, the preprocessing module 001 automatically runs the Python script with the designed preprocessing strategy to perform outlier removal, data cleaning, data enhancement and other processing on the original time storage data, and divides the data into a group of data every 15 minutes, 96 groups of data per day, so that subsequent modules can perform sufficient time series information extraction and feature information learning.

[0125] The feature extraction module 002 is used to extract time series information from the enhanced data to obtain sufficient time series feature information.

[0126] It can be understood that the input time series data is first processed using a one-dimensional convolution and embedding layer module, the data dimension is mapped into a 512-dimensional high-dimensional space, and the time series data obtained by adding a position code and a time encoder to the data.

[0127] Next, the processed data is transferred to the ProbSparse module. By extracting more important data from data information, multi-head sparse self-attention (Probsparse) can significantly improve computational efficiency while ensuring model performance for tasks dealing with long time series or large-scale structures.

[0128] in Represents the time series data that enters the sparse attention mechanism ProbSparse. Represents the output feature data after processing by the ProbSparse (PS) block

[0129] LN represents layer normalization. GELU represents the activation function of the data, FFN represents the feedforward neural network of the model, and PS represents the sparse multi-head attention mechanism (Multi-headSelf-ProbSparse-attetion), which is the core of the time series information feature extraction module and allows the extraction of more important time series features in time series data. Dropout represents the random inactivation module.

[0130] The feature correction module 002 reduces the length of the data sequence, reduces data redundancy and improves model efficiency. It can be understood that it reduces the computational complexity by reducing the length of the input sequence while retaining the key information in the sequence. In this process, the Distilling module first applies a one-dimensional convolution to extract local features, and then downsamples the features through a maximum pooling operation to reduce the sequence length. This processing method helps to retain important information in the input sequence while reducing the amount of calculation, thereby providing a simplified and effective input for subsequent attention mechanisms and feature extraction modules. The sequence is then downsampled through Max Pooling, and a simplified but information-concentrated feature sequence is finally generated, which will be used as the input of subsequent modules for further processing.

[0131] The decoding noise reduction module 004 reduces the error between the prediction result and the real data by applying Kalman filtering to the data, thereby obtaining a more efficient photovoltaic power prediction result.

[0132] It can be understood that the Kalman filter gradually corrects and updates the system state by combining the predicted value of the system model with the actual observed value, thereby reducing the estimation error. First, the Kalman filter predicts the state of the next moment based on the time series and generates a predicted value. Subsequently, the predicted data is corrected by comparing the difference between the predicted value and the observed value and adjusting their weights using their gain. In order to further reduce the error, the filter fully considers the predicted results and actual data during the correction process and optimizes the accuracy of the state estimation. Finally, the Kalman filter continuously updates the state estimate and the error covariance matrix through recursive optimization to ensure that the model can have good prediction results under different conditions, effectively reducing the error between the predicted data and the actual data.

[0133] This example obtains the observation data of photovoltaic stations and NWP (numerical weather forecast data), and processes the preprocessed time series data; uses deep information extraction on the photovoltaic time series data to obtain sufficient feature information; on the basis of fully extracting the time series information, obtains more efficient sequence data by using the coefficient self-attention mechanism; removes redundant sequence data by using the data distillation method, reduces the computational complexity, and retains key information. Through the decoding module, the obtained data is converted into prediction results. By improving the prediction accuracy, it can adapt to data in different scenarios.

[0134] As shown in Table 1 below, it can be observed that the technical solution proposed in the embodiment of the present invention has the lowest results in the three indicators in the numerical performance of the ablation experiment. The lowest mean square error (MSE) and mean absolute error (MAE) mean that the average deviation between the model's prediction result and the true value is the smallest.

[0135] Specifically, in the prediction task, it means that the model has higher reliability and accuracy when dealing with practical problems. In order to further verify the specific performance of this model in the southern microclimate, the evaluation criteria of the Southern Power Grid are added as an important basis in this experiment. The experimental results show that the method proposed in this experiment still has the highest accuracy under this evaluation standard; at the same time, it can be observed in Table 1 that after the ablation experiment, the module proposed in the embodiment of the present invention has practical effects in the three major indicators;

[0136] (Table 1)

[0137]

[0138] Refer to Table 2. It can be observed that by comparing four common and efficient time series algorithm models, the method proposed in the embodiment of the present invention has significantly better numerical performance than other models in the comparative experiment, especially in the two key indicators of MSE (mean square error) and MAE (mean absolute error). The lower the values ​​of the MSE and MAE indicators, the smaller the error of the prediction model, and the stronger the model's ability to fit the actual data, so the deviation between the prediction result and the true value is minimized.

[0139] Specifically, the method proposed in the embodiment of the present invention shows higher accuracy and reliability in the time series prediction task. In terms of the MSE indicator, this method can reduce the square error between the predicted value and the actual value more than other models; and in terms of the MAE indicator, this method further verifies its stability and effectiveness in processing complex time series data by reducing the mean absolute value of the prediction error. This means that in different time series prediction tasks, the method proposed in the embodiment of the present invention can better capture the trends and characteristics of time series data.

[0140] (Table 2)

[0141]

[0142] Reference Figure 5, the photovoltaic power prediction device includes: a solar panel 2001, an inverter 2002, an electric energy collection device 2003, a data recording device 2004, a data storage device 2005, a processor 2006, a network device 2007, and a data bus 2008. The data bus 2008 is used to connect components for data storage and communication, and the data recording device 2004 can be a common photovoltaic power recording device that can record the voltage, current, power, ambient temperature, solar radiation intensity and other parameters of the solar cell. The storage device 2005 can be an SD card, a flash memory, a hard disk, a cloud storage, etc. The storage module needs to store data for a long time to facilitate subsequent analysis or query. The processor 2006 can be a central processing unit (CPU), and the network device 2007 can include a series of devices such as routers, switches, and firewalls. The solar photovoltaic data is obtained through the solar panel 2001, and the solar energy is converted into direct current. The generated direct current is converted into alternating current through the inverter, which is suitable for the power grid or for home use. The power signal generated by the inverter 2002 is captured and monitored by the power collection device 2003, which records the power parameters such as voltage, current, and power in real time. The collected power data will be transmitted to the data recording device 2004, which will further record and store this information. The data recording device 2004 will save the collected data to the data storage device, and the data storage device 2005 is usually used to store data for a long time for analysis and historical query. The data in the data storage device 2005 will be transmitted to the processor 2006 for processing and analysis. The processor uses this data to perform calculations such as photovoltaic power prediction, and transmits the prediction results to related devices. The network device 2007 ensures data communication within the system and between external systems. It uploads the output results of the processor to the cloud or other remote monitoring systems, or exchanges data with other systems.

[0143] The skilled person will understand that the structure shown in the figure is not the only way to construct the photovoltaic power prediction device, it is just an example. In fact, the photovoltaic power prediction device may include more or fewer components, or combine certain components, or even adopt different component arrangements. In other words, the illustration only shows a possible configuration, not a specific limitation on the design of the device.

[0144] The above serial numbers are only for the convenience of description and do not mean that one embodiment is better or worse than other embodiments. The above description provides a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Anyone familiar with the technical field can easily think of various equivalent modifications or alternatives within the technical scope disclosed by the present invention, and these modifications or alternatives should be deemed to be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the scope determined by the claims.

Claims

1. A photovoltaic power generation power prediction method, characterized in that: The method comprises the following steps: Obtain historical photovoltaic operation data and preprocess it to obtain enhanced meteorological data and numerical forecast data of photovoltaic sites; Perform deep feature extraction on enhanced meteorological data and numerical forecast data to capture preliminary feature information; Adaptively weight the captured preliminary feature information to obtain the corrected depth feature information; Decoding the depth feature information, and performing real-time dynamic noise reduction on the decoded output data to obtain meteorological correction data; The meteorological correction data obtained after noise reduction is used as the input data of the time series model to obtain the photovoltaic power prediction results.

2. The photovoltaic power generation prediction method according to claim 1, characterized in that: The deep feature extraction of the enhanced meteorological data to capture preliminary feature information specifically includes: The DC-CNN convolution module is used to process the input meteorological data and numerical forecast data to obtain the time series-based feature matrix of the convolution output: X Out =ReLU(BN(DC(X In ,d)))? Among them, X In represents the input numerical forecast data and meteorological data of the photovoltaic site, ReLU represents the activation function, BN represents the normalized batch operation, DC is the dilated convolution operation, and d is the dilation rate.

3. The photovoltaic power generation prediction method according to claim 2, characterized in that: The adaptive weighted correction of the captured preliminary feature information to obtain the corrected depth feature information specifically includes: ASB is used to combine the time series-based feature matrix obtained by the convolution operation to extract and correct the frequency domain features and obtain deep feature information: in, It represents the result of Fourier transform of the time series feature data output by convolution. represents the inverse Fourier transform, W1 and W2 represent complex weights, and M adaptive represents the adaptive spectrum mask, Represents the deep feature information output by ASB processing.

4. The photovoltaic power generation prediction method according to claim 3, characterized in that: Preferably, the decoding operation of the depth feature information and performing real-time dynamic noise reduction on the decoded output data to obtain the meteorological correction data specifically includes: The Decoder decoding structure in the Informer model is used to decode the meteorological correction data obtained by deep information extraction and information adaptive information analysis using ASB in the Encoder layer, and the data after the decoding operation is denormalized to obtain the target prediction result; Use Kalman filtering to dynamically adjust target prediction results.

5. The photovoltaic power generation prediction method according to claim 1, characterized in that: The method of using the meteorological correction data obtained after noise reduction as the input data of the time series model to obtain the photovoltaic power prediction result specifically includes: Use historical data as training data set to obtain a trained RF model; The RF model is used as the first-layer training model of the time series model to construct a regression relationship between historical data and power values, and the output result of the first layer is used as a new training set for the second layer; Using Informer as the second-layer training model of the time series model; The results of the second-layer training model are used as the input of the first-layer RF model to predict the PV power value on the target date.

6. The photovoltaic power generation prediction method according to claim 5, characterized in that: The method further comprises: Modify the evaluation index of the time series model. The evaluation index in the model is the loss function. By changing the loss function, the original historical operation data is corrected. The changed loss function is formulated as: Among them, P Mi It is expressed as the available power at power-limited time i or the actual power at unrestricted time i, P Pi is the predicted power at 4 hours, and Cap is the total installed capacity of the photovoltaic power station in that month.

7. A photovoltaic power generation power prediction device, characterized in that: The device comprises: A data preprocessing module is used to obtain and preprocess the historical operation data of photovoltaic power plants to obtain enhanced meteorological data and numerical forecast data of photovoltaic sites; Feature extraction module, used to perform deep feature extraction on enhanced meteorological data and numerical forecast data to capture preliminary feature information; A feature correction module is used to perform adaptive weighted correction on the captured preliminary feature information to obtain corrected depth feature information; A decoding and denoising module, used to decode the depth feature information and perform real-time dynamic denoising on the decoded output data to obtain meteorological correction data; The prediction module is used to use the meteorological correction data obtained after noise reduction as the input data of the time series model to obtain the photovoltaic power prediction result.

8. A deep learning photovoltaic power prediction device, characterized in that: The device includes: a solar panel, an inverter, an electric energy collection device, a data recording device, and also includes a memory, a processor, and a prediction program stored in the memory and executable on the processor, wherein the prediction program is to implement the steps of the photovoltaic power prediction method as described in any one of claims 1 to 6.

9. A storage medium for reading, writing and storing by a computer, characterized in that: The storage medium stores a photovoltaic power prediction program based on deep learning, and when the medical image fusion program based on deep learning is executed by the processor, the steps of the photovoltaic power prediction method according to any one of claims 1 to 6 are implemented.

Images