Photovoltaic power generation prediction method based on multivariate neural model and BP algorithm

Abstract

The invention discloses a photovoltaic power generation prediction method based on a multivariate neural model and a BP algorithm, and the method comprises the steps: building a neural network model library through the analysis of environment data and photovoltaic power generation data, carrying out the model training of the intervention of the BP algorithm, and carrying out the data analysis, screening and rectification, and matrix dynamic weight matching. Data training feeding and photovoltaic station power generation data prediction are carried out on the model through a synchronous long and short term memory method and matrix analysis, and power station managers can know future weather conditions in advance and reasonably arrange power generation plans, so that the maximum power generation capacity is achieved under the optimal illumination condition, the power generation efficiency of the power station is improved, and the intelligent management level is enhanced. The photovoltaic prediction result of the power station can provide reference for power grid dispatching, so that the power grid operation is more stable, the photovoltaic power generation utilization rate is higher, and meanwhile, a technical support means is provided for efficient digital management and control.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation, and in particular to a photovoltaic power generation prediction method based on a multivariate neural model and a BP algorithm. Background Art

[0002] As an emerging green energy, solar photovoltaic power generation has attracted more and more attention, and the total power generation is also increasing. It is expected that by 2030, solar photovoltaic power generation will account for more than 10% of the world's total power generation. However, photovoltaic power generation is greatly affected by the weather and natural environment. Whether it is real-time power, daily power generation, monthly power generation, quarterly power generation, or annual power generation, it will change and is unpredictable, which is very unfavorable for stable electricity consumption. Therefore, photovoltaic power generation prediction has become one of the key technologies for efficient use of photovoltaic power generation.

[0003] At present, photovoltaic power generation prediction technology has achieved remarkable results in practical applications, providing strong support for the safe and economic dispatch of power grids, the stable operation of the power market, and the self-management of photovoltaic power generation systems. However, photovoltaic power generation prediction technology still faces some challenges, such as the need to improve prediction accuracy and the need to continuously update and improve prediction models. In addition, since the power generation of photovoltaic power generation systems is affected by a variety of complex factors, it is necessary to further study the influence mechanism of various factors on photovoltaic power generation in order to improve the accuracy and reliability of predictions. Summary of the invention

[0004] The purpose of the present invention is to overcome the above-mentioned shortcomings and provide a photovoltaic power generation prediction method based on a multivariate neural model and a BP algorithm, which uses Internet of Things technology, power parameter sensing technology, automation technology, cloud computing, big data analysis, a multivariate neural model and a BP algorithm to predict photovoltaic power generation data.

[0005] To achieve the above purpose, the present invention designs a photovoltaic power generation prediction method based on a multivariate neural model and a BP algorithm. Based on the learning and processing mechanism of the human nervous system, a multivariate neural model is constructed to analyze and process the input data to achieve classification and prediction. Based on the hierarchical structure of the storage weight matrix, the mutual connection between multiple layers of multiple neurons is used to perform reverse transmission and learning of information, thereby providing the output of relevant prediction data. The prediction method comprises the following steps: S1. Build an environmental model, input environmental data, classify and process the environmental data, and predict the environmental data. By solving the optimal fitness value, further filter and optimize the data to improve data accuracy; the environmental data includes photovoltaic power station location data, sunshine data, meteorological data, photovoltaic installed capacity data, etc.; S2. Electrical energy data encapsulation and parsing. Through the edge computing gateway, parse the electrical energy data of the current photovoltaic site. The electrical energy data includes the voltage, current, and real-time power of the photovoltaic inverter, and cooperate with the complex tariff meter to calculate the power generation in peak and valley periods, and the power generation income in different periods. The different periods include the periods distinguished by electricity prices, that is, calculate the power generation income according to each electricity price period. Steps S1 and S2 have no sequence. S3. Neuron model establishment, including establishing a neuron model based on LSTM (Long Short Term Memory Network) and BP algorithm (Back Propagation Algorithm). S4. Feeding historical data. Input the results of the classification and prediction of the collected historical environmental data and the results of the encapsulation and parsing of historical electrical energy data into the neuron model for model training, that is, feed a large amount of historical data into the local AI library to help the model learn and identify specific patterns and features. S5. Data deviation correction. Screen the prediction data. The screening includes comparing with the prediction accuracy of industry standards, and eliminating the data with large deviations to improve the accuracy of the prediction model data. S6. Generate the final prediction data. According to the BP algorithm of the neuron model and relevant externally collected factors, predict the trend of recent photovoltaic power generation data and store the prediction data.

[0006] Furthermore, the prediction method further includes: S7. Relevant data storage. Write relevant data into the local database and select a suitable destination data entry, such as HTTP service, Redis, Mysql, MongoDB, or MQTT server, etc.

[0007] S8. Provide power station scheduling and operation and maintenance guidance. According to the prediction data and the actual power consumption data of the power station, issue user reports and analysis reports to provide guidance for later operation and maintenance data.

[0008] Furthermore, the location data includes the detailed address of the area where it is located, such as the provincial, municipal, and district addresses and / or longitude and latitude information, etc. The sunshine data includes the annual effective sunshine time and / or irradiance of the local area, etc. The meteorological data includes temperature and humidity data. The photovoltaic installation data includes at least one of the photovoltaic installation capacity of the station, the overall power consumption load of the current power station, the average power factor, etc. Furthermore, the establishment of the environmental model includes parsing and predicting environmental data through the Whale Optimization Algorithm (WOA), and optimizing the prediction data by solving the best fitness value, thereby improving the accuracy of the prediction data, including: S11. Set the initial population of the WOA algorithm and the optimization interval of the parameters [k, a], and take the minimum value of the envelope entropy layout as the fitness value; where k is the decomposition number and a is the sample entropy calculation value. S12. Use VMD (Variational Modal Decomposition, a new time-frequency analysis method) to decompose the original signal, and obtain the fitness values of different parameter combinations through the calculation formula. S13. Use the optimization mechanism of WOA to continuously update the positions of individuals, and at the same time compare the values of irradiance, find the minimum fitness value, so as to find data with large deviations and perform screening and elimination. S14. Loop the steps of S12 - S13. When the fitness is less than or equal to the threshold or the number of iterations reaches the maximum limit, output the optimal parameters. S15. Set the parameters of VMD with the obtained optimal parameter combination [k, a], and decompose and output the original environmental data.

[0009] Further, the historical data is at least the data of the most recent 3 months. In this embodiment, the data of the most recent 3 months is taken as the training data.

[0010] Further, the prediction method further includes introducing a data Gaussian model in step S3 to perform model analysis and induction on the data: Denote N training samples , and its ELM (Extreme Learning Machine) regression model is as follows:

[0011] Among them, x represents the input data vector, y is the output data vector, f(x) represents the output of the neural model, h(x) and H represent the mapping matrices of the hidden layer, represents the weight between the hidden layer and the output layer; Introduce the diagonal matrix and the penalty coefficient C, and then according to the principle of the generalized inverse matrix, we can get:

[0012] In the formula, is the expected output.

[0013] Further, the prediction method further includes a weight adaptive method, that is, the weights can be adaptively adjusted according to the characteristics of different layers and the distribution of training data to improve the generalization ability and search efficiency of the neural model, including: Total output of the model:

[0014] where M is the total number of iterations of the model, is the weight of the m-th iteration, is the output label vector of the model, which is equivalent to the mapping matrix H of the hidden layer; Output of the m-th round of the model:

[0015] Sample Predicted category in the m-th round:

[0016] Optimization objective in the m-th round:

[0017] where, is the weight, and the superscript T represents the transpose of a matrix or vector.

[0018] Furthermore, the data rectification method includes: a data error correction algorithm based on a hash function, namely the hash error correction code (HAC) algorithm, which detects and repairs errors by adding hash error correction codes to the data. The main steps are as follows: S51. Data chunking: Divide the data into multiple chunks, each with a size of a; S52. Hash encoding: Perform hash encoding on each data chunk to generate a hash error correction code with a length of n, where n > a; S53. Transmission: Transmit the data chunks and the hash error correction codes to the receiving end together; S54. Decoding: At the receiving end, decode the received data chunks and hash error correction codes to detect and repair errors.

[0019] The mathematical model of the hash error correction code algorithm can be described by the following formula: H = H(x a ) where H is the hash error correction code, x a is the original data chunk, and H() is the hash function.

[0020] The advantages and beneficial effects of the present invention are as follows: The present invention applies Internet of Things technology, power parameter sensing technology, automation technology, cloud computing and big data analysis technology to provide a photovoltaic power generation prediction method based on a multi-neural model and BP algorithm. By combining the multi-neural network model and BP algorithm, through operations such as data analysis, screening and deviation correction, and matrix dynamic weight ratio, the model is trained and fed with data by synchronizing the long short-term memory method and matrix analysis. The power generation data of relevant photovoltaic sites is predicted to provide a data prediction model based on pure algorithm + historical data for users. The overall prediction accuracy reaches over 97%, and the prediction data is more referenceable and effective. Power station managers can understand the future weather conditions in advance, reasonably arrange the power generation plan, so as to achieve the maximum power generation under the best lighting conditions, improve the power generation efficiency of the power station, and enhance the intelligent management level; at the same time, the photovoltaic prediction results of the power station can provide reference for power grid dispatching, make the power grid operation more stable, the utilization rate of photovoltaic power generation higher, and provide a technical support means for efficient digital control. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is the flowchart of the photovoltaic power generation prediction method of the present invention; Figure 2 is the flowchart of environmental data analysis and prediction; Figure 3 is the training and validation accuracy curve graph in the model experiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] The following describes the specific embodiments of the present invention in conjunction with the drawings and embodiments. The following embodiments are only used to illustrate the technical solutions of the present invention more clearly, and cannot be used to limit the protection scope of the present invention.

[0023] Embodiment 1: A photovoltaic power generation prediction method based on a multi-neural model and BP algorithm of the present invention is based on the learning and processing mechanism of the human nervous system. By constructing a multi-neural model, the input data is analyzed and processed to achieve classification and prediction. Based on the hierarchical structure of the stored weight matrix, the reverse transmission and learning of information are carried out through the mutual connection between multiple neurons in multiple layers, and then the output of relevant prediction data is provided; as Figure 1 shown, the prediction method includes the following steps: S1. Environmental model construction: Input environmental data, classify and process the environmental data, and make predictions. By solving the optimal fitness value, further screen and optimize the data to improve data accuracy. The environmental data includes photovoltaic power station location data, sunshine data, meteorological data, photovoltaic installed capacity data, etc. Essentially, the above is the environmental data classification of this embodiment. According to different forms, different methods can be selected. Therefore, when modeling, it is necessary to first observe the data form, such as the power consumption load of the power station, the capacity construction of the photovoltaic power station, and local environmental factors (irradiance, temperature and humidity, effective sunlight utilization time, etc.). This part of the data needs to be provided by the customer in advance, and a conventional filtering prediction model can be used to analyze and predict the environmental data. S2. Power data encapsulation and parsing: Through the edge computing gateway, parse the power data of the current photovoltaic site. The power data includes the voltage, current, and real-time power of the photovoltaic inverter, and calculates the power generation during peak and valley periods in cooperation with a complex tariff meter, as well as the power generation benefits at different times. The different times include the times distinguished by electricity prices, that is, the power generation benefits are calculated according to each electricity price period. Steps S1 and S2 have no sequence. S3. Neuron model establishment: Include establishing a neuron model based on LSTM (Long Short Term Memory Network) and BP algorithm (Back Propagation Algorithm). The neuron model used in this embodiment based on the long short term memory network and BP algorithm can solve the problem of gradient disappearance or gradient explosion of traditional RNN (Recurrent Neural Network) when processing long sequence data while improving the model learning and training efficiency, enabling the network to learn long-term correlation relationships, so that the prediction results are more accurate. The multi-neuron model consists of an input layer, a hidden layer, and an output layer: Input layer: It is the data feature input layer, and the number of input data feature dimensions corresponds to the number of neurons in the network. Hidden layer: That is, the middle layer of the network, whose function is to accept the output of the previous layer network as the current input value and calculate and output the current result to the next layer. The number of neurons in the hidden network directly affects the fitting ability of the model. The number of neurons in the output layer represents the number of classification categories. For the input layer and output layer of the network, the number of neurons is usually determined. Generally, the more neurons in the hidden layer, the more capacity the model has to achieve a better fitting effect.

[0024] The LSTM (Long Short-Term Memory Network) structure is designed specifically to address the problems of gradient vanishing and explosion in RNN when context information appears. A memory module is added to the structure. The module can be regarded as a memory chip in a computer. Each module contains several recurrently connected memory cells and three gates (input, output, and forget gates, which are equivalent to write, read, and reset). The input of information can only interact with neurons through each gate to prevent gradient explosion or vanishing. With the feeding of long-term data, there will be a large room for improvement in the training of the data model and the accuracy of the data.

[0025] The BP algorithm is a widely used training algorithm in artificial neural networks. The learning process of the algorithm mainly consists of two processes: the forward propagation of signals and the backward propagation of errors. By establishing a multi-neural model based on the BP algorithm and training with historical data, the accuracy of overall data prediction is improved. Through the writing of preset parameters, including environmental factors, photovoltaic capacity, irradiance in the region, temperature and humidity, etc., comprehensive consideration is given to predict and analyze the data of photovoltaic power generation.

[0026] The present invention organically combines LSTM and the BP algorithm, which can establish a more accurate correlation between environmental parameters and electrical energy data. In particular, both short-term and long-term correlations can be taken into account, thus significantly improving the learning efficiency and prediction accuracy of the neuron model.

[0027] S4. Feeding historical data: Input the results of the classification and prediction of the collected historical environmental data by the environmental model and the parsing results of the encapsulated historical electrical energy data into the neuron model for model training, that is, feeding a large amount of historical data into the local AI library to help the model learn and identify specific patterns and features. The historical data in this embodiment includes data such as historical daily power generation, daily average power, daily average irradiance, relative temperature, and relative humidity in the past 3 months. S5. Data deviation correction: Screen the predicted data. The screening includes comparing with the prediction accuracy rate of industry standards, and eliminating the data with large deviations to improve the accuracy of the prediction model data. The specific screening method in this embodiment includes that according to Article 31 of the "Grid Connection Operation Management Rules" of the National Energy Administration, the deviation between the actual photovoltaic output and the short-term 96-point prediction value should be less than 10% of the prediction value; when the deviation is between 10% and 20%, it is evaluated according to the integral electricity of 1 point / kWh; when the deviation exceeds 20%, it is evaluated according to the integral electricity of 3 points / kWh. The deviation of the predicted data within 10% is a reasonable range, and if it is greater than 10%, it is considered a large deviation and relevant assessments are required. The present invention combines the long short-term memory network and the deviation correction algorithm for the learning of the multi-neural model, which can further improve the accuracy of data training. The probability of each sample point appearing can be expressed in the form of a normal distribution, making the prediction result more scientifically reasonable.

[0028] S6. Generate the final predicted data. According to the BP algorithm of the neuron model and relevant externally collected factors, predict the trend of recent photovoltaic power generation data (mainly calculate the photovoltaic power generation power in the past 7 days based on the input data, and the output data frequency is the same as the input data frequency), and store the predicted data.

[0029] As Figure 3 shown, the final prediction accuracy of this embodiment on the test set can reach 97%, which is higher than the general level of about 90% of the current traditional prediction methods. In the figure, the abscissa is the number of training times, and in the ordinate, Training accuracy refers to the prediction accuracy of the model on the training data, Validation accuracy refers to the prediction accuracy of the model on the validation data, Training loss refers to the loss of the model on the training set, and Validation loss refers to the loss of the model on the validation set.

[0030] Preferably, the location data includes the detailed address of the area where it is located, such as the provincial, municipal and district address and / or longitude and latitude information, etc., the sunshine data includes the annual effective sunshine time and / or irradiance of the local area, etc., the meteorological data includes temperature and humidity data, and the photovoltaic installation data includes at least one of the photovoltaic installation capacity of the power station, the overall power consumption load of the current power station, the average power factor, etc.; Selection of input data and output variables of the prediction model in this embodiment: The power load data itself has dynamic characteristics. Taking the historical load data as a unit of day, each day is divided into 288 time points (one point every 5 minutes) as the fluctuation law of the load itself. The historical load data, irradiance, average temperature, relative humidity and annual effective hour data before the prediction day, as well as the highest temperature, lowest temperature and average temperature on the day are used as the input quantities of the model. The specific input quantities, that is, the environmental data, include: 1) longitude and latitude information, 2) the annual effective sunshine time of the local area, 3) irradiance, 4) average temperature, 5) average humidity, 6) the photovoltaic installation capacity of the power station, 7) the overall power consumption load of the current power station, 8) average power factor, 9) highest temperature, 10) lowest temperature, 11) the average value of historical power data in the past 3 months, 12) the average value of historical electricity consumption in the past 3 months. The load data on the prediction day is used as the output quantity for model training and prediction.

[0031] Preferably, as Figure 2 shown, the construction of the environmental model includes parsing and predicting the environmental data through the Whale Optimization Algorithm (WOA, also known as the Whale Optimization Algorithm), optimizing the predicted data by solving the best fitness value, and thus improving the accuracy of the predicted data, including: S11. Set the initial population of the WOA algorithm and the optimization interval of the parameters [k, a], and take the minimum value of the envelope entropy layout as the fitness value; where k is the decomposition number and a is the sample entropy calculation value. S12. Use VMD (Variational Modal Decomposition, a new time-frequency analysis method) to decompose the original signal, and obtain the fitness values of different parameter combinations through the calculation formula. S13. Use the optimization mechanism of WOA to continuously update the positions of individuals, and at the same time compare the values of irradiance to find the minimum fitness value, so as to find the data with large deviations for screening and elimination. S14. Loop the steps of S12~S13. When the fitness is less than or equal to the threshold or the number of iterations reaches the maximum limit, output the optimal parameters. In this embodiment, the fitness threshold is set to 0.01 and the maximum limit of the number of iterations is 1000 times. S15. Set the parameters of VMD with the obtained optimal parameter combination [k, a], and decompose and output the original environmental data.

[0032] VMD is a new time-frequency analysis method that can decompose multi-component signals into multiple component amplitude-modulated and frequency-modulated signals at one time, avoiding the endpoint effect and false component problems encountered in the iterative process. This method can effectively process non-linear and non-stationary signals, but it also has the characteristic of being sensitive to noise. When there is noise, it may cause modal aliasing in the decomposition. The present invention can overcome this aliasing phenomenon based on the WOA algorithm, so as to obtain more accurate prediction results.

[0033] Based on the WOA and VMD methods, this embodiment can overcome the influence of a large amount of noise in many environmental parameters, and can utilize the natural correlation between different environmental parameters, so as to facilitate the screening of more regular environmental parameters. In this way, when establishing the relationship between environmental parameters and photovoltaic power parameters later, the model training parameters can converge more quickly and accurately.

[0034] Preferably, the historical data is at least the data of the most recent 3 months. In this embodiment, the data of the most recent 3 months is taken as the training data and the verification data.

[0035] Preferably, the prediction method further includes introducing a data Gaussian model in step S3 to perform model analysis and induction on the data. Denote N training samples , and its ELM (Extreme Learning Machine) regression model is as follows:

[0036] Among them, \(x\) represents the input data vector, \(y\) is the output data vector, \(f(x)\) represents the output of the neural model, \(h(x)\) and \(H\) represent the mapping matrices of the hidden layer, represents the weight between the hidden layer and the output layer; Introduce the diagonal matrix and the penalty coefficient \(C\), and then according to the principle of the generalized inverse matrix, we can get:

[0037] In the formula, is the expected output.

[0038] Preferably, the prediction method further includes a weight adaptive method, that is, the weights can be adaptively adjusted according to the characteristics of different layers and the distribution of training data to improve the generalization ability and search efficiency of the neural model, including: Total output of the model:

[0039] Among them, \(M\) is the total number of iterations of the model, is the weight of the \(m\)-th iteration, is the model output label vector, which is equivalent to the mapping matrix \(H\) of the hidden layer; Output of the \(m\)-th round of the model:

[0040] Sample Predicted category in the \(m\)-th round:

[0041] Optimization objective in the \(m\)-th round:

[0042] Among them, is the weight, and the superscript \(T\) represents the transpose of the matrix or vector.

[0043] Preferably, the method for data rectification includes: a data error correction algorithm based on a hash function, namely the hash error correction code (HAC) algorithm, which detects and repairs errors by adding hash error correction codes to the data, and the main steps are as follows: S51. Data chunking, dividing the data into multiple chunks, each chunk with a size of \(a\); S52. Hash encoding, performing hash encoding on each data chunk to generate a hash error correction code with a length of \(n\), where \(n > a\); S53. Transmission, transmitting the data chunk and the hash error correction code to the receiving end together; S54. Decoding, at the receiving end, decoding the received data chunk and the hash error correction code to detect and repair errors.

[0044] The mathematical model of the hash error correction code algorithm can be described by the following formula: H = H(x a ) where H is the hash error correction code, x a is the original data block, and H() is the hash function.

[0045] Example 2: The difference from Example 1 is that the historical data takes the historical data of the most recent year as the training data and the validation data.

[0046] Example 3: The difference from Example 1 is that the prediction method further includes the following steps: S7. Store relevant data. Write the relevant data into the local database, and select a suitable destination data entry, such as an HTTP service, Redis, Mysql, MongoDB, or MQTT server, etc.

[0047] S8. Provide power plant scheduling and operation and maintenance guidance. According to the prediction data and the actual power consumption data of the power plant, issue user reports and analysis reports, and provide guidance for later operation and maintenance data.

[0048] Regarding the security of data storage, the system constructed by the method of the present invention transmits the data to the corresponding database location by means of dynamic URL generation, configuring headers and parameters, etc., and at the same time performs encryption settings on the data, such as MD5 and RS verification, to provide the security of data transmission.

[0049] The above are only some relatively systematic and comprehensive embodiments of the photovoltaic power generation prediction method based on the multi - neural model and the BP algorithm of the present invention. In fact, selecting historical data with a longer time, such as historical data of 2 years, can obtain more accurate model parameters, and both long - term and short - term features can be better reflected; other noise - removing methods can also be used for the pre - processing of the original data. In short, retaining regular data as much as possible can achieve twice the result with half the effort for model training; for fixed sites, location data can be not used as input parameters, etc. These combinations or preferred solutions should also be regarded as the protection scope of the present invention and will not be listed one by one here.

Claims

1. A photovoltaic power generation prediction method based on a multivariate neural model and a BP algorithm, which analyzes and processes input data by constructing a multivariate neural model to achieve classification and prediction, and uses the interconnection between multiple layers of multiple neurons based on a hierarchical structure of a storage weight matrix to perform reverse transmission and learning of information, thereby providing output of relevant prediction data; characterized in that: The prediction method comprises the following steps: S1. Build an environmental model, input environmental data, classify and process the environmental data, and predict the environmental data. By solving the optimal fitness value, further screen and optimize the data. The environmental data includes photovoltaic power station location data, sunshine data, meteorological data, and photovoltaic installed capacity data. S2. Power data packaging and parsing: The power data of the current photovoltaic site is parsed through the edge computing gateway. The power data includes the voltage, current, real-time power of the photovoltaic inverter, and the power generation during peak and valley periods, and the power generation income during different periods. The different periods include periods divided by electricity prices. There is no order of precedence for steps S1 and S2; S3, neuron model establishment, including establishing a neuron model based on LSTM and BP algorithms; S4, feeding historical data, inputting the results of the environmental model classification and prediction of the collected historical environmental data, and the results of the historical electric energy data packaging and analysis into the neuron model for model training; S5, data correction, screening the forecast data, the screening includes comparing with the industry standard forecast accuracy, and eliminating the data with large deviation; S6. Generate final prediction data.

2. The photovoltaic power generation prediction method based on a multivariate neural model and a BP algorithm according to claim 1, characterized in that: The prediction method further comprises the following steps: S7, relevant data is stored in the database; S8. Provide power station dispatching, operation and maintenance guidance.

3. The photovoltaic power generation prediction method based on a multivariate neural model and a BP algorithm according to claim 1, characterized in that: The location data includes the province, city, district address and / or longitude and latitude information; the sunshine data includes the local annual effective time of sunlight and / or irradiance; the meteorological data includes temperature and humidity data; and the photovoltaic installed capacity data includes at least one of the photovoltaic installed capacity of the site, the overall power load of the current power station, and the average power factor.

4. The photovoltaic power generation prediction method based on a multivariate neural model and a BP algorithm according to claim 1, characterized in that: The whale optimization variational model is used to analyze and predict environmental data, and the predicted data is optimized by solving the optimal fitness value, thereby improving the accuracy of the predicted data, including: S11, setting the initial population of the WOA algorithm and the optimization interval of the parameter [k, a], and taking the minimum value of the envelope entropy layout as the fitness value; wherein k is the decomposition number, and a is the sample entropy calculation value; S12, using VMD to decompose the original signal, and obtain the fitness values ​​of different parameter combinations through calculation formulas; S13, using the optimization mechanism of WOA, continuously updating the position of individuals, and comparing the irradiance values ​​to find the minimum fitness value; S14, looping steps S12 to S13, when the fitness is less than or equal to the threshold or the number of iterations reaches the maximum threshold, outputting the optimal parameters; S15. Use the obtained optimal parameter combination [k, a] to set the parameters of VMD, and decompose and output the original environmental data.

5. The photovoltaic power generation prediction method based on a multivariate neural model and a BP algorithm according to claim 1, characterized in that: The historical data shall be at least 3 months’ data.

6. The photovoltaic power generation prediction method based on a multivariate neural model and a BP algorithm according to claim 1, characterized in that: The prediction method further includes introducing a data Gaussian model in step S3 to perform model analysis on the data: Remember N training samples The ELM regression model is as follows: ， Among them, x represents the input data vector, y represents the output data vector, f(x) represents the output of the neural model, h(x) and H represent the mapping matrix of the hidden layer, Represents the weight between the hidden layer and the output layer; Introducing the diagonal matrix And the penalty coefficient C, then according to the generalized inverse matrix principle: ， In the formula, is the expected output.

7. The photovoltaic power generation prediction method based on a multivariate neural model and a BP algorithm according to claim 1, characterized in that: The prediction method also includes a weight adaptive method, which adaptively adjusts the weights according to the characteristics of different layers and training data, including: The total output of the model is: , Among them, M is the total number of iterations of the model, is the weight of the mth iteration, is the model output label vector; The output of the model in round m: , sample Prediction category at round m: , The optimization goal of the mth round is: ， in, is the weight, and the superscript T indicates the matrix or vector transpose.

8. The photovoltaic power generation prediction method based on a multivariate neural model and a BP algorithm according to claim 1, characterized in that: The method for data correction includes: a data error correction algorithm based on a hash function, namely a hash error correction code algorithm, which detects and repairs errors by adding a hash error correction code to the data, including the following steps: S51, data block division, dividing the data into multiple blocks, each block size is a; S52, hash coding, performing hash coding on each data block to generate a hash error correction code of length n, where n>a; S53, transmission, transmitting the data block and the hash error correction code together to the receiving end; S54, decoding: At the receiving end, the received data block and hash error correction code are decoded to detect and repair errors.

Images