EMD-BiGRU photovoltaic power generation power prediction method based on FCM clustering and SSA optimization

Abstract

The invention discloses an EMD-BiGRU photovoltaic power generation power prediction method based on FCM clustering and SSA optimization. Photovoltaic related data is obtained and preprocessed, the photovoltaic related data is divided into three types of weather data sets through FCM clustering, EMD decomposition parameters and BiGRU hyper-parameters are optimized through SSA, decomposed subsequences and meteorological characteristics are input into optimized BiGRU prediction, and a time sequence is superposed to obtain a result. According to the method, hyper-parameter empirical configuration is replaced, the data complexity is reduced, the photovoltaic power frequency characteristics are accurately captured, the prediction decision coefficient is higher than 0.96, and the method is suitable for short-term photovoltaic power generation power prediction.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation technology, specifically to an EMD-BiGRU photovoltaic power generation prediction method based on FCM clustering and SSA optimization. Background Technology

[0002] With the rapid development of photovoltaic (PV) clean energy, its share in the global power structure continues to increase. However, PV power generation is easily affected by weather conditions, and the accuracy of short-term forecasts directly affects the dispatch safety and stability of the power system. In existing short-term PV power generation forecasting models, the hyperparameters of the neural networks largely rely on manual experience for configuration, lacking scientific and systematic optimization methods. This experience-based configuration approach is difficult to adapt to the characteristics of PV data under different weather conditions, resulting in insufficient ability of the model to capture the frequency characteristics of power signals, leading to large prediction deviations and failing to meet the actual needs of precise power system dispatch.

[0003] Based on the above problems, there is an urgent need for a technical solution that can replace empirical hyperparameter configuration, improve prediction accuracy and adaptability, and solve the prediction accuracy problem caused by unreasonable parameter configuration in existing models. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies and propose an EMD-BiGRU photovoltaic power generation prediction method based on FCM clustering and SSA optimization, comprising the following steps: S1: Acquire environmental meteorological data and historical photovoltaic power generation data of the photovoltaic power station; preprocess the acquired data; S2: The preprocessed data is divided into three weather datasets: sunny, cloudy, and rainy days using FCM clustering. S3: The SSA optimization process of EMD decomposition is adopted to perform adaptive decomposition on three types of weather datasets to obtain multiple intrinsic mode functions and residuals; S4: The hyperparameters of the BiGRU neural network are optimized using SSA. Multiple intrinsic mode functions, residuals and meteorological features are input into the optimized BiGRU neural network for prediction, and the prediction results of each intrinsic mode function and residual are obtained. S5: The prediction results are superimposed in time sequence to obtain the final photovoltaic power generation prediction value. The feature is that the complexity of photovoltaic data is reduced by FCM clustering, and the EMD decomposition parameters and BiGRU neural network hyperparameters are optimized by SSA respectively, replacing the empirical configuration method of hyperparameters.

[0005] Preferably, each record of environmental meteorological data and photovoltaic power generation historical data includes effective power generation, wind speed, temperature, relative humidity, global solar horizontal radiation, diffuse horizontal radiation, wind direction, daily rainfall, global solar tilt, and solar diffusion tilt. The unit of effective power generation is kW, the unit of wind speed is m / s, the unit of temperature is ℃, the units of global solar horizontal radiation, diffuse horizontal radiation, global solar tilt, and solar diffusion tilt are all W / m², the unit of wind direction is degrees, the unit of daily rainfall is mm, and the unit of relative humidity is not specified. The data sampling interval is 5 min.

[0006] Further preferably, the data preprocessing includes the following steps: removing obvious outliers from the data; filling missing values ​​in the data by averaging the values ​​of the preceding and following 576 data points; retaining valid data from 07:00 to 18:30 each day, with 138 data points retained daily; and normalizing the filled data using the MIN-MAX method, with the normalization formula being... ,in The normalized value. The values ​​are from the original data. The maximum value in the original data. The minimum value in the original data is used; the Pearson correlation coefficient between each environmental feature and the actual power generation is calculated, and temperature, relative humidity, global solar radiation, and global solar tilt are selected as input features for the subsequent neural network. The formula for calculating the Pearson correlation coefficient is as follows: ,in As an environmental feature, This is the actual power. This represents the mean of environmental characteristics. This represents the average power.

[0007] Further preferred, FCM clustering uses solar horizontal radiation as the clustering basis, and the clustering process includes the following steps: constructing an objective function based on Euclidean distance, the expression of which is... ,in Let i be the feature vector of the i-th sample. For the j-th cluster center, For fuzzy index, Let be the membership degree of the i-th sample belonging to the j-th cluster. The total number of samples is given; the cluster centers and membership degrees are iteratively updated using the Lagrange multiplier method, and the cluster center update expression is: ,in Let be the number of iteration steps; the membership update expression is: Set the upper limit of convergence. When satisfied At that time, clustering is completed, among which Let be the membership matrix after the k-th iteration. This is the membership matrix after the (k-1)th iteration.

[0008] A further preferred embodiment of the SSA workflow includes the following steps: Population initialization, assuming the population size of individual sparrows is... Spatial dimension is , To optimize the dimension of the optimal solution to the problem, The position of the i-th sparrow in 3D space is represented as: ; in Let be the position of the i-th sparrow in the j-th dimension; calculate the fitness of each sparrow, expressed as: ,in For individual sparrow fitness values; the discoverer's location is updated when the safety threshold is reached. Less than the alarm threshold At that time, the position update expression is When the safety threshold Greater than or equal to alarm threshold At that time, the position update expression is ,in For the number of iterations, It is a constant. The maximum number of iterations, To obtain random numbers that satisfy a normal distribution, for The matrix has all elements as 1; the follower position is updated when... At that time, the position update expression is ,when At that time, the position update expression is ,in , This is the worst position at present. The optimal position for the (t+1)th iteration; the guard's position is updated when... At that time, the position update expression is ,when At that time, the position update expression is ,in Let i be the fitness value of the i-th sparrow. For optimal fitness, For worst fitness, A random number in the range [-1, 1] To prevent non-zero numbers with a denominator of 0, The step size is a control parameter that follows a normal distribution with a mean of 0 and a variance of 1; based on fitness... Determine if the stopping condition is met. If it is, output the optimal sparrow position; otherwise, return to the discoverer's position update step.

[0009] A further preferred embodiment of the SSA-optimized EMD decomposition process includes the following steps: identifying the original time series. Find all the maxima and minima of the time series; fit the upper and lower envelopes of the time series using a cubic spline interpolation function, and calculate the mean of the upper and lower envelopes. The formula for calculating the mean is: ,in Let be the value of the upper envelope at time t. The value of the lower envelope at time t; using the original sequence Subtract the mean of the envelope A new sequence is obtained The calculation formula is: ;test Whether the Inherent Mode Function (IMF) is satisfied depends on two conditions: the first condition is that the mean of the upper and lower envelopes is zero at any given time; the second condition is that the number of extreme points and zero-crossing points of the component signal is the same or differs by only one. If neither of the above two conditions is met, then... Repeat the above steps as the original sequence until the first intrinsic mode function that satisfies the condition is obtained. ; using the original sequence Subtract the first intrinsic mode function Obtain the remaining sequence The calculation formula is: ; the remaining sequence As a new original sequence, repeat all the above steps until the remaining sequence can no longer be decomposed into intrinsic mode functions, resulting in multiple intrinsic mode functions and the final residual.

[0010] Furthermore, the SSA-optimized BiGRU neural network includes an update gate and a reset gate, with the update gate calculated using the following formula: The formula for resetting the door is: ,in For the Sigmoid function, To update the gate weight matrix, To reset the weight matrix of the gate, This represents the hidden state at time t-1. Let t be the input data at time t; the SSA-optimized BiGRU neural network processes the sequence simultaneously from front to back and from back to front. Processing from front to back generates the positive hidden state, and the formula for calculating the positive hidden state is: The reverse hidden state is generated by processing from back to front. The formula for calculating the reverse hidden state is as follows: The forward hidden state and the reverse hidden state are merged by concatenation to obtain the final hidden state. The formula for calculating the final hidden state is as follows: ,in This indicates a splicing operation.

[0011] Further optimized, the SSA-optimized EMD decomposition parameters are as follows: , The optimization interval is 3 to 12. For three types of weather datasets—sunny, cloudy, and rainy—SSA optimization determines the optimal range. The values ​​are 9, 9, and 10, respectively. The hyperparameters of the BiGRU neural network optimized by SSA include the number of hidden layer neurons, the number of batch samples, and the learning rate. The optimization range for the number of hidden layer neurons is 32, 64, 128, 256, and 512, and the number of hidden layer neurons determined by SSA is 256. The optimization range for the number of batch samples is 32, 64, 128, 256, and 512, and the number of batch samples determined by SSA is 32. The optimization range for the learning rate is 0.01, 0.03, 0.1, 0.3, and 0.5, and the learning rate determined by SSA is 0.01.

[0012] Further preferred processing flow for the three types of weather datasets is as follows: The sunny day dataset is input into the SSA-optimized EMD decomposition process to obtain multiple intrinsic mode functions (IMFs) and residuals corresponding to sunny days. The multiple IMFs, residuals, and four environmental features (temperature, relative humidity, global solar radiation, and global solar tilt) are then input into the SSA-optimized BiGRU neural network to obtain the prediction results for sunny days. The cloudy day dataset is processed according to the above process to obtain the prediction results for cloudy days. The rainy day dataset is processed according to the above process to obtain the prediction results for rainy days.

[0013] In a further optimized manner, the prediction results of each intrinsic mode function and residual are integrated in a strict time-series superposition manner to obtain the final photovoltaic power generation prediction value; the prediction result of the prediction model has a determination coefficient higher than 0.96.

[0014] Technical Effects: This invention innovatively employs a combined approach of FCM clustering dimensionality reduction and SSA two-stage optimization. FCM reduces data complexity by partitioning the weather dataset, while SSA optimizes EMD decomposition parameters and BiGRU hyperparameters separately, replacing traditional empirical configuration methods. This approach accurately addresses prediction bias caused by unreasonable hyperparameter configurations in existing models, effectively captures the frequency characteristics of photovoltaic power, and improves the accuracy and adaptability of short-term forecasts. Attached Figure Description

[0015] Figure 1 This is a flowchart of an EMD-BiGRU photovoltaic power generation prediction method based on FCM clustering and SSA optimization according to the present invention. Figure 2 It is a heatmap of the Pearson coefficients for each item in the dataset; Figure 3 This is a diagram of the sunny day results obtained by FCM clustering in this invention; Figure 4 This is a multi-cloud result image obtained by FCM clustering in this invention; Figure 5 This is a rainy day result image obtained by FCM clustering in this invention; Figure 6 This is a flowchart of the overall prediction model of this invention; Figure 7 This is a diagram showing the SSA-optimized EMD decomposition results under sunny weather conditions according to the present invention; Figure 8 This is a diagram showing the SSA-optimized EMD multi-cloud decomposition results of this invention; Figure 9 This is a diagram showing the SSA-optimized EMD decomposition results for rainy days according to the present invention; Figure 10 This is a comparison chart of the sunny day clustering prediction curve of the model designed in this invention and the traditional machine learning model; Figure 11 This is a comparison chart of the multi-cloud clustering prediction curve of the model designed in this invention and the traditional machine learning model; Figure 12 This is a comparison chart of the rainy day clustering prediction curve of the model designed in this invention and the traditional machine learning model. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] In existing technologies, the hyperparameters of neural network prediction models for photovoltaic power generation rely on empirical configuration, resulting in insufficient model accuracy.

[0018] Based on this, please refer to Figures 1-12This embodiment provides an EMD-BiGRU photovoltaic power generation prediction method based on FCM clustering and SSA optimization, including: acquiring environmental meteorological data and historical photovoltaic power generation data of photovoltaic power plants, with the data sampling interval strictly controlled at 5 minutes to ensure the temporal continuity and integrity of the data. The environmental meteorological data includes effective power generation, wind speed, temperature, relative humidity, global solar horizontal radiation, diffuse horizontal radiation, wind direction, daily rainfall, global solar tilt, and solar diffusion tilt. The historical photovoltaic power generation data mainly records the actual power generation data at different times. The acquired data is preprocessed, including outlier removal, missing value imputation, time filtering, normalization, and feature filtering; specific operations are performed as defined in the subsequent claims. The preprocessed data is input into the FCM clustering module, using solar horizontal radiation as the core clustering criterion. By constructing an objective function and iteratively updating cluster centers and membership degrees, the data is ultimately divided into three weather datasets: sunny, cloudy, and rainy. Each dataset corresponds to similar meteorological characteristics, reducing data complexity. The EMD decomposition process is optimized using SSA (Search-Assisted Algorithm). First, SSA parameters such as population size and spatial dimension are initialized, and individual fitness is calculated. Through dynamic collaboration among discoverers, followers, and watchdogs, the optimal parameter k for EMD decomposition is searched. Adaptive EMD decomposition is performed on three types of weather datasets to obtain multiple intrinsic mode functions (IMFs) and residuals, fully extracting different frequency features of the power signal. Simultaneously, SSA is used to optimize the hyperparameters of the BiGRU neural network, including the number of hidden layer neurons, batch size, and learning rate. Similarly, the optimal hyperparameter combination is determined through SSA population initialization, fitness calculation, and position updates for each role. The multiple IMFs and residuals obtained from EMD decomposition, along with selected meteorological features (temperature, relative humidity, global solar radiation, and global solar tilt), are input into the optimized BiGRU neural network. The network processes the input sequence both forward and backward, generating and merging bidirectional hidden states. Predictions are then made for each IMF and residual, yielding corresponding prediction results. These prediction results are then superimposed in strict temporal order to ensure temporal consistency, ultimately obtaining the predicted value of future photovoltaic power generation. In this technical solution, FCM clustering reduces data redundancy and interference by classifying data according to meteorological characteristics, and the two-stage optimization of SSA ensures the sufficiency of EMD decomposition and the optimal performance of the BiGRU model. The various links are closely connected to form a complete technical link from data processing to model optimization and then to prediction output, ensuring that those skilled in the art can accurately implement the above steps and solve the problem of insufficient prediction accuracy caused by empirical configuration of hyperparameters.

[0019] The lack of clear data types and parameters in existing photovoltaic power prediction methods leads to a lack of targeted data collection.

[0020] Based on this, each record of environmental meteorological data and photovoltaic power generation historical data includes effective power generation, wind speed, temperature, relative humidity, global solar horizontal radiation, diffuse horizontal radiation, wind direction, daily rainfall, global solar tilt, and solar diffusion tilt.

[0021] The effective power generation unit is kW, used to quantify the actual output power of photovoltaic power generation; wind speed is measured in m / s, reflecting the impact of airflow speed on the heat dissipation and power generation efficiency of photovoltaic modules; temperature is measured in °C, characterizing the effect of ambient temperature on the conversion efficiency of photovoltaic cells; relative humidity has no unit, but indirectly affects power generation by influencing the surface cleanliness and heat dissipation of photovoltaic modules; global solar horizontal radiation, diffuse horizontal radiation, global solar tilt, and solar diffusion tilt are all measured in W / m², directly reflecting the intensity and distribution of solar radiation energy, which are core factors affecting photovoltaic power generation; wind direction is measured in degrees, characterizing the direction of wind, and together with wind speed data, it can more comprehensively assess the impact of wind on photovoltaic systems; daily rainfall is measured in mm, reflecting the shading effect of rainfall on solar radiation. The data sampling interval is strictly controlled at 5 minutes. This sampling interval ensures the temporal density of the data, capturing the dynamic changes in photovoltaic power and meteorological conditions, while avoiding data redundancy due to overly dense sampling, thus reducing the burden on storage and computation. Those skilled in the art can, based on this limitation, select the corresponding sensor and data acquisition equipment, and collect data according to the specified data type and sampling interval to ensure that the collected data meets the requirements of subsequent preprocessing and model training, standardize data acquisition standards, and ensure the consistency and effectiveness of data quality.

[0022] The lack of a unified process and quantitative standards for data preprocessing leads to poor data quality, which affects model training. Based on this, data preprocessing was implemented according to the following specific steps. First, outlier detection was performed on the collected raw data. The 3σ criterion was used to identify obvious outliers; that is, when a data value exceeded the mean plus or minus three standard deviations, it was determined to be an obvious outlier and deleted to avoid interference from outlier data on model training. For missing values ​​in the data, the average of the preceding and following 576 data points was used for imputation. Since the data sampling interval was 5 minutes, the 576 data points corresponded to complete data from the preceding and following two days. This imputation method can fully utilize the correlation between adjacent time-series data, ensuring the rationality of the imputed data.

[0023] Considering the absence of solar radiation at night, the photovoltaic system does not generate electricity. Only valid data from 07:00 to 18:30 daily is retained. This time period covers the main period of solar radiation, and the daily data retention is exactly 138 records, effectively reducing the proportion of invalid data. After imputing outliers and missing values, the data is normalized using the MIN-MAX method. The normalization formula is as follows: The formula is designed based on the fundamental principle of data normalization in statistical learning. Its purpose is to eliminate dimensional differences between different data types, mapping all data uniformly to the 0-1 interval, which facilitates the subsequent training and convergence of the neural network. This is a normalized value, without units, and its value ranges from 0 to 1. The value is from the original data, and the unit depends on the data type, such as temperature in °C, effective power generation in kW, etc. The maximum value in the original data; This is the minimum value in the original data. The logical derivation process is as follows: firstly, through... Calculate the offset of the original data relative to the minimum value, eliminate the effect of absolute data offset, and then divide by... The offset is standardized to the 0-1 range to ensure comparability of data with different dimensions and numerical ranges. The theoretical basis of this formula is a standardization method for continuous data preprocessing, applicable to the preprocessing of meteorological and power data in photovoltaic forecasting, effectively improving model training efficiency and prediction accuracy. After normalization, the Pearson correlation coefficient between each environmental feature and actual power generation is calculated, and meteorological features with a significant impact on power generation are selected. The formula for calculating the Pearson correlation coefficient is as follows: This formula measures the degree of linear correlation between two variables, and its design is based on the theories of covariance and standard deviation in probability theory and mathematical statistics. For environmental characteristics, such as temperature and relative humidity, the unit depends on the type of characteristic; The mean of environmental characteristics, in units of... Consistent; This refers to actual power, measured in kW. Average power, in kW; This represents the total number of samples, without units. The logical derivation process is as follows: The numerator... Covariance is used to measure and The combined trend of change; the denominator part for Standard deviation and The product of the standard deviations is used to standardize the covariance, making the correlation coefficient... The value range of is limited to [-1, 1], which facilitates intuitive judgment of the correlation strength. The theoretical basis of this formula is the correlation analysis method of normally distributed variables. There is a significant linear correlation between the meteorological characteristics of photovoltaic systems and power generation, which is consistent with the applicable scenario of this formula. By calculating this coefficient, four environmental features—temperature, relative humidity, global solar radiation, and global solar tilt—are selected as input features for the subsequent neural network. Features with weak correlations, such as diffuse horizontal radiation and wind speed, are eliminated, simplifying the input dimensions and improving the model training efficiency.

[0024] FCM clustering lacks clear computational logic and termination conditions, leading to unstable clustering results.

[0025] Based on this, FCM clustering uses solar horizontal radiation as the clustering basis and is implemented according to the following specific logic. First, an objective function based on Euclidean distance is constructed, with the following expression: ; The objective function is designed to achieve fuzzy clustering of data by minimizing the sum of weighted Euclidean distances from samples to cluster centers, thus grouping data with similar solar radiation characteristics into one class. Let be the feature vector of the i-th sample, which is the solar horizontal radiation data, in W / m². For the j-th cluster center, the average solar horizontal radiation corresponding to the three types of weather is expressed in W / m². This is the fuzziness index, typically with a value of 2, and has no unit. It is used to control the degree of fuzziness in clustering. Let be the membership degree of the i-th sample belonging to the j-th cluster. It is dimensionless and ranges from 0 to 1, representing the probability that the sample belongs to that cluster. The total number of samples is unitless. The logical derivation process is as follows: Euclidean distance. This is used to measure the similarity between the i-th sample and the j-th cluster center; the smaller the distance, the higher the similarity. Membership is introduced. The weights are represented by the power of m, reflecting the characteristic that samples in fuzzy clustering can belong to multiple clusters simultaneously. The objective function calculates the sum of weighted distances from all samples to all cluster centers through double summation, and the minimum value corresponds to the optimal clustering result. This formula is based on fuzzy set theory, breaking through the limitation of samples belonging to only one cluster in traditional hard clustering, and better reflecting the continuous and fuzzy characteristics of photovoltaic meteorological data. During the clustering process, the Lagrange multiplier method is used to iteratively update the cluster centers and membership degrees. The cluster center update expression is: ,in This represents the number of iterations, which is unitless. The derivation logic of this formula is that, for the objective function... Regarding the cluster center Taking the partial derivatives and setting them to zero, and combining this with the constraints of the Lagrange multiplier method, we obtain the iterative update formula for the cluster center. The numerator part... The weighted sum of membership degrees for the (k-1)th iteration, with the denominator being... Let be the sum of the membership weights in the (k-1)th iteration. The ratio of these two weights is the updated cluster center, ensuring that the cluster center always lies at the weighted mean of the samples in that cluster. The theoretical basis is the gradient descent idea in optimization theory. By iteratively adjusting the cluster center position, the objective function value is continuously reduced, improving clustering accuracy. The membership update expression is: The derivation of this formula is also based on the Lagrange multiplier method, for the objective function Regarding membership degree Taking the partial derivative and setting it to zero, combined with the constraint that the sum of membership degrees is 1, we obtain the update formula for membership degrees. The core logic is that the smaller the distance from a sample to the center of the j-th cluster, the greater the membership degree of that sample belonging to the j-th cluster, and vice versa. The membership degree distribution is determined by the fuzzy index *m* and is used to adjust the distribution characteristics of the membership degree. The theoretical basis is the dynamic adjustment mechanism of membership degrees in fuzzy clustering, ensuring a reasonable distribution of samples among clusters. An upper limit for convergence is set during the clustering process. It typically takes a value of 0.001, has no unit, and is valid when the following conditions are met. At that time, clustering is completed, among which This is the membership matrix after the k-th iteration. It is unitless, and the matrix elements are the membership values ​​of each sample. Let be the membership matrix after the (k-1)th iteration. The logic of this convergence criterion formula is that when the difference between the membership matrices of two adjacent iterations is less than the upper convergence limit, it indicates that the clustering result has stabilized, the objective function has approached a local minimum, and further iterations will not significantly improve the clustering effect. The theoretical basis is the convergence criterion of numerical iterative algorithms, ensuring the stability and effectiveness of the clustering process and avoiding invalid iterations.

[0026] The lack of clear execution steps and parameter definitions in the SSA optimization process leads to poor optimization results.

[0027] Based on this, SSA is implemented according to the following specific workflow. First, population initialization is performed. Let the population size of individual sparrows be p, typically 30, without units; let the spatial dimension be m, without units, corresponding to the number of parameters to be optimized. For example, m=1 when optimizing the EMD decomposition parameter k, and m=3 when optimizing the BiGRU hyperparameter, corresponding to the number of hidden layer neurons, the batch size, and the learning rate, respectively. The position of the i-th sparrow in m-dimensional space is represented as... The purpose of this formula is to construct a search space for optimizing parameters, with each individual sparrow corresponding to a set of parameter combinations. Let be the position of the i-th sparrow in the j-th dimension, unitless, and its value corresponds to a candidate value of the j-th optimization parameter. The logical derivation process is as follows: for each optimization parameter, a reasonable search range is set, such as the search range of 3 to 12 for the EMD decomposition parameter k. The positions of individual sparrows are randomly generated within this range, forming an initial population to ensure the comprehensiveness of the search space. The theoretical basis is the population initialization idea of ​​swarm intelligence optimization algorithms, which avoids the algorithm getting trapped in local optima through a diversified initial population. After population initialization, the fitness of individual sparrows is calculated, and the fitness expression is: ,in The fitness value of a sparrow is a unitless value, calculated by substituting the parameter combination corresponding to the sparrow's position into the objective function. The objective function is typically set as the reciprocal of the prediction error; the smaller the prediction error, the larger the fitness value. The logical derivation is that the fitness function establishes a mapping relationship between parameter combinations and the optimization objective, providing a basis for subsequent position updates. The theoretical basis is the design principle of the fitness function in optimization algorithms, ensuring accurate evaluation of the merits of parameter combinations. Next, the discoverer's position is updated. The discoverer is an individual with a fitness value ranking in the top 10%-20%, responsible for the global search. When the safety threshold R2 is less than the alarm threshold T, R2 typically ranges from 0 to 1, and T typically takes a value of 0.8, indicating a safe area, the position update expression is: Where t is the number of iterations (unitless); a is a constant, taking the value 1 (unitless); and t_max is the maximum number of iterations, typically taking the value 100 (unitless). The logic of this formula is based on an exponential function. A decay factor is generated so that the discoverer's position gradually approaches the optimal solution with each iteration, achieving a global search. When the safety threshold R2 is greater than or equal to the alarm threshold T, it indicates that the discoverer is in a dangerous area, and the position update expression is: Where U is a random number following a normal distribution, without units, with a mean of 0 and a variance of 1; L is a 1×m matrix with all elements being 1, without units. The logic of this formula is to introduce a random number U to randomly change the discoverer's position, allowing it to escape the current search region and avoid getting trapped in local optima. The theoretical basis is the global exploration mechanism of swarm intelligence algorithms, balancing the breadth and depth of the search. After the discoverer's position is updated, the follower's position is updated. Followers are individuals with lower fitness values, responsible for local development around the discoverer. When i > p / 2, it indicates that the follower's fitness value is poor, and the position update expression is: Where S_worst is the current worst position, unitless, corresponding to the sparrow position with the smallest fitness value. The logic of this formula is to randomly move followers with poor fitness in the opposite direction of the worst position to find a better combination of parameters. When i≤p / 2, the position update expression is: ,in Let A be a 1×m matrix with elements being random numbers between [-1, 1], without units; S_best^(t+1) is the optimal position in the (t+1)th iteration, also without units. The logic of this formula is to allow followers with better fitness to move closer to the optimal position found by the discoverer, achieving local exploitation. The theoretical basis is the balance between local exploitation and global exploration, improving optimization efficiency. Finally, the positions of the watchdogs are updated; watchdogs comprise 10%-20% of the total population and are responsible for randomly escaping local optima. When the position update expression is given, it indicates that the sparrow's fitness value is better than the current global optimum, but it may be trapped in a local optimum. ,in Let be the optimal solution at time t, which is unitless; β is the step size control parameter, following a normal distribution with a mean of 0 and a variance of 1, and is also unitless. The logic of this formula is to move the sparrow closer to the global optimum by adjusting the step size, thus avoiding local optima. At that time, the position update expression is ,in The worst-case fitness is given by ε, which is unitless; E is a random number in the range [-1, 1], also unitless; ε is a non-zero number to prevent the denominator from being zero, with a value of 1e-6, also unitless. The logic of this formula is to move the sparrow away from the worst-case position and randomly search for new regions. The theoretical basis is a mechanism to avoid premature convergence of the algorithm, ensuring that the globally optimal solution is found. After all positions are updated, the fitness is used to determine the optimal solution. Determine if the stopping condition is met. The stopping condition is usually set to the number of iterations reaching t_max or the fitness value no longer changing significantly. If the condition is met, output the optimal sparrow position and the corresponding optimal parameter configuration. Otherwise, return to the discoverer position update step and continue iterating.

[0028] The lack of unified execution standards and termination rules in the EMD decomposition process leads to inconsistent decomposition results.

[0029] Based on this, the SSA-optimized EMD decomposition is implemented according to the following specific steps. First, the original time series is identified. All maximum and minimum points, the original time series The data represents the clustered photovoltaic power data, where t is time (in minutes) and X is photovoltaic power (in kW). A sliding window method is used to identify maxima and minima. A data point is considered a maximum if it is greater than its immediate neighbors, and a minimum if it is less than its immediate neighbors. Subsequently, a cubic spline interpolation function is used to fit the upper and lower envelopes of the time series. The upper envelope is obtained by fitting all maxima, and the lower envelope by fitting all minima. The cubic spline interpolation function has continuous and smooth characteristics, accurately reflecting the trend of the series. The mean of the upper and lower envelopes is calculated. The formula for calculating the mean is: ,in The power value of the upper envelope at time t is expressed in kW. The lower envelope power value at time t, in kW; The value is the mean, and the unit is kW. The purpose of this formula is to extract the trend term from the original sequence. The logical derivation is as follows: the upper envelope reflects the upper limit trend of the sequence, the lower envelope reflects the lower limit trend, and the mean of both can comprehensively reflect the overall trend of the sequence. The theoretical basis is the core idea of ​​EMD decomposition, which is to extract the intrinsic mode function by separating the trend term and the volatility term. Using the original sequence... Subtract the mean of the envelope A new sequence is obtained The calculation formula is: , The sequence after removing the trend term is expressed in kW. The logic of this formula is to separate the trend term from the original sequence, obtaining a preliminary fluctuation sequence, which prepares for subsequent extraction of the intrinsic mode function (IMF). The theoretical basis is that the IMF needs to satisfy the condition that the mean of the upper and lower envelopes is zero; therefore, the trend term of the original sequence needs to be removed first. Next, we will test... Whether the Inherent Mode Function (IMF) is satisfied depends on two conditions: the first condition is that the mean of the upper and lower envelopes is zero at any given time; the second condition is that the number of extreme points and zero-crossing points of the component signal is the same or differs by only one. Visualization and statistical counting methods are used to verify this. If neither of the above two conditions is met, then... As the original sequence, repeat the steps above—finding the maxima and minima, fitting the envelope, calculating the mean, and obtaining a new sequence—until the first intrinsic mode function that satisfies the conditions is obtained. , This represents the high-frequency intrinsic mode function, expressed in kW, reflecting the high-frequency fluctuation characteristics of photovoltaic power. The original sequence is used. Subtract the first intrinsic mode function Obtain the remaining sequence The calculation formula is: , The remaining sequence after removing high-frequency components, in kW, includes low-frequency components and a trend term. The remaining sequence... As a new original sequence, all the above steps are repeated to extract the second, third, and so on up to the nth intrinsic mode function. The decomposition process stops when the remaining sequence can no longer be decomposed into intrinsic mode functions that meet the conditions. The remaining sequence at this point is the final residual, expressed in kW, reflecting the long-term trend of photovoltaic power. This decomposition process controls the number of decomposition layers through the parameter k optimized by SSA to ensure optimal decomposition results and fully extract different frequency characteristics of the photovoltaic power signal, providing effective input data for subsequent prediction.

[0030] The gating mechanism and hidden state calculation logic of BiGRU neural networks are unclear, resulting in poor model training performance.

[0031] Based on this, the SSA-optimized BiGRU neural network is implemented according to the following specific logic. This neural network includes update and reset gates to control the flow of information and solve the gradient vanishing problem of traditional RNNs. The calculation formula for the update gate is as follows: ,in This is a sigmoid function, which is dimensionless and outputs values ​​between 0 and 1. It is used to map input signals to this range to achieve gated switching control. The weight matrix for updating the gate is unitless and its dimension is the hidden state dimension multiplied by (the hidden state dimension plus the input dimension), which is obtained adaptively through the model training process. It represents the hidden state at time t-1, has no unit, and contains sequence information from the previous time step; The input data at time t is given, and the unit depends on the data type, such as the intrinsic mode function value being kW, the meteorological characteristic temperature being ℃, etc. This means concatenating the hidden state at time t-1 with the input data at time t. The purpose of this formula is to determine how much historical information to retain at the current time. The logical derivation process is as follows: first, the historical hidden state... and current input splicing, through the weight matrix A linear transformation is performed to enhance the model's expressive power, followed by activation using the Sigmoid function. The output value of the update gate is then calculated. When the output value is close to 1, it indicates that most historical information is retained; when the output value is close to 0, it indicates that most historical information is discarded. This is based on the design principle of the gated recurrent unit, which can adaptively adjust the proportion of historical information retained, making it suitable for capturing long-term dependencies in time-series data. The formula for calculating the reset gate is as follows: ,in The weight matrix for resetting the gate is unitless and dimensionless, and the weight matrix for updating the gate is also given. Consistent with the model training, this formula is designed to determine how much historical information to ignore at the current moment. The logical derivation process is similar to that of the update gate, involving concatenation. and After linear transformation and sigmoid activation, the output reset gate value is determined. When the output value is close to 1, it indicates that historical information is retained for use in the current time step; when the output value is close to 0, it indicates that historical information is ignored and only the current input information is used. The theoretical basis is also the design principle of the gating mechanism, which, in conjunction with the update gate, achieves precise control over the flow of information. The SSA-optimized BiGRU neural network processes the sequence both from front to back and from back to front. Processing from front to back generates positive hidden states, and the formula for calculating the positive hidden states is... ,in This represents the positive hidden state at time t, and has no unit. This represents the positive hidden state at time t+1, and has no unit. The logic of this formula is based on the current input. and the positive hidden state at the next moment Calculate the forward hidden state at the current time step to capture the forward temporal dependencies of the sequence. Generate the backward hidden state by processing from the end to the beginning; the formula for calculating the backward hidden state is... ,in This represents the reversed hidden state at time t, and has no unit. This represents the reversed hidden state at time t-1, and has no unit. The logic of this formula is based on the current input. The reverse hidden state of the previous moment The algorithm calculates the reverse hidden state at the current time step to capture the reverse temporal dependencies of the sequence. The theoretical basis is the structure of a bidirectional recurrent neural network, which can fully utilize the global information of the sequence and improve the model's ability to capture temporal features. The forward and reverse hidden states are merged by concatenation to obtain the final hidden state, and the formula for calculating the final hidden state is as follows: ,in This represents a concatenation operation, without units, where the dimension of the concatenated hidden state is the sum of the dimensions of the forward and reverse hidden states. The logic behind this formula is to fuse forward and reverse temporal features, making the final hidden state contain more comprehensive sequence information. The theoretical basis is the concept of feature fusion, combined with bidirectional temporal dependencies, to improve the model's expressive power and prediction accuracy. Finally, a fully connected layer maps the final hidden state to predicted values, yielding the prediction results for each intrinsic mode function and residual.

[0032] The lack of clarity regarding the specific parameter optimization range and results for SSA optimization leads to a lack of basis for the optimization process.

[0033] Based on this, SSA optimization is implemented as follows. For the EMD decomposition parameter k, k is the total number of intrinsic mode functions and residuals obtained from the decomposition, without units. The optimization range for SSA optimization is set to 3 to 12. This range is based on the frequency characteristics of the photovoltaic power signal, typically containing 3-12 different frequency components, which can fully cover high-frequency fluctuations, mid-frequency variations, and low-frequency trends. SSA optimization is performed on three types of weather datasets: sunny, cloudy, and rainy. The optimization process is implemented according to the SSA workflow described in claim 5. The final determined k values ​​are 9, 9, and 10, respectively. The photovoltaic power fluctuations on sunny and cloudy days are relatively regular, and a decomposition layer of 9 is sufficient to extract the frequency features. The photovoltaic power on rainy days is affected by rainfall shading, resulting in more complex fluctuations. Therefore, the decomposition layer is increased to 10 layers to ensure sufficient decomposition. For the hyperparameters of the BiGRU neural network, the hyperparameters optimized by SSA include the number of hidden layer neurons, the batch size, and the learning rate. The number of hidden layer neurons determines the model's expressive power. It is dimensionless, and its optimization range is set at 32, 64, 128, 256, and 512. This range is based on the commonly used neuron count range for recurrent neural networks in deep learning. Considering the scale of the photovoltaic prediction dataset, it avoids underfitting due to too few neurons or overfitting and excessive computation due to too many neurons. After SSA optimization, the final determined number of hidden layer neurons is 256. This number balances model expressive power and computational efficiency, making it suitable for photovoltaic power prediction scenarios. The batch size affects the model's training stability and convergence speed. It is dimensionless, and its optimization range is set at 32, 64, 128, 256, and 512. This range references the commonly used batch sizes in deep learning. After SSA optimization, the final determined batch size is 32. This number ensures the stability of the training process while avoiding excessive memory consumption caused by excessively large batch sizes. The learning rate controls the step size of parameter updates during model training. It is unitless, and the optimization interval is set to 0.01, 0.03, 0.1, 0.3, and 0.5. This interval covers the commonly used learning rate range in deep learning. An excessively large learning rate can lead to training oscillations and non-convergence, while an excessively small learning rate can result in slow training speed. After SSA optimization, the final determined learning rate is 0.01. This learning rate ensures slow and stable model convergence, improving training accuracy. All hyperparameter optimization processes are strictly implemented according to the SSA workflow described in claim 5 to ensure that the obtained parameter combinations are optimal.

[0034] The inconsistent processing procedures for the three types of weather datasets lead to a lack of consistency in prediction results.

[0035] Based on this, the three types of weather datasets are processed according to the following unified process. For the sunny day dataset, it is first input into the SSA-optimized EMD decomposition process. The decomposition parameter k is 9, which is determined by SSA optimization. Following the EMD decomposition steps described in claim 6, all maximum and minimum points of the sunny day photovoltaic power sequence are found, upper and lower envelopes are fitted, the mean is calculated, the trend term is removed, the intrinsic mode function conditions are checked, and the iteration is repeated until 9 intrinsic mode functions and 1 residual are obtained. Then, these 9 intrinsic mode functions, 1 residual, and four selected environmental features—temperature, relative humidity, global solar radiation, and global solar tilt—are input into the SSA-optimized BiGRU neural network. The hyperparameters of the neural network are 256 hidden layer neurons, 32 batch samples, and a learning rate of 0.01, which are determined by optimization. Following the BiGRU calculation logic described in claim 7, the input data is processed simultaneously from front to back and from back to front, generating bidirectional hidden states and concatenating them. The prediction results of each intrinsic mode function and residual are output through a fully connected layer to obtain the prediction results corresponding to sunny days. For the cloudy dataset, the same processing procedure as the sunny dataset is used, with the EMD decomposition parameter k also set to 9. The same EMD decomposition steps are followed to obtain 9 intrinsic mode functions (IMFs) and 1 residual. These are then input into a BiGRU neural network with the same hyperparameter configuration to obtain the prediction results for cloudy days. For the rainy dataset, the processing procedure is consistent with both sunny and cloudy datasets, except that the EMD decomposition parameter k is adjusted to 10, determined through SSA optimization. The EMD decomposition steps are followed to obtain 10 IMFs and 1 residual. These are then input into a BiGRU neural network with the same hyperparameter configuration to obtain the prediction results for rainy days. This unified processing procedure ensures consistency in data processing under different weather conditions. By adjusting the number of decomposition layers only according to weather complexity, the adaptability of the model is guaranteed, and the reliability of the prediction results is improved. Those skilled in the art can follow this procedure to process the three types of weather datasets without additional adjustments, ensuring the standardization and repeatability of the prediction process.

[0036] The lack of clarity in the integration method of the prediction results led to insufficient accuracy in the final prediction values.

[0037] Based on this, the prediction results are integrated in the following specific manner. For the sunny day dataset, the prediction results of each intrinsic mode function and residual are superimposed in strict temporal order. That is, the final predicted value at time t is equal to the algebraic sum of all intrinsic mode function predictions and residual predictions at time t. The temporal order strictly follows the data sampling time, starting from 07:00 every day and superimposing at 5-minute intervals until 18:30 to ensure the accuracy of the time series. For the cloudy day dataset, the same temporal superposition method is used as for the sunny day dataset. The final predicted value at time t is the sum of all intrinsic mode function predictions and residual predictions at that time. The temporal consistency is strictly maintained during the superposition process to avoid prediction errors caused by time misalignment. For the rainy day dataset, the same temporal superposition method is used to integrate the prediction results of each intrinsic mode function and residual to obtain the final predicted value for rainy days. Verified using measured datasets from the Alice Springs 10.5kW subarray of the DKASC photovoltaic power station in Australia, the prediction model achieved a coefficient of determination (COD) higher than 0.96. The COD measures the degree of fit between predicted and actual values, ranging from 0 to 1. A COD higher than 0.96 indicates a very high degree of fit between the predicted and actual power values, accurately reflecting the changing trends of photovoltaic power. This integration method is based on the reversibility principle of EMD decomposition. The original sequence equals the sum of all intrinsic mode functions and residuals; therefore, the superposition of prediction results can reconstruct a complete photovoltaic power sequence. The theoretical basis is the fundamental principle of signal decomposition and reconstruction, ensuring that the integrated prediction results accurately reproduce the true characteristics of photovoltaic power and meet the accuracy requirements for short-term photovoltaic power generation prediction.

[0038] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A photovoltaic power generation prediction method based on FCM clustering and SSA optimization, characterized in that, Includes the following steps: S1: Acquire environmental meteorological data and historical photovoltaic power generation data of the photovoltaic power station; preprocess the acquired data; S2: The preprocessed data is divided into three weather datasets: sunny, cloudy, and rainy days using FCM clustering. S3: The SSA optimization process of EMD decomposition is adopted to perform adaptive decomposition on three types of weather datasets to obtain multiple intrinsic mode functions and residuals; S4: The hyperparameters of the BiGRU neural network are optimized using SSA. Multiple intrinsic mode functions, residuals and meteorological features are input into the optimized BiGRU neural network for prediction, and the prediction results of each intrinsic mode function and residual are obtained. S5: The prediction results are superimposed in time sequence to obtain the final photovoltaic power generation prediction value. The feature is that the complexity of photovoltaic data is reduced by FCM clustering, and the EMD decomposition parameters and BiGRU neural network hyperparameters are optimized by SSA respectively, replacing the empirical configuration method of hyperparameters.

2. The EMD-BiGRU photovoltaic power prediction method based on FCM clustering and SSA optimization according to claim 1, characterized in that, Each record of environmental meteorological data and historical photovoltaic power generation data includes effective power generation, wind speed, temperature, relative humidity, global solar horizontal radiation, diffuse horizontal radiation, wind direction, daily rainfall, global solar tilt, and solar diffusion tilt. The unit for effective power generation is kW, the unit for wind speed is m / s, the unit for temperature is ℃, the units for global solar horizontal radiation, diffuse horizontal radiation, global solar tilt, and solar diffusion tilt are all W / m², the unit for wind direction is degrees, the unit for daily rainfall is mm, and the unit for relative humidity is not specified. The data sampling interval is 5 min.

3. The EMD-BiGRU photovoltaic power prediction method based on FCM clustering and SSA optimization according to claim 1, characterized in that, Data preprocessing includes the following steps: removing obvious outliers; imputing missing values ​​by averaging the previous and next 576 data points; retaining valid data from 07:00 to 18:30 each day, with 138 data points retained daily; and normalizing the imputed data using the MIN-MAX method, with the normalization formula being... ,in The normalized value. The values ​​are from the original data. The maximum value in the original data. The minimum value in the original data is used; the Pearson correlation coefficient between each environmental feature and the actual power generation is calculated, and temperature, relative humidity, global solar radiation, and global solar tilt are selected as input features for the subsequent neural network. The formula for calculating the Pearson correlation coefficient is as follows: ,in As an environmental feature, This is the actual power. This represents the mean of environmental characteristics. This represents the average power.

4. The EMD-BiGRU photovoltaic power prediction method based on FCM clustering and SSA optimization according to claim 1, characterized in that, FCM clustering uses solar horizontal radiation as the clustering basis. The clustering process includes the following steps: constructing an objective function based on Euclidean distance, the expression of which is... ,in Let i be the feature vector of the i-th sample. For the j-th cluster center, For fuzzy index, Let be the membership degree of the i-th sample belonging to the j-th cluster. The total number of samples is given; the cluster centers and membership degrees are iteratively updated using the Lagrange multiplier method, and the cluster center update expression is: ,in Let be the number of iteration steps; the membership update expression is: Set the upper limit of convergence. When satisfied At that time, clustering is completed, among which Let be the membership matrix after the k-th iteration. This is the membership matrix after the (k-1)th iteration.

5. The EMD-BiGRU photovoltaic power prediction method based on FCM clustering and SSA optimization according to claim 1, characterized in that, The SSA workflow includes the following steps: Population initialization, assuming the population size of individual sparrows is... Spatial dimension is , To optimize the dimension of the optimal solution to the problem, The position of the i-th sparrow in 3D space is represented as: ； in Let be the position of the i-th sparrow in the j-th dimension; calculate the fitness of each sparrow, expressed as: ,in For individual sparrow fitness values; the discoverer's location is updated when the safety threshold is reached. Less than the alarm threshold At that time, the position update expression is When the safety threshold Greater than or equal to alarm threshold At that time, the position update expression is ,in For the number of iterations, It is a constant. The maximum number of iterations, To obtain random numbers that satisfy a normal distribution, for The matrix has all elements as 1; the follower position is updated when... At that time, the position update expression is ,when At that time, the position update expression is ,in , This is the worst position at present. The optimal position for the (t+1)th iteration; the guard's position is updated when... At that time, the position update expression is ,when At that time, the position update expression is ,in Let i be the fitness value of the i-th sparrow. For optimal fitness, For worst fitness, A random number in the range [-1, 1] To prevent non-zero numbers with a denominator of 0, The step size is a control parameter that follows a normal distribution with a mean of 0 and a variance of 1; based on fitness... Determine if the stopping condition is met. If it is, output the optimal sparrow position; otherwise, return to the discoverer's position update step.

6. The EMD-BiGRU photovoltaic power prediction method based on FCM clustering and SSA optimization according to claim 1, characterized in that, The SSA-optimized EMD decomposition process includes the following steps: finding the original time series. Find all the maxima and minima of the time series; fit the upper and lower envelopes of the time series using a cubic spline interpolation function, and calculate the mean of the upper and lower envelopes. The formula for calculating the mean is: ,in Let be the value of the upper envelope at time t. The value of the lower envelope at time t; using the original sequence Subtract the mean of the envelope A new sequence is obtained The calculation formula is: ;test Whether the Inherent Mode Function (IMF) is satisfied depends on two conditions: the first condition is that the mean of the upper and lower envelopes is zero at any given time; the second condition is that the number of extreme points and zero-crossing points of the component signal is the same or differs by only one. If neither of the above two conditions is met, then... Repeat the above steps as the original sequence until the first intrinsic mode function that satisfies the condition is obtained. ; using the original sequence Subtract the first intrinsic mode function Obtain the remaining sequence The calculation formula is: ; the remaining sequence As a new original sequence, repeat all the above steps until the remaining sequence can no longer be decomposed into intrinsic mode functions, resulting in multiple intrinsic mode functions and the final residual.

7. The EMD-BiGRU photovoltaic power prediction method based on FCM clustering and SSA optimization according to claim 1, characterized in that, The SSA-optimized BiGRU neural network includes an update gate and a reset gate. The calculation formula for the update gate is as follows: The formula for resetting the door is: ,in For the Sigmoid function, To update the gate weight matrix, To reset the weight matrix of the gate, This represents the hidden state at time t-1. Let t be the input data at time t; the SSA-optimized BiGRU neural network processes the sequence simultaneously from front to back and from back to front. Processing from front to back generates the positive hidden state, and the formula for calculating the positive hidden state is: The reverse hidden state is generated by processing from back to front. The formula for calculating the reverse hidden state is as follows: The forward hidden state and the reverse hidden state are merged by concatenation to obtain the final hidden state. The formula for calculating the final hidden state is as follows: ,in This indicates a splicing operation.

8. The EMD-BiGRU photovoltaic power generation prediction method based on FCM clustering and SSA optimization according to claim 1, characterized in that, The SSA-optimized EMD decomposition parameters are: , The optimization interval is 3 to 12. For three types of weather datasets—sunny, cloudy, and rainy—SSA optimization determines the optimal range. The values ​​are 9, 9, and 10 respectively; The hyperparameters of the BiGRU neural network optimized by SSA include the number of hidden layer neurons, the number of batch samples, and the learning rate. The optimization range for the number of hidden layer neurons is 32, 64, 128, 256, and 512, and the number of hidden layer neurons determined by SSA is 256. The optimization range for the number of batch samples is 32, 64, 128, 256, and 512, and the number of batch samples determined by SSA is 32. The optimization range for the learning rate is 0.01, 0.03, 0.1, 0.3, and 0.5, and the learning rate determined by SSA is 0.

01.

9. The EMD-BiGRU photovoltaic power generation prediction method based on FCM clustering and SSA optimization according to claim 1, characterized in that, The processing flow for the three types of weather datasets is as follows: The sunny day dataset is input into the SSA-optimized EMD decomposition process to obtain multiple intrinsic mode functions (IMFs) and residuals corresponding to sunny days. These IMFs, residuals, and four environmental features (temperature, relative humidity, global solar radiation, and global solar tilt) are then input into the SSA-optimized BiGRU neural network to obtain the prediction results for sunny days. The cloudy day dataset is processed according to the above process to obtain the prediction results for cloudy days. The rainy day dataset is processed according to the above process to obtain the prediction results for rainy days.

10. The EMD-BiGRU photovoltaic power prediction method based on FCM clustering and SSA optimization according to claim 1, characterized in that, The prediction results of each intrinsic mode function and residual are integrated in a strict time-series superposition manner to obtain the final photovoltaic power generation prediction value; the prediction result of the prediction model has a determination coefficient higher than 0.96.

Images