Short-term photovoltaic power prediction method and device based on multi-strategy optimization and data decomposition

Abstract

The short-term photovoltaic power prediction method based on multi-strategy optimization and data decomposition comprises the following steps: collecting meteorological data and historical power data of a target photovoltaic power station region to form a data set; using ICEEMDAN to decompose the historical power data to obtain an intrinsic mode component; inputting the intrinsic mode component group into an improved WLSSVM (Weighted Least Square Support Vector Machine) for prediction training; performing multi-strategy improvement on the snow melting algorithm to obtain an MISAO algorithm; in the training process, the white noise amplitude weight Nstd and the noise addition number NE of the ICEEMDAN are optimized through the MISAO, and the parameters lambda and delta of the WLSSVM are optimized through the MISAO; and after the prediction training is finished, the prediction result of the intrinsic mode component obtained through prediction is superposed to obtain a photovoltaic power prediction value. The device for realizing the method comprises a data acquisition module, a clustering module, a sequence decomposition module and a photovoltaic power prediction model construction and verification module. According to the method, multiple strategies are fused in the MISAO algorithm, the population diversity is increased, the search space is expanded, local optimum is avoided, the prediction efficiency is improved, and the generalization performance and stability of the model are enhanced.

Description

Technical field:

[0001] The present invention belongs to the technical field of photovoltaic power generation, and in particular relates to a short-term photovoltaic power prediction method and device based on multi-strategy optimization and data decomposition. Background technology:

[0002] With the vigorous development of the new energy industry, photovoltaic power generation systems have been developed on a large scale and connected to the grid at a high rate. However, photovoltaic power is significantly affected by weather factors, such as solar radiation, temperature and other meteorological conditions, which will have a more obvious effect on it. Therefore, there is significant randomness and volatility in photovoltaic power output, especially in short-term forecasts. Therefore, designing a photovoltaic power prediction method is of great significance to the sustainable development of the photovoltaic industry.

[0003] At present, the photovoltaic power prediction methods mainly include physical methods, statistical methods and prediction methods based on machine learning. The physical method is based on meteorological data and achieves prediction by building models of solar radiation, temperature and other factors. However, its prediction accuracy is poor under complex meteorological conditions or extreme weather conditions. The statistical method builds a prediction model by analyzing the correlation of historical data, which has the advantages of simple modeling and strong cross-regional adaptability. However, the accuracy of this method is significantly affected by data quality, and it is difficult to meet the real-time requirements of short-term photovoltaic prediction. In recent years, prediction methods based on machine learning, especially deep learning models such as neural networks, have been widely used in the field of photovoltaic prediction, which can effectively extract nonlinear features in data. However, the performance of such models depends to a large extent on hyperparameter settings. In the process of hyperparameter adjustment, traditional optimization algorithms often face problems such as slow convergence speed and easy to fall into local optimal solutions, which leads to unstable photovoltaic prediction effects and unsatisfactory accuracy. Summary of the invention:

[0004] In view of the above problems, the present invention provides a short-term photovoltaic power prediction method and device based on multi-strategy optimization and data decomposition to achieve the purpose of accurately and real-time prediction of photovoltaic power.

[0005] A short-term photovoltaic power forecasting method based on multi-strategy optimization and data decomposition, the method specifically includes:

[0006] Collecting meteorological data and historical power data of the target photovoltaic power station area to form a data set, wherein the meteorological data is a feature of the historical power data; wherein the meteorological data includes global irradiance, diffuse irradiance, temperature, wind speed, wind direction and atmospheric pressure;

[0007] The data set is preprocessed: the data in the data set is divided into three sub-datasets of sunny day, cloudy day and rainy day by using K-means clustering algorithm; each sub-dataset is divided into a training set, a validation set and a test set according to the proportion, wherein the training set and the validation set are used for model training and hyperparameter adjustment, and the test set is used for evaluating the model prediction results;

[0008] The adaptive noise complete empirical mode decomposition algorithm ICEEMDAN is used to decompose the historical power data in the data set to obtain a set of intrinsic mode components: {IMF 1 ,IMF 2 ,…,IMF n}; Input the intrinsic modal component group into the least squares support vector machine model WLSSVM for prediction training;

[0009] The multi-strategy improved snow melting algorithm is obtained by multi-strategy improved snow melting optimization algorithm MISAO. During the training process, the hyperparameters of ICEEMDAN and WLSSVM are optimized by multi-strategy improved snow melting optimization algorithm MISAO, wherein the hyperparameters of ICEEMDAN are white noise amplitude weight Nstd and noise addition number NE, and the hyperparameters of WLSSVM are regularization parameter λ and kernel parameter δ;

[0010] The multi-strategy improvement of the snowmelt algorithm is performed to obtain the multi-strategy improved snowmelt optimization algorithm MISAO, wherein the strategies include a cyclic chaotic mapping method, a diffusion mechanism and a random follow-up search strategy; the specific improvement method is:

[0011] In the snowmelt optimization algorithm, a random generation process is used to determine the starting position of each individual. However, this method cannot ensure the diversity of the population. It tends to locate the starting point quite far away from the ideal solution, thus affecting the accuracy of convergence and the efficiency of search. Therefore, the present invention adopts a cyclic chaotic mapping method to initialize the positions of all search agents, thereby increasing diversity. Its basic principle is explained as follows:

[0012] Among them, w i is the individual in the population, mod is the residual function;

[0013] The present invention introduces a diffusion strategy to expand the exploration entry of the search space and avoid falling into the local optimum. The diffusion mechanism is introduced. Under the Gaussian step size, the current individual is randomly perturbed and the optimal individual is selected to replace the current individual to generate three new individuals. The specific diffusion mechanism is: X i =G(X best ,γ,s)+σ-×(X best -X i-1 ), where G(X best,γ,s) is the generation of random matrices that obey Gaussian distribution, s is the vector of the generated matrix size, σ is a random variable that follows normal distribution, γ is the standard deviation, g is the current iteration number;

[0014] In the original snow melting algorithm SAO, the relocation of population members depends largely on the first member to be achieved, which leads to the corresponding dependence. Nevertheless, this dependence has the potential to trap the algorithm into a local optimal solution, making it difficult to escape. This dilemma arises because the population members tend to follow the current best individual excessively, lacking sufficient diversity and exploration ability. To solve this problem, a random follow search strategy is introduced, which aims to enhance the diversity of the population in the swarm intelligence algorithm and achieve a wider solution space traversal. The specific method of the strategy is as follows:

[0015]

[0016] pick=randperm(dim)

[0017] Among them, ub and lb are the upper and lower bounds of the exploration space of the ant colony search agent, respectively. is the i-th population member s i In the jth dimension after random combination, c 2 is a random number in the range [-1,1], fes is the current evaluation number, Max fes is the maximum number of evaluations; dim is the dimension of the feasible exploration space, k is a random integer in the range of [1, dim], and Temp is the member of the current search agent exchange.

[0018]

[0019] in, is the position of the jth dimension currently evaluated as having the best quality, is the i-1th group member s after the random combination process i-1 The jth dimension of

[0020] After the training is completed, the prediction results of the inherent modal components obtained after prediction are superimposed to obtain the final photovoltaic power prediction value.

[0021] Preferably, the specific method of improving ICEEMDAN to decompose the historical power data in each sub-dataset is:

[0022] Add white noise E to the historical power data X 1 [ω i ], generate auxiliary noise signal X (i) :X (i)=X+β 0 *E 1 [ω i ], where ω i is the i-th white noise added, β 0 is the signal-to-noise ratio;

[0023] The IMF components of each order are extracted by iterative calculation until the energy of the residual signal is lower than the preset threshold or cannot be further decomposed. Specifically, the hth IMF component is:

[0024]

[0025] Among them, R h is the hth order residue, h=1,2,3,…,N, is the hth IMF component, M(﹒) is the local average of the generated modal component, and I is the number of noise disturbances.

[0026] Preferably, the specific method of optimizing the hyperparameters of ICEEMDAN and WLSSVM by MISAO is:

[0027] MISAO initializes the hyperparameters of ICEEMDAN and WLSSVM according to a predetermined strategy;

[0028] Use current parameters to predict short-term PV power and calculate the model error to update the objective function;

[0029] MISAO continuously searches and adjusts hyperparameters until the objective function value is minimized.

[0030] Preferably, the preprocessing of the data set further includes eliminating bad data from the collected historical power data.

[0031] Preferably, the method for removing bad data is data cleaning and data interpolation, and the bad data includes missing data, erroneous data, duplicate data and data beyond a reasonable range.

[0032] Preferably, the data set is normalized after preprocessing, and the specific method is as follows:

[0033]

[0034] Among them, Y i is the normalized data, Y is the unnormalized data, and Y min is the minimum value of the input sequence China, Y max is the maximum value in the input sequence.

[0035] Preferably, the WLSSVM is based on the LSSVM optimization problem for each error Multiply by the coefficient v i To perform weighting.

[0036] Preferably, the objective function is calculated based on the prediction error.

[0037] Preferably, the error indicator is at least one of mean absolute percentage error (MAPE), mean absolute error (MAE) and root mean square error (RMSE). If RMSE is used for error evaluation, the objective function is:

[0038]

[0039] Among them, y i is the actual value, is the model's predicted value.

[0040] A short-term photovoltaic power prediction device based on multi-strategy optimization and data decomposition, the device comprising:

[0041] The data acquisition module is used to collect meteorological data and historical power data of the target photovoltaic power station area to form a data set;

[0042] Clustering module, based on K-means clustering algorithm, divides the data in the dataset into three sub-datasets: sunny, cloudy, and rainy;

[0043] Sequence decomposition module: Apply ICEEMDAN to decompose the datasets under different weather types into different IMF components;

[0044] Photovoltaic power prediction model construction and verification module: used to take meteorological data as input and historical power data of photovoltaic power stations as output, to construct the MISAO-ICEEMDAN-WLSSVM photovoltaic power prediction model: using the multi-strategy improved snowmelt optimization algorithm MISAO to optimize the white noise amplitude weights and noise addition number of the adaptive noise complete empirical mode decomposition ICEEMDAN, and the regularization parameters and kernel parameters of the minimum weighted support vector machine WLSSVM, to construct the MISAO-ICEEMDAN-WLSSVM photovoltaic power prediction model; training the MISAO-ICEEMDAN-WLSSVM photovoltaic power prediction model, and using indicators to evaluate the prediction accuracy of the model, the evaluation indicator is at least one of the mean absolute error MAE, the root mean square error RMSE and the mean absolute percentage error MAPE.

[0045] The present invention designs a short-term photovoltaic power prediction method and device with multi-strategy optimization and data decomposition, integrates multiple strategies in the MISAO algorithm, initializes the search agent position through cyclic chaotic mapping, increases population diversity; introduces diffusion strategy to expand the search space; and adopts random follow-up search strategy to avoid local optimality. These strategies improve the search efficiency and convergence accuracy of the snow melting algorithm, accelerate the model hyperparameter optimization process, and improve prediction efficiency. ICEEMDAN decomposes photovoltaic power data, reduces non-stationarity, and makes it easier for the model to capture data features. Combined with the MISAO algorithm, the WLSSVM model hyperparameters are optimized to enhance the generalization performance and stability of the model. At the same time, the present invention adopts the K-means clustering algorithm to classify the data according to weather type, and trains and predicts respectively according to the characteristics of different weather data, so that the model can better adapt to the changing law of photovoltaic power under different weather conditions, and improves the prediction accuracy and stability. The optimized model can quickly process new data and output prediction results in time, meet the requirements of the power system for rapid response to photovoltaic power, and provide strong support for real-time dispatching and stable operation of the power grid. Description of the drawings:

[0046] Attached Figure 1 This is a prediction flow chart based on the MISAO-ICEEMDAN-WLSSVM model of Example 1 of the present invention.

[0047] Attached Figure 2 This is a flow chart of the multi-strategy improved snow melting optimization algorithm of Example 1 of the present invention.

[0048] Attached Figure 3 This is the sequence decomposition result diagram of ICEEMDAN in Example 1 of the present invention - sunny day.

[0049] Attached Figure 4 This is a comparison chart of the prediction results of the seven algorithms in Example 1 of the present invention - sunny day.

[0050] Attached Figure 5 This is a comparison chart of the prediction results of the seven algorithms in Example 1 of the present invention - cloudy day.

[0051] Attached Figure 6 This is a comparison chart of the prediction results of the seven algorithms in Example 1 of the present invention - rainy day. Specific implementation method:

[0052] In order to make the technical solution of the present invention easier to understand, a short-term photovoltaic power prediction method and device with multi-strategy optimization and data decomposition disclosed in the present invention are clearly and completely described in the form of embodiments and drawings.

[0053] Embodiment 1:

[0054] like Figure 1As shown, a short-term photovoltaic power forecasting method based on multi-strategy optimization and data decomposition, the method specifically includes:

[0055] Step 100: Collect data: Take the power generation data of a photovoltaic power station in North China as an example, with an installed capacity of 20MW. The data collection period is from July 1, 2018 to December 31, 2018, including meteorological data and power data. The meteorological data includes global irradiance, diffuse irradiance, temperature, wind speed, wind direction and atmospheric pressure. The time interval is 15 minutes. Considering that the light intensity is mostly 0 in the evening and early morning, the time selected is 07:00-19:00.

[0056] Step 110: Data preprocessing: Remove bad data from the collected historical power data. The method of removing bad data is data cleaning and data interpolation. The bad data includes missing data, erroneous data, duplicate data and data beyond a reasonable range. Use the K-means clustering algorithm to divide the data in the data set into three sub-data sets: sunny, cloudy and rainy; normalize all sub-data sets. The specific method is as follows: Among them, Y i is the normalized data, Y is the unnormalized data, and Y min is the minimum value of the input sequence China, Y max is the maximum value in the input sequence. Each sub-dataset is divided into training set, validation set and test set according to the proportion. The training set and validation set are used for model training and hyperparameter adjustment, and the test set is used to evaluate the model prediction results.

[0057] Step 120: Use the adaptive noise complete empirical mode decomposition algorithm ICEEMDAN to decompose the historical power data in the data set to obtain a set of intrinsic mode components: {IMF 1 ,IMF 2 ,…,IMF n}; The specific decomposition process is:

[0058] Add white noise E to the historical power data X 1 [ω i ], generate auxiliary noise signal X (i) :X (i) =X+β 0 *E 1 [ω i ], where ω i is the i-th white noise added, β 0 is the signal-to-noise ratio;

[0059] The IMF components of each order are extracted by iterative calculation until the energy of the residual signal is lower than the preset threshold or cannot be further decomposed. Specifically, the hth IMF component is:

[0060]

[0061] Among them, R h is the hth order residue, h=1,2,3,…,N, is the hth IMF component, M(﹒) is the local average of the generated modal component, and I is the number of noise disturbances.

[0062] Step 130: Input the intrinsic modal component group into the least squares support vector machine model WLSSVM for prediction training; the WLSSVM is based on the LSSVM optimization problem to calculate the error of each item. Multiply by the coefficient v i To perform weighting, the optimization problem is:

[0063]

[0064] Among them, J is the objective function, ω is the weight coefficient vector, γ is the regularization parameter, and v i is the weight, v i Calculated based on sample training error; where y i is the true output of the i-th sample, x i is the ith input sample, is a function that maps input samples to a high-dimensional feature space, and b is a bias term;

[0065] Introducing the Lagrange multiplier, we get:

[0066]

[0067] Among them, a i is the Lagrange multiplier (i=1,2,…,n,a i ≥0);

[0068] According to the KTT condition and Mercer condition, the regression function of WLSSVM is:

[0069]

[0070] Among them, K(x i ,x j )=exp(-|x i -x j | 2 / 2σ 2 ) is the radial basis function.

[0071] Step 140: performing multi-strategy improvement on the snow melting algorithm to obtain a multi-strategy improved snow melting optimization algorithm MISAO. During the training process, the multi-strategy improved snow melting optimization algorithm MISAO is used to optimize the hyperparameters of ICEEMDAN and WLSSVM, wherein the hyperparameters of the ICEEMDAN are the white noise amplitude weight Nstd and the noise addition number NE, and the hyperparameters of the WLSSVM are the regularization parameter λ and the kernel parameter δ;

[0072] The specific method of optimizing the hyperparameters of ICEEMDAN and WLSSVM by MISAO is:

[0073] MISAO initializes the hyperparameters of ICEEMDAN and WLSSVM according to a predetermined strategy;

[0074] Using current parameters to predict short-term photovoltaic power, and calculating the error of the model to update the objective function; the objective function is calculated based on the prediction error, and the error index is at least one of mean absolute percentage error (MAPE), mean absolute error (MAE) and root mean square error (RMSE);

[0075] MISAO continuously searches and adjusts hyperparameters until the objective function value is minimized.

[0076] in:

[0077]

[0078]

[0079] If RMSE is used for error evaluation, the objective function is:

[0080]

[0081] Where n is the number of prediction results, y i is the actual value, is the model's predicted value.

[0082] like Figure 2 As shown in the figure, the multi-strategy improved snow melting algorithm is obtained by multi-strategy improved snow melting optimization algorithm MISAO, the strategies include cyclic chaotic mapping method, diffusion mechanism and random follow-up search strategy; the specific improvement methods are as follows:

[0083] The Snow Melt Algorithm (SAO) is divided into two stages: exploration and development. In the detection stage, after snow or liquid water sublimates or evaporates to form water vapor, the water vapor moves irregularly in space, and detection is performed through irregular motion. The position update formula of the detection process in the SAO algorithm is as follows:

[0084]

[0085] in, is the updated position of the individual, RB i (j) represents the Brownian motion random number vector, r 1 is a random number in [0,1], B(j) is the optimal solution for the current population, is the centroid position of the entire group in this position update, X i (j) is the position of the current individual, k is a random integer between [1,4], and Elite_pool(k) indicates that an individual is randomly selected from the set Elite_pool, where:

[0086]

[0087] Elite_pool(k)∈[B(j),X second (j),X third (j),Z c (j)]

[0088] Where, X second (j) and X third (j) is the second and third individuals in the current population in terms of fitness value, Z c (j) is the average value of the top 50% of individuals in the current population in terms of fitness.

[0089] At each position update during exploration, a value is randomly drawn from this set to assist in the position update.

[0090] The SAO algorithm calls the top 50% of individuals in the population fitness elite individuals to facilitate the calculation of Z c (j)

[0091]

[0092] Where N 1 is the number of elite individuals, so N 1 Numerically equal to half of N.

[0093] During the development phase of the SAO algorithm, it was mainly introduced that when snow is converted into water through melting behavior, in the SAO algorithm, the snow melting process is expressed by the degree-day method, and the formula is as follows:

[0094]

[0095] Where M(j) is the snowmelt rate, T(j) is the average daily temperature, and max are the current and maximum iterations respectively, and DDF(j) refers to the degree-day coefficient, which ranges from 0.35 to 0.6.

[0096] The position update equation at this stage is as follows:

[0097]

[0098] r 1 is a random number in [-1,1]. This form gives non-optimal individuals in the population a better chance to develop more feasible solutions based on the current optimal position.

[0099] In the snowmelt optimization algorithm, a random generation process is used to determine the starting position of each individual. However, this method cannot ensure the diversity of the population. It tends to locate the starting point quite far away from the ideal solution, thus affecting the accuracy of convergence and the efficiency of search. Therefore, the present invention adopts a cyclic chaotic mapping method to initialize the positions of all search agents, thereby increasing diversity. The basic principle is explained as follows:

[0100] Among them, w i is the individual in the population, mod is the residual function;

[0101] The present invention introduces a diffusion strategy to expand the exploration entry of the search space and avoid falling into the local optimum. The diffusion mechanism is introduced, and the current individual is randomly perturbed under the Gaussian step size and the optimal individual is selected to replace the current individual to generate three new individuals. The specific diffusion mechanism is:

[0102] X i =G(X best ,γ,s)+σ-×(X best -X i-1 );

[0103] Among them, G(X best ,γ,s) is the generation of random matrices that obey Gaussian distribution: s is the vector of the generated matrix size, σ is a random variable that follows normal distribution, γ is the standard deviation, In order to facilitate the use of individual positioning and approach the optimal solution, the parameter log(g) / g is used to reduce the Gaussian step size as the number of iterations increases, where g is the current number of iterations.

[0104] In the original snow melting algorithm SAO, the relocation of population members depends largely on the first member to be achieved, which leads to the corresponding dependence. Nevertheless, this dependence has the potential to trap the algorithm into a local optimal solution, making it difficult to escape. This dilemma arises because the population members tend to follow the current best individual excessively, lacking sufficient diversity and exploration ability. To solve this problem, a random follow search strategy is introduced, which aims to enhance the diversity of the population in the swarm intelligence algorithm and achieve a wider solution space traversal. The specific method of the strategy is as follows:

[0105]

[0106] pick=randperm(dim)

[0107] Among them, ub and lb are the upper and lower bounds of the exploration space of the ant colony search agent, respectively. is the i-th population member s i In the jth dimension after random combination, c 2 is a random number in the range [-1,1], fes is the current evaluation number, Max fes is the maximum number of evaluations; dim is the dimension of the feasible domain exploration space, k is a random integer in the range of [1, dim], Temp is the member of the current search agent exchange,

[0108]

[0109] in, is the position of the jth dimension currently evaluated as having the best quality, is the i-1th group member s after the random combination process i-1 The jth dimension of

[0110] Step 150: After the training is completed, the prediction results of the inherent modal components obtained after the prediction are superimposed to obtain the final photovoltaic power prediction value.

[0111] In order to verify the effectiveness of the improvement of the present invention, the three evaluation indicators of MAPE, RMSE and MAE were used to calculate the errors of the prediction effects of TCN, GRU, SVM, WLSSVM, SAO-WLSSVM and MISAO-ICEEMDAN-WLSSVM prediction models under three weather types, and the evaluation results are shown in Table 1.

[0112] Table 1 Prediction results of photovoltaic power by different models

[0113]

[0114] Under three weather conditions, the MAPE values ​​of the proposed model were reduced by at least 25.3%, 29.1% and 32.8% respectively compared with the other six models, the MAE values ​​were reduced by at least 39.0%, 17.8% and 32.8% respectively, and the RMSE values ​​were reduced by 37.2%, 16.0% and 13.3% respectively. The results show that the MISAO-ICEEMDAN-WLSSVM prediction model has high prediction accuracy and stability.

[0115] In order to strongly demonstrate the necessity of classifying PV forecasts according to different weather conditions, we selected the same model as before and made forecasts based on the data from July 1 to July 15, 2018. The indicators are recorded in Table 2.

[0116] Table 2 Necessity of classifying PV forecasts under different weather conditions

[0117]

[0118] The data in Table 2 show that key indicators such as MAE, MAPE, and RMSE are significantly higher when the classification is not based on weather conditions compared to the previous experiments based on weather classification. It can be observed that compared with the relatively smooth and accurate prediction curve in the previous classification case, the prediction curve in the unclassified case is more volatile and deviates more significantly from the actual value.

[0119] from Figure 3 It can be seen that ICEEMDAN decomposition obtains 12 groups of IMF components and one group of RES components. The sequences are arranged in order from high to low frequency, and the fluctuations of sequences with different frequencies have a certain regularity, which avoids mode aliasing.

[0120] Figures 4 to 6 The predicted values ​​and true values ​​of each comparison method are given. The closer the predicted value is to the true value, the higher the prediction accuracy of the method. The predicted values ​​of this method are better than those of the comparison method in multiple weather scenarios, indicating that the method disclosed in the present invention has better prediction stability.

[0121] Embodiment 2:

[0122] A short-term photovoltaic power prediction device based on multi-strategy optimization and data decomposition, the device comprising:

[0123] The data acquisition module is used to collect meteorological data and historical power data of the target photovoltaic power station area to form a data set;

[0124] Clustering module, based on K-means clustering algorithm, divides the data in the dataset into three sub-datasets: sunny, cloudy, and rainy;

[0125] Sequence decomposition module: Apply ICEEMDAN to decompose the datasets under different weather types into different IMF components;

[0126] Photovoltaic power prediction model construction and verification module: used to take meteorological data as input and historical power data of photovoltaic power stations as output, to construct the MISAO-ICEEMDAN-WLSSVM photovoltaic power prediction model: using the multi-strategy improved snowmelt optimization algorithm MISAO to optimize the white noise amplitude weights and noise addition number of the adaptive noise complete empirical mode decomposition ICEEMDAN, and the regularization parameters and kernel parameters of the minimum weighted support vector machine WLSSVM, to construct the MISAO-ICEEMDAN-WLSSVM photovoltaic power prediction model; training the MISAO-ICEEMDAN-WLSSVM photovoltaic power prediction model, and using indicators to evaluate the prediction accuracy of the model, the evaluation indicator is at least one of the mean absolute error MAE, the root mean square error RMSE and the mean absolute percentage error MAPE.

[0127] It should be pointed out that for ordinary technicians in this technical field, several improvements, substitutions, modifications and embellishments can be made without departing from the principles and purpose of the present invention. These improvements, substitutions, modifications and embellishments should also be regarded as the scope of protection of the present invention.

Claims

1. A short-term photovoltaic power forecasting method based on multi-strategy optimization and data decomposition, characterized in that: The method specifically comprises: Collecting meteorological data and historical power data of the target photovoltaic power station area to form a data set, wherein the meteorological data is a feature of the historical power data; wherein the meteorological data includes global irradiance, diffuse irradiance, temperature, wind speed, wind direction and atmospheric pressure; The data set is preprocessed: the data in the data set is divided into three sub-datasets of sunny day, cloudy day and rainy day by using K-means clustering algorithm; each sub-dataset is divided into a training set, a validation set and a test set according to the proportion, wherein the training set and the validation set are used for model training and hyperparameter adjustment, and the test set is used for evaluating the model prediction results; The adaptive noise complete empirical mode decomposition algorithm ICEEMDAN is used to decompose the historical power data in the data set to obtain a set of intrinsic mode components: {IMF1,IMF2,…,IMF n }; Input the intrinsic modal component group into the least squares support vector machine model WLSSVM for prediction training; The multi-strategy improved snow melting algorithm is obtained by multi-strategy improved snow melting optimization algorithm MISAO. During the training process, the hyperparameters of ICEEMDAN and WLSSVM are optimized by multi-strategy improved snow melting optimization algorithm MISAO, wherein the hyperparameters of ICEEMDAN are white noise amplitude weight Nstd and noise addition number NE, and the hyperparameters of WLSSVM are regularization parameter λ and kernel parameter δ; The multi-strategy improvement of the snowmelt algorithm is performed to obtain the multi-strategy improved snowmelt optimization algorithm MISAO, wherein the strategies include a cyclic chaotic mapping method, a diffusion mechanism and a random follow-up search strategy; the specific improvement method is: The cyclic chaos mapping method is introduced to initialize the positions of all search agents in the snow melting algorithm. The cyclic chaos mapping method is: Among them, w i is the individual in the population, mod is the residual function; Introduce a diffusion mechanism, randomly perturb the current individual under Gaussian step size and select the best individual to replace the current individual to generate three new individuals. The specific diffusion mechanism is: X i =G(X best ,γ,s)+σ-×(X best -X i-1 ), where G(X best ,γ,s) is the generation of random matrices that obey Gaussian distribution, s is the vector of the generated matrix size, σ is a random variable that follows normal distribution, γ is the standard deviation, g is the current iteration number; A random follow-up search strategy is introduced, and the specific method of the strategy is as follows: pick=randperm(dim) Among them, ub and lb are the upper and lower bounds of the exploration space of the ant colony search agent, respectively. is the i-th population member s i In the jth dimension after random combination, c2 is a random number in the range [-1,1], fes is the current evaluation number, Max fes is the maximum number of evaluations; dim is the dimension of the feasible domain exploration space, k is a random integer in the range of [1, dim], Temp is the member of the current search agent exchange, in, is the position of the jth dimension currently evaluated as having the best quality, is the i-1th group member s after the random combination process i-1 The jth dimension of After the prediction training is completed, the prediction results of the inherent modal components obtained after the prediction are superimposed to obtain the final photovoltaic power prediction value.

2. The short-term photovoltaic power prediction method based on multi-strategy optimization and data decomposition as claimed in claim 1 is characterized in that: The specific method of improving ICEEMDAN to decompose the historical power data in each sub-dataset is as follows: Add white noise E1[ω i ], generate auxiliary noise signal X (i) :X (i) =X+β0*E1[ω i ], where ω i is the i-th white noise added, β0 is the signal-to-noise ratio; The IMF components of each order are extracted by iterative calculation until the energy of the residual signal is lower than the preset threshold or cannot be further decomposed. Specifically, the hth IMF component is: Among them, R h is the hth order residue, h=1,2,3,…,N, is the hth IMF component, M(﹒) is the local average of the generated modal component, and I is the number of noise disturbances.

3. The short-term photovoltaic power prediction method based on multi-strategy optimization and data decomposition as claimed in claim 1, characterized in that: The specific method of optimizing the hyperparameters of ICEEMDAN and WLSSVM by MISAO is: MISAO initializes the hyperparameters of ICEEMDAN and WLSSVM according to a predetermined strategy; Use current parameters to predict short-term PV power and calculate the model error to update the objective function; MISAO continuously searches and adjusts hyperparameters until the objective function value is minimized.

4. The short-term photovoltaic power forecasting method based on multi-strategy optimization and data decomposition as claimed in claim 1, characterized in that: The preprocessing of the data set further includes eliminating bad data from the collected historical power data.

5. The short-term photovoltaic power prediction method based on multi-strategy optimization and data decomposition as claimed in claim 4, characterized in that: The method for eliminating bad data is data cleaning and data interpolation. The bad data includes missing data, erroneous data, duplicate data and data beyond a reasonable range.

6. The short-term photovoltaic power prediction method based on multi-strategy optimization and data decomposition as claimed in claim 1 is characterized in that: The data set is normalized after preprocessing, and the specific method is as follows: Among them, Y i is the normalized data, Y is the unnormalized data, and Y min is the minimum value of the input sequence China, Y max is the maximum value in the input sequence.

7. The short-term photovoltaic power forecasting method based on multi-strategy optimization and data decomposition as claimed in claim 1, characterized in that: The WLSSVM is based on the LSSVM optimization problem and calculates the error of each item. Multiply by the coefficient v i To perform weighting.

8. The short-term photovoltaic power forecasting method based on multi-strategy optimization and data decomposition as claimed in claim 3, characterized in that: The objective function is calculated based on the prediction error.

9. The short-term photovoltaic power prediction method based on multi-strategy optimization and data decomposition as claimed in claim 8, characterized in that: The error indicator is at least one of mean absolute percentage error (MAPE), mean absolute error (MAE) and root mean square error (RMSE).

10. A short-term photovoltaic power prediction device based on multi-strategy optimization and data decomposition, characterized in that: The device comprises: Data acquisition module: used to collect meteorological data and historical power data of the target photovoltaic power station area to form a data set; Clustering module: used to divide the data in the dataset into three sub-datasets: sunny, cloudy, and rainy based on the K-means clustering algorithm; Sequence decomposition module: used to decompose the datasets under different weather types into different IMF components using ICEEMDAN; Photovoltaic power prediction model construction and verification module: used to take meteorological data as input and historical power data of photovoltaic power stations as output, to construct the MISAO-ICEEMDAN-WLSSVM photovoltaic power prediction model: using the multi-strategy improved snowmelt optimization algorithm MISAO to optimize the white noise amplitude weights and noise addition number of the adaptive noise complete empirical mode decomposition ICEEMDAN, and the regularization parameters and kernel parameters of the minimum weighted support vector machine WLSSVM, to construct the MISAO-ICEEMDAN-WLSSVM photovoltaic power prediction model; training the MISAO-ICEEMDAN-WLSSVM photovoltaic power prediction model, and using indicators to evaluate the prediction accuracy of the model, the evaluation indicator is at least one of the mean absolute error MAE, the root mean square error RMSE and the mean absolute percentage error MAPE.

Images