Short-term wind power combination probability forecasting method and system considering meteorological classification

Abstract

The present application relates to the field of wind power prediction technology, and is a short-term wind power combination probability prediction method, system, terminal and storage medium considering meteorological classification in short term, the former is based on the meteorological condition classification method of subtraction clustering and GK fuzzy clustering algorithm, the division of meteorological type is realized. The advantages of parameter and non-parameter probability prediction model are comprehensively considered, the combination prediction model which integrates the advantages of three different probability prediction models of kernel density estimation, sparse Bayesian learning and beta distribution estimation is constructed under each meteorological condition, and the weight of each sub-model is optimized by using improved particle swarm algorithm. The Gaussian similarity method is used to determine the meteorological condition type of the time to be predicted, and the corresponding combination probability prediction model is used for prediction, so as to obtain the probability density function and fluctuation interval of the wind power at the time to be predicted. The present application improves the effectiveness of wind power probability prediction results by using the idea of meteorological classification and combination model.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction technology, specifically to a short-term wind power combination probability prediction method, system, terminal equipment, and computer-readable storage medium that takes into account meteorological classification. Background Technology

[0002] As the contradiction between global warming and the energy crisis intensifies, new energy sources are gradually replacing traditional fossil fuels as the cornerstone of societal energy. Wind power, as a crucial component of new energy, has received significant attention from countries worldwide. While power system operation requires a certain degree of stability, the inherent randomness and volatility of wind power create substantial uncertainty when large-scale wind power is integrated into the grid. This increases the difficulty of balancing power generation and consumption, posing a significant challenge to the safe and stable operation of the power system. High-precision wind power forecasting is therefore crucial for helping power system dispatching departments formulate reasonable dispatch plans and improve wind power absorption capacity.

[0003] Current research on wind power forecasting is developing rapidly. Most studies utilize single models to predict wind power under different meteorological conditions, failing to achieve universality across various weather types. Alternatively, they rely on manually determining the number of typical meteorological conditions and using a single model for different conditions, offering limited improvement in accuracy and resulting in significant prediction errors. Furthermore, most studies employ single-value prediction methods that cannot quantify the uncertainty of wind power forecasts, thus being constrained by deterministic prediction. Probabilistic prediction research can provide the fluctuation range and probability density function of future wind power, offering more comprehensive information for wind power forecasting. However, parametric methods in probabilistic prediction require pre-assumed wind power probability density functions, while non-parametric probabilistic prediction models are highly dependent on data quality. Therefore, relying solely on a single probabilistic model, whether parametric or non-parametric, makes it difficult to accurately quantify the uncertainty of wind power forecast results. Summary of the Invention

[0004] This invention provides a short-term wind power combination probability prediction method that takes into account meteorological classification, which overcomes the shortcomings of the above-mentioned prior art. It can effectively solve the problem that it is difficult to accurately and quantitatively describe the wind power prediction results when using a single probability model.

[0005] One of the technical solutions of this invention is achieved through the following measures: a short-term wind power combination probability prediction method considering meteorological classification, which is carried out according to the following steps: S1: Acquire historical wind power data and weather forecast data, and perform data filtering, missing data filling and data normalization on the raw data. S2. Through correlation analysis, the input variables of the wind power prediction model are determined, and meteorological feature vectors are constructed. S3. Meteorological condition classification is achieved by using a meteorological condition classification method based on subtractive clustering and GK fuzzy clustering algorithm. The optimal number of clusters is obtained through a clustering validity function. S4. Construct a wind power probability combination prediction model under various meteorological conditions, and train three sub-probability prediction models under each meteorological condition: kernel density estimation, sparse Bayesian learning, and beta distribution estimation. S5. The meteorological condition identification method based on Gaussian similarity determines the type of meteorological condition at the time to be predicted, and uses the combined probability prediction model corresponding to the meteorological condition to make a prediction, thereby obtaining the probability density function and fluctuation range of wind power at the time to be predicted. S6 utilizes an improved particle swarm optimization algorithm to optimize the parameters and corresponding weights of each sub-model.

[0006] The following are further optimizations and / or improvements to one of the above-mentioned technical solutions: In step S2 above, the Pearson correlation coefficient is used to analyze the correlation between meteorological data and wind power. Four meteorological data with high correlation, namely wind speed s at 10 meters height, wind direction d at 10 meters height, wind speed S at 100 meters height, and wind direction D at 100 meters height, are selected as input variables of the model.

[0007] In step S5 above, a Gaussian function is introduced in the similarity calculation, and a meteorological condition identification method is proposed.

[0008] In step S6 above, an improved particle swarm optimization algorithm is used with the objective function of minimizing the successive ranking probability score to optimize the model parameters of each sub-model and the weights of each sub-model, thereby obtaining the combined probability prediction model corresponding to each meteorological condition type.

[0009] In step S4 above, a combined prediction model based on parametric and non-parametric probability prediction sub-models is established under different meteorological conditions; wherein, the combined prediction sub-models include: kernel density estimator sub-probability prediction model, sparse Bayesian learning sub-probability prediction model, and beta distribution estimator sub-probability prediction model.

[0010] The above-mentioned normalized mean absolute error (NMAE) and normalized root mean square error (NRMSE) are selected as evaluation indicators for the expected value of the prediction, and PICP, PINAW, and CRPS are selected as evaluation indicators for the probability prediction results.

[0011] The second technical solution of the present invention is achieved through the following measures: a short-term wind power combination probability prediction system considering meteorological classification, comprising: The data acquisition module is used to acquire wind farm operation data and weather forecast data; The power prediction module is used to input the data acquired by the data acquisition module into the trained wind power combination probability prediction model to predict wind power. The prediction weight module is used to optimize the weights of each sub-model based on the improved particle swarm optimization algorithm, thereby determining the prediction weights of each sub-prediction model. The data output module is used to fuse the prediction results of the above sub-probability prediction models based on the weights to obtain the final wind power prediction result.

[0012] The third technical solution of the present invention is achieved through the following measures: a terminal device, including a processor and a memory, wherein the processor is used to implement various instructions; the memory is used to store multiple instructions, the instructions being adapted to be loaded by the processor and executed by the processor to perform the short-term wind power combination probability prediction method considering meteorological classification as described in any one of claims 1 to 6.

[0013] The fourth technical solution of the present invention is achieved by the following measures: a computer-readable storage medium storing a plurality of instructions, the instructions being adapted to be loaded and executed by a processor of a terminal device, the short-term wind power combination probability prediction method considering meteorological classification as described in any one of claims 1 to 6.

[0014] The beneficial effects of this invention are: 1. The impact of meteorological conditions on the model's prediction performance was considered, and corresponding wind power probability prediction models were established for different meteorological conditions.

[0015] 2. Based on the averaging principle of Bayesian models, and taking into account the advantages of both parametric and non-parametric probabilistic prediction models, a combined model integrating the advantages of three sub-probabilistic prediction models—kernel density estimation, sparse Bayesian learning, and beta distribution estimation—is constructed under various meteorological conditions. This overcomes the shortcoming that a single probabilistic prediction model is difficult to accurately and quantitatively describe the uncertainty of wind power prediction results.

[0016] 3. An improved particle swarm optimization algorithm is used to optimize the parameters of the three sub-probability prediction models and the weights of each sub-model, overcoming the shortcomings of the traditional particle swarm optimization algorithm that gets trapped in local optima and improving the model prediction accuracy.

[0017] 4. The proposed method can not only obtain the expected value of wind power at the time to be predicted, but also the probability density function and fluctuation range of wind power at the time to be predicted, thus providing more useful information for power system dispatching, etc. Attached Figure Description

[0018] Appendix Figure 1 This is a flowchart of the wind power probability prediction method based on meteorological classification and combination methods in Embodiment 10 of the present invention.

[0019] Appendix Figure 2 This is a time series diagram of wind power in Embodiment 10 of the present invention.

[0020] Appendix Figure 3 This is a correlation coefficient diagram between different meteorological variables and wind power in Embodiment 10 of the present invention.

[0021] Appendix Figure 4 This is a flowchart of the meteorological condition classification method in Embodiment 10 of the present invention.

[0022] Appendix Figure 5 This is a flowchart of the meteorological condition identification method in Embodiment 10 of the present invention.

[0023] Appendix Figure 6 These are the probability density function and cumulative distribution function corresponding to different beta distribution estimation models in Embodiment 10 of the present invention.

[0024] Appendix Figure 7 This is a comparison chart of the NMAE values ​​of each model in Embodiment 10 of the present invention.

[0025] Appendix Figure 8 This is a comparison chart of the NRMSE values ​​of each model in Embodiment 10 of the present invention.

[0026] Appendix Figure 9 The following is a comparison chart of the prediction results of each model in Embodiment 10 of the present invention. Figure a shows the prediction results of each model when considering meteorological classification, Figure b shows the prediction results of each model when not considering meteorological classification, and Figure c shows the prediction results based on the combined model under different methods.

[0027] Appendix Figure 10 This is a reliability comparison chart of the various models in Embodiment 10 of the present invention.

[0028] Appendix Figure 11 This is a comparison chart of the acuity of each model in Embodiment 10 of the present invention.

[0029] Appendix Figure 12 This is a graph showing the probability density function prediction results at each time point in Embodiment 10 of the present invention.

[0030] Appendix Figure 13 This is a graph showing the prediction results of the cumulative distribution function at each time point in Embodiment 10 of the present invention.

[0031] Appendix Figure 14 This is a comparison chart of the probability density function prediction results of each model in Embodiment 10 of the present invention.

[0032] Appendix Figure 15 This is a graph showing the probability prediction results of wind farms three days ahead of schedule in Example 10 of the present invention (spring).

[0033] Appendix Figure 16This is a graph showing the probability prediction results of a wind farm three days ahead of schedule in Embodiment 10 of the present invention (summer).

[0034] Appendix Figure 17 This is a graph showing the probability prediction results of a wind farm three days ahead of schedule in Autumn in Embodiment 10 of the present invention.

[0035] Appendix Figure 18 This is a graph showing the probability prediction results of a wind farm three days ahead of schedule in Example 10 of the present invention (winter). Detailed Implementation

[0036] The present invention is not limited to the following embodiments, and the specific implementation can be determined according to the technical solution of the present invention and the actual situation.

[0037] The present invention will be further described below with reference to embodiments: Example 1: The short-term wind power combination probability prediction method considering meteorological classification is carried out according to the following steps: S1: Acquire historical wind power data and weather forecast data, and perform data filtering, missing data filling and data normalization on the raw data. S2. Through correlation analysis, the input variables of the wind power prediction model are determined, and meteorological feature vectors are constructed. S3. Meteorological condition classification is achieved by using a meteorological condition classification method based on subtractive clustering and GK fuzzy clustering algorithm. The optimal number of clusters is obtained through a clustering validity function. S4. Construct a wind power probability combination prediction model under various meteorological conditions, and train three sub-probability prediction models under each meteorological condition: kernel density estimation, sparse Bayesian learning, and beta distribution estimation. S5. The meteorological condition identification method based on Gaussian similarity determines the type of meteorological condition at the time to be predicted, and uses the combined probability prediction model corresponding to the meteorological condition to make a prediction, thereby obtaining the probability density function and fluctuation range of wind power at the time to be predicted. S6 utilizes an improved particle swarm optimization algorithm to optimize the parameters and corresponding weights of each sub-model.

[0038] Example 2: As an optimization of the above example, in step S2, the Pearson correlation coefficient is used to analyze the correlation between meteorological data and wind power. The four meteorological data with high correlation, namely wind speed s at 10 meters height, wind direction d at 10 meters height, wind speed S at 100 meters height, and wind direction D at 100 meters height, are selected as the input variables of the model.

[0039] Example 3: As an optimization of the above example, in step S5, a Gaussian function is introduced in the similarity calculation, and a meteorological condition identification method is proposed.

[0040] Example 4: As an optimization of the above example, in step S6, an improved particle swarm optimization algorithm is used with the objective function of minimizing the successive ranking probability score to optimize the model parameters of each sub-model and the weights of each sub-model, so as to obtain the combined probability prediction model corresponding to each meteorological condition type.

[0041] Example 5: As an optimization of the above example, in step S4, a combined prediction model based on parametric and non-parametric probability prediction sub-models is established under different meteorological conditions; wherein, the combined prediction sub-model includes: kernel density estimator sub-probability prediction model, sparse Bayesian learning sub-probability prediction model, and beta distribution estimator sub-probability prediction model.

[0042] Example 6: As an optimization of the above example, the Normalized Mean Absolute Error (NMAE) and the Normalized Root Mean Square Error (NRMSE) are selected as two indicators for evaluating the expected value of the prediction, and PICP, PINAW, and CRPS are selected as three indicators for evaluating the probability prediction results.

[0043] This invention proposes a short-term wind power combination probability prediction method that considers meteorological classification. It classifies meteorological conditions based on subtractive clustering and Gustafson-Kessel (GK) fuzzy clustering algorithm and identifies meteorological conditions using Gaussian similarity. Under each meteorological condition, a combined model is constructed that integrates the advantages of three sub-probability prediction models: kernel density estimation, sparse Bayesian learning, and beta distribution estimation, thereby realizing wind power probability prediction.

[0044] Example 7: The short-term wind power combination probability prediction system considering meteorological classification includes: The data acquisition module is used to acquire wind farm operation data and weather forecast data; The power prediction module is used to input the data acquired by the data acquisition module into the trained wind power combination probability prediction model to predict wind power. The prediction weight module is used to optimize the weights of each sub-model based on the improved particle swarm optimization algorithm, thereby determining the prediction weights of each sub-prediction model. The data output module is used to fuse the prediction results of the above sub-probability prediction models based on the weights to obtain the final wind power prediction result.

[0045] Example 8: The terminal device includes a processor and a memory. The processor is used to implement various instructions; the memory is used to store multiple instructions, which are adapted to be loaded by the processor and executed by the processor to perform the short-term wind power combination probability prediction method considering meteorological classification as described in any one of claims 1 to 6.

[0046] Example 9: The computer-readable storage medium stores a plurality of instructions adapted for loading and execution by a processor of a terminal device of the short-term wind power combination probability prediction method considering meteorological classification as described in any one of claims 1 to 6.

[0047] Example 10: This short-term wind power combination probability prediction method considering meteorological classification, such as... Figure 1 As shown, proceed with the following steps: Historical wind power data and weather forecast data are acquired, and the raw data is processed by data filtering, missing data filling, and data normalization.

[0048] The Pearson correlation coefficient was used to analyze the correlation between seven meteorological quantities—wind speed s at 10 meters height, wind direction d at 10 meters height, wind speed S at 100 meters height, wind direction D at 100 meters height, temperature T, air pressure A, and humidity H—and wind power P. Meteorological quantities with high correlation were selected as input variables for the model. A meteorological feature vector of historical sample set was constructed.

[0049] The concept of clustering is used to classify meteorological conditions. A clustering validity function based on probability distribution is used to determine the rationality of the number of clusters, and subtractive clustering and GK fuzzy clustering algorithms are used to classify meteorological conditions.

[0050] Three sub-probability prediction models—kernel density estimation, sparse Bayesian learning, and beta distribution estimation—were trained under the defined meteorological categories. An improved particle swarm optimization algorithm was then used to optimize the model parameters and weights of each sub-model with the objective function of minimizing the Continuous Ranked Probability Score (CRPS), resulting in a combined probability prediction model for each meteorological condition type.

[0051] The Gaussian similarity method is used to determine the type of meteorological conditions at the time to be predicted, and the combined probability prediction model corresponding to the meteorological conditions is used to make a prediction, so as to obtain the probability density function and fluctuation range of the final wind power.

[0052] This invention establishes a combined probabilistic prediction model for wind power under different weather types. It integrates the advantages of single probabilistic prediction models while mitigating their larger errors, thus improving the accuracy and effectiveness of wind power prediction. By combining subtractive clustering and GK fuzzy clustering algorithms, weather types are categorized, and a clustering validity function is used to determine the rationality of the number of meteorological condition clusters, avoiding the clusters from getting trapped in local optima. First, a Gaussian function is used to determine the meteorological condition type of the time to be predicted. Based on the Bayesian model averaging principle, a combined prediction model is constructed for each meteorological condition type, incorporating three sub-probabilistic models: kernel density estimation, sparse Bayesian learning, and beta distribution estimation. An improved particle swarm optimization algorithm based on a linear decreasing weight strategy is used to optimize the weights of the three sub-models. Wind power probabilistic prediction is achieved using the meteorological classification and combined model approach.

[0053] Specifically, the detailed implementation process of this embodiment is as follows: The accuracy of wind power forecasting is closely related to the quality of NWP data and historical power observation data. This example improves data quality by performing data filtering, missing data imputation, and data normalization on the raw data. A five-point cubic smoothing filter is used to filter the wind power observation data, thereby reducing noise interference with the raw data; Lagrange interpolation is used to interpolate missing data, thus improving data integrity.

[0054] Figure 2 The graph shows the fluctuation of wind power output at a wind farm over a week, revealing strong randomness and volatility. Wind power output is closely related to meteorological parameters such as wind speed. To investigate the influence of various factors on wind power output, the input variables for the model are determined. The Pearson correlation coefficient is selected. This study analyzes the correlation between seven meteorological quantities—wind speed s at 10 meters height, wind direction d at 10 meters height, wind speed S at 100 meters height, wind direction D at 100 meters height, temperature T, air pressure A, and humidity H—and wind power P. Meteorological quantities with high correlation are selected as input variables for the model. The calculation method is as follows: in, The correlation coefficient between time series is represented by the coefficient of correlation between time series. The larger the value, the stronger the correlation. and They represent and The average of two time series, where n is the number of samples in the time series.

[0055] Figure 3The correlation coefficient diagram between seven meteorological quantities (s, d, S, D, T, A, H) and wind power P is shown. Through correlation analysis, four meteorological quantities at the target time—wind speed s at 10 meters, wind direction d at 10 meters, wind speed S at 100 meters, and wind direction D at 100 meters—were selected as the input variables for the model.

[0056] This example uses clustering to classify meteorological conditions, proposing a meteorological condition classification method based on subtractive clustering and GK fuzzy clustering algorithms. The flowchart of this method is shown below. Figure 4 As shown.

[0057] Based on historical datasets, meteorological parameters such as wind speed at various times are selected to construct a meteorological feature vector X, let... Perform subtractive clustering on X to obtain the upper bound of the clustering. .Pick Initialize the fuzzy partitioning matrix and solve for the ideal partitioning matrix using the GK fuzzy clustering algorithm. The solution yields... Corresponding clustering validity function If the number of clusters If the fuzzy partition matrix meets the validity evaluation criteria, then the optimal validity clustering has been obtained. This represents the optimal number of meteorological condition types. If the matrix... In the j-th column If the maximum value is found, then sample j belongs to the i-th meteorological condition type. This process continues to obtain the sample set corresponding to each typical meteorological condition. The cluster center for the i-th typical meteorological condition is... It can be obtained according to formula (2). In addition, the cluster centers of all meteorological conditions are... As shown in formula (3).

[0058] in, Let m be the number of samples contained in the i-th meteorological condition, and m be the number of features of each sample.

[0059] in, denoted as the number of typical meteorological condition types, and m as the number of features for each sample.

[0060] Historical sample sets can be used to cluster meteorological conditions into several typical types. However, prediction requires determining the meteorological condition type at the time of prediction, and then using the power prediction model corresponding to that meteorological condition. This invention, based on traditional Euclidean distance, introduces a Gaussian function in similarity calculation, proposing a meteorological condition identification method based on Gaussian similarity. The flowchart of this method is shown below. Figure 5 As shown.

[0061] Suppose we select m meteorological variables to construct a meteorological feature vector X. After clustering using a meteorological condition classification method based on subtractive clustering and GK fuzzy clustering algorithms, we obtain... Given a typical meteorological condition type, the meteorological feature vector for the predicted time is first constructed based on m meteorological variables at the predicted time.

[0062] Assume the weights of each meteorological variable are as follows: ,but ,and The Euclidean distance between the j-th meteorological feature at the time to be predicted b and the j-th meteorological feature under typical meteorological condition a. As shown in formula (4).

[0063] in, Let j be the j-th meteorological feature value of the cluster center for typical meteorological condition a. Let be the j-th meteorological characteristic value at the time b to be predicted.

[0064] A nonlinear Gaussian transform method is used to convert the Euclidean distance between the predicted time b and the j-th feature of typical meteorological condition a into a Gaussian index. The Gaussian index is then used to represent the similarity between the predicted time b and each feature of each typical meteorological condition. The Gaussian transform function of the j-th meteorological feature is then used. As shown in formula (5).

[0065] in, Let be the Euclidean distance between the j-th meteorological feature at the time to be predicted (b) and the j-th meteorological feature under typical meteorological condition (a). It is the flexure point. It is a constant and satisfies This invention selects the optimal empirical value. , It is the maximum value of the j-th meteorological feature. It is the minimum value of the j-th meteorological feature.

[0066] Finally, the Gaussian similarity between the time to be predicted (b) and the typical meteorological condition (a) is obtained. , The larger the value, the higher the similarity between the predicted time and the meteorological conditions. Finally, the meteorological condition with the highest similarity is determined as the meteorological condition type to which the predicted time belongs. The expression is shown in formula (7).

[0067] in, Let be the Gaussian transformation function between the j-th meteorological feature at time b to be predicted and the j-th meteorological feature under meteorological condition a. To determine the weights corresponding to the j-th meteorological feature, this invention employs the equal weight method to determine the weights corresponding to each meteorological feature. .

[0068] Then, a wind power prediction sub-model is constructed. Using a single probability model for prediction makes it difficult to accurately quantify the uncertainty of wind power forecast results. However, combined prediction models, by combining multiple individual models, can obtain combined prediction results that integrate the advantages of each individual model, thereby improving prediction accuracy. Therefore, this invention, based on the Bayesian model averaging method, comprehensively considers the advantages of parametric and non-parametric probability prediction models. Under various meteorological conditions, it constructs combined prediction models that integrate the advantages of three sub-probability models: kernel density estimation, sparse Bayesian learning, and beta distribution estimation. An improved particle swarm optimization algorithm is then used to optimize the parameters and corresponding weights of each sub-model.

[0069] The Bayesian model averaging principle involves weighted averaging of the probability density functions of multiple individual models and using this average as the prediction result of the probability density function of the target variable. The weights of the sub-models are their posterior probabilities. Essentially, Bayesian model averaging is a combined prediction method, where the posterior probabilities of each individual model are used as weights to perform a weighted average of the predicted values ​​of all models, thereby obtaining the predicted value of the target variable. The probability density function prediction result of the target variable y obtained based on the Bayesian model averaging method is shown in formula (8).

[0070] in, Let be the predicted value of the probability density function of the target variable at time t, and D be the given dataset. Let K be the j-th sub-model, and K be the number of sub-models. For the model The obtained probability density function prediction value, For the model The posterior probability, For the model The prior distribution, For the model The likelihood function, For the model The corresponding prior probability distribution of the parameters.

[0071] Methods such as Expectation Maximization (EM) are used to estimate the weights of individual models in the Bayesian model average. While the EM method is simple and time-saving, it assumes the target variable follows a normal distribution, and when the combined model contains sub-models following different distributions, the sub-model weights calculated using the EM method are prone to getting trapped in local optima. Therefore, this invention is based on the idea of ​​a single-valued combined prediction model, which uses sub-models... posterior probability Transform into An improved particle swarm optimization algorithm is used to optimize the weights of each sub-model. satisfy ,and Therefore, when using the combined model for prediction, the probability density function prediction result of the target variable y is shown in formula (11).

[0072] in, This represents the prediction result of the probability density function of the target variable at time t. For sub-model The weight, Indicates the use of sub-models The probability density function prediction results are obtained.

[0073] This embodiment uses different types of data to train three sub-prediction models.

[0074] 1) Principle of kernel density estimation model Kernel density estimation, as a non-parametric probabilistic prediction model, does not require pre-assuming the distribution of the target variable. Instead, it studies the distribution of the data sample entirely based on the characteristics of the data sample itself. Because wind power has strong randomness and volatility, it is difficult to obtain a definite distribution of the wind power probability density function. Therefore, kernel density estimation can yield a more realistic probability prediction result. When estimating the probability density value of a point, the kernel density estimation model first calculates a density value based on the distance between the point to be estimated and each sample point. If the distance between the point to be estimated and the sample points is relatively short, the density value obtained is relatively large, and vice versa. Finally, a weighted average of all density values ​​is taken to obtain the probability density value of the point to be estimated in the sample distribution.

[0075] The probability density function obtained based on the kernel density estimation model is closely related to the data itself.

[0076] If the input variable is Then, the objective function obtained using the kernel density estimation model The probability density function is shown in Equation (12).

[0077] Where n is the number of samples. , and These are the bandwidth parameters that control the smoothness of the input and output variables, respectively. This is the kernel function.

[0078] The predictive performance of the kernel density estimation model is closely related to three factors: explanatory variables, kernel function, and bandwidth parameter. To improve the prediction accuracy of the kernel density estimation model, variables with strong correlation to the target variable should be used as explanatory variables. Based on the correlation analysis results, this invention selects four meteorological quantities (s, S, d, and D) corresponding to each time point as key explanatory variables. This invention uses the commonly used Gaussian kernel function as the kernel function of the kernel density estimation model. Furthermore, the choice of bandwidth also has a significant impact on the probability density function prediction results. If the bandwidth is too small, the probability density distribution of the target variable will be limited to the vicinity of the power observation value, resulting in many peaks in the probability density function prediction curve. If the bandwidth value is larger, the resulting probability density function prediction curve will be smoother, which may cause the prediction results to miss some important features. This invention chooses an improved particle swarm optimization algorithm to optimize the bandwidth parameter of the kernel density estimation model.

[0079] 2) Principles of Sparse Bayesian Learning Models Sparse Bayesian learning models possess good generalization ability and sparsity characteristics, along with probabilistic prediction capabilities. This model utilizes kernel functions to extend traditional linear feature extraction methods to the nonlinear domain. When solving regression problems, the relationship between the target variable y and the explanatory variable x can be expressed as: in, and Let these be the explanatory variables and the target variable for the nth data set, respectively. For the weight vector, the error term Follows a mean of 0 and a variance of The target variable follows a normal distribution, therefore its mean is . variance is It follows a normal distribution.

[0080] Based on the kernel method concept, in formula (13) It can be represented as: in, The kernel function is represented by the commonly used Gaussian kernel function in this invention. This represents the number of kernel functions.

[0081] Because the weight variable w is linear and the kernel function Because of its highly nonlinear nature, sparse Bayesian learning models have the ability to solve nonlinear problems. In particular, when multiple kernel functions are included, sparse Bayesian learning models have greater flexibility, which is also an advantage of using kernel methods to construct nonlinear prediction models.

[0082] The probability of the target variable y appearing in N tests can be expressed as: in, Represented as , for .

[0083] If we directly use maximum likelihood estimation... as well as When solving the problem, overlearning may occur. To ensure the sparsity and generalization ability of the model, the sparse Bayesian learning model will... as well as Treat it as a random variable, and assume It follows a mean of 0 and a variance of . It follows a normal distribution. Furthermore, assume the hyperparameters... as well as It follows a gamma distribution with both shape and scale parameters of 0. Therefore, based on Bayes' theorem, we can obtain: Because it is difficult to find the analytical solution on the right side of formula (4-9), the left side of formula (4-9) is decomposed, and the decomposition result is shown in formula (4-10).

[0084] Therefore, the original problem is transformed into solving w. The posterior distribution in the given case and the solution, and the posterior distribution in the given case. To simplify the calculation, it is usually assumed that... The posterior distribution is the Dirac distribution. ,in They are respectively as well as The maximum likelihood estimate is obtained. Therefore, based on Bayesian inference, the mean and variance of the weight vector can be obtained as shown in formulas (18) and (19), respectively.

[0085] in, .

[0086] For new explanatory variables Target variable Still follows the mean The variance is Since the target variable y follows a normal distribution, the probability density function obtained when using a sparse Bayesian learning model to predict the target variable y is... It can be represented as: 3) Principle of Beta Distribution Estimation Model The beta distribution estimation model is a parametric probability prediction model, and the probability density function obtained based on this model can be expressed as: in, and All are shape parameters greater than 0. As a normalization constant, it guarantees that the sum of all probabilities is 1, and .

[0087] Assuming that the target variable is obtained from historical data statistics during model training. The mean and standard deviation are respectively and ,So and The expressions are shown in formulas (22) and (23) respectively.

[0088] This invention sets parameters and The values ​​of are 0 and 1 respectively. At this time, the model is a standard beta distribution model, and the probability density function is shown in formula (4-17).

[0089] when and The beta distribution estimation model follows different distribution forms when different values ​​are taken. Figure 6 It shows the probability density function and cumulative distribution function corresponding to four mainstream distribution forms: uniform distribution, trapezoidal distribution, Rayleigh distribution, and Gaussian distribution.

[0090] If shape parameters and If both sums are 1, then the distribution is uniform; if both sums are 2, then the distribution is trapezoidal; if sums are 2 and 3.4 respectively, then the distribution is Rayleigh; if both sums are 4, then the distribution follows a Gaussian pattern. Furthermore, from... Figure 6It can be seen that the probability density function and cumulative distribution function corresponding to different distributions are very different. However, the beta distribution estimation model, as a parametric regression statistical model, can fit the probability distribution of random variables well. By adjusting the parameters of the model, the target variable can be made to approximate the uniform distribution, Gaussian distribution, and other forms.

[0091] This embodiment evaluates the prediction performance of each model from two aspects: the predicted expected value and the probabilistic prediction results, thus verifying the effectiveness of the method proposed in this invention. Specifically, the Normalized Mean Absolute Error (NMAE) and the Normalized Root Mean Square Error (NRMSE) are selected as evaluation indicators for the predicted expected value, while PICP, PINAW, and CRPS are selected as evaluation indicators for the probabilistic prediction results.

[0092] The dataset contains n sample points, and the confidence level is... The predicted interval obtained at that time is The Prediction Interval Coverage Probability (PICP) metric reflects the reliability of the prediction results. The PICP calculation method is as follows: Wherein, if the observed value of sample i Falling within the prediction range Inside, then the characteristic function The value is 1 otherwise. The smaller the reliability deviation (Error, Er) E between PICP and the confidence level, the higher the reliability of the probability prediction result. Er can be expressed as: The Prediction Interval Normalized Average Width (PINAW) metric reflects the sensitivity of the prediction results. PINAW can be expressed as: Where n is the number of test samples, and R is the parameter for interval width normalization, typically the difference between the maximum and minimum values ​​of the target variable. and These represent the lower and upper bounds of the prediction interval, respectively. A smaller PINAW value indicates higher model sensitivity.

[0093] The prediction results are as follows: This example uses data from eight wind farms in Ningxia as a case study. The dataset for each wind farm includes NWP and wind power observations from 2017 to 2019, with a time resolution of 15 minutes. The dataset is divided into training and testing sets.

[0094] This example selects four models for prediction: KDE, SBL, BDE, and MMC. The symbolic variables of each model are shown in Table 1. The model used in this invention is MMCy.

[0095] Figure 7 and Figure 8 The figure shows the average NMAE and NRMSE values ​​of the prediction results for each model when predicting eight wind farms. As can be seen from the figure, the NMAE and NRMSE values ​​of the model in this example are significantly lower than those of the other models, demonstrating that it effectively improves the prediction accuracy.

[0096] Figure 9 The presentation compares the expected values ​​of wind power forecasts obtained by using various models to predict wind power output three days in advance for wind farms. Figure 9 (a) and Figure 9 (b) Presents the prediction results obtained by using four models, MMC, SBL, KDE and BDE, in two cases: considering meteorological classification and not considering meteorological classification. Figure 9 (c) Presents the prediction results obtained by using the combined model with and without considering meteorological classification. The results show that the wind power prediction obtained by using the model proposed in this example is closer to the actual power, which verifies that the model proposed in this example has good predictive performance.

[0097] Figure 10 and Figure 11 The Er and PINAW values ​​at different confidence levels are presented when using each model to predict eight wind farms. It can be found that the absolute Er value of the proposed model at different confidence levels is smaller than that of other models, and the PINAW value is lower at all confidence levels, indicating that the model proposed in this invention has high reliability and sensitivity.

[0098] Table 2 shows the CRPS values ​​of the prediction results for each of the eight stations, as well as the average CRPS value for the eight stations. Based on the data in the table, it can be found that the CRPS values ​​of the prediction results for each station using the model proposed in this invention are lower than those of other models, effectively verifying the effectiveness of the model proposed in this invention in improving prediction accuracy.

[0099] Figure 12 and Figure 13The figures show the wind power probability density function and cumulative distribution function prediction results obtained by the model of this invention when predicting wind power at 12 specific times from 1:00 to 12:00 on June 1, 2019. The blue curves represent the wind power probability density function and cumulative distribution function prediction results, respectively. The red dots represent observed wind power values, and the two red lines represent the upper and lower limits of the 80% confidence interval, respectively. All observed wind power values ​​at each time point fall within the 80% confidence interval. Figure 14 The results of probability density function predictions obtained by using various models to predict two different times, 00:00 on June 1, 2019 and 00:00 on December 1, 2019, are presented. This intuitively demonstrates the good reliability and accuracy of the model of this invention.

[0100] Figures 15 to 18 The paper presents the probabilistic prediction results obtained by using the proposed method to predict wind power output three days in advance for Wind Farm No. 6 in four different seasons: spring, summer, autumn, and winter. The probabilistic prediction intervals obtained using the model of this invention can well cover the observed wind power values, with most wind power values ​​falling within the 80% confidence interval and almost all wind power values ​​falling within the 90% confidence interval. This verifies that the proposed method can obtain effective probabilistic prediction results for wind power output.

[0101] In summary, this invention provides a short-term wind power combination probability prediction method that considers meteorological classification. It classifies meteorological conditions based on subtractive clustering and the Gustafson-Kessel (GK) fuzzy clustering algorithm and identifies meteorological conditions using Gaussian similarity. Under each meteorological condition, a combined model is constructed that integrates the advantages of three sub-probabilistic prediction models: kernel density estimation, sparse Bayesian learning, and beta distribution estimation. This overcomes the shortcomings of single probabilistic prediction models, which struggle to accurately quantify the uncertainty of wind power prediction results, thus improving the model's prediction accuracy and achieving probabilistic wind power prediction. This provides more useful information for power system dispatching and other applications.

[0102] The above technical features constitute the embodiments of the present invention, which have strong adaptability and implementation effect. Unnecessary technical features can be added or removed according to actual needs to meet the needs of different situations.

Claims

1. A short-term wind power combination probability prediction method considering meteorological classification, characterized in that... Follow these steps: S1: Acquire historical wind power data and weather forecast data, and perform data filtering, missing data filling and data normalization on the raw data; S2. Through correlation analysis, the input variables of the wind power prediction model are determined, and meteorological feature vectors are constructed. S3. Meteorological condition classification is achieved by using a meteorological condition classification method based on subtractive clustering and GK fuzzy clustering algorithm. The optimal number of clusters is obtained through a clustering validity function. S4. Construct a wind power probability combination prediction model under various meteorological conditions, and train three sub-probability prediction models under each meteorological condition: kernel density estimation, sparse Bayesian learning, and beta distribution estimation. S5. The meteorological condition identification method based on Gaussian similarity determines the type of meteorological condition at the time to be predicted, and uses the combined probability prediction model corresponding to the meteorological condition to make a prediction, thereby obtaining the probability density function and fluctuation range of wind power at the time to be predicted. S6 employs an improved particle swarm optimization algorithm with the objective function of minimizing the successive ranking probability score. This optimizes the model parameters of each sub-probability prediction model and the weights of each sub-model, resulting in a combined probability prediction model corresponding to each meteorological condition type.

2. The short-term wind power combination probability prediction method considering meteorological classification according to claim 1, characterized in that... In step S2, the Pearson correlation coefficient is used to analyze the correlation between meteorological data and wind power. Four meteorological data with high correlation, namely wind speed s at 10 meters height, wind direction d at 10 meters height, wind speed S at 100 meters height, and wind direction D at 100 meters height, are selected as input variables for the model.

3. The short-term wind power combination probability prediction method considering meteorological classification according to claim 1 or 2, characterized in that... In step S5, a Gaussian function is introduced in the similarity calculation, and a meteorological condition identification method is proposed.

4. The short-term wind power combination probability prediction method considering meteorological classification according to claim 1 or 2, characterized in that... In step S4, parametric and non-parametric sub-probability prediction models are established under different meteorological conditions to form a probability prediction combination model. The sub-probability prediction models in the probability prediction combination model include: kernel density estimation sub-probability prediction model, sparse Bayesian learning sub-probability prediction model, and beta distribution estimation sub-probability prediction model.

5. The short-term wind power combination probability prediction method considering meteorological classification according to claim 4, characterized in that... The Normalized Mean Absolute Error (NMAE) and Normalized Root Mean Square Error (NRMSE) are selected as evaluation metrics for the expected value of the prediction. The PICP, PINAW, and CRPS are selected as evaluation metrics for the probability prediction results. PICP is the prediction interval coverage, PINAW is the prediction interval bandwidth, and CRPS is the continuous ranking probability score.

6. A short-term wind power combination probability prediction system considering meteorological classification, wherein the short-term wind power combination probability prediction system considering meteorological classification uses the method as described in claim 1, characterized in that... include: The data acquisition module is used to acquire wind farm operation data and weather forecast data; The power prediction module is used to input the data acquired by the data acquisition module into the trained wind power combination probability prediction model to predict wind power. The prediction weight module is used to optimize the weights of each sub-probability prediction model based on the improved particle swarm optimization algorithm, thereby determining the prediction weight of each sub-probability prediction model. The data output module is used to fuse the prediction results of the above sub-probability prediction models based on the weights to obtain the final wind power prediction result.

7. A terminal device, characterized in that... It includes a processor and a memory, the processor being used to implement various instructions; the memory being used to store multiple instructions, the instructions being adapted to be loaded by the processor and executed by the processor as described in any one of claims 1 to 5, the short-term wind power combination probability prediction method considering meteorological classification.

8. A computer-readable storage medium, characterized in that... It stores multiple instructions, which are adapted to be loaded and executed by the processor of the terminal device, according to any one of claims 1 to 5, the short-term wind power combination probability prediction method considering meteorological classification.

Images