Short-term wind power prediction method based on clustering optimization analysis

Abstract

The invention relates to a short-term wind power prediction method based on clustering optimization analysis, and the method comprises the steps: obtaining a prediction data set and a field group historical data set of an original wind power plant (one-field one-measurement), and carrying out the data preprocessing, and obtaining two types of data sets with higher credibility; obtaining an optimal clustering wind field division region by using the processed field group historical data set and combining with an AP clustering algorithm after PSO particle swarm optimization; according to the division result of the optimal clustering wind field, the prediction data sets of the sub-region wind fields are loaded into a BP neural network model, meanwhile, the correlation coefficients of the wind power plant output and the wind power plant group output are utilized, values are assigned to the corresponding wind power plants to serve as weights, and finally wind power plant refined group short-term wind power prediction is obtained. Compared with a traditional superposition method and a single BP neural network, the method has the advantage that the prediction precision is improved in 24-hour day-ahead prediction. It can be indicated that the model has excellent prediction performance and has certain reference significance in future development of wind power cluster power prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of short-term wind power prediction, and in particular to a short-term wind power prediction method based on cluster optimization analysis. Background Art

[0002] In recent years, my country's wind power forecasting technology has developed rapidly, wind power generation capacity has increased year by year, and wind farms in many regions have developed in the direction of clustering and scale. However, most of the research focuses on the accurate power forecasting of a single wind turbine and a single wind farm, and there is relatively little research on the overall power forecasting of wind farm clusters. Therefore, it is necessary to further carry out in-depth research on wind farm cluster forecasting.

[0003] Accurate cluster prediction can reduce the impact of wind power's spatiotemporal characteristics on grid operation, which is beneficial to the bidding for access to the grid on the power supply side. However, different cluster prediction methods can mine the original data set to different degrees, and their economy and accuracy will also be reflected differently.

[0004] At present, wind power prediction methods are mainly divided into physical methods that calculate wind farm output with physical parameters, statistical methods based on a large amount of historical data training, intelligent methods and composite methods, based on the relationship between numerical weather forecast (NWP) data and wind power output. Prediction methods with strong nonlinear mapping, generalization and adaptability, such as support vector machines and back propagation neural networks, have been widely used in wind power prediction. Summary of the invention

[0005] In view of the problems existing in the background technology, the present invention provides a method for classifying wind farm groups by adopting the PSO-AP clustering optimization algorithm, weighting the wind farms according to the correlation coefficient between the output of a single wind farm and the wind farm group, then performing BP prediction on each sub-area respectively, and finally adding the results of each component to obtain the final result.

[0006] In order to solve the above technical problems, the present invention adopts the following technical solution: a PSO-AP-BP composite model based on correlation coefficient weights, comprising the following steps:

[0007] S1: Use data processing methods to handle outliers, clean missing values, and generate wind farm groups with high confidence and hourly power data series for each wind farm.

[0008] S2: Take the historical output data sets of each wind farm as the clustering object, and apply the AP clustering algorithm combined with PSO particle swarm optimization to obtain the best clustering result.

[0009] S3: Obtain wind farm data for training, divide it into a training set and a test set, and complete data normalization as data preparation for BP neural network prediction.

[0010] S4: Input the data in S3 into the PSO-AP-BP composite model based on the correlation coefficient weight, and use the clustering results to apply the BP neural network to the clustering areas respectively to obtain the sub-area prediction value.

[0011] S5: Complete the prediction of each cluster area and the denormalization of the actual output data.

[0012] S6: Select the optimal wind farm group power prediction model using corresponding evaluation indicators to obtain the final wind farm group wind power output.

[0013] In the above-mentioned PSO-AP-BP composite model prediction method based on correlation coefficient weights, as described in S1, in the actual site-level wind farm output data, the negative effects of abnormal output values ​​of a single or multiple wind turbines are relatively small, and the possible abnormal values ​​can be approximately treated as normal data.

[0014] In the above-mentioned PSO-AP-BP composite model prediction method based on correlation coefficient weight, the power data sequence described in S1 is a predicted power data sequence and an actual historical output data sequence.

[0015] In the above-mentioned PSO-AP-BP composite model prediction method based on correlation coefficient weight, in the PSO-AP optimization algorithm model described in S2, the PSO particle swarm optimization algorithm is adopted to optimize P and λ at a dynamic scale to obtain the optimal dual parameter combination, and a particle swarm with a population size of N and a search space and search capability is constructed by initializing the AP clustering dual parameters. The position and speed attributes are assigned using the silhouette coefficient as the fitness objective function to realize the update iteration of the particle swarm, and realize the information sharing and collaboration within the particle swarm; the PSO algorithm expression is:

[0016] V i =V i +c1×rand()×(pbest i -X i )+c2×rand()×(gbest i -X i )

[0017] Where: V is the particle update speed; c1, c2 are learning factors; X i ——Current location; pbest i ——The best historical parameters of the particle; gbest i ——The historical optimal parameters of the population.

[0018] Xi =X i +V i

[0019] Where: x i ——Current location.

[0020] In the above PSO-AP-BP composite model prediction method based on correlation coefficient weight, the wind farm data used for training in S3 is the prediction data obtained by the wind farm prediction system as the input of the BP neural network, and the actual total output of the wind farm group is used as the output.

[0021] In the above-mentioned PSO-AP-BP composite model prediction method based on correlation coefficient weight, the PSO-AP-BP composite model based on correlation coefficient weight described in S4 adopts the correlation coefficient between the wind farm output and the wind farm group output, and assigns values ​​to the corresponding wind farms as weights respectively. The correlation coefficient mechanism is as follows:

[0022]

[0023] Where: ρ is the Pearson correlation coefficient, ρ>0, indicating that the two vectors are positively correlated. The larger the absolute value, the higher the degree of correlation; x i --actual power sequence of wind farm; Y i ——The actual total output power of the field group.

[0024] In the above PSO-AP-BP composite model prediction method based on correlation coefficient weight, S6 uses root mean square error (RMSE) and mean absolute percentage error (MAPE) to evaluate the prediction performance of the PSO-AP-BP composite model based on correlation coefficient weight.

[0025] In the above PSO-AP-BP composite model prediction method based on correlation coefficient weight, S6 uses evaluation indicators such as prediction accuracy (PA) and correlation (CC) to quantify the wind power prediction model to more clearly compare the prediction performance between the models. The calculation formula is as follows:

[0026] PA=1-MAPE

[0027]

[0028] Where: len is the number of predicted values ​​of the wind farm group; P, p are the predicted value sequence and the actual value sequence, p k 、p k ——Power prediction and actual value.

[0029] In the above-mentioned PSO-AP-BP composite model prediction method based on correlation coefficient weight, the final output of the wind farm group described in S6 is the superposition of the prediction values ​​obtained by the BP neural network based on the correlation coefficient for each clustering area.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] (1) The AP clustering algorithm is improved by using the PSO particle swarm algorithm, which improves the situation where the algorithm may not be able to find the optimal solution, effectively increases the global search capability, and obtains the optimal combination solution of the bias parameter and the damping coefficient. In the total power prediction of the wind farm group, the adaptive optimization AP clustering method is more effective than the traditional AP clustering method. Using the silhouette coefficient as the evaluation index of clustering quality, the parameter adaptive AP clustering method based on the PSO algorithm proposed in this paper can improve the coefficient compared with the static parameter AP clustering method.

[0032] (2) The correlation coefficient between the fluctuation of wind power output of a single wind farm and the fluctuation of total output of a cluster wind farm is used as a weighted optimization method. Compared with the corresponding single BP prediction and PSO-AP-BP prediction, the correlation coefficient composite model proposed in this paper can achieve higher short-term wind power forecasting accuracy. The correlation coefficient weighting can improve the short-term wind power forecasting of 24 hours ahead by 0.33% and 0.36% respectively. Compared with the currently more commonly used superposition method, the method proposed in this paper can improve the 24-hour day-ahead forecast accuracy by 1.35%.

[0033] (3) Compared with the point prediction results of a single wind farm, the power prediction of a wind farm cluster is more practical in the daily scheduling process. It can reasonably plan the unit switching plan on the power supply side and the allocation on the load side, so as to achieve a more reasonable supply and demand balance on the power supply side and the load side. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A flow chart of wind power prediction model design according to an embodiment of the present invention

[0035] Figure 2 The number of clusters and the silhouette coefficient diagram for a certain month in one embodiment of the present invention are shown in FIG.

[0036] Figure 3 A comparison diagram of BP prediction results and true values ​​according to an embodiment of the present invention

[0037] Figure 4 A graph showing the correlation coefficient of power between a single wind farm and a wind farm group according to an embodiment of the present invention

[0038] Figure 5 The prediction curve trends of different prediction models in one embodiment of the present invention are shown as follows:

[0039] Figure 6 This is a bar chart of the superposition method and PSO-AP-BP prediction in one embodiment of the present invention DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present invention will be described clearly and completely below in combination with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0042] The present invention will be further described below in conjunction with specific embodiments, but the present invention is not limited thereto.

[0043] Figure 1 The wind power prediction model design process diagram of the present invention is as follows: Figure 1 As shown, the wind power of the present invention

[0044] The prediction method includes the following steps:

[0045] S1: Use data processing methods to handle outliers, clean missing values, and generate wind farm groups with high confidence and hourly power data series for each wind farm.

[0046] S2: Take the historical output data sets of each wind farm as the clustering object, and apply the AP clustering algorithm combined with PSO particle swarm optimization to obtain the best clustering result.

[0047] S3: Obtain wind farm data for training, divide it into a training set and a test set, and complete data normalization as data preparation for BP neural network prediction.

[0048] S4: Input the data in S3 into the PSO-AP-BP composite model based on the correlation coefficient weight, and use the clustering results to apply the BP neural network to the clustering areas respectively to obtain the sub-area prediction value.

[0049] S5: Complete the prediction of each cluster area and the denormalization of the actual output data.

[0050] S6: Select the optimal wind farm group power prediction model using corresponding evaluation indicators to obtain the final wind farm group wind power output.

[0051] The detailed description of the specific steps is as follows:

[0052] The power data sequence in step 1 is the predicted power data sequence and the actual historical output data sequence of 13 wind farms. In the actual site-level wind farm output data, the negative effects of abnormal output values ​​of a single or multiple wind turbines are relatively small, and the possible abnormal values ​​can be approximately treated as normal data.

[0053] In the PSO-AP optimization algorithm model described in step 2, the AP clustering algorithm is a clustering algorithm based on representative points. According to the principle of the representativeness of the data points themselves, it mainly uses the negative Euclidean distance of each data point as the similarity between the data points, and constructs the corresponding matrix based on the self-similarity and mutual similarity of the data points. The optimal cluster center is determined by iterative refinement, and the clusters are classified according to the degree of similarity to ensure the clustering quality and the number of clusters. In traditional AP clustering, the bias parameter value P is used as the element on the main diagonal of simi(i, k), and p is generally set to the median of simi(i, k), and λ takes a fixed value. Unlike traditional AP clustering, this paper adopts the PSO particle swarm optimization algorithm to optimize P and λ on a dynamic scale to obtain the optimal dual parameter combination. By initializing the AP clustering dual parameters, a particle swarm with a population size of N and a search space and search capability is constructed. The silhouette coefficient is used as the fitness objective function to assign position and velocity attributes to achieve particle swarm update iteration, and realize information sharing and collaboration within the particle swarm; such as Figure 2 As shown, the PSO algorithm expression is:

[0054] V i =V i +c1×rand()×(pbest i -X i )+c2×rand()×(gbest i -X i )

[0055] Where: V is the particle update speed; c1, c2 are learning factors; X i ——Current position; pbest——Particle historical optimal parameters; gbest i ——The historical optimal parameters of the population.

[0056] X i =X i +V i

[0057] Where: x i ——Current location.

[0058] The wind farm data used for training in step 3 is the prediction data obtained by the wind farm prediction system as the input of the BP neural network, and the actual total output of the wind farm group is used as the output.

[0059] In the PSO-AP-BP composite model based on correlation coefficient weights described in step 4, the BP neural network is a three-layer, forward-transfer and feedback-combined feedforward neural network, including an input layer, a hidden layer and an output layer. It has the characteristics of high cognitive ability and adaptive ability, as well as strong nonlinear mapping and generalization ability, and can efficiently predict wind power. The training process of the BP neural network consists of two-way propagation of characteristic parameter data. Forward propagation, that is, obtaining an output based on the input, and back propagation, comparing the predicted value with the true value, calculating the loss, and then back propagating the loss to update the weights and biases, and optimize the parameters to minimize the loss and optimize the model. However, in the actual short-term power forecasting process, the prediction results based on the original data time series and the output power of the field group often have large errors in some prediction time nodes, such as Figure 3 As shown in the figure, this paper proposes a BP neural network model based on correlation coefficient weights. In order to obtain the correlation between the actual output of a single wind farm and the total output power of the farm group, according to the correlation principle between the measured power of a single wind farm and the measured total power of multiple regions, the wind farm with greater correlation will obtain a greater weight in the prediction process. Figure 4 As shown, the correlation coefficient mechanism is as follows:

[0060]

[0061] Where: ρ is the Pearson correlation coefficient, ρ>0, indicating that the two vectors are positively correlated. The larger the absolute value, the higher the degree of correlation; x i ——actual power sequence of wind farm; Y i ——The actual total output power of the field group.

[0062] Step 6 uses the root mean square error (RMSE) and mean absolute percentage error (MAPE) to evaluate the prediction performance of the PSO-AP-BP composite model based on the correlation coefficient weight. At the same time, evaluation indicators such as prediction accuracy (PA) and correlation (CC) are used to quantify the wind power prediction model to more clearly compare the prediction performance of each model. The calculation formula is as follows:

[0063] PA=1-MAPE

[0064]

[0065] Where: len is the number of predicted values ​​of the wind farm group; P, p are the predicted value sequence and the actual value sequence, p k 、p k ——Power prediction and actual value.

[0066] Example

[0067] The output data of 13 wind farms in Northern Europe are used as the original data set, with a time span from February 1, 2022 to December 31, 2022, a sampling resolution of 1 hour, and a total of 7944 original wind power data. The wind farm data types used are station-level output prediction data sets and station-level output data sets. There are 4 wind farms in Norway and Sweden, 2 wind farms in Denmark, and 1 wind farm each in Estonia, France and Lithuania. The prediction model uses Matlab to build a virtual environment to complete model training and data prediction. BP prediction, BP prediction based on correlation coefficient weights, PSO-AP-BP prediction, and PSO-AP-BP prediction methods based on correlation coefficient weights are used to predict the total output of the wind farm group. The prediction results of different models are shown as follows. Figure 5 As shown, compared with the first three single predictions or composite models such as BP prediction, the prediction results of the PSO-AP-BP prediction model based on correlation coefficient weights proposed in this paper are basically consistent with the output trend of the actual data, the prediction error is not large, it basically fits the actual output curve and can better reflect the change trend of the power output on the day before. Compared with the traditional BP single prediction model, the PSO-AP-BP model prediction based on correlation coefficient weights mentioned in this paper significantly reduces the root mean square error, improves the prediction accuracy by 2.62%, and the correlation coefficient can reach 0.936. Figure 6 The comparison between the superposition method and the PSO-AP-BP composite model based on correlation coefficient weights is shown. The prediction results obtained by the model proposed in this paper are closer to the true value and show good performance in prediction accuracy.

[0068] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A short-term wind power prediction method based on cluster optimization analysis, characterized in that: The wind power prediction method of the present invention comprises the following steps: S1: Use data processing methods to handle outliers, clean missing values, and generate wind farm groups with high confidence and hourly power data series for each wind farm. S2: Take the historical output data sets of each wind farm as the clustering object, and apply the AP clustering algorithm combined with PSO particle swarm optimization to obtain the best clustering result. S3: Obtain wind farm data for training, divide it into a training set and a test set, and complete data normalization as data preparation for BP neural network prediction. S4: Input the data in S3 into the PSO-AP-BP composite model based on the correlation coefficient weight, and use the clustering results to apply the BP neural network to the clustering areas respectively to obtain the sub-area prediction value. S5: Complete the prediction of each cluster area and the denormalization of the actual output data. S6: Select the optimal wind farm group power prediction model using corresponding evaluation indicators to obtain the final wind farm group wind power output.

2. A short-term wind power prediction method based on cluster optimization analysis according to claim 1, characterized in that As described in S1, in the actual site-level wind farm output data, the negative effects of abnormal output values ​​of a single or multiple wind turbines are relatively small, and the possible abnormal values ​​can be approximately treated as normal data.

3. The short-term wind power prediction method based on cluster optimization analysis according to claim 1 is characterized in that The power data sequence described in S1 is a predicted power data sequence and an actual historical output data sequence.

4. The short-term wind power prediction method based on cluster optimization analysis according to claim 1 is characterized in that In the PSO-AP optimization algorithm model described in S2, the PSO particle swarm optimization algorithm is used to optimize P and λ at a dynamic scale to obtain the optimal dual parameter combination. By initializing the AP clustering dual parameters, a particle swarm with a population size of N and a search space and search capability is constructed. The silhouette coefficient is used as the fitness objective function to assign position and velocity attributes to achieve particle swarm update iteration and realize information sharing and collaboration within the particle swarm. The PSO algorithm expression is: V i =V i +c1×rand()×(pbest i -X i )+c2×rand()×(gbest i -X i ) Where: V is the particle update speed; c1, c2 are learning factors; X is the current position; pbest is the historical optimal parameter of the particle; gbest is the historical optimal parameter of the population. X i =X i +V i Where: X i ——Current location.

5. The short-term wind power prediction method based on cluster optimization analysis according to claim 1 is characterized in that The wind farm data used for training in S3 is the prediction data obtained by the wind farm prediction system as the input of the BP neural network, and the actual total output of the wind farm group is used as the output.

6. The short-term wind power prediction method based on cluster optimization analysis according to claim 1 is characterized in that The PSO-AP-BP composite model based on correlation coefficient weight described in S4 adopts the correlation coefficient between the wind farm output and the wind farm group output, and assigns values ​​to the corresponding wind farms as weights. The correlation coefficient mechanism is as follows: Where: ρ is the Pearson correlation coefficient, ρ>0, indicating that the two vectors are positively correlated. The larger the absolute value, the higher the degree of correlation; X i ——actual power sequence of wind farm; Y i ——The actual total output power of the field group.

7. The short-term wind power prediction method based on cluster optimization analysis according to claim 1 is characterized in that S6 uses root mean square error (RMSE) and mean absolute percentage error (MAPE) to evaluate the prediction performance of the PSO-AP-BP composite model based on correlation coefficient weights.

8. The short-term wind power prediction method based on cluster optimization analysis according to claim 1 is characterized in that S6 uses evaluation indicators such as prediction accuracy (PA) and correlation (CC) to quantify the wind power prediction model to more clearly compare the prediction performance of each model. The calculation formula is as follows: PA=1-MAPE Where: len——the number of predicted values ​​of the wind farm group; P, p——predicted value sequence and actual value sequence, p k 、p k ——Power prediction and actual value.

9. The short-term wind power prediction method based on cluster optimization analysis according to claim 1 is characterized in that The final output of the wind farm group described in S6 is the superposition of the predicted values ​​of each clustering area obtained by the BP neural network based on the correlation coefficient.

Images