CN106650784A - Feature clustering comparison-based power prediction method and device for photovoltaic power station - Google Patents

Abstract

The invention relates to a method and device for predicting the power of a photovoltaic power station based on feature clustering comparison, and belongs to the technical field of photovoltaic power generation. The present invention first acquires the three most important feature quantities that affect the prediction accuracy of photovoltaic power, and accumulates them as historical meteorological data; then clusters the obtained historical meteorological data as data samples, divides the samples into k categories with high similarity, and Get the cluster centers of various types; use various historical data to establish corresponding prediction models; select the prediction model corresponding to the cluster center closest to the current object for prediction. The invention divides different meteorological data into different sample types, and establishes photovoltaic output prediction models under different meteorological conditions, so that the training of the prediction models is more targeted. Power prediction accuracy.

Description

A photovoltaic power plant power prediction method and device based on feature cluster comparison

technical field

The invention relates to a method and device for predicting the power of a photovoltaic power station based on feature clustering comparison, and belongs to the technical field of photovoltaic power generation.

Background technique

As a kind of clean energy, photovoltaic power generation has received extensive attention worldwide. There are more and more large-scale ground photovoltaic power stations and distributed rooftop photovoltaic power stations in my country. However, due to the volatility and intermittent nature of photovoltaic power generation, power grid Scheduling has brought great difficulties, and the development of photovoltaic power forecasting has effectively alleviated this problem. However, the photovoltaic power forecasting errors currently on the market are relatively large, and it is difficult to provide effective reference for grid dispatching.

At present, my country has carried out some technical research on the power prediction of photovoltaic power plants. The existing prediction models include neural network models, radial basis function models, and multi-layer perception models. Among them, the neural network model is the most widely used. The neural network model generates output through the action of various neurons in the input layer, hidden layer and output layer, and then uses the error as the objective function to continuously correct the network weight until the error meets the requirements. , the trained network can predict the output power of photovoltaic modules. However, the current neural network model can only control the prediction error at about 20%. If there is a thunderstorm and strong wind, the prediction accuracy will be even worse. To further improve the prediction accuracy, it is necessary to do a reasonable preprocessing of the input data. Accuracy of historical data for training samples.

A prediction method is given in the paper titled "PV Output Prediction Based on Weather Type Clustering and LS-SVM", which clusters historical meteorological data by seasonal type and obtains four different types of clusters for each season samples to form a corresponding prediction model; in actual prediction, first determine the season according to the date to be predicted, and find the corresponding prediction sub-model based on the meteorological characteristics of the day to be predicted, and make a prediction. However, since the date to be predicted is divided into seasons according to time and the forecasting sub-model is sought, such a division method has a certain degree of subjectivity and compulsion, which in turn limits the process of finding the optimal forecasting model , part of the sample data, especially the sample data in the period of seasonal alternation, will inevitably reduce the prediction accuracy due to the failure to find the best prediction model.

Contents of the invention

The purpose of the present invention is to provide a method for predicting the power of photovoltaic power plants based on feature clustering and comparison, and solve the problems of low prediction accuracy and low accuracy of the huge amount of current photovoltaic power plant power. At the same time, it also provides a photovoltaic power plant power forecasting device based on feature cluster comparison.

In order to solve the above technical problems, the present invention provides a photovoltaic power plant power prediction method based on feature cluster comparison, the steps of the prediction method are as follows:

1) Collect the characteristic quantities that affect the prediction accuracy of photovoltaic power, total radiation, wind speed and temperature, and use the historical data of the characteristic quantities to form a sample set;

2) Perform feature clustering on the obtained sample set, divide the sample set into k categories with high similarity, and obtain the cluster centers C _i of various sample data, i=1, 2, ..., k;

3) Use various sample data to establish k-type prediction models correspondingly;

4) Calculate the distance between the current prediction object and the cluster centers of various sample data, and select the prediction model corresponding to the class of the cluster center closest to the current prediction object for prediction.

Further, the step 2) uses the K-means algorithm to perform feature clustering, and the initial clustering centers of various types are obtained using the idea of Huffman tree construction.

Further, the prediction model of the step 3) adopts a BP neural network prediction model, and the prediction model adopts a three-layer structure including an input layer, a hidden layer and an output layer, and the input layer adopts three characteristic quantities of total radiation, temperature and wind speed , the output layer is the output power of the photovoltaic power station.

Further, in the step 4), the distance between the current predicted object and the cluster centers of each category is calculated by using the weighted Euclidean distance.

Further, the step 2) divides the sample set into three categories, representing rainy, cloudy and sunny days respectively.

The present invention also provides a photovoltaic power station power forecasting device based on feature clustering comparison, the forecasting device includes an acquisition module, a feature clustering module, a prediction model building module and a prediction module,

The collection module is used to collect feature quantities that affect the prediction accuracy of photovoltaic power, total radiation, wind speed and temperature, and use historical data of feature quantities to form a sample set;

The feature clustering module is used to perform feature clustering on the obtained sample set, divide the sample set into k categories with high similarity, and obtain the cluster centers C _i of various sample data, i=1, 2 ,...,k;

Described prediction module establishment module utilizes various sample data to establish corresponding k class prediction models respectively;

The prediction module is used to calculate the distance between the current prediction object and the cluster centers of various sample data, and select the prediction model corresponding to the class of the cluster center closest to the current prediction object for prediction.

Further, the feature clustering module uses the K-means algorithm to perform feature clustering, and various initial cluster centers are obtained using the idea of Huffman tree construction.

Further, the building block of the prediction module adopts a BP neural network prediction model, and the prediction model adopts a three-layer structure including an input layer, a hidden layer and an output layer, and the input layer adopts three characteristic quantities of total radiation, temperature and wind speed, The output layer is the output power of the photovoltaic power station.

Further, the prediction module calculates the distance between the current prediction object and the cluster center of each category by using the Euclidean distance calculation of the weighted method.

Further, the feature clustering module divides the sample set into three categories, representing rainy, cloudy and sunny days respectively.

The beneficial effects of the present invention are: the present invention at first obtains the three most important feature quantities that affect the prediction accuracy of photovoltaic power, and accumulates as historical meteorological data; then clusters the obtained historical meteorological data as data samples, and divides the samples into similarity Higher class k, and get the cluster center of each type; use each type of historical data to establish the prediction model of the corresponding class; select the prediction model corresponding to the cluster center closest to the current object for prediction. The invention divides different meteorological data into different sample types, and establishes photovoltaic output prediction models under different meteorological conditions, so that the training of the prediction models is more targeted. Power prediction accuracy.

Description of drawings

Fig. 1 is the flowchart of the method for predicting power of photovoltaic power plants based on feature clustering comparison in the present invention;

Figure 2 is a schematic diagram of the construction process of the Huffman tree;

Fig. 3 is the flow chart of initial clustering center selection based on Huffman tree;

Fig. 4 is a prediction model diagram based on BP neural network.

detailed description

The specific embodiments of the present invention will be further described below in conjunction with the accompanying drawings.

Embodiment of the method for predicting the power of photovoltaic power plants based on feature clustering comparison in the present invention

The photovoltaic power plant power prediction method of the present invention firstly obtains the characteristic quantity that affects the photovoltaic power prediction accuracy, and uses the historical data of the characteristic quantity to form a sample set; then uses the feature clustering algorithm to gather the samples into k categories, and uses various historical data to establish the corresponding category. Prediction model; finally calculate the distance between the current object and various cluster centers, and select the prediction model corresponding to the class of the cluster center closest to the current object to predict the current object, so as to realize the prediction of the power of the photovoltaic power station. The implementation process of this method is shown in Figure 1, and the specific steps are as follows.

1. Obtain the three most important characteristic quantities that affect the prediction accuracy of photovoltaic power, and use the historical data of the characteristic quantities to form a sample set.

The three most important characteristic quantities affecting the prediction accuracy of photovoltaic power selected by the present invention are total radiation, wind speed and temperature respectively, and these three characteristic quantities can be obtained from NWP numerical weather prediction.

2. Perform feature clustering on the obtained sample set, divide the samples into k classes with high similarity, and obtain the cluster centers C _i of each class, i=1, 2, ..., k.

There are many algorithms for realizing sample clustering, such as partition-based clustering algorithms, hierarchical-based clustering algorithms, grid-based clustering algorithms, density-based clustering algorithms, etc., all of which can realize the clustering process in the present invention , the following is an example based on the most typical K-means algorithm in the partition clustering algorithm, the clustering process of the present invention is described in detail, and the specific implementation is as follows:

(1) Establish feature matrix.

Taking the historical meteorological data of photovoltaic power plants as samples, the characteristic quantities are total radiation, temperature, and wind speed, and the sample composition is as follows:

Each sample data covers various feature quantities

Among them, u _i1 , u _i2 , and u _i3 respectively represent three characteristic quantities of total radiation, wind speed and temperature.

(2) Select the initial cluster center.

In this embodiment, the initial cluster center is selected based on the Huffman tree construction idea. The number of vectors is n, the dimension of the sample set is 3, and the type of clustering is 3 types. The type of clustering can be adjusted according to the actual situation. In this embodiment, only 3 types of cases are given. The specific steps for selecting the initial center point are as follows .

Step 1: Calculate the Euclidean distance between two vectors.

where i=1,2...n; j=1,2...n; and i≠j.

The dissimilarity matrix is obtained, and the matrix is as follows:

Step 2: find the minimum value d _ij in the matrix, calculate the average value of the two vectors u _i and u _j , and obtain the vector V ₁ .

Step 3: Delete the two vectors u _i and u _j from the original feature matrix, and add vector V ₁ to obtain a new feature matrix.

Step 4: Repeat step 1, step 2, and step 3 until there is only one vector left in the feature matrix. The process of constructing a tree in the iterative process is shown in Figure 2.

Step 5: Cluster the data into 5 categories, and subtract the nodes from the tree sequentially from the vertex J _m to 4, forming 5 trees, as shown in Figure 3.

The sixth step is to calculate the average value of the node vectors at the bottom of the three trees to obtain the vector is the initial cluster center.

(3) Calculate the average error criterion function under the initial cluster center.

Where k=3, represents the number of divided clusters; u _i ⁽⁰⁾ represents the centers of the three clusters.

(4) Calculate the new cluster center.

N ₁ means belong to number of elements

N ₂ means belong to number of elements

N ₃ means belong to number of elements

Where N ₁ +N ₂ +N ₃ =nn represents the total amount of sample data

(5) Repeat the above steps (2), (3) and (4) until the following relationship is satisfied.

And there is E ^(s+1) ≤ E ^(s) , which means that after s clustering iterations, the average error criterion function is in a state of convergence, and the cluster center does not change any more, so the iteration is stopped, and the clustering result is completed.

3. Establish a photovoltaic power prediction model.

In this embodiment, the BP neural network prediction model is used as the optical power prediction model, and the BP neural network model is trained by using various types of historical data after clustering, so as to obtain prediction models corresponding to each type. As shown in Figure 4, the BP neural network prediction model adopts a three-layer structure, input layer, hidden layer, and output layer. The input layer uses three characteristic quantities of total radiation, temperature, and wind speed. The number of layer neurons is obtained through actual engineering verification. According to the clustering results obtained in step 2, the three types of sample data are respectively brought into the BP neural network prediction model for training, and three different types of prediction models can be obtained. The specific implementation process is as follows:

The first step is network initialization: assign a random number in the interval (-1, 1) to each connection weight, set the error function e, and give the calculation accuracy value Δ and the maximum number of learning times.

In the second step, select k input samples (historical data of the last 30 days) and the corresponding expected output (the expected output value corresponds to the actual power generated in 30 days).

The third step is to calculate the input and output of each neuron in the hidden layer.

The fourth step is to use the expected output of the network (actual power generation) and the actual output after training to calculate the partial derivative of the error function to each output neuron.

The fifth step is to use the error partial derivative of each neuron in the output layer and the output of each neuron in the hidden layer to modify the connection weight.

The sixth step is to use the error partial derivative of each neuron in the hidden layer and the output of each neuron in the input layer to modify the connection weight.

The seventh step is to calculate the global error.

The eighth step is to judge whether the network error meets the requirements. When the error reaches the preset accuracy or the number of learning times is greater than the set maximum number of times, the algorithm ends. Otherwise, select the next learning sample and the corresponding expected output, and return to the third step. Go to the next round of study.

4. Calculate the distance between the current prediction object and various cluster centers, and select the prediction model corresponding to the class of the cluster center closest to the current prediction object for prediction.

The first step is to obtain the meteorological data of the current forecast object, and select the forecast model according to the output results of the clustering algorithm.

Calculate the weighted Euclidean distance between the meteorological data of the current object and various cluster centers. If the distance to the i-th cluster center C _i is the smallest, use the i-th type forecasting model for prediction.

The second step is to normalize the meteorological data.

The third step is to use the current meteorological data as input, load the corresponding prediction model, and output the prediction results.

Through the above process, the present invention can establish a power prediction model according to different meteorological conditions, which can effectively improve the problem of low prediction accuracy of photovoltaic output power under cloudy and rainy weather conditions.

Embodiment of the photovoltaic power plant power prediction device based on feature clustering comparison of the present invention

The prediction device in this embodiment includes an acquisition module, a feature clustering module, a prediction model building module, and a prediction module; the acquisition module is used to collect feature quantities that affect the prediction accuracy of photovoltaic power, total radiation, wind speed and temperature, and use the feature quantities Historical data constitutes a sample set; the feature clustering module is used to perform feature clustering on the obtained sample set, divide the sample set into k categories with high similarity, and obtain the cluster centers C _i of various sample data, i= 1, 2, ..., k; the prediction module building module uses various sample data to establish corresponding k-type prediction models; the prediction module is used to calculate the distance between the current prediction object and the cluster centers of various sample data, and select the The current prediction object is predicted by the prediction model corresponding to the class of the nearest cluster center. The specific implementation means of each module has been described in the example of the method, and will not be repeated here.

Claims (10)

1. the predicting power of photovoltaic plant method that a kind of feature based cluster compares, it is characterised in that the step of the Forecasting Methodology It is as follows：

1) characteristic quantity of collection impact photovoltaic power precision of prediction, global radiation, wind speed and temperature, and using the history number of characteristic quantity According to composition sample set；

2) sample set for obtaining is carried out into feature clustering, sample set is divided into into the higher k classes of similitude, and obtain Different categories of samples The cluster centre C of data_i, i=1,2 ... ..., k；

3) Different categories of samples data correspondence establishment k class forecast models are respectively adopted；

4) the distance between cluster centre of current predictive object and Different categories of samples data is calculated, is chosen with current predictive to image distance It is predicted from the nearest corresponding forecast model of cluster centre place class.

2. the predicting power of photovoltaic plant method that feature based cluster according to claim 1 compares, it is characterised in that institute State step 2) feature clustering is carried out using K-means algorithms, all kinds of initial cluster centers constructs the thought of tree using Huffman Obtain.

3. the predicting power of photovoltaic plant method that feature based cluster according to claim 1 and 2 compares, its feature exists In the step 3) forecast model adopt BP neural network forecast model, the forecast model is using including input layer, hidden layer With the three-decker of output layer, using global radiation, temperature, three characteristic quantities of wind speed, output layer is photovoltaic plant output to input layer Power.

4. the predicting power of photovoltaic plant method that feature based cluster according to claim 1 compares, it is characterised in that institute State step 4) in the distance between current predictive object and all kinds of cluster centres be calculated using the Euclidean distance of weighting method.

5. the predicting power of photovoltaic plant method that feature based cluster according to claim 2 compares, it is characterised in that institute State step 2) sample set is divided into into three classes, overcast and rainy, cloudy and fine day is represented respectively.

6. the predicting power of photovoltaic plant device that a kind of feature based cluster compares, it is characterised in that the prediction meanss include adopting Collection module, feature clustering module, forecast model set up module and prediction module,

Described acquisition module is used for collection affects the characteristic quantity of photovoltaic power precision of prediction, global radiation, wind speed and temperature, and profit Sample set is constituted with the historical data of characteristic quantity；

Described feature clustering module is used to for the sample set for obtaining to carry out feature clustering, sample set is divided into into similitude higher K classes, and obtain the cluster centre C of Different categories of samples data_i, i=1,2 ... ..., k；

Described prediction module sets up module and is utilized respectively Different categories of samples data foundation correspondence k class forecast models；

Described prediction module is used to calculate the distance between cluster centre of current predictive object and Different categories of samples data, chooses The corresponding forecast model of the cluster centre place class nearest with current predictive object distance is predicted.

7. the predicting power of photovoltaic plant device that feature based cluster according to claim 6 compares, it is characterised in that institute The feature clustering module stated carries out feature clustering using K-means algorithms, and all kinds of initial cluster centers are constructed using Huffman The thought of tree is obtained.

8. the predicting power of photovoltaic plant device that the feature based cluster according to claim 6 or 7 compares, its feature exists In described prediction module sets up module using BP neural network forecast model, and the forecast model is adopted to be included input layer, imply The three-decker of layer and output layer, input layer adopts global radiation, temperature, three characteristic quantities of wind speed, and output layer is that photovoltaic plant is defeated Go out power.

9. the predicting power of photovoltaic plant device that feature based cluster according to claim 6 compares, it is characterised in that institute The prediction module stated is calculating Euclidean distance of the distance between current predictive object and all kinds of cluster centres using weighting method It is calculated.

10. the predicting power of photovoltaic plant device that feature based cluster according to claim 7 compares, it is characterised in that Sample set is divided into three classes by described feature clustering module, and overcast and rainy, cloudy and fine day is represented respectively.