A method and device for predicting wind farm power

Abstract

The present invention discloses a method and device for predicting wind farm power, the method comprising: obtaining prediction day information; the prediction day information includes the prediction day date; determining the season of the prediction day according to the prediction day date, extracting similar days according to the prediction day season and wind speed, and obtaining similar day data; the similar day data includes similar day wind speed data; performing wind speed prediction based on similar days according to the similar day data, and obtaining prediction day wind speed data; performing wind farm power prediction according to the similar day data and the prediction day wind speed data, and obtaining the power of the wind farm. Establishing a wind power prediction combination model based on GA-BP multi-hidden layer deep neural network can greatly improve the accuracy of wind power prediction and make the model fitting more stable.

Description

Technical Field

[0001] The invention relates to a method and a device for predicting wind farm power, belonging to the technical field of new energy power prediction. Background Art

[0002] With the continuous consumption of fossil fuels, countries around the world are responding to the global energy crisis and environmental degradation through global agreements. my country must vigorously develop new energy and renewable energy. Wind energy, as one of the renewable energy sources, has low costs and abundant reserves. Accurate prediction of wind speed and wind power can, to a certain extent, assist the power system in timely and effective dispatch and ensure the reliable operation of the power system.

[0003] Wind power prediction technology is particularly important, which can reduce operating costs. Accurately predicting wind power can not only reduce the relevant configuration of spare capacity, but also reduce operating costs, and achieve the optimal scheduling of wind farm generators; it can also enhance market competitiveness. The difference between wind power generation and traditional power generation methods is that the power output of wind farms changes all the time with the volatility and intermittency of wind. It is impossible to formulate a power generation plan before generating electricity, and the reliability of power supply is low. If the accuracy of wind power prediction can be improved and the reliability of wind power generation can be improved, the competitiveness of the wind power market can be improved. The wind power market economy can be developed, and social benefits can also be increased. Some remote areas in my country, especially the poorer mountainous areas, are rich in wind energy resources and have large available space. If good wind power prediction can be achieved, the development of related industries and the establishment of supporting facilities can be carried out locally, which can promote local development, drive economic growth, and accelerate the pace of national development. Wind power prediction technology ensures the stable operation of grid connection, understands and grasps the direction and change trend of wind power in advance, and can reduce the fluctuation of grid frequency and voltage flicker when the wind turbine is suddenly cut off, and ensure the quality of power.

[0004] Currently, the mainstream methods of wind power forecasting include BP model, GRNN model and GA-BP traditional model, which generally have the problem of low accuracy and are unable to cope with increasingly demanding wind power forecasting scenarios. Summary of the invention

[0005] The purpose of the present invention is to overcome the deficiencies in the prior art and to provide a method and device for predicting wind farm power, which can improve the accuracy of wind power prediction.

[0006] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0007] In a first aspect, the present invention provides a method for predicting wind farm power, comprising the following steps:

[0008] Obtaining forecast day information; the forecast day information includes the forecast day date;

[0009] Determine the season of the forecast day according to the date of the forecast day, extract similar days according to the season and wind speed of the forecast day, and obtain similar day data; the similar day data includes similar day wind speed data;

[0010] According to the data of similar days, the wind speed forecast based on similar days is carried out to obtain the wind speed data of the forecast day;

[0011] The wind farm power is predicted based on the similar day data and the predicted day wind speed data to obtain the wind farm power.

[0012] Further, the method for extracting similar days according to the season and wind speed of the forecast day includes:

[0013] Acquire historical wind speed data of the wind farm, and organize the historical wind speed data into four seasonal wind speed sets according to seasons;

[0014] The seasonal wind speed sets are clustered by K-means using the Weka platform, and the sum of squared errors corresponding to different cluster numbers K are selected as the cluster number evaluation criterion;

[0015] Draw the corresponding elbow diagram to get the most suitable clustering number K, use different colors to distinguish the clusters formed by different clustering numbers K and correspond to the number of similar days;

[0016] According to the season of the forecast day, the cluster where the forecast day is located is selected for similar days, and the Pearson coefficient similar days are extracted again to obtain similar days with strong correlation coefficients within the cluster.

[0017] Furthermore, the method for performing K-means clustering includes the following steps:

[0018] The objective function is determined to minimize the square error E, and the expression is as follows:

[0019]

[0020] where μ i is the mean vector, expressed as follows:

[0021]

[0022] In the formula, μ i It is cluster C i The cluster center (i.e., centroid) of is also the mean vector.

[0023] Furthermore, based on the data of similar days, the wind speed forecast based on similar days is performed, and the method for obtaining the wind speed data of the forecast day includes:

[0024] Get the original wind speed sequence of similar days;

[0025] The complete empirical mode decomposition of adaptive white noise (CEEMDAN) is used to decompose the original wind speed sequence of similar days to obtain a decomposed wind speed sequence.

[0026] The empirical mode decomposition (VMD) is used to perform secondary decomposition on the primary decomposition wind speed sequence to obtain the secondary decomposition wind speed sequence.

[0027] The secondary decomposed wind speed sequence is processed according to the empirical mode decomposition method and the echo state network prediction algorithm to obtain the predicted daily wind speed data.

[0028] Furthermore, the method of using the complete empirical mode decomposition of adaptive white noise (CEEMDAN) to decompose the original wind speed series of similar days includes:

[0029] Add K times of white noise to the signal sequence x(t) to construct K times of decomposition sequence x i (t)(i=1,2,…,K), the expression is as follows:

[0030] x i (t) = x(t) + εδ i (t) (3)

[0031] Where ε is the white noise weight coefficient, δ i (t) is the white noise generated by the i-th processing;

[0032] For the above sequence x to be decomposed i (t) (i=1,2,…,K) performs K times of empirical mode decomposition, and takes the average value of all the first modal components after solution IMF1(t) as the first modal component of CEEMDAN decomposition, which is expressed as follows:

[0033]

[0034] In the formula, For x i (t) The first modal component obtained after empirical mode decomposition;

[0035] The residual signal of the signal sequence x(t) is solved and the expression is shown as follows:

[0036] r1(t)=x(t)-IMF1(t) (5)

[0037] r1(t) represents the residual signal after the first CEEMDAN decomposition;

[0038] After adding specific noise to the j-th stage residual signal obtained after decomposition, continue to perform EMD decomposition. j The expression is as follows:

[0039]

[0040] r j (t) = r j-1 (t)-IMF j (t) (7)

[0041] In the formula, IMF j (t) is the jth IMF obtained by CEEMDAN decomposition; E j-1 (·) represents the j-1th IMF component after EMD decomposition of the sequence, ε j-1 represents the weight coefficient of noise added by CEEMDAN to the residual signal of the j-1th stage; r j (t) represents the residual signal of the jth stage;

[0042] If the residual signal becomes monotonic after a certain number of decompositions, the decomposition stop condition is met and the algorithm ends.

[0043] Furthermore, the method of using the empirical mode decomposition (VMD) to perform a secondary decomposition of the primary decomposition wind speed sequence to obtain the secondary decomposition wind speed sequence includes:

[0044] Step a: Initialization λ 1 and the maximum number of iterations N;

[0045] Step b: Update and ω k ;ω k The expressions of and λ are as follows:

[0046]

[0047]

[0048]

[0049] Step c: If satisfied The iteration is completed and the final output is and ω k If not, repeat steps ab until it is satisfied

[0050] Furthermore, the method of processing the secondary decomposition wind speed sequence according to the empirical mode decomposition method and the echo state network prediction algorithm to obtain the predicted daily wind speed data includes:

[0051] The secondary decomposition wind speed series is decomposed using the empirical mode decomposition method (EMD) to obtain the modal components.

[0052] The echo state network (ESN) is used to predict the predicted values ​​of all modal components, and the predicted values ​​are added together to obtain the predicted results of the wind speed data on the predicted day.

[0053] Furthermore, the secondary decomposition wind speed sequence is decomposed using the empirical mode decomposition method EMD to obtain the modal components, including:

[0054] Step A: Find all extreme points of the secondary decomposition wind speed sequence, connect the local maximum points into an upper envelope by using the cubic spline curve method, and connect the local minimum points into a lower envelope by using the cubic spline curve method;

[0055] Step B: respectively obtain the upper and lower envelope average curves of the original wind speed sequence of each similar day, and subtract the average curve from the original signal to obtain the suspected IMF curve;

[0056] Step C: Determine whether the suspected IMF curve meets the IMF conditions. If so, take the curve as the IMF component. If not, take the curve as the original data set and repeat steps AB until it meets the IMF component and obtains the IMF signal.

[0057] Step D: Subtract the IMF signal from the original signal to obtain a new original signal. If the new original signal is a monotonic function, the process ends. If the new original signal is not a monotonic function, steps AB are repeated.

[0058] Further, the method of predicting the predicted values ​​of all modal components using the echo state network ESN includes:

[0059] Determine the size of the reserve pool, scale the matrix so that the spectral radius is less than 1, and randomly generate input connections and output feedback weights;

[0060] The samples are loaded into the input and output in sequence, the state record reservoir state is updated, and the output connection weights are determined using linear regression;

[0061] The input vector is loaded into the input node and runs idle, and the predicted value is obtained from the output node.

[0062] Furthermore, the similar day data includes similar day wind speed data and wind farm power data;

[0063] The method for predicting wind farm power according to the similar day data and the predicted day wind speed data includes:

[0064] Perform data preprocessing and data standardization on similar day wind speed data and wind farm power data;

[0065] Using similar day wind speed data and wind farm power data to train the GA-BP neural network, and obtaining a trained GA-BP neural network;

[0066] Perform data preprocessing and data standardization on the forecast day wind speed data;

[0067] The processed predicted daily wind speed data is input into the trained GA-BP neural network to obtain the wind farm power prediction data;

[0068] The wind farm power prediction data is denormalized to obtain the model prediction result, that is, the wind power every 15 minutes within 24 hours on the prediction day.

[0069] In a second aspect, the present invention provides a wind speed prediction device for a wind farm, comprising:

[0070] Acquisition module: used to acquire forecast day information; the forecast day information includes the forecast day date;

[0071] Similar day extraction module: used to determine the season of the forecast day according to the date of the forecast day, extract similar days according to the season and wind speed of the forecast day, and obtain similar day data; the similar day data includes similar day wind speed data;

[0072] Wind speed prediction module: used to predict wind speed based on similar days according to similar day data and obtain wind speed data for the predicted day;

[0073] Power prediction module: used to predict the power of the wind farm based on the similar day data and the predicted day wind speed data, and obtain the power of the wind farm.

[0074] Compared with the prior art, the present invention has the following beneficial effects:

[0075] 1. The present invention compares the wind speed distribution in four seasons and finds that there is a large seasonal distribution difference in wind speed in four seasons. Based on this factor, cluster analysis is performed to extract similar days by season, and the result is more accurate;

[0076] 2. The present invention uses the data mining platform Weka as a clustering platform because Weka has the advantages of visualization and is more suitable for small-scale machine learning. In subsequent clustering applications, it can improve efficiency. The Weka platform is selected as the clustering platform for subsequent similar day extraction;

[0077] 3. The wind speed in four seasons was clustered by K-means using the Weka platform. Different parameters were set for clustering operations. The sum of squared errors corresponding to different cluster numbers were selected as the cluster number evaluation criterion. The corresponding elbow diagram was drawn to select the best cluster number. The cluster scatter point number was generated and the components of each cluster were recorded, which effectively improved the accuracy of the prediction results.

[0078] 4. The present invention ultimately predicts wind power based on similar days. It improves the structure of the GA-BP neural network and proposes a wind power prediction model based on similar days and GA-BP deep neural network to predict the wind power of the wind farm. Compared with the BP model, GRNN model and GA-BP traditional model, it can fit more stably, with higher prediction accuracy and better feasibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 is a flow chart of the wind speed prediction method of the present invention;

[0080] Figure 2 It is a flow chart of a K-means clustering algorithm provided by the present invention;

[0081] Figure 3 It is a flow chart of an EMD algorithm provided by the present invention;

[0082] Figure 4 It is a CEEADAN algorithm flow chart provided by the present invention;

[0083] Figure 5 It is an approximate day-GA-BP-DNN model design diagram provided by the present invention. DETAILED DESCRIPTION

[0084] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0085] Embodiment 1:

[0086] like Figure 1 As shown, an embodiment of the present invention provides a wind farm power prediction method, comprising the following steps:

[0087] Obtain forecast day information; the forecast day information includes the forecast day date;

[0088] The season of the forecast day is determined according to the date of the forecast day, and similar days are extracted according to the season and wind speed of the forecast day to obtain similar day data; the similar day data includes similar day wind speed data;

[0089] According to the data of similar days, the wind speed forecast based on similar days is carried out to obtain the wind speed data of the forecast day;

[0090] Based on similar day data and predicted day wind speed data, the wind farm power is predicted to obtain the wind farm power.

[0091] Specifically, the method for extracting similar days according to the season and wind speed of the forecast day includes:

[0092] Obtain historical wind speed data of wind farms, and organize the historical wind speed data into four seasonal wind speed sets according to seasons;

[0093] The seasonal wind speed sets were clustered by K-means using the Weka platform, and the sum of squared errors corresponding to different cluster numbers K were selected as the evaluation criteria for the number of clusters.

[0094] Draw the corresponding elbow diagram to get the most suitable clustering number K, use different colors to distinguish the clusters formed by different clustering numbers K and correspond to the number of similar days;

[0095] According to the season of the forecast day, the cluster where the forecast day is located is selected for similar days, and the Pearson coefficient similar days are extracted again to obtain similar days with strong correlation coefficients within the cluster.

[0096] Specifically, Figure 2 The figure shows the K-means clustering algorithm flow chart. The specific steps are as follows:

[0097] Step 1: Determine its objective function as minimizing the square error E, as shown in the following formula:

[0098]

[0099] Where μi is the mean vector, expressed as follows:

[0100]

[0101] In the formula, μ i It is cluster C i The cluster center (i.e., centroid) of is also the mean vector.

[0102] Step 2: Analyze the historical wind speed data of the wind farm, draw probability distribution diagrams for different seasons, and compare the differences in wind speed distribution in the four seasons.

[0103] Step 3: Cluster the wind speeds in four seasons using the Weka platform, and select the optimal number of clusters by comparing the squared error and the elbow method.

[0104] Step 4: Select the cluster where the predicted day is located to perform secondary extraction of similar days based on the Pirson coefficient, and select similar days with strong correlation coefficients within the cluster.

[0105] Specifically, the method of performing wind speed forecasting based on similar days according to similar day data and obtaining wind speed data on the forecast day includes:

[0106] Get the original wind speed sequence of similar days;

[0107] The complete empirical mode decomposition of adaptive white noise (CEEMDAN) is used to decompose the original wind speed sequence of similar days to obtain a decomposed wind speed sequence.

[0108] The empirical mode decomposition (VMD) is used to perform secondary decomposition on the primary decomposition wind speed sequence to obtain the secondary decomposition wind speed sequence.

[0109] The secondary decomposed wind speed sequence is input into the wind speed prediction model to obtain the predicted daily wind speed data; the wind speed prediction model is composed of the empirical mode decomposition (EMD) method and the echo state network (ESN) prediction algorithm.

[0110] Specifically, Figure 4 As shown in Figure 1, the method of using the complete empirical mode decomposition of adaptive white noise (CEEMDAN) to decompose the original wind speed series of similar days includes:

[0111] Step 1: First, add K white noises to the signal sequence x(t) to construct K decomposition sequences x i (t)(i=1,2,…,K), the expression is as follows:

[0112] x i (t) = x(t) + εδ i (t) (3)

[0113] Where ε is the white noise weight coefficient, δ i (t) is the white noise generated by the i-th processing.

[0114] Step 2: For the above sequence x to be decomposed i (t) (i=1,2,…,K) performs K times of empirical mode decomposition, and takes the average value of all the first modal components after solution IMF1(t) as the first modal component of CEEMDAN decomposition, which is expressed as follows:

[0115]

[0116] In the formula, For x i (t) The first modal component obtained after empirical mode decomposition.

[0117] Step 3: Solve the residual signal of the signal sequence x(t), and the expression is as follows:

[0118] r1(t)=x(t)-IMF1(t) (5)

[0119] r1(t) represents the residual signal after the first CEEMDAN decomposition.

[0120] Step 4: After adding specific noise to the j-th stage residual signal obtained after decomposition, continue to perform EMD decomposition. j The expression is as follows:

[0121]

[0122] r j (t) = r j-1 (t)-IMF j (t) (7)

[0123] In the formula, IMF j (t) is the jth IMF obtained by CEEMDAN decomposition; E j-1 (·) represents the j-1th IMF component after EMD decomposition of the sequence, ε j-1 represents the weight coefficient of noise added by CEEMDAN to the residual signal of the j-1th stage; r j (t) represents the residual signal of the jth stage.

[0124] If the residual signal becomes monotonic after a certain number of decompositions, and the decomposition stop condition is met, the algorithm ends.

[0125] Specifically, the method of using the empirical mode decomposition (VMD) to perform a secondary decomposition on the primary decomposition wind speed sequence to obtain the secondary decomposition wind speed sequence includes:

[0126] Step a: Initialization λ 1 and the maximum number of iterations N;

[0127] Step b: Update and ω k ;ω k The expressions of and λ are as follows:

[0128]

[0129]

[0130]

[0131] Step c: If satisfied The iteration is completed and the final output is and ω k If not, repeat steps ab until it is satisfied

[0132] Specifically, the method of inputting the secondary decomposed wind speed sequence into the wind speed prediction model to obtain the predicted daily wind speed data includes:

[0133] The secondary decomposition wind speed series is decomposed using the empirical mode decomposition method (EMD) to obtain the modal components.

[0134] The echo state network (ESN) is used to predict the predicted values ​​of all modal components, and the predicted values ​​are added together to obtain the predicted results of the wind speed data on the predicted day.

[0135] Specifically, the secondary decomposition wind speed sequence is decomposed using the empirical mode decomposition method EMD, such as Figure 3 As shown, the methods for obtaining the modal components include:

[0136] Step 1: Use the Empirical Mode Decomposition (EMD) method. First, find all extreme points, connect the local maximum points into the upper envelope using the cubic spline method, and obtain the lower envelope in the same way.

[0137] Step 2: Obtain the average curve of the upper and lower envelopes, and subtract the average curve from the original signal to obtain the suspected IMF curve, and determine whether the suspected IMF curve meets the IMF conditions. If it does, it is used as the first IMF component. If it does not meet the IMF curve, use the curve as the original data set, repeat the first two steps, and determine whether it meets the IMF component. If it still does not meet the requirements, repeat the above two steps again until it meets the requirements.

[0138] Step 3: Subtract the IMF signal from the original signal to obtain the new original signal. If the new original signal is a monotonic function, the process ends. If the new original signal is not a monotonic function, the process returns to the previous two steps.

[0139] Specifically, the method of predicting the predicted values ​​of all modal components using the echo state network ESN includes:

[0140] Determine the size of the reserve pool, scale the matrix so that the spectral radius is less than 1, and randomly generate input connections and output feedback weights;

[0141] The samples are loaded into the input and output in sequence, the state record reservoir state is updated, and the output connection weights are determined using linear regression;

[0142] The input vector is loaded into the input node and runs idle, and the predicted value is obtained from the output node.

[0143] Specifically, the method for predicting wind farm power based on similar day data and predicted day wind speed data includes:

[0144] The wind speed data of the forecast day is input into the GA-BP neural network to obtain the wind power every 15 minutes within 24 hours of the forecast day;

[0145] The training and prediction methods of GA-BP neural network are as follows Figure 5 As shown, including:

[0146] Obtain historical wind speed data, obtain historical wind speed data of similar days in different seasons based on the K-means clustering algorithm process, and obtain seasonal forecast day wind speed data based on the empirical mode decomposition (EMD) and echo state network (ESN) prediction algorithm;

[0147] Data preprocessing and data standardization are performed on seasonal similar day wind speed data and wind farm power data;

[0148] Construct GA-BP neural network;

[0149] Use seasonal similar day wind speed data and wind farm power data to train the GA-BP neural network, and obtain a trained GA-BP neural network;

[0150] Data preprocessing and data standardization are performed on seasonal forecast daily wind speed data;

[0151] The processed predicted daily wind speed data is input into the trained GA-BP neural network to obtain the wind farm power prediction data;

[0152] The wind farm power prediction data is denormalized to obtain the model prediction result, that is, the wind power every 15 minutes within 24 hours on the prediction day.

[0153] GA-BP neural network is a random search algorithm optimized by simulating the genetic evolution mechanism of survival of the fittest. GA-BP is divided into three parts, namely, determining the structure of BP network according to the function input and output, determining the number of parameters to be optimized by GA, and prediction of BP network. The basic structure of BP neural network includes the following four parts: determining the number of input layer nodes, selecting the number of hidden layer nodes, selecting the number of hidden layer layers, and determining the number of output layer nodes. The number of input layer nodes is generally determined according to the dimension of the input sample vector or the number of extracted features, the number of output layer nodes is determined according to the number of test classifications required, and the number of hidden layer layers is usually selected. Increasing the number of layers can further reduce the error and improve the accuracy, but it will make the network more complex. Back propagation learning with enough hidden layers can approximate any nonlinear function to arbitrary accuracy. Design the input, hidden layer, and output of the BP neural network, set the corresponding parameters, input the learning sample, and the BP neural network will train and learn by itself. After the threshold and weight are continuously changed, the learning condition is reached and the result is finally output.

[0154] The similar daily wind speed data and predicted daily wind speed data in the four seasonal clusters are input into the constructed GA-BP model to predict the wind power every 15 minutes within 24 hours a day. Currently, only deep neural networks with up to 10 layers are considered. When the number of hidden layers increases, even if the prediction accuracy may be higher, the convergence speed will be slower and the calculation time will be longer.

[0155] To improve the accuracy of model selection, Figure 1 As shown, the complete process of this embodiment includes the following steps:

[0156] Analyze and select Weka, RapidMiner, and Apache Mahout data mining platforms;

[0157] Choose a suitable similar day extraction platform;

[0158] Summarize the evaluation indicators of wind power prediction model.

[0159] A model combining K-means clustering algorithm and Pearson correlation coefficient method is constructed to extract seasonal similar days for modal decomposition and short-term power prediction of wind farms. The wind speeds of the four seasons are clustered, and the optimal number of clusters is selected according to the elbow method. The cluster where the prediction day is located is selected for secondary extraction of Pearson coefficient similar days to obtain similar days with strong correlation coefficients within the cluster.

[0160] The original wind speed series of similar days is decomposed by CEEMDAN to obtain several modal components and a residual component;

[0161] Since a single CEEMDAN decomposition can only reduce the non-stationarity of the modal components to a limited extent, the most complex component among the modal components is decomposed by VMD;

[0162] The echo state network (ESN) was used to predict all modal components, and the EMD-ESN model and the ESN model were compared to analyze the error of the prediction model. The predicted values ​​were added together to obtain the final wind power prediction result.

[0163] In order to improve the accuracy of wind power forecasting, a wind power forecasting combined model based on similar days and GA-BP deep neural network is proposed.

[0164] A multi-hidden layer deep neural network model based on GA-BP is established. By comparing the prediction errors corresponding to different hidden layer networks, the best deep neural network is selected to obtain the final prediction value of wind power.

[0165] The original wind speed series is decomposed once using the complete empirical mode decomposition of adaptive white noise (CEEMDAN);

[0166] Secondary decomposition of wind speed series using empirical mode decomposition (VMD);

[0167] A wind speed prediction model that combines the empirical mode decomposition (EMD) method with the echo state network (ESN) prediction algorithm;

[0168] Compare with EMD-ESN model and ESN model to analyze the error of prediction model;

[0169] A wind power forecasting combined model based on similar days and GA-BP deep neural network is used;

[0170] Establish a multi-hidden layer deep neural network model based on GA-BP;

[0171] Compare the prediction errors corresponding to different hidden layer networks and select the best deep neural network;

[0172] The final predicted value of wind power is obtained.

[0173] The present invention selects the most suitable clustering platform through comparative analysis of the data mining platforms Weka, Rapid Miner and Apache Mahout. Since Weka has the advantages of visualization and is more suitable for small-scale machine learning, it can improve efficiency in subsequent clustering applications. The Weka platform is selected as the clustering platform for subsequent similar day extraction.

[0174] The present invention analyzes the historical wind speed of the wind farm, draws probability distribution diagrams of different seasons, and compares the wind speed distribution of the four seasons. It is found that there is a large seasonal distribution difference in the wind speed of the four seasons.

[0175] Based on this factor, cluster analysis is performed to extract similar days by season, and K-means clustering is performed on the wind speed in four seasons through the Weka platform. numClusters is set to 2, 3, …, 10 for clustering operations. The sum of squared errors corresponding to different cluster numbers is selected as the cluster number evaluation criterion, and the corresponding elbow diagram is drawn to select the best cluster number. The cluster scatter point number is generated and the components of each cluster are recorded.

[0176] Select the forecast day of each season, find the cluster where the forecast day is located, perform secondary extraction of similar days with Pearson coefficients, record the Pearson coefficients between similar days, and select similar days with strong correlation coefficients within the cluster.

[0177] After the similar day extraction is completed, the wind speed prediction based on similar days is performed. The empirical mode decomposition (VMD) and the complete empirical mode decomposition of adaptive white noise (CEEMDAN) are summarized, and the modeling process is analyzed in detail. The CEEMDAN-VMD-ESN model is used to perform modal decomposition and short-term prediction of the short-term wind speed of the wind farm, and compared with the EMD-ESN model and the ESN model, and the error of the prediction model is analyzed;

[0178] The present invention finally predicts wind power based on similar days, improves the structure of GA-BP neural network, and proposes a wind power prediction model based on similar days and GA-BP deep neural network to predict the wind power of wind farms. The example analysis shows that the K-means-GA-BP-DNN-5 (when the number of hidden layers is 5) in autumn and the K-means-GA-BP-DNN-4 (when the number of hidden layers is 4) in spring, summer and winter can fit more stably, have higher prediction accuracy and better feasibility than the BP model, GRNN model and GA-BP traditional model.

[0179] Embodiment 2:

[0180] This embodiment provides a wind speed prediction device for a wind farm, comprising:

[0181] Acquisition module: used to acquire forecast day information; the forecast day information includes the forecast day date;

[0182] Similar day extraction module: used to determine the season of the forecast day according to the date of the forecast day, extract similar days according to the season and wind speed of the forecast day, and obtain similar day data; the similar day data includes similar day wind speed data;

[0183] Wind speed prediction module: used to predict wind speed based on similar days according to similar day data and obtain wind speed data for the predicted day;

[0184] Power prediction module: used to predict the power of the wind farm based on the similar day data and the predicted day wind speed data, and obtain the power of the wind farm.

[0185] The device of this embodiment can be used to implement the method described in the first embodiment.

[0186] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0187] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0188] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0189] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0190] The above are only preferred embodiments of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A wind farm power prediction method, characterized in that: The following steps are involved: Obtaining forecast day information; the forecast day information includes the forecast day date; Determine the season of the forecast day according to the date of the forecast day, extract similar days according to the season and wind speed of the forecast day, and obtain similar day data; the similar day data includes similar day wind speed data; According to the data of similar days, the wind speed forecast based on similar days is carried out to obtain the wind speed data of the forecast day; According to the similar day data and the predicted day wind speed data, the wind farm power is predicted to obtain the power of the wind farm; The methods for extracting similar days based on the season and wind speed of the forecast day include: Acquire historical wind speed data of the wind farm, and organize the historical wind speed data into four seasonal wind speed sets according to seasons; The seasonal wind speed sets are clustered by K-means using the Weka platform, and the sum of squared errors corresponding to different cluster numbers K are selected as the cluster number evaluation criterion; Draw the corresponding elbow diagram to get the most suitable clustering number K, use different colors to distinguish the clusters formed by different clustering numbers K and correspond to the number of similar days; According to the season of the forecast day, the cluster where the forecast day is located is selected to perform secondary extraction of similar days based on the Pearson coefficient, and similar days with strong correlation coefficients within the cluster are obtained; The method of performing K-means clustering includes the following steps: The objective function is determined to minimize the square error E, and the expression is as follows: where μ i is the mean vector, expressed as follows: In the formula, μ i It is cluster C i The cluster center (i.e., centroid) of is also the mean vector; Based on the data of similar days, the wind speed forecast based on similar days is carried out. The method of obtaining the wind speed data of the forecast day includes: Get the original wind speed sequence of similar days; The original wind speed sequence of similar days is decomposed once using complete empirical mode decomposition of adaptive white noise to obtain a decomposed wind speed sequence; The empirical mode decomposition is used to perform secondary decomposition on the primary decomposition wind speed sequence to obtain the secondary decomposition wind speed sequence; Processing the secondary decomposed wind speed sequence according to the empirical mode decomposition method and the echo state network prediction algorithm to obtain predicted daily wind speed data; The method of using complete empirical mode decomposition of adaptive white noise to decompose the original wind speed series of similar days includes: Add K times of white noise to the signal sequence x(t) to construct K times of decomposition sequence x i (t)(i=1,2,…,K), the expression is as follows: x i (t)=x(t)+εδ i (t) (3) Where ε is the white noise weight coefficient, δ i (t) is the white noise generated by the i-th processing; For the above sequence x to be decomposed i (t) (i=1,2,…,K) performs K times of empirical mode decomposition, and takes the average value of all the first modal components after solution IMF1(t) as the first modal component of CEEMDAN decomposition, which is expressed as follows: In the formula, For x i (t) The first modal component obtained after empirical mode decomposition; The residual signal of the signal sequence x(t) is solved and the expression is shown as follows: r1(t)=x(t)-IMF1(t) (5) r1(t) represents the residual signal after the first CEEMDAN decomposition; After adding specific noise to the j-th stage residual signal obtained after decomposition, continue to perform EMD decomposition; IMF j The expression is as follows: r j (t)=r j-1 (t)-IMF j (t) (7) In the formula, IMF j (t) is the jth IMF obtained by CEEMDAN decomposition; E j-1 (·) represents the j-1th IMF component after EMD decomposition of the sequence, ε j-1 represents the weight coefficient of noise added by CEEMDAN to the residual signal of the j-1th stage; r j (t) represents the residual signal of the jth stage; If the residual signal is monotonic when the decomposition reaches a certain number of times, the decomposition stop condition is met and the algorithm ends; The method of using empirical mode decomposition to perform secondary decomposition of the primary decomposition wind speed sequence to obtain the secondary decomposition wind speed sequence includes: Step a: Initialization λ 1 and the maximum number of iterations N; Step b: Update and ω k ;ω k The expressions of and λ are as follows: Step c: If satisfied The iteration is completed and the final output is and ω k If not, repeat steps ab until it is satisfied The method of processing the secondary decomposition wind speed sequence according to the empirical mode decomposition method and the echo state network prediction algorithm to obtain the predicted daily wind speed data includes: The secondary decomposition wind speed series is decomposed using the empirical mode decomposition method (EMD) to obtain the modal components. The predicted values ​​of all modal components are predicted using the echo state network (ESN), and the predicted values ​​are added together to obtain the predicted results of the wind speed data on the predicted day. The secondary decomposition wind speed series is decomposed using the empirical mode decomposition method EMD to obtain the modal components, including: Step A: Find all extreme points of the secondary decomposition wind speed sequence, connect the local maximum points into an upper envelope by using the cubic spline curve method, and connect the local minimum points into a lower envelope by using the cubic spline curve method; Step B: respectively obtain the upper and lower envelope average curves of the original wind speed sequence of each similar day, and subtract the average curve from the original signal to obtain the suspected IMF curve; Step C: Determine whether the suspected IMF curve meets the IMF conditions. If so, take the curve as the IMF component. If not, take the curve as the original data set and repeat steps AB until it meets the IMF component and obtains the IMF signal. Step D: Subtract the IMF signal from the original signal to obtain a new original signal. If the new original signal is a monotonic function, the process ends. If the new original signal is not a monotonic function, steps AB are repeated. Methods for predicting the predicted values ​​of all modal components using the echo state network ESN include: Determine the size of the reserve pool, scale the matrix so that the spectral radius is less than 1, and randomly generate input connections and output feedback weights; The samples are loaded into the input and output in sequence, the state record reservoir state is updated, and the output connection weights are determined using linear regression; The input vector is loaded into the input node and runs idle, and the predicted value is obtained from the output node.

2. The wind farm power prediction method according to claim 1, characterized in that: The similar day data include similar day wind speed data and wind farm power data; The method for predicting wind farm power according to the similar day data and the predicted day wind speed data includes: Perform data preprocessing and data standardization on similar day wind speed data and wind farm power data; Using similar day wind speed data and wind farm power data to train the GA-BP neural network, and obtaining a trained GA-BP neural network; Perform data preprocessing and data standardization on the forecast day wind speed data; The processed predicted daily wind speed data is input into the trained GA-BP neural network to obtain the wind farm power prediction data; The wind farm power prediction data is denormalized to obtain the model prediction result, that is, the wind power every 15 minutes within 24 hours on the prediction day.

3. A wind farm power prediction device for executing the method according to any one of claims 1 to 2, characterized in that: include: Acquisition module: used to acquire forecast day information; the forecast day information includes the forecast day date; Similar day extraction module: used to determine the season of the forecast day according to the date of the forecast day, extract similar days according to the season and wind speed of the forecast day, and obtain similar day data; the similar day data includes similar day wind speed data; Wind speed prediction module: used to predict wind speed based on similar days according to similar day data and obtain wind speed data for the predicted day; Power prediction module: used to predict the power of the wind farm based on the similar day data and the predicted day wind speed data, and obtain the power of the wind farm.

Images