A short-term wind speed correction method for wind farms based on cloud model and RBF neural network
By combining cloud models and RBF neural networks, the problems of wind speed uncertainty and weather forecast errors have been solved, achieving higher accuracy in short-term wind speed prediction for wind farms and meeting the stability requirements for wind farm operation and grid connection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ELECTRIC POWER RES INST OF STATE GRID ZHEJIANG ELECTRIC POWER COMAPNY
- Filing Date
- 2022-11-11
- Publication Date
- 2026-07-03
AI Technical Summary
Existing technologies fail to effectively account for wind speed uncertainties and weather forecast errors, resulting in insufficient accuracy in short-term wind speed prediction for wind farms, which affects the safety and stability of wind farm operation and grid connection.
By employing a cloud model-based and RBF neural network approach, a training set is constructed, a single-point prediction error cloud model is established, and an RBF neural network is trained. This approach, combined with wind speed uncertainty characteristics and weather forecast data, improves the accuracy of wind speed prediction.
It improves the accuracy and precision of short-term wind speed forecasting, meeting the requirements of wind power forecasting and power system stability.
Smart Images
Figure CN116599026B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electrical digital data processing technology for wind farms, and more specifically, to a method for short-term wind speed correction in wind farms. Background Technology
[0002] Wind speed forecasting is fundamental to wind power forecasting. Accurate wind speed prediction can effectively address the issues of large-scale wind farm operation and grid connection. Short-term wind speed forecasting, typically measured in hours or days, is a major focus of wind speed forecasting research. In recent years, numerical weather prediction (NWP) has continuously developed, establishing relatively mature weather forecasting systems. However, due to the inherent characteristics of NWP and limitations imposed by factors such as imperfect parameterization schemes, its accuracy is not high and its errors are relatively large. To further improve the accuracy of wind speed forecasting and enhance the safety and stability of large-scale wind power grid connection, it is essential to correct the short-term wind speeds in NWP.
[0003] For example, Chinese patent document CN104734175B describes an intelligent correction method for the wind speed-power curve of a wind turbine. This method acquires actual wind speed and power operating data of the wind turbine, establishes a Beansfield interval relationship between wind speed and power, obtains the wind speed-power polynomial, and corrects the coefficients of the polynomial to reflect the wind speed-power characteristics of the wind turbine. However, this existing technical solution only considers the measured wind speed value and does not account for the uncertainty of predicted wind speed. It cannot solve the forecasting errors caused by wind speed changes during periods of drastic weather variation in wind farms, does not fully analyze the characteristics of wind speed, and lacks consideration for wind speed uncertainty. Summary of the Invention
[0004] Given that existing technologies do not consider prediction errors, weather forecast wind speeds, and the uncertainty of wind speed, which affects the accuracy of wind speed prediction for wind farms, this invention provides a short-term wind speed correction method for wind farms based on cloud models and RBF neural networks. This method fully considers factors such as wind speed uncertainty, weather forecast wind speeds, and prediction errors. It fully integrates the fast, easy-to-implement, and globally optimal characteristics of radial basis function (RBF) neural networks with the advantages of cloud models in handling uncertainty problems, resulting in a more comprehensive and accurate short-term wind speed correction method, thereby improving the accuracy of short-term wind speed prediction.
[0005] The technical solution adopted in this invention is: a short-term wind speed correction method for wind farms based on cloud models and RBF neural networks, comprising the following steps:
[0006] S1: Construct the training set.
[0007] S2: Divide all samples in the training set by time, establish a single-point prediction error cloud model corresponding to the wind speed segment in each time period, and calculate the feature value corresponding to the single-point prediction error cloud model.
[0008] S3: Train the RBF neural network based on each sample and its corresponding feature value in the training set, and use the trained RBF neural network as a short-term wind speed correction model.
[0009] S4: Use the short-term wind speed correction model to correct the wind speed for the time period to be predicted for the wind farm.
[0010] By using the RBF neural network, the best approximation can be achieved globally, cleverly eliminating the drawbacks of local optima, and significantly improving training speed. The cloud model can better describe the uncertainty of wind speed, and can also improve the temporal resolution and relevance of training sample selection, which can meet the high requirements of the RBF neural network for training sample data selection. Combining the cloud model with the RBF neural network improves the accuracy of short-term wind speed correction to a certain extent, retains the RBF's good ability to track wind speed change trends, and fully incorporates the consideration of wind speed uncertainty, making the mapping relationship more comprehensive and the short-term wind speed correction results more accurate, better meeting the high requirements of subsequent wind power prediction and maintaining power system stability.
[0011] Furthermore, each sample in the training set in step S1 includes: the forecast wind speed data of the wind farm to be predicted in the numerical weather forecast at any time in the selected historical time period and the measured wind speed value of the wind farm to be predicted at the meteorological tower at that time.
[0012] By incorporating the forecast wind speed data from numerical weather predictions of the wind farm to be predicted at any point in a selected historical period into the training set samples, the training sample data becomes more targeted and accurate, resulting in more accurate short-term wind speed correction results.
[0013] Furthermore, the forecast wind speed data in the numerical weather forecast includes: forecast wind speed value, wind direction sine, and wind direction cosine.
[0014] Further, step S2 includes:
[0015] S21: When the wind speed is v at any time within the time period, select samples of the predicted wind speed segment in (v±a) m / s from that time period to form the corresponding wind speed segment sample set.
[0016] S22: Calculate the forecast wind speed error value of each sample selected in step S21, and use the forecast wind speed error value as the cloud droplet sample value of the sample in the single-point prediction error cloud model corresponding to the wind speed segment sample set.
[0017] xi =U i -u i (Equation 1)
[0018] In Equation 1, x i The predicted wind speed error value, U, for the i-th sample in this sample set. i Let u be the measured wind speed of the i-th sample in the sample set. i Let be the predicted wind speed value of the i-th sample in this sample set;
[0019] S23: Input all cloud droplet sample values obtained in step S22 into the reverse cloud generator to obtain the feature values of the single-point prediction error cloud model corresponding to the wind speed segment sample set.
[0020] S24: Input the feature values obtained in step s23 into the forward cloud generator to obtain virtual cloud droplets that conform to the distribution law of the feature values. All virtual cloud droplets constitute the single-point prediction error cloud model corresponding to the wind speed segment sample set.
[0021] S25: Adjust the coverage of the wind speed segment by using the forecast wind speed values for different time periods; repeat steps S21 to S24 to obtain the single-point prediction error cloud model and corresponding feature values for the samples in each time period.
[0022] Furthermore, the single-point prediction error cloud model includes:
[0023] The expected value E of the single-point prediction error cloud model x The calculation expression is as follows:
[0024]
[0025] In Equation 2, n represents the cloud droplet sample x in the sample set. i Quantity;
[0026] The entropy E of the single-point prediction error cloud model n The calculation expression is as follows:
[0027]
[0028] In Equation 3, n represents the cloud droplet sample x in the sample set. i Quantity;
[0029] The hyperentropy H of the single-point prediction error cloud model e The calculation expression is as follows:
[0030]
[0031]
[0032] In equations 4 and 5, S2 Let V be the variance of the cloud droplet samples in the sample set.
[0033] Further, step S3 includes:
[0034] S31: Normalize the predicted wind speed data, measured wind speed values, and corresponding feature values in each sample of the training set.
[0035]
[0036] In Equation 6, X i Let x be the normalized value of the i-th sample in any class of data in the training set, and let x be the original value of the i-th sample in that class of data. min x max These are the minimum and maximum values of this type of data in the training set, respectively.
[0037] S32: Train the RBF neural network using the training set;
[0038] The input of the RBF neural network is a time series consisting of the predicted wind speed data of each sample in the training set and the feature values of the single-point prediction error cloud model to which the sample belongs. The output of the RBF neural network is the measured wind speed value of each sample in the training set. After the RBF neural network is trained, the trained RBF neural network is used as a short-term wind speed correction model.
[0039] Further, step S4 includes:
[0040] S41: Obtain the forecast wind speed data of the numerical weather forecast for the wind farm to be predicted at each moment of the predicted time period;
[0041] S42: Re-establish the single-point prediction error cloud model. The method is the same as that used to establish the single-point prediction error cloud model for the training set. Divide all the forecast wind speed data within the time period to be predicted into time periods, establish the single-point prediction error cloud model corresponding to the wind speed in each time period, and calculate the feature value corresponding to the single-point prediction error cloud model.
[0042] S43: Combine the forecast wind speed data obtained in step S41 and the feature values obtained in step S42 into a time series, input the time series into the short-term wind speed correction model, and obtain the normalized wind speed correction results for each moment of the time period to be predicted.
[0043] S44: Perform inverse normalization on the wind speed correction result of step S43 to obtain the final wind speed correction result for each moment of the time period to be predicted.
[0044] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention establishes a single-point prediction error cloud model by using weather forecast wind speed data and measured wind speed values of the wind farm to be predicted as samples, trains an RBF neural network, and modifies the wind speed. This solves the problem that existing technologies lack consideration of wind speed uncertainty, which affects the accuracy of wind speed prediction for wind farms. Furthermore, it fully integrates the fast and easy-to-implement characteristics and global optimality of the RBF neural network with the advantages of cloud models in handling uncertainty problems, thereby improving the accuracy of short-term wind speed correction results and improving the accuracy of short-term wind speed prediction. Attached Figure Description
[0045] Figure 1 This is a flowchart of the short-term wind speed correction method of the present invention;
[0046] Figure 2 This is a cloud map of the prediction error cloud model 1 after the training set is divided in the embodiment;
[0047] Figure 3 This is a cloud map of the prediction error cloud model 2 after the training set is divided in the embodiment;
[0048] Figure 4 This is a cloud map of the prediction error cloud model three after the training set is divided in the embodiment;
[0049] Figure 5 This is a cloud map of the prediction error cloud model four after the training set is divided in the embodiment;
[0050] Figure 6 This is a cloud map of the prediction error cloud model five after the training set is divided in the embodiment;
[0051] Figure 7 This is the cloud map of prediction error cloud model six after the training set is divided in the embodiment;
[0052] Figure 8 This is a cloud map of the prediction error cloud model seven after the training set is divided in the embodiment.
[0053] Figure 9 This is a cloud map of prediction error cloud model eight after the training set is divided in the embodiment.
[0054] Figure 10 This is a cloud map of the prediction error cloud model nine after the training set is divided in the embodiment. Detailed Implementation
[0055] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0056] Example 1.
[0057] like Figure 1As shown, this embodiment 1 provides a short-term wind speed correction method for wind farms based on cloud models and RBF neural networks, including the following steps:
[0058] Step 1: Collect the forecast wind speed data from numerical weather forecasts for the wind farm to be predicted during a selected historical period, typically one month to one year. The forecast wind speed data includes the forecast wind speed value, wind direction sine, and wind direction cosine. Collect the measured wind speed value of the wind farm at the corresponding time of the historical period. Combine the forecast wind speed data and measured wind speed value at each time point to form a sample, and combine all samples into a training set.
[0059] In this embodiment, a training set was constructed using 8,000 samples of NWP forecast wind speed data and corresponding measured wind speed values from meteorological towers at 8,000 times between June 24, 2011 and July 28, 2011.
[0060] Step 2: Divide all samples in the training set by time, with each time period containing at least 1000 samples. Establish a single-point prediction error cloud model for the wind speed segment in each time period and calculate the corresponding feature value of the model.
[0061] The more specific steps are as follows:
[0062] Step 1: When the wind speed is v at any time during the time period, select samples with predicted wind speeds within the range of (v±a) m / s from the time period to form the corresponding wind speed segment sample set; where v and a can be randomly selected.
[0063] Step 2: Calculate the forecast wind speed error value of each sample selected in Step 1 using Equation (1), and use the error value as the cloud droplet sample value of the sample in the single-point prediction error cloud model corresponding to the wind speed segment sample set.
[0064] x i =U i -u i (Equation 1)
[0065] In Equation 1, x i The predicted wind speed error value, U, for the i-th sample in this sample set. i Let u be the measured wind speed of the i-th sample in the sample set. i Let be the predicted wind speed value of the i-th sample in this sample set.
[0066] Step 3: Input all cloud droplet sample values obtained in Step 2 into the reverse cloud generator to obtain the feature values of the single-point prediction error cloud model corresponding to the wind speed segment sample set, including: the expected value E of the single-point prediction error cloud model. x The calculation expression is as follows:
[0067]
[0068] In Equation 2, n represents the cloud droplet sample x in the sample set. i The number of values for n varies depending on the wind speed range.
[0069] The entropy E of the single-point prediction error cloud model n The calculation expression is as follows:
[0070]
[0071] The hyperentropy H of the single-point prediction error cloud model e The calculation expression is as follows:
[0072]
[0073]
[0074] In equations 4 and 5, S 2 Let V be the variance of the cloud droplet samples in the sample set.
[0075] Step 4, take the eigenvalues obtained in Step 3, i.e., E x E n H e Input the positive cloud generator to obtain n virtual cloud droplets that match the distribution pattern of the feature values. All virtual cloud droplets constitute the cloud map of the single-point prediction error cloud model corresponding to the wind speed segment sample set.
[0076] Step 5: By adjusting the coverage of the wind speed segments, the predicted wind speed values corresponding to all samples in the training set are divided into different wind speed segments. There should be no overlapping coverage of predicted wind speed data in adjacent wind speed segments. Repeat steps 1 to 4 to obtain the single-point prediction error cloud model and corresponding feature values corresponding to the sample set of each wind speed segment.
[0077] In this embodiment, the predicted wind speed of each sample in the training set is divided into 9 wind speed segments. The cloud map of the single-point prediction error cloud model corresponding to each wind speed segment is as follows: Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 , Figure 8 , Figure 9 and Figure 10 As shown, the cloud map reflects the error distribution between the predicted and actual wind speed values in the NWP (Non-Wave Power) forecast.
[0078] Step 3: Train the RBF neural network based on each sample and its corresponding feature value in the training set, and use the trained RBF neural network as a short-term wind speed correction model.
[0079] The specific steps are as follows:
[0080] Step 1: Normalize the predicted wind speed data (including predicted wind speed value, wind direction sine and cosine), measured wind speed value and corresponding feature value in each sample in the training set.
[0081]
[0082] In Equation 6, X i Let x be the normalized value of the i-th sample in any class of data in the training set, and let x be the original value of the i-th sample in that class of data. min x max These are the minimum and maximum values of this type of data in the training set, respectively.
[0083] Step 2: Train the RBF neural network using the training set; wherein the input of the RBF neural network is a time series consisting of the forecast wind speed data of each sample in the training set and the feature value of the single-point prediction error cloud model to which the sample belongs.
[0084] In this embodiment, the feature value includes E x E n H e The output of the RBF neural network is the measured wind speed value corresponding to each sample in the training set. Training is complete when the number of neurons in the RBF neural network increases to the upper limit, which is equal to the number of input samples. After the RBF neural network is trained, it is used as a short-term wind speed correction model.
[0085] Step four: Use a short-term wind speed correction model to correct wind speed. Specific steps include:
[0086] Step 1: Obtain the NWP forecast wind speed data for the wind farm to be predicted during the prediction period. Short-term predictions are generally four days or less.
[0087] Step 2: Rebuild the cloud model. The method is the same as that used to build the cloud model for the training set. Divide all the forecast wind speed data within the time period to be predicted into time segments, build a single-point prediction error cloud model for the wind speed in each time segment, and calculate the feature value corresponding to the single-point prediction error cloud model.
[0088] Step 3: Combine the forecast wind speed data obtained in Step 1 and the feature values obtained in Step 2 into a time series, and input the time series into the short-term wind speed correction model. The model outputs the normalized wind speed correction results for each moment of the time period to be predicted.
[0089] Step 4: Perform inverse normalization on the results of Step 3 to obtain the final wind speed correction results for each moment of the time period to be predicted.
[0090] In this embodiment, the error indices of short-term wind speed correction using the CM-RBF method of the present invention and the traditional RBF method are compared in Table 1:
[0091] Table 1. Comparison of model error indices in a specific embodiment of the present invention.
[0092]
[0093] Where SSE is the sum of squared errors, MSE is the mean squared error, MSPE is the mean squared percentage error, MAE is the mean absolute error, and RSME is the root mean square error.
[0094] As shown in Table 1, the present invention has a good tracking effect on wind speed changes, and the fluctuation pattern is roughly the same as the actual wind speed, which effectively improves the prediction accuracy of a single model.
[0095] As can be seen from the above embodiments, the wind farm short-term wind speed correction method based on cloud model and RBF neural network proposed in this invention can effectively improve the accuracy of wind speed prediction in wind farms, improve the accuracy of short-term wind speed correction results, and better meet the needs of subsequent wind power prediction and maintaining stable power operation.
Claims
1. A method for short-term wind speed correction in wind farms based on cloud models and RBF neural networks, characterized in that, Includes the following steps: S1: Construct the training set; S2: Divide all samples in the training set by time, establish a single-point prediction error cloud model corresponding to the wind speed segment within each time period, and calculate the feature values corresponding to the single-point prediction error cloud model; the feature values include the cloud droplet sample expectation E. x Entropy E n and hyperentropy H e ; Step S2 further includes: S21: When the wind speed is v at any moment within the time period, select a predicted wind speed segment from that time period. The samples form the corresponding wind speed segment sample set; S22: Calculate the forecast wind speed error value of each sample selected in step S21, and use the forecast wind speed error value as the cloud droplet sample value of the sample in the single-point prediction error cloud model corresponding to the wind speed segment sample set. S23: Input all cloud droplet sample values obtained in step S22 into the reverse cloud generator to obtain the feature values of the single-point prediction error cloud model corresponding to the wind speed segment sample set. S24: Input the feature values obtained in step S23 into the forward cloud generator to obtain virtual cloud droplets that conform to the distribution law of the feature values. All virtual cloud droplets constitute the cloud map of the single-point prediction error cloud model corresponding to the wind speed segment sample set. S3: Train the RBF neural network based on each sample and its corresponding feature value in the training set, and use the trained RBF neural network as a short-term wind speed correction model. S4: Use the short-term wind speed correction model to correct the wind speed for the time period to be predicted for the wind farm.
2. The method for short-term wind speed correction of wind farms based on cloud models and RBF neural networks according to claim 1, characterized in that, Each sample in the training set in step S1 includes: the forecast wind speed data of the wind farm to be predicted in the numerical weather forecast at any time in the selected historical time period and the measured wind speed value of the wind farm to be predicted at the time of the meteorological tower.
3. The method for short-term wind speed correction in wind farms based on cloud models and RBF neural networks according to claim 2, characterized in that, The forecast wind speed data in the numerical weather prediction includes: forecast wind speed value, wind direction sine, and wind direction cosine.
4. The method for short-term wind speed correction in wind farms based on cloud models and RBF neural networks according to claim 3, characterized in that, Step S2 further includes: The forecast wind speed error value for each sample is: (Equation 1) In Equation 1, X i Let U be the predicted wind speed error value for the i-th sample in this sample set. i Let u be the measured wind speed of the i-th sample in the sample set. i Let be the predicted wind speed value of the i-th sample in this sample set; S25: Adjust the coverage of the wind speed segment by using the forecast wind speed values for different time periods; repeat steps S21 to S24 to obtain the single-point prediction error cloud model and corresponding feature values for the samples in each time period.
5. A short-term wind speed correction method for wind farms based on a cloud model and an RBF neural network according to claim 2 or 4, characterized in that, The single-point prediction error cloud model includes: The expected value E of the single-point prediction error cloud model x The calculation expression is as follows: (Equation 2) In Equation 2, n represents the cloud droplet sample X in the sample set. i Quantity; The entropy E of the single-point prediction error cloud model n The calculation expression is as follows: (Equation 3) In Equation 3, n represents the cloud droplet sample X in the sample set. i quantity The hyperentropy H of the single-point prediction error cloud model e The calculation expression is as follows: (Equation 4) (Equation 5) In equations 4 and 5, Let V be the variance of the cloud droplet samples in the sample set.
6. The method for short-term wind speed correction in wind farms based on cloud models and RBF neural networks according to claim 2, characterized in that, Step S3 includes: S31: Normalize the predicted wind speed data, measured wind speed values, and corresponding feature values in each sample of the training set. (Equation 6) In Equation 6, X i Let X be the normalized value of the i-th sample in any class of data in the training set, and let X be the original value of the i-th sample in that class of data. min X max These are the minimum and maximum values of this type of data in the training set, respectively.
7. The method for short-term wind speed correction of wind farms based on cloud models and RBF neural networks according to claim 6, characterized in that, Step S3 further includes: S32: Train the RBF neural network using the training set; The input of the RBF neural network is a time series consisting of the predicted wind speed data of each sample in the training set and the feature values of the single-point prediction error cloud model to which the sample belongs. The output of the RBF neural network is the measured wind speed value of each sample in the training set. After the RBF neural network is trained, the trained RBF neural network is used as a short-term wind speed correction model.
8. The method for short-term wind speed correction of wind farms based on cloud models and RBF neural networks according to claim 7, characterized in that, Step S4 includes: S41: Obtain the forecast wind speed data of the numerical weather forecast for the wind farm to be predicted at each moment of the predicted time period; S42: Re-establish the single-point prediction error cloud model. The method is the same as that used to establish the single-point prediction error cloud model for the training set. Divide all the forecast wind speed data within the time period to be predicted into time periods, establish the single-point prediction error cloud model corresponding to the wind speed in each time period, and calculate the feature value corresponding to the single-point prediction error cloud model.
9. A short-term wind speed correction method for wind farms based on cloud models and RBF neural networks according to claim 8, characterized in that, Step S4 further includes: S43: Combine the forecast wind speed data obtained in step S41 and the feature values obtained in step S42 into a time series, input the time series into the short-term wind speed correction model, and obtain the normalized wind speed correction results for each moment of the time period to be predicted. S44: Perform inverse normalization on the wind speed correction result of step S43 to obtain the final wind speed correction result for each moment of the time period to be predicted.