A method for predicting the output range of wind power

Abstract

A method for predicting the wind power output interval first takes the historical wind power output data as training samples and inputs them into an LSTM model to obtain the simulated data of the historical wind power output and the predicted data of the future wind power output points. Then, the K-means clustering algorithm is used to segment the simulated data of the historical wind power output and its relative error. Subsequently, three distribution functions, namely Gaussian, beta, and Weibull, are respectively selected for each wind power output segment to fit the relative error, and the relative error fitting function for each wind power output segment is obtained. The fitting function with the best performance is selected from them to determine the confidence interval of the relative error of the wind power output. Finally, the wind power output prediction interval is obtained by combining the predicted data of the future wind power output and the confidence interval of the relative error of the wind power output. This method effectively improves the quality of the wind power output interval prediction.

Description

Technical Field

[0001] The present invention belongs to the field of new energy output prediction, and particularly relates to a method for predicting wind power output interval. Background Art

[0002] As an important technology in the field of power system optimal dispatching, wind power output prediction technology plays an important role in the high-proportion new energy grid connection and consumption and the safe and stable operation of the power system. For wind power output prediction, the most concerned index is the model prediction accuracy, which depends on many factors, such as the uncertainty of wind power output input data, model structure, and model parameters. Wind power output interval prediction can give the range or probability of possible deviation from the actual output, and can be used to evaluate the volatility of the deterministic single-point prediction results of wind power output.

[0003] From the existing research results, short-term wind power output interval prediction generally includes three categories: physical mechanism prediction models, statistical prediction models, and machine learning prediction models. In physical mechanism prediction models, it is generally based on numerical weather prediction models to predict the wind power output interval according to meteorological conditions and wind speed laws. The randomness of short-term wind power output undoubtedly increases the difficulty of wind power output interval prediction. Statistical prediction models aim to statistically infer the future wind power output change trend based on the relationship between historical sample data of wind power output. The methods applied to the wind power output prediction interval mainly include Bayesian, mean-variance, Markov chain, and regression analysis methods. However, these methods also have some deficiencies: for Bayesian and Markov chain methods, the calculation amount is large, and the obtained wind power output prediction interval is insufficient; the empirical coverage probability of the mean-variance estimation method is low; although the regression analysis method is relatively simple in principle implementation, the prediction result is not very satisfactory. Therefore, in recent years, there is also the LUBE method that directly determines the upper and lower bounds of the wind power output prediction interval using neural networks or machine learning models. This method usually takes support vector machine models, extreme learning machines, etc. as the basis, and uses indicators such as coverage width, average bandwidth, interval coverage rate, and cumulative bandwidth deviation as the cost function of the prediction model for single-objective or multi-objective iterative optimization, so as to establish a non-parametric prediction interval for wind power output. Although the LUBE method has been widely used, it also has problems such as being unable to handle the internal fluctuation characteristics of time series and capturing the periodicity of wind power output sequences, resulting in poor quality of the obtained wind power output prediction interval. Summary of the Invention

[0004] Based on the above purpose, the present invention provides a method for predicting wind power output interval that can effectively improve the prediction quality.

[0005] To achieve the above purpose, the present invention provides the following technical solutions:

[0006] A method for predicting wind power output interval successively includes the following steps:

[0007] Step A: Use the historical wind power output data as training samples and input them into the LSTM model to obtain the simulated historical wind power output data and the predicted future wind power output data;

[0008] Step B: Use the K-means clustering algorithm to segment the simulated historical wind power output data and its relative error;

[0009] Step C: Fit the relative error for each wind power output segment respectively to obtain the relative error fitting function for each wind power output segment, and then determine its confidence interval of the relative error of wind power output;

[0010] Step D: Combine the predicted future wind power output data and the confidence interval of the relative error of wind power output to obtain the prediction interval of wind power output.

[0011] The said Step B includes the following steps in sequence:

[0012] Step B1: Determine the clustering sample set Spw:

[0013] Spw = {Spw1, Spw2,..., Spw i ,..., Spw nwt}

[0014]

[0015]

[0016] In the above formula, Spw i is the i-th clustering sample, nwt is the total number of clustering samples, is the simulated value of the historical wind power output of the i-th clustering sample, Rew i is the relative error of the wind power output of the i-th clustering sample, w i is the value of the historical wind power output corresponding to the i-th clustering sample, pw e is the rated power of the wind power;

[0017] Step B2: First determine the optimal number of clusters Kw, and then randomly select Kw individuals from the clustering sample set as the initial clustering centers;

[0018] Step B3: Calculate the Euclidean distance from each remaining individual in the clustering sample set to each initial clustering center, and classify it into the category of the nearest centroid;

[0019] Step B4: Determine the new clustering centers for each category;

[0020] Step B5: Repeat Steps B3 - B4 in a loop until the difference between the new clustering center and the clustering center of the previous loop is less than or equal to the threshold. At this time, the simulated historical wind power output data and its relative error are divided into Kw segments.

[0021] In step B2, the Euclidean distance is calculated using the following formula:

[0022]

[0023] In the above formula, is the Euclidean distance from the i-th remaining individual to the k-th cluster center, is the simulated wind power output value of the k-th cluster center, is the relative error of the k-th cluster center;

[0024] In step B3, the new cluster centers of each category are calculated using the following formula:

[0025]

[0026] In the above formula, Zw′ k is the new cluster center of the k-th category, N k is the number of individuals in the k-th category, Yw k is the sample set of the k-th category.

[0027] Step C sequentially includes the following steps:

[0028] Step C1: For each wind power output segment, use the Gaussian, beta, and Weibull distribution functions to fit the relative error respectively to obtain three relative error fitting functions;

[0029] Step C2: Determine the wind power output confidence intervals of the three relative error fitting functions respectively;

[0030] Step C3: Select the wind power output confidence interval corresponding to the optimal relative error fitting function of each wind power output segment as its wind power output confidence interval, and determine the wind power output error confidence interval of each wind power output segment.

[0031] In step C2, the wind power output confidence intervals of the three relative error fitting functions are calculated using the following formula:

[0032]

[0033] In the above formula, is the wind power output confidence interval of the i-th sample in the k-th output segment with a confidence level of α, are the upper and lower limits of the confidence interval respectively, is the simulated historical wind power output value of the i-th sample in the k-th output segment, is the inverse function of the relative error fitting function at a confidence level of α, pw e is the rated power of the wind power;

[0034] In step C3, the confidence interval of the wind power output error for each wind power output segment is calculated using the following formula:

[0035]

[0036]

[0037] In the above formula, is the confidence interval of the wind power output error for the k-th output segment, are the upper and lower limits of the confidence interval of the wind power output error, respectively.

[0038] In step C3, the optimal relative error fitting function is determined by evaluating the following indicators:

[0039]

[0040]

[0041] In the above formula, PICP is the interval coverage rate index, and the larger the value of this index, the better. N k is the number of samples in the k-th output segment, c i is the state variable of the historical wind power output value corresponding to the i-th sample. If the historical wind power output value corresponding to the i-th sample is within the prediction confidence interval with a confidence level of α, c i = 1, otherwise c i = 0, is the interval average bandwidth index, and the smaller the value of this index, the better.

[0042] In step D, the wind power output prediction interval is calculated using the following formula:

[0043]

[0044] In the above formula, is the predicted value of the future wind power output corresponding to the prediction interval at a confidence level of α, are respectively the upper and lower limits of the confidence interval of the wind power output error for the k-th output segment to which it belongs.

[0045] In step A, the LSTM model is the LSTM-GWO model optimized by using the grey wolf algorithm for updating the number of hidden layers HN and the learning rate α of the weight coefficient.

[0046] The construction of the LSTM-GWO model includes the following steps in sequence:

[0047] Step A1: Randomly generate multiple initial grey wolf individuals;

[0048] Step A2: Construct the LSTM model corresponding to each initial grey wolf individual, where the grey wolf individual is a parameter combination (HN, α).

[0049] Step A3: Use the constructed LSTM model to train the wind power output training data XT and output the training results:

[0050]

[0051] In the above formula, M is the number of training samples, n is the step length of the sample, and xt n+1 is the (n + 1)-th training sample data;

[0052] Step A4: Taking the root mean square error of the wind power output training data set as the optimization criterion, calculate the fitness value of the grey wolf individual based on the training results, and successively select the three grey wolf individuals with the best fitness values in the wolf pack as the α, β, and δ leader wolves:

[0053]

[0054] In the above formula, fit i (gen) is the fitness value of the i-th grey wolf individual x i (gen) in the gen-th iteration, is the training result corresponding to the (n + 1)-th training data;

[0055] Step A5: Repeat steps A2 - A4 to determine the fitness values of the α, β, and δ wolves in the next iteration, and compare them one by one with the fitness values of the α, β, and δ wolves in the previous iteration, and replace the individual with a poor fitness value with an individual with a better fitness value, so as to realize the update of the α, β, and δ wolves;

[0056] Step A6: Loop and repeat step A5 until the iteration is completed, and take the α wolf in the final generation as the optimal individual, that is, the optimal parameter combination.

[0057] Compared with the prior art, the beneficial effects of the present invention are:

[0058] 1. A method for predicting the wind power output interval of the present invention first uses the historical wind power output data as training samples and inputs them into the LSTM model to obtain the simulated data of the historical wind power output and the predicted data of the future wind power output. Then, the K-means clustering algorithm is used to segment the simulated data of the historical wind power output and its relative error. Subsequently, the relative error fitting is performed on each wind power output segment respectively to obtain the relative error fitting function of each wind power output segment, and further determine the confidence interval of the relative error of the wind power output. Finally, the predicted interval of the wind power output is obtained by combining the predicted data of the future wind power output and the confidence interval of the relative error of the wind power output. This method takes the error of the wind power sample data as the core, performs K-means clustering through the relationship between the error and the corresponding wind power output, can eliminate the influence of the wind power output fluctuation on the error distribution to a certain extent, thereby improving the fitting accuracy of the relative error distribution of the training samples, and further improving the quality of the wind power output interval prediction. Therefore, the present invention effectively improves the quality of the wind power output interval prediction.

[0059] 2. The LSTM model adopted by the method for predicting the wind power output interval of the present invention is the LSTM-GWO model optimized by using the grey wolf algorithm for updating the number of hidden layers HN and the weight coefficient update learning rate α. This design takes advantage of the respective advantages of the LSTM model and the GWO algorithm, and proposes a method for predicting the wind power output point by organically integrating the two, which can improve the prediction accuracy of the model and avoid the influence of artificially setting the model parameters on the prediction accuracy. Therefore, the present invention improves the prediction accuracy of the LSTM model. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is the flow chart of the present invention.

[0061] Figure 2 is the K-means clustering result of Example 1.

[0062] Figure 3 is the schematic diagram of the relative error fitting of the first wind power output segment in Example 1. DETAILED DESCRIPTION OF THE INVENTION

[0063] The present invention will be further described below in conjunction with the drawings and the specific embodiments.

[0064] The present invention proposes a method for predicting the wind power output interval, aiming to solve the problems of adaptive optimization of model parameters and considering the influence of different wind power output prediction error distributions and grades on the interval prediction quality. Among them, for the adaptive optimization of model parameters, the present invention uses the GWO algorithm to automatically search for the optimal values of HN and α of the LSTM model, so as to obtain the LSTM-GWO prediction model; in the LSTM-GWO model, the individuals of grey wolves and wolf packs are composed of parameters (HN, α). The root mean square error (RMSE) of the wind power output training sample set is used as the fitness of the grey wolf individuals, and the population is updated, and finally the optimal parameters are found, and thus the optimal LSTM-GWO wind power output point prediction model is obtained. For the influence of different wind power output prediction error distributions and grades on the interval prediction quality, the present invention proposes to segment the wind power output by using the K-means clustering algorithm, and then fit the errors corresponding to the predicted values of the segmented wind power output to obtain the error distribution function of each segmented output segment, and then the prediction interval of the wind power output can be obtained.

[0065] Embodiment 1:

[0066] See Figure 1 、 Figure 2 A method for predicting the wind power output interval is carried out in the following steps in sequence:

[0067] 1. Randomly generate multiple initial grey wolf individuals;

[0068] 2. Construct an LSTM model corresponding to each initial grey wolf individual, where the grey wolf individual is a parameter combination (HN, α);

[0069] 3. Use the constructed LSTM model to train the wind power output training data XT and output the training result:

[0070]

[0071] In the above formula, M is the number of training samples, n is the time step of the sample, and xt n+1 The (n + 1)-th training sample data;

[0072] 4. Based on the root mean square error of the wind power output training data set as the optimization criterion, calculate the fitness value of the grey wolf individuals based on the training result, and sequentially select the three grey wolf individuals with the best fitness values in the wolf pack as the α, β, and δ leaders:

[0073]

[0074] In the above formula, fit i (gen) is the fitness value of the i-th grey wolf individual x i (gen) in the gen-th iteration, is the training result corresponding to the (n + 1)-th training data;

[0075] 5. Repeat steps 2 - 4 to determine the fitness values of the α, β, and δ wolves for the next iteration, and compare them one by one with the fitness values of the α, β, and δ wolves in the previous iteration. Replace the individuals with poorer fitness values with those with better fitness values to achieve the update of the α, β, and δ wolves;

[0076] 6. Repeat step 5 in a loop until the iteration is completed, and take the α wolf in the final generation as the optimal individual, that is, the optimal parameter combination, so as to obtain the LSTM - GWO model;

[0077] 7. Input the historical wind power output data as training samples into the LSTM - GWO model to obtain the simulated data of the historical wind power output and the predicted data of the future wind power output;

[0078] 8. Determine the clustering sample set Spw:

[0079] Spw = {Spw1, Spw2,..., Spw i ,..., Spw nwt}

[0080]

[0081]

[0082] In the above formula, Spw i is the i-th clustering sample, nwt is the total number of clustering samples, is the simulated value of the historical wind power output of the i-th clustering sample, Rew i is the relative error of the wind power output of the i-th clustering sample, w i is the historical wind power output value corresponding to the i-th clustering sample, pw e is the rated power of the wind power;

[0083] 9. First, determine the optimal number of clusters Kw = 6 according to the AIC criterion, and then randomly select 6 individuals from the clustering sample set as the initial clustering centers;

[0084] 10. Calculate the Euclidean distance from each remaining individual in the clustering sample set to each initial clustering center, and assign it to the category of the nearest centroid:

[0085]

[0086] In the above formula, is the Euclidean distance from the i-th remaining individual to the k-th clustering center, is the simulated value of the wind power output of the k-th clustering center, is the relative error of the k-th clustering center;

[0087] 11. Determine the new cluster centers of each category using the following formula:

[0088]

[0089] In the above formula, Zw′ k is the new cluster center of the k-th category, N k is the number of individuals in the k-th category, and Yw k is the sample set of the k-th category;

[0090] 12. Repeat steps 10 - 11 in a loop until the difference between the new cluster center and the cluster center of the previous loop is less than or equal to the threshold. At this time, the historical wind power output simulation data and its relative error are divided into 6 segments, as Figure 2 shown;

[0091] 13. For each wind power output segment, use the Gaussian, beta, and Weibull distribution functions to fit the relative error respectively to obtain three relative error fitting functions. Among them, the relative error fitting of the first wind power output segment is as Figure 3 shown;

[0092] 14. Use the following formula to determine the wind power output confidence intervals of the three relative error fitting functions respectively:

[0093]

[0094] In the above formula, is the wind power output confidence interval of the i-th sample in the k-th output segment with a confidence level of α, are the upper and lower limits of the confidence interval respectively, is the simulated value of the historical wind power output of the i-th sample in the k-th output segment, is the inverse function of the relative error fitting function at a confidence level of α, and pw e is the rated power of the wind power;

[0095] 15. Evaluate and determine the optimal relative error fitting function for each wind power output segment through the following indicators:

[0096]

[0097]

[0098] In the above formula, PICP is the interval coverage rate index, and the larger the value of this index, the better. N k is the number of samples in the k-th output segment, c i is the state variable of the wind power historical output value corresponding to the i-th sample. If the wind power historical output value corresponding to the i-th sample is within the prediction confidence interval with a confidence level of α, c i = 1, otherwise ci = 0, is the interval average bandwidth index, and the smaller the index value, the better;

[0099] 16. Select the wind power output confidence interval corresponding to the optimal relative error fitting function of each wind power output segment as its wind power output confidence interval, and use the following formula to determine the wind power output error confidence interval of each wind power output segment:

[0100]

[0101]

[0102] In the above formula, is the wind power output error confidence interval of the k-th output segment, are the upper and lower limits of the wind power output error confidence interval respectively;

[0103] 17. Use the following formula to calculate the wind power output prediction interval:

[0104]

[0105] In the above formula, is the predicted value of the future wind power output is the prediction interval corresponding to the confidence level of α, are respectively the upper and lower limits of the wind power output error confidence interval of the k-th output segment to which it belongs.

Claims

1. A method for predicting the wind power output interval, characterized in that: The method sequentially includes the following steps: Step A: Input the historical wind power output data as training samples into the LSTM model to obtain the historical wind power output simulation data and the predicted future wind power output data; Step B: Use the K-means clustering algorithm to segment the historical wind power output simulation data and its relative error; Step C: Fit the relative error for each wind power output segment respectively to obtain the relative error fitting function for each wind power output segment, and then determine its wind power output relative error confidence interval, which sequentially includes the following steps: Step C1: For each wind power output segment, use the Gaussian, beta, and Weibull distribution functions to fit the relative error respectively to obtain three relative error fitting functions; Step C2: Determine the wind power output confidence intervals of the three relative error fitting functions respectively; Step C3: Select the wind power output confidence interval corresponding to the optimal relative error fitting function for each wind power output segment as its wind power output confidence interval, and determine the wind power output relative error confidence interval for each wind power output segment, where the optimal relative error fitting function is determined by evaluating the following indicators: In the above formula, PICP is the interval coverage rate index, and the larger the value of this index, the better. N k is the number of samples in the k-th output segment, c i is the state variable of the historical wind power output value corresponding to the i-th sample. If the historical wind power output value corresponding to the i-th sample is within the prediction confidence interval with a confidence level of α, c i = 1; otherwise c i = 0. is the interval average bandwidth index, and the smaller the value of this index, the better. Step D: Combine the predicted future wind power output data and the wind power output relative error confidence interval to obtain the wind power output prediction interval.

2. A method for predicting the wind power output interval according to claim 1, characterized in that: The step B sequentially includes the following steps: Step B1: Determine the clustering sample set Spw: Spw = {Spw1, Spw2, …, Spw i , …, Spw nwt} In the above formula, Spw i is the i-th clustering sample, and nwt is the total number of clustering samples. is the simulated value of the historical wind power output of the i-th clustering sample, and Rew i is the relative error of the wind power output of the i-th clustering sample, and w i is the historical wind power output value corresponding to the i-th clustering sample, and pw e is the rated power of the wind power. Step B2: First determine the optimal number of clusters Kw, and then randomly select Kw individuals from the clustering sample set as the initial cluster centers; Step B3: Calculate the Euclidean distances from the remaining individuals in the clustering sample set to each initial cluster center, and assign them to the category of the nearest centroid; Step B4: Determine the new cluster centers for each category; Step B5: Repeat steps B3 - B4 in a loop until the difference between the new cluster center and the cluster center of the previous loop is less than or equal to the threshold. At this time, the historical wind power output simulation data and its relative error are divided into Kw segments.

3. A method for predicting the wind power output interval according to claim 2, characterized in that: In step B2, the Euclidean distance is calculated using the following formula: In the above formula, is the Euclidean distance from the i-th remaining individual to the k-th cluster center, is the simulated wind power output value of the k-th cluster center, is the relative error of the k-th cluster center; In step B3, the new cluster centers for each category are calculated using the following formula: In the above formula, Zw′ k is the new clustering center of the k-th category, N k is the number of individuals in the k-th category, and Yw k is the sample set of the k-th category.

4. A method for predicting the wind power output interval according to claim 1, characterized in that: In step C2, the wind power output confidence intervals of the three relative error fitting functions are calculated using the following formula: In the above formula, is the confidence interval of wind power output for the i-th sample in the k-th output segment with a confidence level of α, are the upper and lower limits of the confidence interval, respectively, is the simulated value of the historical wind power output of the i-th sample in the k-th output segment, is the inverse function of the relative error fitting function at a confidence level of α, pw e is the rated power of wind power; In step C3, the wind power output relative error confidence intervals for each wind power output segment are calculated using the following formula: In the above formula, is the confidence interval of the relative error of wind power output in the k-th output segment, are the upper and lower limits of the confidence interval of the relative error of wind power output, respectively.

5. A method for predicting the wind power output interval according to claim 1, characterized in that: In step D, the wind power output prediction interval is calculated using the following formula: In the above formula, is the predicted value of the future wind power output The prediction interval corresponding to the confidence level of α, are respectively The upper and lower limits of the confidence interval of the relative error of wind power output in the k-th output segment to which it belongs.

6. A wind power output interval prediction method according to claim 1, characterized in that: In step A, the LSTM model is an LSTM-GWO model optimized by using the grey wolf optimization algorithm for updating the number of hidden layers HN and the learning rate α of the weight coefficient.

7. A method for predicting the wind power output interval according to claim 6, characterized in that: The construction of the LSTM-GWO model successively includes the following steps: Step A1: Randomly generate multiple initial gray wolf individuals; Step A2: Construct the LSTM model corresponding to each initial gray wolf individual, where the gray wolf individual is a parameter combination (HN, α); Step A3: Use the constructed LSTM model to train the wind power output training data XT and output the training results: In the above formula, M is the number of training samples, n is the step size of the samples, and xt n+1 is the data of the (n + 1)-th training sample; Step A4: Taking the root mean square error of the wind power output training data set as the optimization criterion, calculate the fitness value of the gray wolf individuals based on the training results, and successively select the three gray wolf individuals with the best fitness values in the wolf pack as the α, β, and δ leader wolves: In the above formula, fit i (gen) is the fitness value of the i-th gray wolf individual x i (gen) in the gen-th iteration, is the training result corresponding to the (n + 1)-th training data; Step A5: Repeat steps A2 - A4 to determine the fitness values of the α, β, and δ wolves for the next iteration, and compare them one by one with the fitness values of the α, β, and δ wolves in the previous iteration, and replace the individuals with poor fitness values with the individuals with better fitness values, so as to realize the update of the α, β, and δ wolves; Step A6: Loop and repeat step A5 until the iteration is completed, and take the α wolf in the final generation as the optimal individual, that is, the optimal parameter combination.

Images