A wind power ultra-short-term probability prediction method based on an improved bootstrap method

Abstract

The application discloses a kind of wind power ultra-short-term probability prediction method based on improved bootstrap method, comprising the following steps: step 1, wavelet threshold function is designed, wind power time series is denoised pretreatment, and wind power data after denoising is normalized;Step 2, the power data is divided into training set and test set, and long short-term memory network is trained;Step 3, the optimal hyperparameter of long short-term memory network is obtained by using improved particle swarm optimization algorithm, on this basis, the point prediction of wind power is realized;Step 4, the training set is resampled using improved bootstrap method, the prediction error is analyzed, the probability prediction interval is calculated combining point prediction model, and the ultra-short-term probability prediction of wind power is realized.The model proposed by the application is obviously superior to the prediction performance of traditional bootstrap model, and has high reliability, which is an effective improvement to traditional wind power prediction method.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation technology, and in particular to an ultra-short-term probability prediction method for wind power based on an improved bootstrap method. Background Technology

[0002] With the continuous depletion of fossil fuels worldwide, developing renewable energy has become a necessity. Wind power, as the most efficient renewable energy generation technology in modern power systems, has experienced rapid development globally in recent years. However, compared to traditional power generation, the significant intermittency of wind energy causes fluctuations in wind turbine power output, and the resulting uncertainty is substantial. Therefore, accurate and effective wind power forecasting methods are crucial for improving the predictability of wind farm output power.

[0003] Current improvements to wind power forecasting methods largely focus on deterministic forecasting, specifically improving point forecasting models. However, in actual grid regulation, random factors often exist, making point forecasting insufficient to reflect the uncertainty of wind power output. Probabilistic forecasting, on the other hand, allows us to determine the degree of uncertainty for each forecast, thus providing a more accurate assessment of wind power forecasting performance. Generally, probabilistic forecasting methods include parametric and non-parametric methods. Parametric methods require prior assumptions that the forecasting error follows a known distribution, such as Gaussian or Cauchy distributions, and perform well in ideal conditions. However, in reality, while parametric methods reflect the uncertainty of wind power forecasting to some extent, the forecasting error generally cannot perfectly match the known distribution, leading to underfitting. Conversely, non-parametric methods make no assumptions about the overall distribution, directly estimating the error distribution based on the forecasting error itself, resulting in good fitting and practicality.

[0004] Bootstrap nonparametric methods have the advantage of not requiring any assumptions about the probability distribution of data samples and are currently widely used in wind power forecasting. However, traditional bootstrap methods are limited to repeated sampling from the original sample and cannot obtain information beyond the observation points. Therefore, they are likely to produce bootstrap samples that are very similar to the original sample. This problem is particularly pronounced when the sample size is small. When we use such bootstrap samples to predict future observations, the prediction results often become biased because they may not reflect the true distribution. Therefore, it is necessary to improve traditional bootstrap methods by extending the resampling range beyond the observation points, making the distribution of bootstrap samples closer to the true distribution. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a wind power ultra-short-term probabilistic prediction method based on an improved bootstrap method, which uses the improved bootstrap method to construct a nonparametric probabilistic prediction model and improve the overall performance of probabilistic prediction.

[0006] To address the aforementioned technical problems, this invention provides a method for ultra-short-term probabilistic prediction of wind power based on an improved bootstrap method, comprising the following steps:

[0007] Step 1: Design a wavelet threshold function to perform noise reduction preprocessing on the wind power time series, and normalize the noise-reduced wind power data;

[0008] Step 2: Divide the power data into training and testing sets, and train the Long Short-Term Memory network;

[0009] Step 3: The optimal hyperparameters of the long short-term memory network are obtained by using an improved particle swarm optimization algorithm. Based on this, point prediction of wind power is realized.

[0010] Step 4: Resample the training set using an improved bootstrapping method, analyze the prediction error, and calculate the probability prediction interval using a point prediction model to achieve ultra-short-term probability prediction of wind power.

[0011] Preferably, in step 1, the wind power data includes both true values ​​and noise, expressed as follows:

[0012] P(t) = s(t) + n(t)

[0013] Where s(t) represents the actual wind power data; n(t) represents the noise interference; and P(t) represents the wind power data containing noise.

[0014] Preferably, in step 1, the designed new wavelet threshold function is:

[0015]

[0016] Among them, w, These are the original wavelet coefficients and the denoised wavelet coefficients, respectively; λ is the threshold, calculated using the following formula: N is the signal length; σ is the signal standard deviation, taken as σ = median(w) / 0.6745. The median function is used to return the median of a set of data. α is the adjustment factor, ranging from [0, +∞), and is generally taken as α = 1.

[0017] Preferably, in step 2, the wind power data used has a resolution of 15 min, and the ratio of training set to test set is 9:1.

[0018] Preferably, in step 3, the particle swarm optimization algorithm parameters are improved as follows:

[0019]

[0020]

[0021]

[0022] Where ω is the inertia weight; c1 and c2 are the individual learning factor and the group learning factor, respectively; ω max ω min To adjust the parameters, ω can generally be taken as... max =0.9,ω min =0.4; k is the current iteration number; T max This represents the maximum number of iterations.

[0023] Preferably, in step 3, the optimal hyperparameters of the Long Short-Term Memory network are obtained using an improved particle swarm optimization algorithm, specifically including the following steps:

[0024] Step 31: Initialize the particle population by setting parameters such as population size and maximum number of iterations.

[0025] Step 32: Train the Long Short-Term Memory network, and select the Mean Squared Error (MSE) as the fitness value of each particle in the algorithm. The calculation formula is as follows:

[0026]

[0027] Where, N t , y i These represent the total number of samples in the training set, the predicted value of the i-th sample, and the actual value of the i-th sample, respectively.

[0028] Step 33: Compare the particle fitness values ​​and update the optimal fitness of individual particles and the population.

[0029] Step 34: Update particle velocity and position.

[0030] Step 35: When the iteration termination condition is met, output the optimized hyperparameters, including the number of hidden layer neurons L1 and L2, the number of training iterations K, and the learning rate lr.

[0031] Preferably, in step 4, the training set is resampled using an improved bootstrapping method, and the prediction error is calculated to obtain the prediction interval. This specifically includes the following steps:

[0032] Step 41: Let the original training set D be of size n. n =(x1,x2,…,x n The training set is rearranged in ascending order as follows:

[0033]

[0034] Step 42: In the rearranged training set, for the minimum and maximum values, calculate their neighborhoods respectively:

[0035]

[0036]

[0037] Where β is the adjustment factor, ranging from [2, +∞), and it is generally possible to take β = 2.

[0038] Step 43, take Together with the original training set, they form a new sample set of size n+2, denoted as:

[0039]

[0040] Step 44, from Sampling with replacement is performed to obtain a new sample of size M, forming a bootstrap sample set Y1. This process is repeated B times to obtain B bootstrap sample sets Y1. B =(Y1,Y2,…,Y) B For the i-th sample set Y i Take one of the subsamples The sample estimate means and variance are calculated using the following formulas:

[0041]

[0042]

[0043] in, For the sample mean, σ *2 This represents the variance of the prediction error.

[0044] Step 45: Substitute the calculation result into the following formula to obtain the upper limit of the prediction interval for the sample at confidence level 1-α. and lower limit

[0045]

[0046]

[0047] in, Let z be the predicted value of the i-th sample in the test set. 1-α / 2 The quantile is the 1-α confidence level of the standard Gaussian distribution.

[0048] The beneficial effects of this invention are as follows: The wind power ultra-short-term probabilistic prediction method based on the improved bootstrap method proposed in this invention combines wavelet threshold denoising, improved particle swarm optimization, long short-term memory network and improved bootstrap method to obtain a new probabilistic prediction model; compared with the prediction model of the traditional bootstrap method, it has higher prediction performance and better comprehensive index of prediction interval, which is an effective improvement of the traditional wind power prediction model. Attached Figure Description

[0049] Figure 1 This is a flowchart of the process of the method of the present invention.

[0050] Figure 2 This is a graph showing the wind power probability prediction interval at a 90% confidence level based on January data from this invention.

[0051] Figure 3 This is a graph showing the wind power probability prediction interval at a 90% confidence level based on the May data of this invention. Detailed Implementation

[0052] like Figure 1 As shown, a method for ultra-short-term probabilistic prediction of wind power based on an improved bootstrapping method includes the following steps:

[0053] Step 1: Design a wavelet threshold function to perform noise reduction preprocessing on the wind power time series, and normalize the noise-reduced wind power data;

[0054] Step 2: Divide the power data into training and testing sets, and train the Long Short-Term Memory network;

[0055] Step 3: The optimal hyperparameters of the long short-term memory network are obtained by using an improved particle swarm optimization algorithm. Based on this, point prediction of wind power is realized.

[0056] Step 4: Resample the training set using an improved bootstrapping method, analyze the prediction error, and calculate the probability prediction interval using a point prediction model to achieve ultra-short-term probability prediction of wind power.

[0057] In step 1, the wind power data includes true values ​​and noise, which can be expressed as:

[0058] P(t) = s(t) + n(t)

[0059] Where s(t) represents the actual wind power data; n(t) represents the noise interference; and P(t) represents the wind power data containing noise.

[0060] The wavelet thresholding noise reduction process is shown below:

[0061] (1) Decomposition. The wavelet basis is determined to be db4, and the decomposition level is 3. The power signal is decomposed by dimensionality reduction.

[0062] (2) Threshold processing. Choosing a threshold function. Process the coefficients of each decomposed layer to obtain the denoised wavelet coefficients;

[0063] (3) Reconstruction. The power signal is reconstructed by inverse wavelet transform.

[0064] The wind power data used in step 2 has a resolution of 15 minutes, and the ratio of training set to test set is 9:1. The parameters of the Long Short-Term Memory network are set as follows: number of neurons L1 = 32, L2 = 64, number of training iterations K = 100, and learning rate lr = 0.005.

[0065] The hyperparameters for particle swarm optimization in step 3 are the number of neurons, the number of training iterations, and the learning rate. The velocity and position update equations for the i-th particle are as follows:

[0066]

[0067] x i (k+1)=x i (k)+v i (k+1)

[0068] Where k is the current iteration number; ω is the inertia weight; c1 and c2 are the individual learning factor and the group learning factor, respectively; v is the particle velocity; x is the particle position; r1 and r2 follow a uniform distribution on [0,1]; p i For the individual's optimal position; p g This is the optimal position for the group.

[0069] Step 4 involves constructing the prediction interval using an improved bootstrapping method, and includes the following steps:

[0070] (1) Let the original training set D be of size n. n =(x1,x2,…,x n The training set is rearranged in ascending order as follows:

[0071]

[0072] (2) In the rearranged training set, for the minimum and maximum values, their neighborhoods are calculated as follows:

[0073]

[0074]

[0075] Where β is the adjustment factor, ranging from [2, +∞), and it is generally possible to take β = 2.

[0076] (3) Take Together with the original training set, they form a new sample set of size n+2, denoted as:

[0077]

[0078] (4) From Sampling with replacement is performed to obtain a new sample of size M, forming a bootstrap sample set Y1. This process is repeated B times to obtain B bootstrap sample sets Y1. B =(Y1,Y2,…,Y) B For the i-th sample set Y i Take one of the subsamples The sample estimate means and variance are calculated using the following formulas:

[0079]

[0080]

[0081] in, For the sample mean, σ *2 This represents the variance of the prediction error.

[0082] (5) Substituting the calculation result into the following formula, we can obtain the upper limit of the prediction interval for the sample at confidence level 1-α. and lower limit

[0083]

[0084]

[0085] in, Let z be the predicted value of the i-th sample in the test set. 1-α / 2 The quantile is the 1-α confidence level of the standard Gaussian distribution.

[0086] To verify the effectiveness of the method of this invention, the following experiment was conducted: Based on the power data of a wind farm in a region of Northwest my country, the time period was selected as January and May 2021, the data resolution was 15 minutes, and the ratio of training set to test set was 9:1. Historical power and numerical weather forecast data were used after normalization. Meanwhile, Prediction Interval Coverage (PICP), Average Interval Width (MPIW), and Interval Score (IS) were used as evaluation indicators for probabilistic prediction, and the calculation formulas are as follows:

[0087]

[0088]

[0089] Where PICP represents the coverage of the true values, which measures the reliability of the predicted interval. N is the total number of samples in the test set; c i For coverage; y i The actual value; and Let $\begin{cases}$ be the lower and upper bounds of the prediction interval for the $i$-th point, respectively.

[0090]

[0091] MPIW represents the average width of the prediction interval, which measures the sharpness of the prediction interval; the smaller the value, the better the prediction performance.

[0092]

[0093] in, The interval score for the i-th point at a confidence level of 1-α; The prediction interval width for the i-th point at confidence level 1-α is calculated using the following formula: The expression for the interval score IS at a confidence level of 1-α is:

[0094]

[0095] This metric considers both reliability and sharpness, providing a comprehensive evaluation of the overall performance of probabilistic prediction. The interval score IS is negative; the closer it is to zero, the better the prediction quality.

[0096] To verify the effectiveness of the proposed wind power ultra-short-term probabilistic prediction model based on the improved bootstrapping method, several existing methods—Long Short-Term Memory Network combined with Bootstrapping (PLSTM-B), Long Short-Term Memory Network combined with Gaussian Distribution (PLSTM-G), Unoptimized Long Short-Term Memory Network combined with Bootstrapping (LSTM-B), and Unoptimized Long Short-Term Memory Network combined with Gaussian Distribution (LSTM-G)—were used as comparative reference models. The probabilistic prediction indices of different models are shown in Tables 1 and 2. The experimental results clearly demonstrate that the method proposed in this invention has the best prediction performance.

[0097] Table 1 Probability Forecasting Indicators for January

[0098]

[0099] Table 2 Probability Forecasting Indicators for May

[0100]

[0101]

[0102] Figure 2 , Figure 3 The forecast ranges for wind power output in January and May are shown at the 90% confidence level. Figure 2 , Figure 3 As can be seen, the prediction interval of the method of the present invention has good reliability and comprehensive performance, and is an effective probabilistic prediction method. Therefore, the present invention can realize ultra-short-term prediction of wind power and can be applied to practical engineering.

Claims

1. A method for ultra-short-term probabilistic prediction of wind power based on an improved bootstrapping method, characterized in that, Includes the following steps: Step 1: Design a wavelet threshold function to perform noise reduction preprocessing on the wind power time series, and normalize the noise-reduced wind power data; Step 2: Divide the power data into training and testing sets, and train the Long Short-Term Memory network; Step 3: The optimal hyperparameters of the Long Short-Term Memory (LSTM) network are obtained using an improved particle swarm optimization (PSO) algorithm. Based on this, point prediction of wind power is achieved. The improved PSO algorithm parameters are as follows: in, Inertial weight; , These are individual learning factors and group learning factors, respectively. , To adjust the parameters; This represents the current iteration number; This represents the maximum number of iterations. Step 4: Resample the training set using an improved bootstrapping method, analyze the prediction error, and calculate the probability prediction interval using a point prediction model to achieve ultra-short-term probability prediction of wind power; specifically, this includes the following steps: Step 41: Assume the original capacity is... training set The training set is rearranged in ascending order as follows: Step 42: In the rearranged training set, for the minimum and maximum values, calculate their neighborhoods respectively: in, The adjustment factor has a range of [value missing]. ; Step 43, take Together with the original training set, it forms a new set with a capacity of The sample is denoted as: Step 44, from Sampling with replacement is performed to obtain a new sample of size M, forming a bootstrap sample set. Repeat this process. Next, get A self-bootstrapping sample set For the first a sample set Take one of the subsamples The sample estimate means and variance are calculated using the following formulas: in, Estimate the mean of the sample. The variance of the prediction error; Step 45: Substitute the calculation result into the following formula to obtain the sample's confidence level. Upper limit of the prediction interval and lower limit : in, For the test set The predicted value for each sample, For standard Gaussian distribution confidence quantiles.

2. The wind power ultra-short-term probabilistic prediction method based on the improved bootstrapping method as described in claim 1, characterized in that, In step 1, the wind power data includes both true values ​​and noise, represented as follows: in, This is real wind power data; For noise interference; This is wind power data that includes noise.

3. The wind power ultra-short-term probabilistic prediction method based on the improved bootstrapping method as described in claim 1, characterized in that, In step 1, the designed wavelet threshold function is: in, , These are the original wavelet coefficients and the denoised wavelet coefficients, respectively. The threshold is calculated using the following formula: , The signal length; Let be the standard deviation of the signal. , The function is used to return the median of a set of data. The adjustment factor has a range of [value missing]. .

4. The wind power ultra-short-term probabilistic prediction method based on the improved bootstrapping method as described in claim 1, characterized in that, In step 2, the wind power data used has a resolution of 15 minutes, and the ratio of training set to test set is 9:

1.

5. The wind power ultra-short-term probabilistic prediction method based on the improved bootstrapping method as described in claim 1, characterized in that, Step 3 involves using an improved particle swarm optimization algorithm to obtain the optimal hyperparameters of the Long Short-Term Memory (LSTM) network. This includes the following steps: Step 31: Initialize the particle population, setting the population size and maximum number of iterations; Step 32: Train the Long Short-Term Memory network and select mean squared error. The fitness value of each particle in the algorithm is calculated using the following formula: in, , , These represent the total number of training set samples and the number of samples in the training set, respectively. The predicted value of the first sample and the predicted value of the second sample The true value of each sample; Step 33: Compare particle fitness values ​​and update the optimal fitness of individual particles and the population. Step 34: Update particle velocity and position; Step 35: When the iteration termination condition is met, output the optimized hyperparameters, including the number of hidden layer neurons. , Number of training sessions With learning rate .

Images