A wind speed prediction method based on error compensation
By combining ARMA and SVR with ELM error compensation, the problem of unsatisfactory wind speed prediction results during periods of drastic wind speed changes is solved, achieving higher accuracy wind speed prediction, especially improving stability and accuracy during periods of large wind speed fluctuations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- QINGDAO UNIV OF TECH
- Filing Date
- 2023-02-23
- Publication Date
- 2026-06-30
Smart Images
Figure CN116341717B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wind power prediction technology and can be used to predict wind speed time series data of a single variable. In particular, it relates to a wind speed prediction method based on ELM error compensation, ARMA and SVR. Background Technology
[0002] Since the beginning of the 21st century, human demand for energy has become increasingly strong. However, the combustion of fossil fuels such as coal, oil, and natural gas still dominates the world's energy structure. But the combustion of fossil fuels inevitably involves the emission of greenhouse gases and harmful gases, causing serious environmental problems. Moreover, due to their non-renewable nature, the Earth's reserves of fossil energy will become increasingly scarce and will eventually be exhausted.
[0003] To address the energy crisis, scientists have begun researching emerging energy sources. Currently, common new energy technologies include wind power, hydropower, photovoltaic power, tidal energy, nuclear energy, and geothermal energy. Among these, wind power, as one of the fastest-growing new technologies in terms of installed capacity in recent years, has surpassed photovoltaic power and secured a place in the global energy market, becoming the second largest renewable energy source after hydropower.
[0004] However, wind generation is often unstable and fluctuating, which makes wind power utilization difficult, resulting in large fluctuations in power generation, difficulty in grid connection, and even damage to wind turbines when wind speed changes rapidly, causing large-scale grid disconnection, affecting power generation quality, and endangering personnel safety. In order to stabilize power generation and improve power generation quality, wind speed prediction is necessary.
[0005] Wind speed forecasting is crucial in the field of wind power generation. Early wind speed forecasting typically employed linear models, such as the Autoregressive Moving Average (ARMA) model, which identified relationships between historical values and established regression equations to predict future wind speeds. Since the 1990s, with the rise of artificial intelligence, many nonlinear methods have been used for wind speed forecasting, such as Artificial Neural Networks (ANN), Support Vector Machines (SVM), Extreme Learning Machines (ELM), Long Short-Term Memory Networks (LSTM), and Convolutional Neural Networks (CNN). These methods generally use historical wind speeds and other features influencing wind speed as input to build predictive models and thus predict future wind speeds. While these methods generally achieve good overall forecasting results, their effectiveness may be less than ideal during periods of high frequency and rapid speed changes.
[0006] To further improve the prediction accuracy of the model, this application proposes a prediction method based on ELM error compensation. First, historical wind speed data is divided using the ARMA modeling method and used as input to build an SVR model to predict future wind speed. Then, the error of the historical wind speed prediction values is used to build an ELM model to predict the future error value. Finally, this error value is added to the wind speed prediction value to obtain the wind speed prediction value after error compensation. Finally, verification experiments demonstrate that this method can improve the overall accuracy of wind speed prediction. Summary of the Invention
[0007] Generally, the more features a target object has, the more accurate the prediction will be. In wind speed prediction, wind speed, as the target object, is affected by factors such as ambient temperature, atmospheric pressure, ambient humidity, and air density. The more data collected containing these features, the more accurate the prediction will be. However, in actual measurements, due to cost and other limitations, many features are ignored, leaving only the target object. For wind speed prediction, many wind farms only collect historical wind speed data and rarely collect other factors. Therefore, correctly selecting the amount of historical wind speed data to input the prediction model becomes crucial. Furthermore, traditional methods, such as ARMA, often suffer from decreased wind speed prediction accuracy over time. To address this, this invention provides a wind speed prediction method based on error compensation, which includes the following steps:
[0008] (1) Obtain historical wind speed data of a wind farm and perform a stationarity test on the sequence. If the stationarity test is passed, it is used for modeling in the ARMA model. If the test is not passed, the difference method is used to transform the non-stationary sequence into a stationary sequence.
[0009] (2) Use the stationary sequence obtained in step (1) for ARMA modeling, adopt the AIC criterion, and use an iterative algorithm to solve for the partial autocorrelation coefficient p and autocorrelation coefficient q that minimize the AIC value;
[0010] The ARMA model is as follows:
[0011]
[0012] Where, α i It is the autoregressive coefficient, β j These are the moving average coefficients; both are undetermined coefficients and are not zero. ε t Let x be the error term at time t. t ε represents the value at time t. t-j Let x be the error term at time tj. t-i The value at time ti;
[0013] (3) Use the p value obtained in step (2) as the basis to divide the dataset;
[0014] (4) Perform SVR modeling, use manual trial and error method, select Gaussian kernel function and determine the value of penalty factor, then train the model and use test set to check the accuracy of model prediction;
[0015] (5) Use the trained SVR model to predict the training set to obtain the wind speed prediction value of the training set. Then, subtract the predicted wind speed value from the actual wind speed value of the training set to obtain the error training set.
[0016] (6) The same processing method as the wind speed training set is used for the error training set, that is, repeat steps (1) and (2) for the error sequence. The ARMA modeling method is used to first determine the stability of the error sequence. If it is unstable, it needs to be differentially processed. Then, the ARMA model is established. The AIC criterion is used and the iterative algorithm is used to solve the partial autocorrelation coefficient p′ and autocorrelation coefficient q′ that minimize the AIC value of the ARMA model of the error sequence.
[0017] (7) Divide the error dataset using the partial autocorrelation coefficient p′ from step (6);
[0018] The partitioning process in steps (3) and (7) is the same: starting from the first value of the sequence to the p′th value, every p′ units forms a data group. Then, starting from the second value of the sequence, every p′ units again forms a second data group until the p′+1th value, and so on, until the dataset is completely partitioned. For each data group, the first p′-1 values are used as input, and the p′th value is used as output; that is, the first p′-1 values are used to predict the p′th value.
[0019] (8) Train the ELM model using the error training set from step (7), and then use the model to predict the error to obtain the error prediction value. Finally, add the error prediction value to the wind speed prediction value from step (4) to obtain the wind speed prediction value after error compensation.
[0020] ELM output y i for:
[0021]
[0022] Where, x i This is the input to the ELM; N represents the input dimension, i.e., the number of neurons in the input layer; K is the number of neurons in the hidden layer; ω j and b j These represent the weights and biases between the j-th hidden layer neuron and the input, respectively; β jis the weight between the j-th hidden layer neuron and the output; g(·) is the activation function, which can be either the Sigmoid function or the RBF function. In this invention, the Sigmoid function is used.
[0023] (9) Finally, the root mean square error is used to evaluate the accuracy of the model prediction.
[0024] The root mean square error (RMSE) is defined as follows:
[0025]
[0026] in x represents the predicted value. i This represents the true value, and N represents the total number of samples.
[0027] Preferably, in step (1), a stationary sequence refers to:
[0028]
[0029] Wherein, E(X) t ) represents sequence X t The expectation, Var(X) t ) represents variance, r t,t-i and r 0,i Let μ and σ represent the autocorrelation coefficients of the values at time t and ti, and at time 0 and i, respectively. 2 These two numbers are constants.
[0030] Preferably, in step (1), a stationary sequence refers to a sequence whose mean and variance do not change with time, and whose autocorrelation coefficients are the same at equal time intervals, i.e.:
[0031]
[0032] Wherein, E(X) t ) represents sequence X t The expectation, Var(X) t ) represents variance, r t,t-i and r 0,i Let μ and σ represent the autocorrelation coefficients of the values at time t and ti, and at time 0 and i, respectively. 2 It is a constant;
[0033] The finite difference method refers to:
[0034] y t =Y t -Y t-1
[0035] Among them, y t Y represents the value after difference at time t. t and Yt-1 These represent the values of the original data at time t and time t-1, respectively. Preferably, in step (2), the AIC criterion refers to:
[0036] AIC = 2k - 2ln(L)
[0037] Where k is the number of unknown parameters in the model, and L is the maximum likelihood function value of the model.
[0038] Preferably, in step (2), the ARMA model is:
[0039]
[0040] Where, α i It is the autoregressive coefficient, β j These are the moving average coefficients; both are undetermined coefficients and are not zero. ε t Let x be the error term at time t. t ε represents the value at time t. t-j Let x be the error term at time tj. t-i This is the value at time ti.
[0041] Preferably, in step (3), the autocorrelation function refers to:
[0042]
[0043] Where, ρ t,t-k X represents t and X t-k The autocorrelation function, Represents sequence X t The mean.
[0044] Preferably, in step (3), the training set and the test set are divided as follows: the first three-quarters of the divided data are used as the training set, and the last quarter is used as the test set.
[0045] Preferably, in step (4), the original model of SVR is:
[0046]
[0047]
[0048] Where st represents the constraint condition, C is the penalty factor, m is the number of samples, and ξ i and It is a slack variable, ε is the tolerance bias, and y is the slack variable. i This is the output; w and b are the hyperplane parameters that need to be optimized.
[0049] Preferably, in step (4), the Gaussian kernel function is:
[0050]
[0051] Where x' represents the origin of the kernel function, x represents the sampling point, and ‖x-x'‖ 2 σ represents the Euclidean distance between the sampling point and the origin, and σ controls the correlation between sampling points at the same interval of the Gaussian kernel function.
[0052] Preferably, in step (4), the root mean square error (RMSE) is defined as:
[0053]
[0054] in x represents the predicted value. i This represents the true value, and N represents the total number of samples.
[0055] Compared with the prior art, the advantages of the present invention are as follows:
[0056] 1. This invention provides a wind speed prediction method based on error compensation. This method combines ARMA and SVR based on ELM error compensation, which can divide a univariate wind speed time series into multiple segments for model training. Compared with trial-and-error data partitioning, this method optimizes the SVR model training set, makes more reasonable selection of the number of inputs, and compensates for prediction errors in periods of large wind speed fluctuations by combining ELM with error compensation prediction, ultimately achieving excellent prediction results.
[0057] 2. By selecting different amounts of data as input, this invention demonstrates that, compared to the trial-and-error method, the SVR model trained using the data selected according to this invention as input has a better prediction effect.
[0058] 3. This invention uses ELM as an error compensation method to predict historical errors and add them to the predicted wind speed value, thereby improving the prediction effect during periods of high frequency and rapid wind speed change. Attached Figure Description
[0059] The invention will now be further described with reference to the accompanying drawings.
[0060] Figure 1 This is wind speed data for a wind farm in China.
[0061] Figure 2 The sequence is the original wind speed difference sequence from the training set, and at this point, the sequence is a stationary sequence.
[0062] Figure 3 The graph shows the autocorrelation and partial autocorrelation functions after differencing the original data.
[0063] Figure 4 Flowcharts are designed for ARMA and SVR wind speed prediction methods based on ELM error compensation;
[0064] Figure 5 The prediction performance of the trained model when p is 4;
[0065] Figure 6 The prediction performance of the trained model when p is 2;
[0066] Figure 7 The prediction performance of the trained model when p is 6;
[0067] Figure 8 This is the error training set data;
[0068] Figure 9 The differenced error training set data;
[0069] Figure 10 The graph shows the autocorrelation and partial autocorrelation functions of the error training set.
[0070] Figure 11 ELM error prediction results;
[0071] Figure 12 This is the final wind speed prediction result based on ELM after error compensation. Detailed Implementation
[0072] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0073] For wind speed prediction, wind speed is generally affected by factors such as ambient temperature, atmospheric pressure, ambient humidity, and air density. These features can all be used as objects of machine learning. However, in actual data collection, considering cost or other factors, some wind farms generally do not detect these factors and only collect the object to be predicted, i.e., wind speed. The resulting data becomes a univariate time series containing only historical wind speeds. This type of data cannot be directly used as input for methods such as SVR. Therefore, to use more advanced machine learning methods to predict wind speed, the original wind speed series must be processed. Since the processed series essentially only contains multiple historical wind speed values, the prediction performance may deteriorate in areas with large frequency fluctuations. To compensate for this error, error compensation methods can be employed.
[0074] The purpose of this invention is to process a set of univariate wind speed sequences into sequences that can be used as input to an SVR model using ARMA modeling, then use the SVR model to predict future wind speeds, and use ELM to process historical errors to obtain the predicted error value. Finally, the predicted future wind speed is added to the predicted error to obtain the future wind speed after error compensation.
[0075] This invention utilizes historical data to establish an ARMA model. The order of the ARMA model is used to partition the dataset. Then, an SVR model is trained using the partitioned data to predict wind speed. The actual wind speed values in the training set are subtracted from the predicted wind speed values to obtain an error training set. An ARMA model is then established on the error training set using the aforementioned method, and its partial autocorrelation coefficient is used as the basis for partitioning the error training set. After processing the error series, an ELM is used to predict the errors. The resulting error prediction is then added to the wind speed prediction to obtain the error-compensated wind speed prediction value.
[0076] Before partitioning the dataset, the original dataset was ordered using the ARMA modeling method. The dataset was then partitioned using the obtained partial autocorrelation coefficient p, resulting in multiple time series with strong correlations. The partitioned data was further divided into training and test sets. The training set was used to train the SVR model, and the test set was used to test the prediction performance of the trained model. Finally, the error sequence was obtained by subtracting the actual value from the predicted value. The processing of the error sequence was the same as that of the original wind speed data, using the partial autocorrelation coefficient of the ARMA model to partition the data. Then, error prediction was performed using ELM, and the error prediction result was added to the wind speed prediction result to obtain the error-compensated wind speed prediction value.
[0077] like Figure 4 As shown, the present invention employs a combined prediction method of ARMA and SVR with ELM error compensation, specifically including the following steps:
[0078] Step 1: Collect wind speed data from a wind farm and perform an ADF stationarity test. If the test passes, ARMA modeling can proceed; otherwise, differencing is required. The wind speed data for the wind farm can be found here. Figure 1 It is clearly a non-stationary sequence because its mean and variance are constantly changing. Therefore, differencing is performed, see... Figure 2 At this point, the sequence is stationary and can be modeled using ARMA.
[0079] A stationary sequence is defined as:
[0080]
[0081] Wherein, E(X) t ) represents sequence X tThe expectation, Var(X) t ) represents variance, r t,t-i and r 0,i Let μ and σ represent the autocorrelation coefficients of the values at time t and ti, and at time 0 and i, respectively. 2 These two numbers are constants.
[0082] The finite difference method refers to:
[0083] y t =Y t -Y t-1
[0084] Among them, y t Y represents the value after difference at time t. t and Y t-1 These represent the values of the original data at time t and time t-1, respectively.
[0085] Step 2: The stationary sequence obtained in Step 1 is used for ARMA modeling. The AIC criterion is adopted, and an iterative algorithm is used to solve for the partial autocorrelation coefficient p and autocorrelation coefficient q that minimize the AIC value. In this invention, the partial autocorrelation coefficient p of the model is mainly used because p reflects the correlation between the current value and historical values. The value of p indicates that the current value is highly correlated with all historical values lagged by p. Using these p values to form a set of data can provide a reference for dividing the data. This avoids data waste when selecting the number of data segments and also avoids machine performance loss due to selecting too much data.
[0086] In this embodiment, the ARMA model is:
[0087]
[0088] Where, α i It is the autoregressive coefficient, β j These are the moving average coefficients; both are undetermined coefficients and are not zero. ε t Let x be the error term at time t. t ε represents the value at time t. t-j Let x be the error term at time tj. t-i Let x be the value at time ti. From the above formula, we can see that the current value x... t Compared with the first p values
[0089] Next, using the AIC criterion and iterative methods, the p and q values that minimize the AIC are found. The AIC criterion also provides the best estimate of the ARMA model order and parameters, and is suitable for problems with limited sample data. The expression for the AIC criterion is:
[0090] AIC = 2k - 2ln(L)
[0091] Where k is the number of unknown parameters in the model, and L is the maximum likelihood function value of the model.
[0092] Using the steps described above, the calculated p-value can be used as the basis for dividing the dataset.
[0093] Step 3: Use the p obtained in Step 2 to extract data. In this embodiment, p is 4. The specific extraction method is as follows: From the data at the first time step to the data at the p-th time step, each p value is taken as a new data set. The first p-1 numbers are selected as input data, and the p-th number is taken as output. Then, following the same rule, the same division operation is performed on the data from the second time step to the p+1-th time step. This process is repeated until the entire dataset is divided.
[0094] Step 4: Perform SVR modeling. Use a manual trial-and-error method to determine the ideal penalty factor and kernel function parameters. Then, train the model using the pre-divided dataset and test the model's accuracy using the test set. The prediction results are as follows: Figure 5 As shown. The original model of SVR is as follows:
[0095]
[0096]
[0097] Where st represents the constraint condition, C is the penalty factor, m is the number of samples, and ξ i and It is a slack variable, ε is the tolerance bias, and y is the slack variable. i This is the output; w and b are the hyperplane parameters that need to be optimized. The kernel function is very important in SVR. In this embodiment, the kernel function is implemented using the Gaussian kernel function (radial basis function), and its expression is:
[0098]
[0099] Where x' represents the origin of the kernel function, x represents the sampling point, and ‖x-x'‖ 2 σ represents the Euclidean distance between the sampling point and the origin, and σ controls the correlation between sampling points at the same interval of the Gaussian kernel function.
[0100] Step 5: To demonstrate that the division criteria chosen in this invention are relatively good, a trial-and-error method is used to retrain the model using different p-values as the division criteria. The prediction results are then compared with the results from Step 4, and the root mean square error is used to determine the accuracy of the prediction. Figure 6 p = 2 Figure 7 p = 6.
[0101] The comparison revealed that the best prediction performance was achieved when p = 4, i.e., the p value obtained in step 2, with a root mean square error (RMSE) of only 0.91226. In contrast, the RMS error was 0.91539 when p = 2 and 0.91718 when p = 6 using the trial-and-error method, both higher than the model that achieved the optimal p value. Therefore, the comparison demonstrates that the SVR model trained using data selected based on the p value determined by the ARMA model as input exhibits superior prediction performance and improves prediction accuracy.
[0102] Step 6: Use the model trained in Step 4 to test the training set again. The purpose of this step is to use the prediction model to obtain the predicted values of the training set, and then subtract the predicted wind speed values from the actual wind speed values of the training set. This will give you the error training set, which will be used to establish the error prediction model.
[0103] Step 7: For the error training set obtained in Step 6, the same processing method as for the wind speed training set is used, that is, repeating Steps 1 and 2 on the error sequence, using the ARMA modeling method to first determine the stability of the error sequence. If it is unstable, differencing is required. Figure 8 As shown, the error sequence is clearly a non-stationary time series, so it needs to be differencing to make it stationary. The differrated error sequence is as follows. Figure 9 As shown, the sequence is stationary at this point.
[0104] Step 8: Next, establish the ARMA model, use the AIC criterion and an iterative algorithm to solve for the partial autocorrelation coefficient p′ and autocorrelation coefficient q′ that minimize the AIC value of the error sequence ARMA model;
[0105] Step 9: Divide the error dataset using the partial autocorrelation coefficient p′ obtained in Step 8. This step is the same as the partitioning process in Step 3, that is, starting from the first value of the sequence to the p-th value, every p units form a data group. Then, starting from the second value of the sequence, every p units again forms a second data group, until the p+1-th value, and so on, until the dataset is completely partitioned. For each data group, the first p-1 values are used as input, and the p-th value is used as output, that is, the p-th value is predicted using the first p-1 values.
[0106] Step 10: For the partitioned error training set obtained in Step 8, use ELM to build a prediction model to predict future errors. The output of ELM is:
[0107]
[0108] Where, x i This is the input to the ELM; N represents the input dimension, i.e., the number of neurons in the input layer; K is the number of neurons in the hidden layer; ω j and bj These represent the weights and biases between the j-th hidden layer neuron and the input, respectively; β j Let g be the weights between the j-th hidden layer neuron and the output; g(·) is the activation function. Here, the sigmoid function is chosen as the activation function, and its expression is:
[0109]
[0110] Comparison of the error predicted by ELM with the actual error, for example Figure 11 As shown, the prediction error generally fits the true error well, with only occasional spikes. However, considering the small unit of error, the final prediction result will not be significantly affected.
[0111] Finally, the obtained error prediction value is added to the wind speed prediction value to obtain the final wind speed prediction value, as shown in the figure. Figure 12 As shown, the root mean square error of its prediction model is 0.89609, which is lower than... Figure 5 The root mean square error (RMSE) is 0.91226, indicating that the wind speed prediction after error compensation is better than the direct prediction. Figure 5 .
[0112] Verification experiments have demonstrated that, during periods of high frequency and rapid wind speed change, the error-compensated wind speed prediction method employed in this invention has a better fitting effect than the traditional direct prediction method, thus improving the overall accuracy of wind speed prediction.
[0113] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A wind speed prediction method based on error compensation, characterized in that, Includes the following steps: (1) Obtain historical wind speed data of a wind farm and perform stationarity test on the historical wind speed data of the wind farm. If the stationarity test is passed, it is used for modeling in the ARMA model. If the test is not passed, the difference method is used to transform the non-stationary sequence into a stationary sequence. (2) Use the stationary sequence obtained in step (1) for ARMA modeling, adopt the AIC criterion, and use an iterative algorithm to solve for the partial autocorrelation coefficient p and autocorrelation coefficient q that minimize the AIC value; The ARMA model is as follows: ； in, It is the autoregressive coefficient. These are the moving average coefficients; both are undetermined coefficients and are not zero. Let be the error term at time t. Represents the value at time t. This is the error term at time tj. The value at time ti; (3) Use the p value obtained in step (2) as the basis to divide the dataset. When calculating the p value, it is necessary to first calculate the autocorrelation function. (4) Perform SVR modeling, use manual trial and error method, select Gaussian kernel function and determine the value of penalty factor, then train the model and use test set to check the accuracy of model prediction; (5) Use the trained SVR model to predict the training set to obtain the wind speed prediction value of the training set. Then, subtract the predicted wind speed value from the actual wind speed value of the training set to obtain the error training set. (6) The same processing method as the wind speed training set is used for the error training set, that is, repeat steps (1) and (2) for the error sequence. The ARMA modeling method is used to first determine the stability of the error sequence. If it is unstable, it needs to be differentially processed. Then, the ARMA model is established, and the partial autocorrelation coefficient that minimizes the AIC value of the ARMA model of the error sequence is solved using the AIC criterion and an iterative algorithm. and autocorrelation coefficient ; (7) Using the partial autocorrelation coefficient from step (6) Divide the error dataset; The division process in step (3) and step (7) is the same, that is, starting from the first value of the sequence to the... Each value, every Each unit is grouped into a set of data. Then, starting from the second value of the sequence, skipping a certain number of units... Unit, up to the The first value represents the second set of data, and so on, until the dataset is completely divided; for each set of data, the first... The nth value is used as input, the nth The values are used as output, i.e., before use The value is predicted. One value; (8) Use the error training set in step (7) to train the ELM model, and then use the model to make error predictions to obtain the error prediction value; finally, add the error prediction value to the wind speed prediction value in step (4) to obtain the wind speed prediction value after error compensation. ELM output for: ； in, This is the input to ELM; N represents the input dimension, which is the number of neurons in the input layer; K is the number of neurons in the hidden layer. and These are the weights and biases between the j-th hidden layer neuron and the input, respectively; represents the weights between the j-th hidden layer neuron and the output; g(·) is the activation function, which can be either the Sigmoid function or the RBF function. In this step, the Sigmoid function is used. (9) Finally, the root mean square error is used to evaluate the accuracy of the model prediction; The root mean square error (RMSE) is defined as follows: ； in Indicates the predicted value. This represents the true value, and N represents the total number of samples. In step (1), a stationary sequence refers to: ； in, Represents a sequence Expectations Represents variance. and Let represent the autocorrelation coefficients of the values at time t and ti, and at time 0 and i, respectively. These two numbers are constants.
2. The wind speed prediction method based on error compensation according to claim 1, characterized in that, In step (1), a stationary sequence is one in which the mean and variance do not change over time, and the autocorrelation coefficients at equal time intervals are the same, i.e.: ； in, Represents a sequence Expectations Represents variance. and Let represent the autocorrelation coefficients of the values at time t and ti, and at time 0 and i, respectively. These two numbers are constants; The finite difference method refers to: ； in, This represents the value after differencing at time t. These represent the values of the original data at time t and time t-1, respectively.
3. The wind speed prediction method based on error compensation according to claim 1, characterized in that, In step (2), the AIC criterion refers to: ； Where k is the number of unknown parameters in the model, and L is the maximum likelihood function value of the model.
4. The wind speed prediction method based on error compensation according to claim 1, characterized in that, In step (3), the autocorrelation function refers to: ； in, express and The autocorrelation function, Represents a sequence The mean.
5. The wind speed prediction method based on error compensation according to claim 1, characterized in that, In step (3), the training set and test set are divided as follows: the first three-quarters of the divided data are used as the training set and the last quarter is used as the test set.
6. The wind speed prediction method based on error compensation according to claim 1, characterized in that, In step (4), the original model of SVR is: ； Where st represents the constraint, C is the penalty factor, and m is the number of samples. It is a slack variable. It is tolerance bias. This is the output; w and b are the hyperplane parameters that need to be optimized.
7. The wind speed prediction method based on error compensation according to claim 1, characterized in that, In step (4), the Gaussian kernel function is: ； in, Represents the origin of the kernel function. Indicates the sampling point. This represents the Euclidean distance between the sampling point and the origin. It controls the correlation between sampled points at the same interval of the Gaussian kernel function.