Wind power plant ultra-short-term power prediction optimization method based on laser radar

Abstract

The invention provides a wind power plant ultra-short-term power prediction optimization method based on a laser radar, and the method employs the laser radar to scan a wind power plant region, and obtains the wind speed and direction data based on a coherent Doppler wind measurement principle. The wind speed extrapolation prediction is carried out by adopting an algorithm combining EMD and ARCH, the EMD is firstly used for decomposing a wind speed time sequence, then an ARCH model is established for each intrinsic mode function, and finally the ultra-short-term wind speed is obtained by combining prediction results. Then, an algorithm combining SVR and GA is used for power preliminary prediction, the SVR establishes a nonlinear relation between input parameters and power output, and the GA optimizes SVR parameters; then, the preliminary prediction result is processed and corrected based on an FNN algorithm, and the laser radar prediction wind speed and the preliminary power prediction value are input into a fuzzy neural network, so that the prediction accuracy is improved. And performing precision evaluation on the correction result according to specification requirements, and if the result does not reach the standard, adjusting parameters or improving the method. And when the precision reaches the standard, outputting a wind power forecast time sequence power value in the future 3-5 hours.

Description

Technical Field

[0001] The present invention relates to the field of ultra-short-term power prediction of wind farms, and in particular to an ultra-short-term power prediction optimization method for wind farms based on laser radar. Background Art

[0002] According to the requirements of the Wind Power Forecast Function Specification, wind farms need to provide ultra-short-term power forecasts, that is, the wind power forecast time series power value within the next 4 hours, and the accuracy requirement is to achieve the monthly root mean square error of the 4th hour forecast should be less than 15%. At present, the mainstream optimization of ultra-short-term power forecasting methods are based on the optimization of models and algorithms, and few start from the source of data to improve the accuracy of forecasts. At present, the data source used for ultra-short-term power forecasts is the wind speed and direction data measured by the wind tower of the wind farm, or the wind measurement data of the wind speed and direction instrument on the wind turbine. The forecast data is difficult to grasp the sudden changes in wind speed caused by the weather system in the local area. Summary of the invention

[0003] The main purpose of the present invention is to provide a laser radar-based ultra-short-term power forecasting optimization method for wind farms, so as to solve the problem that the current ultra-short-term power forecasting uses the data source of wind speed and direction data measured by the wind farm wind tower, or the wind speed and direction meter on the wind turbine, and the forecast data is difficult to grasp the problem of sudden changes in wind speed caused by the weather system in the local area.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for optimizing ultra-short-term power prediction of wind farms based on laser radar, the method comprising: S1. Scan the wind farm area using laser radar and obtain wind speed and direction data based on the coherent Doppler wind measurement principle of laser radar; S2, wind speed extrapolation prediction based on the algorithm combining EMD and ARCH; S3, using the algorithm combining SVR and GA to make preliminary power prediction; S4, processing and correcting the preliminary power prediction results based on the FNN algorithm; S5. According to the relevant specifications, the accuracy of the revised power prediction results shall be evaluated and the prediction error and other indicators shall be calculated; If the accuracy is not up to standard, return to the previous step to adjust the algorithm parameters or improve the method; S6. When the accuracy meets the requirement, output the wind power forecast time series power value within the next 3-5 hours.

[0005] In the preferred embodiment, in step S2: Firstly, the wind speed time series collected by the lidar is decomposed into multiple intrinsic mode functions (IMFs) and a residual term using EMD. Then, an ARCH model is built for each IMF to capture the volatility characteristics and autocorrelation of wind speed; Finally, the prediction results of each IMF and residual term are combined to obtain the online prediction results of ultra-short-term plane wind speed.

[0006] In the preferred solution, the specific steps of performing wind speed extrapolation prediction based on the algorithm combining EMD and ARCH in step S2 are: S21, collecting wind speed time series data collected by the laser radar over a period of time; S22. The EMD algorithm construction method is: Identifying Local Maxima and Minima: Finding Wind Speed ​​Time Series All local maxima and minima in ; Calculate the upper and lower envelopes: Connect the local maximum and minimum points respectively through cubic spline interpolation to form the upper envelope and lower envelope ; Calculate mean envelope: Calculate the mean of the upper and lower envelopes ; Extract the first IMF: Subtract the mean envelope from the original data to get , repeat the above steps until The IMF condition is satisfied, that is, the number of extreme points and zero crossing points of the function is equal or differs by at most one, and at any point, the mean of the envelope defined by the local maximum point and the local minimum point is zero. At this time, This is the first IMF, denoted by ; Calculate the residual term by subtracting the first IMF from the original data to obtain ; Repeat the above step S22: Repeat the above process to get the second IMF and the new residual , and continue to repeat until the residual term can no longer be decomposed. Finally, the wind speed time series is decomposed into IMF and a residual ,Right now ; S23. The steps for building an ARCH model for each IMF are: , satisfies the ARCH(p) model, that is ,in is the mean, is the error term, and , are independent and identically distributed random variables, usually assumed to be standard normal distribution, ,here is the order of the model, is the parameter to be estimated; The parameters of the ARCH model are estimated using methods such as maximum likelihood estimation. It follows a standard normal distribution, then The conditional probability density function is ; By maximizing the log-likelihood function To estimate the parameters; and , used to describe the structure of the error term and the dynamic changes of the variance in the ARCH model; S24. Forecast: Use the estimated ARCH model parameters to forecast each IMF for a period of time in the future and obtain the forecast value ; For the residual Perform simple trend extrapolation or other appropriate forecasting methods to obtain the forecast value ; The final wind speed forecast is ; Through the above S21-S24 algorithm steps, the wind speed extrapolation prediction based on the combination of EMD and ARCH models can be realized, thereby improving the accuracy of ultra-short-term planar wind speed prediction in wind farms.

[0007] In the preferred embodiment, in step S3: First, the nonlinear relationship model between input parameters such as wind speed and wind direction and power output is established using SVR; Then, GA is used to optimize the parameters of SVR to improve the prediction performance of the model; In this way, a preliminary wind power forecast is made based on the data from the existing wind measuring equipment in the wind farm.

[0008] In the preferred solution, step S3 uses an algorithm combining SVR and GA to perform preliminary power prediction. The specific steps are: S31. Collect data from existing wind measuring equipment in the wind farm, including wind speed, wind direction, and corresponding actual power output data; divide the data into a training set and a test set, denoted as ,in Indicates Wind speed of samples and wind direction , Indicates the corresponding power output; The steps of constructing S32 and SVR models are as follows: Select the generalized exponential kernel function, whose expression is ,in and is a parameter to be determined; The basic form of the SVR model: Assume that the training sample set is , the goal of SVR is to find a function , so that for all sample points ,satisfy ,in is the weight vector, is a function that maps input data to a high-dimensional feature space, is the bias term, is a pre-set insensitive loss parameter; Construction of optimization problem: by introducing slack variables and , the optimization problem of SVR can be expressed as: ; subject to ; ; ,in is the penalty parameter; S33, GA optimization SVR parameter method is: SVR parameters , , kernel function parameters and Encoding and forming chromosomes; A set of chromosomes is randomly generated as the initial population. The population size can be adjusted according to the complexity of the problem and the computing resources. The prediction performance of SVR on the training set is used as the fitness function, and the inverse of the square error and mean absolute error indicators are used as the fitness value, that is, the higher the fitness value, the better the model performance; Roulette wheel selection and other methods are used to select excellent chromosomes to enter the next generation according to their fitness values. Chromosomes with high fitness values ​​are more likely to be selected. Perform crossover operation on the selected chromosome to generate new chromosome; When a certain number of iterations is met or the fitness value reaches a certain threshold, the algorithm terminates, otherwise, it returns to the previous step to continue the genetic operation; S34. Use the optimized SVR parameters to establish a power prediction model, predict the test set data, and obtain preliminary wind power prediction results.

[0009] In the preferred embodiment, in step S4: A fuzzy neural network structure is constructed, and the wind speed predicted by the lidar is used as one of the inputs, together with the preliminary power prediction value, into the network; Through fuzzy reasoning and the learning ability of neural networks, the power forecast results are automatically adjusted to improve the accuracy of ultra-short-term power forecasts for wind farms.

[0010] In the preferred solution, the specific method for processing and correcting the preliminary power prediction result based on the FNN algorithm in step S4 is: S41. Collect the wind speed data predicted by the laser radar, recorded as ,in Indicates The predicted wind speed value at a time point; the preliminary power prediction result is obtained, recorded as ,in Indicates wind speed The corresponding preliminary power prediction value; S42, the input layer has two nodes, which input the wind speed value predicted by the lidar and preliminary power prediction , for wind speed input , let its fuzzy set be , the Gaussian membership function is used for fuzzification, that is, ,in and They are fuzzy sets about the center and width parameters of wind speed; For the preliminary power prediction , the fuzzy set is , the Gaussian membership function is also used for fuzzification, that is, ,in and They are fuzzy sets with respect to the center and width parameters of power; The number of nodes in this layer is , each node represents a fuzzy rule, and the output of the node is the excitation intensity of the corresponding rule, which is calculated by multiplying the membership of each input fuzzy set, that is, ; This layer has only one node, which is used to output the corrected power value and use the weighted average method for defuzzification, that is, ,in is the output power value of the corresponding rule obtained through neural network learning; S42, using mean square error as the error function, that is ,in It is The corrected power value of samples, It is The actual power value of each sample; The weights and biases of the neural network are randomly initialized, and for each training sample, the forward propagation calculates the output power value , calculate the error according to the error function, and then back-propagate to update the weights and biases; The update formula is and ,in is the learning rate, and are the gradients of the error with respect to the weight and bias respectively; S43, input the new laser radar predicted wind speed value and the preliminary power prediction value into the trained fuzzy neural network; after fuzzification, rule excitation, defuzzification and forward propagation of the neural network, the revised power prediction result is obtained.

[0011] In the preferred solution, the specific method for evaluating the accuracy of the corrected power prediction result in step S5 is: S51, using WS-MAPE for accuracy evaluation, the power prediction value after correction is The actual power value is , the calculation formula of WS-MAPE is: ,in It is the weight determined based on factors such as the importance or time sequence of the data; S52. According to the specific requirements and relevant specifications of the wind farm, determine the accuracy evaluation standard and set the threshold of WS-MAPE as ,like , then the accuracy is considered to meet the standard; otherwise, the accuracy is not up to standard; S53, calculating WS-MAPE: calculating the WS-MAPE value of the corrected power prediction result according to the formula; S54. Compare the calculated WS-MAPE value with the set threshold Make comparisons; like , then the accuracy meets the standard and the final power prediction result is output; If the accuracy does not meet the standard, proceed to the next step; S55, if the degree does not meet the standard, you can return to the steps of S2-S5 and adjust the parameters in the algorithm. Specifically, in the wind speed extrapolation prediction based on the combination of EMD and ARCH models, you can adjust the order of the ARCH model and the parameters of the EMD decomposition; In the preliminary power prediction combining SVR and GA, the penalty parameter and kernel function parameter of SVR as well as the population size, crossover probability and mutation probability parameters of genetic algorithm can be adjusted; In the power result correction of FNN, the number, center and width parameters of fuzzy sets, as well as the learning rate, number of layers and number of neurons of the neural network can be adjusted; S56. After adjusting the parameters or improving the method, re-predict the power and re-evaluate the accuracy. Repeat the above steps S2-S5 until the accuracy reaches the target.

[0012] The present invention provides a method for optimizing ultra-short-term power forecasting of wind farms based on lidar to improve data accuracy: by using lidar to scan the wind farm area to obtain wind speed and direction data, the distance and range of the predicted wind speed can be extrapolated, and the sudden change of wind speed caused by the weather system in the local area can be better grasped, thus making up for the deficiency of the wind speed and direction instrument data on the traditional wind farm wind tower and wind turbine.

[0013] Optimize wind speed prediction: The wind speed is extrapolated and predicted by combining EMD and ARCH algorithms, decomposing the wind speed time series into multiple intrinsic mode functions and residual terms, and establishing an ARCH model for each IMF. This can effectively capture the fluctuation characteristics and autocorrelation of wind speed, and obtain more accurate ultra-short-term plane wind speed online prediction results.

[0014] Improve power prediction performance: Use an algorithm that combines SVR and GA to perform preliminary power prediction, use SVR to establish a nonlinear relationship model, and then use GA to optimize parameters, thereby improving the model's prediction performance for wind power.

[0015] Enhanced prediction accuracy: The preliminary power prediction results are processed and corrected based on the FNN algorithm. With the help of fuzzy reasoning and the learning ability of neural networks, the power prediction results are automatically adjusted, further improving the accuracy of ultra-short-term power prediction of wind farms.

[0016] Meet the requirements of the specification: Carry out accuracy assessment according to the requirements of relevant specifications to ensure that the output of the wind power forecast time series power value within the next 4 hours meets the accuracy requirements and reaches the standard of the monthly root mean square error of the 4th hour forecast being less than 15%. At the same time, if the accuracy does not meet the standard, the algorithm parameters can be adjusted or the method can be improved to ensure the reliability and sustainable optimization of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 is the prediction optimization flow chart of the present invention; DETAILED DESCRIPTION Example 1 like Figure 1 As shown, a method for optimizing ultra-short-term power prediction of a wind farm based on laser radar comprises: S1. Scan the wind farm area using laser radar and obtain wind speed and direction data based on the coherent Doppler wind measurement principle of laser radar; S2, wind speed extrapolation prediction based on the algorithm combining EMD and ARCH; Firstly, the wind speed time series collected by the lidar is decomposed into multiple intrinsic mode functions (IMFs) and a residual term using EMD. Then, an ARCH model is built for each IMF to capture the volatility characteristics and autocorrelation of wind speed; Finally, the prediction results of each IMF and residual term are combined to obtain the online prediction results of ultra-short-term plane wind speed; S3, using the algorithm combining SVR and GA to make preliminary power prediction; First, the nonlinear relationship model between input parameters such as wind speed and wind direction and power output is established using SVR; Then, GA is used to optimize the parameters of SVR to improve the prediction performance of the model; In this way, preliminary wind power forecasts are made based on data from existing wind measurement equipment at the wind farm; S4, processing and correcting the preliminary power prediction results based on the FNN algorithm; A fuzzy neural network structure is constructed, and the wind speed predicted by the lidar is used as one of the inputs, together with the preliminary power prediction value, into the network; Through fuzzy reasoning and the learning ability of neural networks, the power forecast results are automatically adjusted to improve the accuracy of ultra-short-term power forecasts for wind farms; S5. According to the relevant specifications, the accuracy of the revised power prediction results shall be evaluated and the prediction error and other indicators shall be calculated; If the accuracy is not up to standard, return to the previous step to adjust the algorithm parameters or improve the method; S6. When the accuracy meets the requirement, output the wind power forecast time series power value within the next 3-5 hours.

[0018] In the preferred solution, the specific steps of performing wind speed extrapolation prediction based on the algorithm combining EMD and ARCH in step S2 are: S21, collecting wind speed time series data collected by the laser radar over a period of time; S22. The EMD algorithm construction method is: Identifying Local Maxima and Minima: Finding Wind Speed ​​Time Series All local maxima and minima in ; Calculate the upper and lower envelopes: Connect the local maximum and minimum points respectively through cubic spline interpolation to form the upper envelope and lower envelope ; Calculate mean envelope: Calculate the mean of the upper and lower envelopes ; Extract the first IMF: Subtract the mean envelope from the original data to get , repeat the above steps until The IMF condition is satisfied, that is, the number of extreme points and zero crossing points of the function is equal or differs by at most one, and at any point, the mean of the envelope defined by the local maximum point and the local minimum point is zero. At this time, This is the first IMF, denoted by ; Calculate the residual term by subtracting the first IMF from the original data to obtain ; Repeat the above step S22: Repeat the above process to get the second IMF and the new residual Continue to repeat until the residual term can no longer be decomposed. Finally, the wind speed time series is decomposed into IMF and a residual ,Right now .

[0019] S23. The steps for building an ARCH model for each IMF are: , satisfies the ARCH(p) model, that is ,in is the mean, is the error term, and , are independent and identically distributed random variables, usually assumed to be standard normal distribution, ,here is the order of the model, is the parameter to be estimated; The parameters of the ARCH model are estimated using methods such as maximum likelihood estimation. It follows a standard normal distribution, then The conditional probability density function is ; By maximizing the log-likelihood function To estimate the parameters; and , used to describe the structure of the error term and the dynamic changes of the variance in the ARCH model; S24. Forecast: Use the estimated ARCH model parameters to forecast each IMF for a period of time in the future and obtain the forecast value ; For the residual Perform simple trend extrapolation or other appropriate forecasting methods to obtain the forecast value ; The final wind speed forecast is ; Through the above S21-S24 algorithm steps, the wind speed extrapolation prediction based on the combination of EMD and ARCH models can be realized, thereby improving the accuracy of ultra-short-term planar wind speed prediction in wind farms.

[0020] In the preferred solution, step S3 uses an algorithm combining SVR and GA to perform preliminary power prediction. The specific steps are: S31. Collect data from existing wind measuring equipment in the wind farm, including wind speed, wind direction, and corresponding actual power output data; divide the data into a training set and a test set, denoted as ,in Indicates Wind speed of samples and wind direction , Indicates the corresponding power output; The steps of constructing S32 and SVR models are as follows: Select the generalized exponential kernel function, whose expression is ,in and is a parameter to be determined; The basic form of the SVR model: Assume that the training sample set is , the goal of SVR is to find a function , so that for all sample points ,satisfy ,in is the weight vector, is a function that maps input data to a high-dimensional feature space, is the bias term, is a pre-set insensitive loss parameter; Construction of optimization problem: by introducing slack variables and , the optimization problem of SVR can be expressed as: ; subject to ; ; ,in is the penalty parameter; S33, GA optimization SVR parameter method is: SVR parameters , , kernel function parameters and Encoding and forming chromosomes; A set of chromosomes is randomly generated as the initial population. The population size can be adjusted according to the complexity of the problem and the computing resources. The prediction performance of SVR on the training set is used as the fitness function, and the inverse of the square error and mean absolute error indicators are used as the fitness value, that is, the higher the fitness value, the better the model performance; Roulette wheel selection and other methods are used to select excellent chromosomes for the next generation based on fitness values. Chromosomes with high fitness values ​​are more likely to be selected; Perform crossover operation on the selected chromosome to generate new chromosome; When a certain number of iterations is met or the fitness value reaches a certain threshold, the algorithm terminates, otherwise, it returns to the previous step to continue the genetic operation; S34. Use the optimized SVR parameters to establish a power prediction model, predict the test set data, and obtain preliminary wind power prediction results.

[0021] In the preferred solution, the specific method for processing and correcting the preliminary power prediction result based on the FNN algorithm in step S4 is: S41. Collect the wind speed data predicted by the laser radar, recorded as ,in Indicates The predicted wind speed value at a time point; the preliminary power prediction result is obtained, recorded as ,in Indicates wind speed The corresponding preliminary power prediction value.

[0022] S42, the input layer has two nodes, which input the wind speed value predicted by the lidar and preliminary power prediction , for wind speed input , let its fuzzy set be , the Gaussian membership function is used for fuzzification, that is, ,in and They are fuzzy sets about the center and width parameters of wind speed; For the preliminary power prediction , the fuzzy set is , the Gaussian membership function is also used for fuzzification, that is, ,in and They are fuzzy sets with respect to the center and width parameters of power; The number of nodes in this layer is , each node represents a fuzzy rule, and the output of the node is the excitation intensity of the corresponding rule, which is calculated by multiplying the membership of each input fuzzy set, that is, ; This layer has only one node, which is used to output the corrected power value and use the weighted average method for defuzzification, that is, ,in is the output power value of the corresponding rule obtained through neural network learning; S42, using mean square error as the error function, that is ,in It is The corrected power value of samples, It is The actual power value of each sample; The weights and biases of the neural network are randomly initialized, and for each training sample, the forward propagation calculates the output power value , calculate the error according to the error function, and then back-propagate to update the weights and biases; The update formula is and ,in is the learning rate, and are the gradients of the error with respect to the weight and bias respectively; S43, input the new laser radar predicted wind speed value and the preliminary power prediction value into the trained fuzzy neural network; after fuzzification, rule excitation, defuzzification and forward propagation of the neural network, the revised power prediction result is obtained.

[0023] In the preferred solution, the specific method for evaluating the accuracy of the corrected power prediction result in step S5 is: S51, using WS-MAPE for accuracy evaluation, the power prediction value after correction is The actual power value is , the calculation formula of WS-MAPE is: ,in It is a weight determined based on factors such as the importance or time sequence of the data; S52. According to the specific requirements and relevant specifications of the wind farm, determine the accuracy evaluation standard and set the threshold of WS-MAPE as ,like , then the accuracy is considered to meet the standard; otherwise, the accuracy is not up to standard; S53, calculating WS-MAPE: calculating the WS-MAPE value of the corrected power prediction result according to the formula; S54. Compare the calculated WS-MAPE value with the set threshold Make comparisons; like , then the accuracy meets the standard and the final power prediction result is output; If the accuracy does not meet the standard, proceed to the next step; S55, if the degree does not meet the standard, you can return to the steps of S2-S5 and adjust the parameters in the algorithm. Specifically, in the wind speed extrapolation prediction based on the combination of EMD and ARCH models, you can adjust the order of the ARCH model and the parameters of the EMD decomposition; In the preliminary power prediction combining SVR and GA, the penalty parameter and kernel function parameter of SVR as well as the population size, crossover probability and mutation probability parameters of genetic algorithm can be adjusted; In the power result correction of FNN, the number, center and width parameters of fuzzy sets, as well as the learning rate, number of layers and number of neurons of the neural network can be adjusted; S56. After adjusting the parameters or improving the method, re-predict the power and re-evaluate the accuracy. Repeat the above steps S2-S5 until the accuracy reaches the target.

[0024] Example 2 Further illustrate with reference to Example 1, Figure 1 The structure shown is the detailed steps of the wind speed extrapolation prediction algorithm based on the combination of empirical mode decomposition (EMD) and autoregressive conditional heteroskedasticity (ARCH) model: 1. Data Preparation Collect the wind speed time series data collected by the lidar over a period of time, recorded as ,in Indicates To ensure the accuracy and completeness of the data, possible outliers can be detected and processed. An outlier detection algorithm can be used, such as a method based on the mean and standard deviation. If the deviation of a data point from the mean exceeds a certain multiple of the standard deviation, it is considered an outlier and can be interpolated or processed by other appropriate methods.

[0025] 2. Empirical Mode Decomposition (EMD) 1. Determine local maxima and minima: Traversing the wind speed time series , by comparing the sizes of adjacent data points, find all local maximum and minimum points. This step can be achieved through a simple comparison operation. For each data point ,like and ,but is a local maximum point; if and ,but is a local minimum point.

[0026] 2. Calculate the upper and lower envelopes: The local maximum and minimum points are connected by cubic spline interpolation to form the upper envelope and lower envelope Cubic spline interpolation is a numerical analysis method that constructs a piecewise cubic polynomial function between given data points so that the function has a second-order continuous derivative over the entire interval. The specific implementation can use the cubic spline interpolation function in the numerical computing library.

[0027] 3. Calculate the mean envelope: Calculate the mean of the upper and lower envelopes This step is to obtain the local average trend of the original data.

[0028] 4. Extract the first intrinsic mode function (IMF): Subtracting the mean envelope from the original data yields Repeat the above steps until The IMF condition is satisfied, that is, the number of extreme points and zero crossing points of the function is equal or differs by at most one, and at any point, the mean of the envelope defined by the local maximum point and the local minimum point is zero. At this time, This is the first IMF, denoted by .

[0029] judge Whether the IMF conditions are met can be calculated by the following methods: The number of extreme points and zero crossings of , if they differ by no more than one, and calculate If the local envelope mean is close to zero within an acceptable error range, the condition is considered to be met.

[0030] 5. Calculate the residual term: Subtract the first IMF from the original data to get .

[0031] 6. Repeat steps: right Repeat the above process to get the second IMF and the new residual Continue to repeat until the residual term can no longer be decomposed. Finally, the wind speed time series is decomposed into IMF and a residual ,Right now .

[0032] The calculation of the mean of the upper and lower envelopes is a key step in extracting IMFs. This formula averages the upper and lower envelopes formed by local maximum and minimum points to obtain the local average trend of the original data, thereby gradually extracting IMFs with different frequency characteristics from the original data in subsequent steps.

[0033] 3. Building an ARCH model for each IMF 1. For each , assuming that it satisfies the ARCH(p) model, that is, ,in is the mean, is the error term, and , are independent and identically distributed random variables, usually assumed to be standard normal distribution, ,here is the order of the model, is the parameter to be estimated.

[0034] First, determine the order of the ARCH model , the appropriate order can be selected by some information criteria, AIC (Akaike Information Criterion) or BIC (Bayesian Information Criterion). These criteria weigh the goodness of fit and complexity of the model, and select the order that minimizes the criterion value as the optimal order.

[0035] 2. Parameter estimation: The parameters of the ARCH model are estimated using methods such as maximum likelihood estimation. It follows a standard normal distribution, then The conditional probability density function is .

[0036] By maximizing the log-likelihood function To estimate the parameters. This can be achieved through numerical optimization algorithms such as Newton's method or quasi-Newton's method. Specifically, for a given initial parameter value, the gradient of the log-likelihood function and the Hessian matrix are calculated and the parameter values ​​are continuously updated until they converge to the optimal solution.

[0037] and Used to describe the structure of the error term and the dynamic changes of the variance in the ARCH model. The first formula shows that the error term consists of the product of the variance and a standard normally distributed random variable, and the second formula describes the dynamic changes of the variance depending on the square of the past error terms. These two formulas together form the core of the ARCH model, which can capture the autocorrelation and heteroskedasticity of fluctuations in the wind speed time series.

[0038] IV. Forecasting and Combination 1. Using the estimated ARCH model parameters, forecast each IMF for a period of time in the future and obtain the forecast value .

[0039] According to the prediction method of the ARCH model, for a given current moment and past observations, the estimated parameters are used to calculate the variance prediction value at the future moment. , and then combined with the random variable The assumed distribution of (usually the standard normal distribution) generates the forecast error term , and finally get the IMF's forecast value .

[0040] 2. For the residual term Perform simple trend extrapolation or other appropriate forecasting methods to obtain the forecast value .

[0041] For the prediction of residual items, linear trend extrapolation, exponential smoothing and other methods can be used. If the residual items show a linear trend, the least squares method can be used to fit a straight line, and then the future moments can be predicted based on the straight line equation.

[0042] 3. The final wind speed forecast is .

[0043] The predicted values ​​of each IMF and the predicted value of the residual term are added together to obtain the final wind speed prediction value. This is based on the principle of empirical mode decomposition, that is, the original signal can be decomposed into the sum of multiple IMFs and a residual term. When predicting, the prediction results of each part are combined to obtain a more accurate overall prediction.

[0044] Through the above complex algorithm steps, the wind speed extrapolation prediction based on the combination of EMD and ARCH models can be realized, and the accuracy of ultra-short-term planar wind speed prediction of wind farms can be improved.

[0045] Example 3 In conjunction with Example 1, the detailed steps of the power preliminary prediction algorithm based on the combination of support vector machine regression (SVR) and genetic algorithm (GA) are further described: 1. Data Preparation Collect data from existing wind measurement equipment in wind farms, including wind towers, wind speed and direction meters on wind turbines, including wind speed, wind direction, and corresponding actual power output data. Divide the data into training and test sets, denoted as ,in Indicates Wind speed of samples and wind direction , Indicates the corresponding power output.

[0046] In the data preparation stage, data preprocessing can also be performed to normalize the wind speed and direction data so that their range is within a specific range to improve the performance and stability of the algorithm. The following formula can be used for normalization: ; in, is the original data, and are the minimum and maximum values ​​in the data, respectively. is the normalized data.

[0047] 2. Support Vector Machine Regression (SVR) Model Construction 1. Choose the kernel function: Commonly used kernel functions include linear kernel function, polynomial kernel function, radial basis kernel function (RBF), etc. Here we choose the generalized exponential kernel function (Generalized Exponential Kernel), whose expression is ,in and is a parameter to be determined. This kernel function can better capture the complex nonlinear relationship of data, but the selection of parameters is more complicated.

[0048] Analyze the characteristics of the generalized exponential kernel function: As the parameter and The shape of the kernel function will change. When it increases, the value of the kernel function approaches zero faster, which means that the influence range between data points becomes smaller; when When increases, the kernel function becomes more sensitive to the distance changes between data points. Therefore, the reasonable selection of these two parameters is crucial to the performance of the model.

[0049] 2. The basic form of the SVR model: Assume that the training sample set is , the goal of SVR is to find a function , so that for all sample points ,satisfy ,in is the weight vector, is a function that maps input data to a high-dimensional feature space, is the bias term, is a pre-set insensitive loss parameter.

[0050] Explain the meaning of the function form: Representing a linear function in a high-dimensional feature space can better capture the nonlinear relationship in the data by mapping the input data to a high-dimensional space. It is used to control the tolerance of the model to errors, that is, when the difference between the predicted value and the actual value is less than When the loss is not calculated.

[0051] 3. Construction of optimization problem: By introducing slack variables and , the optimization problem of SVR can be expressed as: ; subject to ; ; ,in is the penalty parameter; Analyze the various parts of the optimization problem: Represents the complexity of the model, hoping to find a weight vector Keep the model as simple as possible.

[0052] represents the penalty for training error, is a penalty parameter that controls the tolerance to error. This means that the error is punished more severely, and the model is more inclined to fit the training data, but it may lead to overfitting. The error penalty is lighter and the model is more focused on keeping it simple, but it may lead to underfitting.

[0053] Constraints and Ensure that the difference between the predicted value and the actual value is within the insensitive loss parameter and the range of slack variables.

[0054] ,The generalized exponential kernel function can better capture the complex nonlinear relationship of the data, but the selection of parameters is more complicated. and Different values ​​of will affect the shape of the kernel function and the performance of the model, and need to be adjusted through optimization algorithms.

[0055] and constraints, the goal of this optimization problem is to minimize the complexity of the model (given by ) and the training error (denoted by The penalty parameter Controls the tolerance to error, the larger This means that the penalty for error is heavier and the model tends to fit the training data more, but it may lead to overfitting.

[0056] 3. Genetic Algorithm (GA) Optimization of SVR Parameters 1. Chromosome encoding: The parameters of SVR , , kernel function parameters and Encode to form chromosomes. Real number encoding can be used, that is, each parameter is represented as a real number. The parameter range can be mapped to a specific interval, and then the real number is used to represent the value of the parameter in the interval.

[0057] Explain the choice of encoding method: Real number encoding has higher precision and flexibility, and can better represent continuous parameter space. Compared with binary encoding, real number encoding does not require decoding operations, which reduces computational complexity.

[0058] 2. Initialize the population: A set of chromosomes is randomly generated as the initial population. The population size can be adjusted according to the complexity of the problem and the computing resources. Generally speaking, a larger population size can increase the search space of the algorithm, but it will also increase the computing time.

[0059] Analyze the choice of fitness function: Mean square error is a common indicator to measure the difference between the predicted value and the actual value, which gives a greater penalty for larger errors. By taking the inverse of the mean square error as the fitness value, the optimization problem can be transformed into a problem of maximizing the fitness value, allowing the genetic algorithm to search in the direction of improving model performance.

[0060] 4. Genetic Operations: Selection operation: Roulette selection method and other methods are used to select excellent chromosomes to enter the next generation according to the fitness value. Chromosomes with high fitness values ​​are more likely to be selected.

[0061] Explain the principle of roulette wheel selection: the proportion of each chromosome's fitness value to the total fitness value is taken as its probability of being selected. , assuming that Chromosomes, The fitness value of a chromosome is , then the probability of being selected is Then, through random selection, chromosomes are selected to enter the next generation according to the probability distribution.

[0062] Crossover operation: Perform a crossover operation on the selected chromosome to generate a new chromosome. A real number crossover method, such as linear crossover, can be used.

[0063] Explain the significance of mutation operation: mutation operation can introduce new parameter values ​​to prevent the algorithm from falling into the local optimal solution. By randomly perturbing the parameters within a certain range, the algorithm can have the opportunity to explore different parameter combinations and increase the possibility of finding the global optimal solution.

[0064] 5. Termination condition judgment: When a certain number of iterations is met or the fitness value reaches a certain threshold, the algorithm terminates. Otherwise, return to step 4 to continue genetic operations. An example of setting the termination condition: the maximum number of iterations can be set to 100 times, or the algorithm can be terminated when the fitness value has not been significantly improved for several consecutive times. This can prevent the algorithm from running infinitely and ensure that good results are obtained within a reasonable time.

[0065] and It is used to measure the prediction performance of the SVR model on the training set and as the fitness function of the genetic algorithm. The mean square error (MSE) reflects the difference between the predicted value and the actual value. The smaller the MSE, the better the model performance. Therefore, the fitness value is the inverse of the MSE. The higher the fitness value, the better the model performance.

[0066] 4. Preliminary Power Prediction The power prediction model is established using the optimized SVR parameters, and the test set data is predicted to obtain preliminary wind power prediction results.

[0067] The specific steps are as follows: 1. The optimized parameters , , and Substitute into the SVR model.

[0068] 2. For each sample point in the test set , calculate its mapping in high-dimensional feature space .

[0069] 3. According to the functional form of the SVR model , calculate the predicted power output value .

[0070] Through the above complex algorithm steps, the preliminary power prediction based on the combination of SVR and GA can be realized, and the accuracy of wind farm power prediction can be improved.

[0071] Example 4 In combination with Example 1, the detailed algorithm steps for correcting power results based on fuzzy neural network (FNN) are further described as follows: 1. Data Preparation 1. Collect the wind speed data predicted by the lidar, recorded as ,in Indicates The predicted wind speed value at each time point. When collecting wind speed data, ensure the accuracy and reliability of the data. The LiDAR can be calibrated and maintained regularly to reduce measurement errors. At the same time, the data is detected and processed for outliers. An outlier detection based on statistical methods is used. If the deviation of the wind speed value from the mean exceeds a certain multiple of the standard deviation, it is considered an outlier and can be interpolated or processed by other appropriate methods.

[0072] 2. Obtain the preliminary power prediction result, recorded as ,in Indicates wind speed The corresponding preliminary power prediction value. The preliminary power prediction can be obtained by combining the support vector machine regression (SVR) and genetic algorithm (GA) mentioned above.

[0073] 2. Fuzzy neural network structure design 1. Input layer: The input layer has two nodes, which input the wind speed value predicted by the lidar and preliminary power prediction The role of the input layer is to pass the original data to the subsequent fuzzification layer.

[0074] 2. Fuzzification layer: for wind speed input , let its fuzzy set be , the Gaussian membership function is used for fuzzification, that is, ,in and They are There are three fuzzy sets about the center and width parameters of wind speed.

[0075] Center Parameters Determines the position of the fuzzy set on the wind speed axis, the width parameter Controls the range of the fuzzy set. Different fuzzy sets can cover different wind speed ranges. By adjusting these parameters, the fuzzification process can be better adapted to the distribution of data.

[0076] If the wind speed data are mainly distributed in a specific range, the center of the fuzzy set can be set within this range, and the width parameter can be adjusted according to the discreteness of the data to ensure that each fuzzy set can effectively cover a certain number of data points.

[0077] For the preliminary power prediction , let its fuzzy set be , the Gaussian membership function is also used for fuzzification, that is, ,in and They are There are fuzzy sets with respect to the center and width parameters of the power.

[0078] Similarly, the center parameter and width parameters Determines the location and range of the power fuzzy set. According to the distribution of the preliminary power prediction value, reasonable setting of these parameters can improve the fuzzification effect.

[0079] 3. Rule layer: The number of nodes in this layer is , each node represents a fuzzy rule. The rule form is "If the wind speed is And the initial power is ,So……".

[0080] A rule could be “If the wind speed is moderate and the initial power is high, then the power should be slightly reduced.” These rules are developed based on experience and understanding of the wind farm system and can be determined through expert knowledge or data analysis.

[0081] The output of the node is the excitation intensity of the corresponding rule, which is calculated by multiplying the membership of each input fuzzy set, that is, .

[0082] The excitation intensity indicates the degree to which the input data satisfies the rule. If the wind speed and preliminary power are both in the corresponding fuzzy set, the excitation intensity of the rule will be high; conversely, if the input data does not match the conditions of the rule, the excitation intensity will be low.

[0083] 4. Defuzzification layer: This layer has only one node, which is used to output the corrected power value.

[0084] The weighted average method is used for defuzzification, that is, ,in It is the output power value of the corresponding rule obtained through neural network learning.

[0085] In the defuzzification process, the excitation strength of each rule is first calculated The corresponding output power value The product of the rules is then divided by the sum of the products of all rules by the sum of the excitation intensities to obtain the corrected power value. This allows for a comprehensive consideration of the contribution of each rule to the final result, and for adaptive adjustments to be made based on the different conditions of the input data.

[0086] and , these two formulas are used to calculate the membership of the input variables (wind speed and preliminary power forecast) to each fuzzy set. The Gaussian membership function can better reflect the transition of variables between different fuzzy sets. and width parameters , the range and shape of the fuzzy set can be controlled.

[0087] When the input variable is close to the center of the fuzzy set, the membership degree is close to 1, indicating that the variable completely belongs to this fuzzy set; when the input variable is far away from the center, the membership degree gradually decreases, indicating that the degree of association between the variable and this fuzzy set decreases.

[0088] ,This formula calculates the activation strength of the rule, i.e., the degree to which the input variables simultaneously satisfy the corresponding fuzzy sets. The product operation ensures that the rule is strongly activated only when both the wind speed and the preliminary power are within the corresponding fuzzy range.

[0089] If the membership of one of the input variables is low, then the excitation intensity of the entire rule will also be low, which reflects the conditionality of the rule.

[0090] This formula is defuzzified by weighted average, and the output power value of each rule is According to its excitation intensity The weighted sum is performed to obtain the corrected power value, which can comprehensively consider the contribution of different rules to the final result.

[0091] Rules with high excitation intensity have a greater impact on the final result, while rules with low excitation intensity have a smaller impact. This approach allows the corrected power value to be adaptively adjusted according to the specific circumstances of the input data.

[0092] 3. Neural Network Training 1. Define the error function: Use the mean square error (MSE) as the error function, that is ,in It is The corrected power value of samples, It is The actual power value of each sample.

[0093] The mean square error measures the difference between the corrected power value and the actual power value. It gives a greater penalty for larger errors, which helps the optimization algorithm converge to a better result faster.

[0094] 2. Back propagation algorithm: Randomly initialize the weights and biases of the neural network. The initial values ​​of weights and biases can be randomly generated within a certain range, which provides a starting point for the learning of the neural network.

[0095] For each training sample, the forward propagation calculates the output power value .

[0096] Forward propagation refers to the process of calculating the input data through each layer of the neural network to obtain the output result. In the fuzzy neural network, the input data is first fuzzified, and then the corrected power value is obtained based on the excitation intensity of the rule layer and the calculation of the defuzzification layer.

[0097] The error is calculated according to the error function, and then the weights and biases are updated by backpropagation. The update formula is and ,in is the learning rate, and are the gradients of the error with respect to the weights and biases, respectively.

[0098] Learning Rate Controls the update step size. If the learning rate is too large, the algorithm may become unstable and fail to converge to the optimal solution; if the learning rate is too small, the algorithm will converge very slowly. Therefore, it is necessary to choose a suitable learning rate to balance the stability and convergence speed of the algorithm.

[0099] gradient and Represents the rate of change of the error function with respect to weights and biases. By updating the weights and biases in the opposite direction of the gradient, the error can be gradually reduced. During the back propagation process, starting from the output layer, the gradient of the error with respect to weights and biases of each node is calculated in turn, and these gradients are propagated back to the previous layer to update the parameters of the entire neural network.

[0100] This formula measures the difference between the corrected power value and the actual power value. The mean square error is a commonly used error metric that gives a greater penalty for larger errors, helping the optimization algorithm converge to better results faster.

[0101] and These two formulas are used to update the weights and biases of the neural network. Learning rate Controls the update step size, gradient and It represents the rate of change of the error function with respect to the weights and biases. By updating the weights and biases in the opposite direction of the gradient, the error can be gradually reduced.

[0102] 4. Power result correction 1. Input the new lidar predicted wind speed value and preliminary power prediction value into the trained fuzzy neural network.

[0103] 2. After fuzzification, rule excitation, defuzzification and forward propagation of neural network, the corrected power prediction result is obtained.

[0104] Through the above algorithm steps, the fuzzy neural network can be used to effectively process and correct the preliminary power prediction results, thereby improving the accuracy of ultra-short-term power prediction of wind farms.

[0105] Example 5 Further explanation in conjunction with Example 1, the accuracy assessment steps are: 1. Selection of accuracy evaluation indicators 1. Select the weighted sum of mean absolute percentage error (WS-MAPE) evaluation metric. The corrected power prediction value is The actual power value is .

[0106] First, we analyze why WS-MAPE is chosen as the accuracy evaluation indicator. Compared with the traditional mean absolute percentage error (MAPE), WS-MAPE can be weighted according to the importance of different data points, reflecting the accuracy of the prediction results more flexibly. In the actual application of wind farms, the weights can be determined according to factors such as power demand in different time periods and the importance of weather conditions.

[0107] The calculation formula of WS-MAPE is: ,in The weight is determined based on factors such as the importance or time sequence of the data.

[0108] Further explain the meaning of each part of the formula. Represents the absolute percentage error of each data point, that is, the absolute difference between the predicted value and the actual value divided by the actual value. This ratio reflects the degree of deviation of the predicted value from the actual value. Adjustments can be made based on specific circumstances. If electricity demand is higher during a certain time period, or if certain weather conditions have a greater impact on power forecasts, higher weights can be given to the corresponding data points to emphasize the importance of the accuracy of these data points in the overall assessment.

[0109] 2. Determine the evaluation criteria 1. According to the specific requirements and relevant specifications of the wind farm, determine the accuracy evaluation standard. Set the threshold of WS-MAPE to ,like , the accuracy is considered to meet the standard; otherwise, the accuracy is not up to standard.

[0110] Explain how the threshold is determined . You can refer to industry standards, historical data performance, or determine a suitable threshold through multiple tests and adjustments. You can first select a more conservative threshold, and then gradually adjust it according to the actual application needs and algorithm performance. If the wind farm has high requirements for the accuracy of power forecasting, the threshold can be set lower; if a certain degree of error is allowed, the threshold can be appropriately increased.

[0111] Other factors can also be considered to determine the evaluation criteria. For example, the accuracy requirements may be different in different time periods. During peak power demand periods, the accuracy requirements for power forecasts may be higher, so a stricter threshold can be set. During periods of low power demand, the threshold can be appropriately relaxed to balance the performance of the algorithm and the needs of actual applications.

[0112] 3. Accuracy Assessment Process 1. Calculate WS-MAPE: Calculate the WS-MAPE value of the corrected power prediction result according to the formula.

[0113] The calculation process is described in detail. First, for each data point , calculate the absolute percentage error Then, multiply this error by the corresponding weight Finally, the weighted errors of all data points are added together to obtain the value of WS-MAPE. This calculation process can be implemented using programming languages ​​or data analysis software to ensure the accuracy and efficiency of the calculation.

[0114] 2. Comparison and judgment: Compare the calculated WS-MAPE value with the set threshold For comparison. , the accuracy meets the standard and the final power prediction result is output. If the accuracy does not meet the standard, proceed to the next step.

[0115] Explain the comparison and judgment process. If the value of WS-MAPE is less than or equal to the threshold , indicating that the power prediction result is within the acceptable accuracy range and can be considered to be up to standard. At this point, the final power prediction result can be output for the operation and scheduling of the wind farm. If the value of WS-MAPE is greater than the threshold , it means that the accuracy is not up to standard and parameter adjustment or method improvement is needed to improve the accuracy of the prediction.

[0116] 4. Parameter Adjustment and Method Improvement 1. Parameter adjustment: If the accuracy is not up to standard, you can return to the previous steps and adjust the parameters in the algorithm. In the wind speed extrapolation prediction based on the combination of empirical mode decomposition (EMD) and autoregressive conditional heteroskedasticity (ARCH) model, you can adjust the order of ARCH model, the parameters of EMD decomposition, etc.

[0117] For the order adjustment of the ARCH model, you can try different orders and observe the changes in WS-MAPE. Generally speaking, the higher the order, the stronger the model's ability to fit the data, but it may also lead to overfitting. You can select the appropriate order through methods such as cross-validation.

[0118] The parameters of EMD decomposition include the number of IMFs, screening stop criteria, etc. These parameters can be adjusted to observe the impact of the decomposition results on the wind speed extrapolation forecast, and thus affect the accuracy of power forecast. Increasing the number of IMFs may improve the ability to capture wind speed changes, but it will also increase the computational complexity.

[0119] In the preliminary prediction of power combining support vector machine regression (SVR) with genetic algorithm (GA), the penalty parameter and kernel function parameter of SVR as well as the population size, crossover probability, mutation probability and other parameters of GA can be adjusted.

[0120] The penalty parameter of SVR controls the tolerance of the model to errors. A larger penalty parameter will make the model more likely to fit the training data, but may lead to overfitting; a smaller penalty parameter will make the model smoother, but may sacrifice some fitting accuracy. You can find a suitable balance by experimenting with different penalty parameter values ​​and observing the changes in WS-MAPE.

[0121] The kernel function parameters determine how the SVR is mapped in the high-dimensional space. Different kernel function parameters will affect the performance of the model. You can try different kernel function parameter combinations, such as the parameters of the radial basis function (RBF) kernel. , observe the impact on power prediction.

[0122] The parameters of the genetic algorithm include population size, crossover probability, mutation probability, etc. The population size determines the size of the search space. A larger population size may increase the chance of finding the optimal solution, but it will also increase the calculation time. The crossover probability and mutation probability control the search strategy of the genetic algorithm. These parameters can be adjusted to balance the algorithm's exploration ability and convergence speed.

[0123] In the power result correction of fuzzy neural network (FNN), the number, center and width parameters of fuzzy sets, as well as the learning rate, number of layers, number of neurons and other parameters of the neural network can be adjusted.

[0124] The number, center and width parameters of the fuzzy sets determine the effect of the fuzzification. You can try different combinations of parameters to observe the effect on the power result correction. Increasing the number of fuzzy sets can improve the resolution of the input data, but it will also increase the computational complexity. Adjusting the center and width parameters can make the fuzzy sets better fit the distribution of the data.

[0125] The learning rate of a neural network controls the step size of parameter updates. A larger learning rate may speed up the convergence of the algorithm, but it may also easily lead to oscillation or non-convergence; a smaller learning rate will slow down the convergence of the algorithm, but it may be more stable. You can find a suitable learning rate value by adjusting the learning rate and observing the changes in WS-MAPE.

[0126] The number of layers and neurons in a neural network also affects the performance of the model. Increasing the number of layers and neurons can improve the expressiveness of the model, but it will also increase the computational complexity and the risk of overfitting. You can find a structure that strikes a balance between accuracy and computational efficiency by experimenting with different combinations of the number of layers and neurons.

[0127] 2. Method improvement: In addition to adjusting parameters, you can also consider improving the algorithm. Try different kernel functions, fuzzification methods, neural network structures, etc. Or combine the advantages of other algorithms to improve the existing algorithm.

[0128] For the selection of kernel functions, you can try uncommon kernel functions, such as high-order forms of polynomial kernel functions, exponential kernel functions, etc. Different kernel functions are suitable for different data distributions and problem characteristics. You can observe the impact on power prediction by experimenting with different kernel functions.

[0129] The fuzzification method can try different membership functions, such as triangular membership function, trapezoidal membership function, etc. Different membership function shapes will affect the fuzzification effect, and the appropriate membership function can be selected according to the characteristics of the data.

[0130] The neural network structure can try different numbers of layers, neuron connection methods, etc. More complex structures such as deep neural networks and convolutional neural networks can be used to improve the performance of the model. Or the advantages of other algorithms can be combined, such as genetic algorithms combined with neural networks, particle swarm optimization algorithms combined with fuzzy neural networks, etc., to improve the algorithm's search ability and convergence speed.

[0131] New data preprocessing methods, such as principal component analysis (PCA) and independent component analysis (ICA), can be introduced to reduce the dimension and noise of the data and improve the performance of the algorithm.

[0132] Principal component analysis (PCA) can project high-dimensional data into a low-dimensional space, retaining the main features of the data while reducing the dimension and noise of the data. Before power prediction, PCA processing can be performed on the input data before subsequent algorithm steps.

[0133] Independent component analysis (ICA) can separate mixed signals into independent components and is suitable for processing multi-source data or data with noise interference. You can try to apply ICA to multi-source data processing of wind farms, such as the fusion of wind speed, wind direction, temperature and other data to improve the accuracy of power prediction.

[0134] Conduct deeper analysis on the data to discover potential patterns and regularities in the data in order to better adjust algorithms and parameters.

[0135] Time series analysis methods such as autoregressive moving average (ARIMA) and seasonal ARIMA can be used to model and predict power data. By analyzing the time series characteristics of power data, trends, seasonality and other regularities in the data can be discovered, thereby better adjusting algorithms and parameters.

[0136] Data clustering analysis can also be performed to divide the data into different categories, and different algorithms and parameters can be used for different categories of data. Data can be divided into different weather modes based on factors such as wind speed and wind direction, and then power prediction and accuracy evaluation can be performed for different weather modes.

[0137] 5. Re-evaluate accuracy 1. After adjusting the parameters or improving the method, re-predict the power and re-evaluate the accuracy.

[0138] Detailed description of the re-evaluation process. Use the adjusted parameters or improved method to perform power forecasting and obtain new forecast results. Then, follow the previous steps to calculate the WS-MAPE value and compare it with the threshold If the accuracy is up to standard, the final power prediction result is output; if the accuracy is not up to standard, the parameters are adjusted or the method is improved until the accuracy is up to standard.

[0139] 2. Repeat the above steps until the accuracy reaches the target.

[0140] The importance of repeated steps is emphasized. Since power forecasting is a complex problem, it may take multiple adjustments to parameters and improvements to achieve satisfactory accuracy. By continuously repeating the evaluation and improvement process, the accuracy of power forecasting can be gradually improved to meet the actual application needs of wind farms.

[0141] Through the algorithm steps, a more comprehensive and accurate accuracy evaluation can be carried out on the revised power forecast results, and effective parameter adjustments and method improvements can be made based on the evaluation results to improve the accuracy of ultra-short-term power forecasting of wind farms.

[0142] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A method for optimizing ultra-short-term wind farm power forecasting based on lidar, characterized by: The method includes: S1. Scan the wind farm area using laser radar and obtain wind speed and direction data based on the coherent Doppler wind measurement principle of laser radar; S2, wind speed extrapolation prediction based on the algorithm combining EMD and ARCH; S3, using the algorithm combining SVR and GA to make preliminary power prediction; S4, processing and correcting the preliminary power prediction results based on the FNN algorithm; S5. According to the relevant specifications, the accuracy of the revised power prediction results shall be evaluated and the prediction error and other indicators shall be calculated; If the accuracy is not up to standard, return to the previous step to adjust the algorithm parameters or improve the method; S6. When the accuracy meets the requirement, output the wind power forecast time series power value within the next 3-5 hours.

2. According to the method for optimizing ultra-short-term wind farm power forecasting based on laser radar according to claim 1, it is characterized in that: in step S2: Firstly, the wind speed time series collected by the lidar is decomposed into multiple intrinsic mode functions (IMFs) and a residual term using EMD. Then, an ARCH model is built for each IMF to capture the volatility characteristics and autocorrelation of wind speed; Finally, the prediction results of each IMF and residual term are combined to obtain the online prediction results of ultra-short-term plane wind speed.

3. According to claim 2, a method for optimizing ultra-short-term wind farm power forecasting based on laser radar is characterized by: The specific steps of wind speed extrapolation prediction based on the algorithm combining EMD and ARCH in step S2 are: S21, collecting wind speed time series data collected by the laser radar over a period of time; S22. The EMD algorithm construction method is: Identifying Local Maxima and Minima: Finding Wind Speed ​​Time Series All local maxima and minima in ; Calculate the upper and lower envelopes: Connect the local maximum and minimum points respectively through cubic spline interpolation to form the upper envelope and lower envelope ; Calculate mean envelope: Calculate the mean of the upper and lower envelopes ; Extract the first IMF: Subtract the mean envelope from the original data to get , repeat the above steps until The IMF condition is satisfied, that is, the number of extreme points and zero crossing points of the function is equal or differs by at most one, and at any point, the mean of the envelope defined by the local maximum point and the local minimum point is zero. At this time, This is the first IMF, denoted by ; Calculate the residual term by subtracting the first IMF from the original data to obtain ; Repeat the above S22 step: Repeat the above process to get the second IMF and the new residual , and continue to repeat until the residual term can no longer be decomposed. Finally, the wind speed time series is decomposed into IMF and a residual ,Right now ; S23. The steps for building an ARCH model for each IMF are: , satisfies the ARCH(p) model, that is ,in is the mean, is the error term, and , are independent and identically distributed random variables, usually assumed to be standard normal distribution, ,here is the order of the model, is the parameter to be estimated; The parameters of the ARCH model are estimated using methods such as maximum likelihood estimation. It follows a standard normal distribution, then The conditional probability density function is ; By maximizing the log-likelihood function To estimate the parameters; and , used to describe the structure of the error term and the dynamic changes of the variance in the ARCH model; S24. Forecast: Use the estimated ARCH model parameters to forecast each IMF for a period of time in the future and obtain the forecast value ; For the residual Perform simple trend extrapolation or other appropriate forecasting methods to obtain the forecast value ; The final wind speed forecast is ; Through the above S21-S24 algorithm steps, the wind speed extrapolation prediction based on the combination of EMD and ARCH models can be realized, thereby improving the accuracy of ultra-short-term planar wind speed prediction in wind farms.

4. The method for optimizing ultra-short-term wind farm power forecasting based on laser radar according to claim 1, characterized in that: in step S3: First, the nonlinear relationship model between input parameters such as wind speed and wind direction and power output is established using SVR; Then, GA is used to optimize the parameters of SVR to improve the prediction performance of the model; In this way, a preliminary wind power forecast is made based on the data from the existing wind measuring equipment in the wind farm.

5. The method for optimizing ultra-short-term wind farm power forecasting based on laser radar according to claim 4 is characterized by: Step S3 uses an algorithm combining SVR and GA to perform preliminary power prediction. The specific steps are: S31. Collect data from existing wind measuring equipment in the wind farm, including wind speed, wind direction, and corresponding actual power output data; divide the data into a training set and a test set, denoted as ,in Indicates Wind speed of samples and wind direction , Indicates the corresponding power output; The steps of constructing S32 and SVR models are as follows: Select the generalized exponential kernel function, whose expression is ,in and is a parameter to be determined; The basic form of the SVR model: Assume that the training sample set is , the goal of SVR is to find a function , so that for all sample points ,satisfy ,in is the weight vector, is a function that maps input data to a high-dimensional feature space, is the bias term, is a pre-set insensitive loss parameter; Construction of optimization problem: by introducing slack variables and , the optimization problem of SVR can be expressed as: ； subject to ; ； ,in is the penalty parameter; S33, GA optimization SVR parameter method is: SVR parameters , , kernel function parameters and Encoding and forming chromosomes; A set of chromosomes is randomly generated as the initial population. The population size can be adjusted according to the complexity of the problem and the computing resources. The prediction performance of SVR on the training set is used as the fitness function, and the inverse of the square error and mean absolute error indicators are used as the fitness value, that is, the higher the fitness value, the better the model performance; Roulette wheel selection and other methods are used to select excellent chromosomes to enter the next generation according to their fitness values. Chromosomes with high fitness values ​​are more likely to be selected. Perform crossover operation on the selected chromosome to generate new chromosome; When a certain number of iterations is met or the fitness value reaches a certain threshold, the algorithm terminates, otherwise, it returns to the previous step to continue the genetic operation; S34. Use the optimized SVR parameters to establish a power prediction model, predict the test set data, and obtain preliminary wind power prediction results.

6. According to claim 1, a method for optimizing ultra-short-term wind farm power forecasting based on laser radar is characterized by: In S4: A fuzzy neural network structure is constructed, and the wind speed predicted by the lidar is used as one of the inputs, together with the preliminary power prediction value, into the network; Through fuzzy reasoning and the learning ability of neural networks, the power forecast results are automatically adjusted to improve the accuracy of ultra-short-term power forecasts for wind farms.

7. The method for optimizing ultra-short-term wind farm power forecasting based on laser radar according to claim 6 is characterized by: The specific method for processing and correcting the preliminary power prediction result based on the FNN algorithm in step S4 is: S41. Collect the wind speed data predicted by the laser radar, recorded as ,in Indicates The predicted wind speed value at a time point; the preliminary power prediction result is obtained, recorded as ,in Indicates wind speed The corresponding preliminary power prediction value; S42, the input layer has two nodes, which input the wind speed value predicted by the lidar and preliminary power prediction , for wind speed input , let its fuzzy set be , the Gaussian membership function is used for fuzzification, that is, ,in and They are fuzzy sets about the center and width parameters of wind speed; For the preliminary power prediction , the fuzzy set is , the Gaussian membership function is also used for fuzzification, that is, ,in and They are fuzzy sets with respect to the center and width parameters of power; The number of nodes in this layer is , each node represents a fuzzy rule, and the output of the node is the excitation intensity of the corresponding rule, which is calculated by multiplying the membership of each input fuzzy set, that is, ; This layer has only one node, which is used to output the corrected power value and use the weighted average method for defuzzification, that is, ,in is the output power value of the corresponding rule obtained through neural network learning; S42, using mean square error as the error function, that is ,in It is The corrected power value of samples, It is The actual power value of each sample; The weights and biases of the neural network are randomly initialized, and for each training sample, the forward propagation calculates the output power value , calculate the error according to the error function, and then back-propagate to update the weights and biases; The update formula is and ,in is the learning rate, and are the gradients of the error with respect to the weight and bias respectively; S43, input the new laser radar predicted wind speed value and the preliminary power prediction value into the trained fuzzy neural network; after fuzzification, rule excitation, defuzzification and forward propagation of the neural network, the revised power prediction result is obtained.

8. The method for optimizing ultra-short-term wind farm power forecasting based on laser radar according to claim 1 is characterized by: The specific method for evaluating the accuracy of the corrected power prediction result in step S5 is: S51, using WS-MAPE for accuracy evaluation, the power prediction value after correction is The actual power value is , the calculation formula of WS-MAPE is: ,in It is the weight determined based on factors such as the importance or time sequence of the data; S52. According to the specific requirements and relevant specifications of the wind farm, determine the accuracy evaluation standard and set the threshold of WS-MAPE as ,like , then the accuracy is considered to meet the standard; otherwise, the accuracy is not up to standard; S53, calculating WS-MAPE: calculating the WS-MAPE value of the corrected power prediction result according to the formula; S54. Compare the calculated WS-MAPE value with the set threshold Make comparisons; like , then the accuracy meets the standard and the final power prediction result is output; If the accuracy does not meet the standard, proceed to the next step; S55, if the degree does not meet the standard, you can return to the steps of S2-S5 and adjust the parameters in the algorithm. Specifically, in the wind speed extrapolation prediction based on the combination of EMD and ARCH models, you can adjust the order of the ARCH model and the parameters of the EMD decomposition; In the preliminary power prediction combining SVR and GA, the penalty parameter and kernel function parameter of SVR as well as the population size, crossover probability and mutation probability parameters of genetic algorithm can be adjusted; In the power result correction of FNN, the number, center and width parameters of fuzzy sets, as well as the learning rate, number of layers and number of neurons of the neural network can be adjusted; S56. After adjusting the parameters or improving the method, re-predict the power and re-evaluate the accuracy. Repeat the above steps S2-S5 until the accuracy reaches the target.

Images