The preferred embodiments will be described in detail below in conjunction with the accompanying drawings. It should be emphasized that the following description is only exemplary and not intended to limit the scope of the invention and its application.
 figure 1 It is an overall flow chart of short-term power prediction of wind farms provided by the present invention. Wind farm short-term power prediction includes three steps: a. Ultra-short-term (0-4h) power prediction; taking the measured wind speed, power and other data provided by the wind farm SCADA (Supervisory Control And Data Acquisition, data acquisition and monitoring) system as input, through Data processing methods such as the time series method and the power prediction method based on the combination of wavelet transform and neural network, output the predicted power of 0-4h, and the time resolution is 15 minutes; b. Short-term (0-72h) power prediction; NWP ( Numerical weather forecast) data provided by the system at different heights such as wind speed, wind direction, temperature, humidity and atmospheric pressure as input, through BP neural network, adaptive fuzzy neural inference system (ANFIS), least squares support vector machine (LS-SVM) and other data processing methods, output the forecast power of 0-72h, and the time resolution is 15 minutes; c. Combination forecast; take ultra-short-term (0-4h) forecast power and short-term (0-72h) forecast power as input, through BP-based The combined prediction method of the neural network is used to output the final short-term forecast power of the wind field within 0-72 hours, and the time resolution is 15 minutes.
 figure 1 Among them, the specific implementation process of the wind farm short-term power prediction method is as follows:
 Step 1: Calculate the ultra-short-term forecast power.
 figure 2 Is the flow chart of ultra-short-term power forecasting. Ultra-short-term power forecast refers to the power forecast within 0-4 hours, figure 2 In , the specific process of the ultra-short-term power forecasting process is:
 Step 101: Acquire wind speed and power data of the wind farm in real time, and perform preprocessing on the acquired data; the preprocessing includes eliminating erroneous data and normalizing data.
 The data required by the present invention are divided into dynamic data and static data. The dynamic data is the real-time wind speed and power value of the wind field, which comes from the SCADA system of the wind field and is obtained through the data communication interface; the static data is the start and stop of the wind field unit and the power curve of the fan. etc., can be realized through human-computer interaction and manual entry.
 In order to improve the learning accuracy and efficiency of the prediction model, it is necessary to preprocess the acquired data and remove the wrong data, mainly including the fault data of the wind measuring instrument of the wind turbine, the normal or abnormal shutdown data of the wind turbine, and the communication failure data, etc. At the same time, because of the saturation problem in neuron training, it is necessary to normalize the acquired data (including the acquired data and the entered data). In this embodiment, the data is normalized to the [0.1, 0.9] interval, which is realized by the following formula :
 Y = 0.1 + 0.8 × X - X min X max - X min
 In the formula: X is the acquired data, such as the measured power value; X minis the minimum value of the acquired data; X max is the maximum value of the obtained data; Y is the normalized result of the obtained data.
 Step 102: Predict the power using the differential autoregressive moving average model.
 Commonly used time series methods include continuous method, Kalman filter method, autoregressive moving average method (ARMA) and differential autoregressive moving average method (ARIMA). For the characteristics of frequent wind speed fluctuations and obvious non-stationarity, this embodiment ARIMA method is used for power prediction. ARIMA is composed of three parts: autoregressive item (AR), difference item (I), moving average item (MA), through historical data analysis and model parameter identification, the present embodiment establishes the model as ARIMA(3,1,2) . Among them, 3 is the order of the autoregressive model, 1 is the order of the moving average model, and 2 is the order of the difference.
 Step 103: Using wavelet transform combined with neural network technology to predict power.
 The main purpose of wavelet transform is to analyze the local characteristics of nonlinear and non-stationary signals. After a known basic function ψ(t) is translated and scaled, it is compared with the analyzed signal (realized by integration), and the signal can be analyzed in each Time, local characteristics of various local ranges; neural network has a strong nonlinear mapping ability, through the learning of sample pairs, it can realize the mapping from the input n-dimensional space to the output m-dimensional space.
 In the present invention, the wavelet analysis is used as the pre-processing means of the BP neural network to provide the input feature vector for the neural network, that is, the signal is input to the BP neural network after wavelet transformation. The implementation steps are:
 Step A: performing n-layer wavelet decomposition on the time series composed of the acquisition time corresponding to the preprocessed power data. In this embodiment, n=6 to obtain n detail signal components and 1 approximation signal component.
 Step B: use the BP neural network to build models for the decomposed n+1 signals respectively, and make predictions.
 Step C: The prediction data of n+1 signal components are superimposed to obtain a final prediction result.
 Step 104: Using the linear combination forecasting method, perform weighted optimization on the predicted power results of steps 102 and 103 to obtain ultra-short-term predicted power.
 The combined forecasting method is to combine different forecasting models and methods, comprehensively utilize the information provided by various forecasting methods, and obtain a combined forecasting model in an appropriate weighted average way, which can maximize the usefulness of various single forecasting methods. information that can increase the predictive accuracy of the system.
 The core of combined forecasting methods is how to properly combine various forecasting methods, so the key lies in how to obtain the weighted average coefficients of various forecasting methods. In this method, the minimum root mean square error is used as the optimization target, and the prediction results of the ARIMA model in step 102 and the method of combining wavelet transform and neural network in step 103 are combined for prediction, and the combined prediction value and weight coefficient are obtained as follows: in i=1,2,...,m,f c is the combination forecast value; m is the number of forecast methods, f i is the predicted value of the i-th method; w i is the weighting coefficient of the i-th prediction method; e i and Var(e i ) are the prediction error and variance of the i-th prediction method, respectively.
 image 3 It is a flow chart of short-term power forecasting. Short-term power forecasting refers to power forecasting within 0-72 hours. image 3 Among them, the short-term forecast power is calculated through the weather forecast value provided by the numerical weather forecast system, which specifically includes:
 Step 201: Obtain weather forecast value and perform preprocessing.
 The present invention obtains the 72h weather forecast value output by the numerical weather forecast system (NWP) through the data communication interface, including wind speed, wind direction, air temperature at different heights such as 0 meters, 30 meters, 50 meters, 70 meters, 100 meters, 120 meters, etc. , humidity and atmospheric pressure and other data, the time resolution is 15 minutes.
 In order to improve the learning accuracy and efficiency of the prediction model, it is necessary to preprocess the acquired data. The measures taken are: (1) With the help of the actual observation data of wind resources, check and sum the mapping relationship between the meteorological data and the weather at representative observation points Adjustment; (2) Combining with the operation data of the built wind farm, correct the mapping relationship between the weather of the representative observation point and the output of the wind farm. At the same time, invalid or erroneous data needs to be eliminated, mainly including wind farm power cut data, abnormal shutdown data of wind turbines, communication failure data, etc.
 Step 202: For each type of fan type in the wind farm, three prediction models, BP neural network, adaptive fuzzy neural reasoning system and least squares support vector machine, are used to predict the power.
 There may be multiple types of wind turbines in the wind farm, and the power prediction based on multiple prediction models is performed for each type of wind turbine. The prediction models used include BP neural network, adaptive fuzzy neural inference system (ANFIS), least squares support vector There are three kinds of machines (LS-SVM).
 (1) BP neural network
 BP neural network refers to a multi-layer forward neural network based on the error backpropagation algorithm. It can approach any nonlinear mapping with arbitrary precision, has a distributed information storage and processing structure, and has certain fault tolerance and good robustness. It has been widely used in the field of wind power forecasting.
 In this method, the BP neural network adopts a three-layer network structure containing only one hidden layer, and the number of neurons in the input layer is 12. According to the NWP forecast data, the wind speed at 30 meters, the sine of the wind direction, the cosine of the wind direction, and the wind speed of 50 meters , wind direction sine, wind direction cosine, wind speed at 70 meters, wind direction sine, wind direction cosine, temperature, air pressure, humidity. The number of neurons in the output layer is 1, which is the predicted power value. The number of neurons in the middle layer is determined by a trial algorithm with the goal of minimizing the root mean square error of the training samples, and it is 27. The transfer function of hidden layer neurons adopts S-type tangent function, the transfer function of output layer neurons adopts S-type logarithmic function, and the training algorithm adopts LM algorithm.
 (2) Adaptive Fuzzy Neural Inference System (ANFIS)
 ANFIS is a fuzzy reasoning system based on the Sugeno model. Its core is a neuron-fuzzy model, which organically combines the self-learning function of the artificial neural network and the fuzzy language expression ability of the fuzzy reasoning system to complement each other. The fuzzy membership function and fuzzy rules are completed through the learning of a large number of known data, and do not need to rely on expert experience to manually determine in advance.
 In the present invention, ANFIS is a five-layer first-order Sugeno fuzzy system, and the concrete construction process is as follows: the first layer is a fuzzy layer, and the membership function of the fuzzy set is a triangular function; the second layer performs the calculation of the fuzzy rule excitation intensity; the third layer performs The normalized calculation of the applicability of each rule; the fourth layer is used to calculate the output of each rule, and the subsequent item (conclusion) output function is a linear function; the fifth layer is used to calculate the total output of the system. Its parameter learning adopts a hybrid learning algorithm to shorten the training time of the network. The hybrid learning algorithm is an algorithm that adds a least square estimator to the original MN algorithm.
 (3) Least squares support vector machine (LS-SVM)
 The biggest feature of support vector machine (SVM) is that it can effectively overcome the problems of too much deviation of the prediction results of common prediction methods and the existence of over-learning, dimensionality disaster and local extremum. LS-SVM is an improvement to SVM, which combines traditional SVM The inequality constraint in is changed to an equality constraint, and the error square sum loss function is used as the empirical loss of the training set, so that the solution of the quadratic programming problem is transformed into the problem of solving the linear equation system, thereby improving the speed of solving the problem and the convergence accuracy.
 In the case of linear inseparability, the selection of kernel function in LS-SVM is very critical, and the choice of kernel function directly affects the realization and effect of the algorithm. Commonly used kernel functions mainly include: polynomial kernel function, RBF radial basis kernel, Gaussian radial basis function and Sigmoid kernel function etc., the present invention is respectively calculated for four kinds of kernel functions, through the analysis and comparison to calculation result, selected Gaussian The radial basis function is used as the kernel function, and the formula is:
 K ( x k , x j ) = exp ( - | x - x i | 2 δ 2 )
 After selecting the kernel function, there are two parameters that LS-SVM needs to choose, namely the hyperparameter γ and the kernel parameter σ, where γ determines the size of the training error and the strength of the generalization ability, and σ determines the width of the local field . In the parameter selection process, by comparing the three-step search method with the global search, it is found that the three-step search method can find the optimal [γ, σ] in different value ranges, so the present invention adopts the three-step search method to Determine two parameters.
 Step 203: Using the power combination prediction model based on the principle of maximum information entropy, weighted optimization is performed on the prediction results of step 202.
 From the perspective of information theory, the combined forecasting process is a comprehensive process of information, that is, the statistical characteristics of the predicted quantity are obtained from the forecasting results of various single forecasting models, and as the information provided to the combined forecasting model, the principle of maximum information entropy is applied Based on this information, objective forecasts can be made about future forecast values.
 The present invention adopts the power combination prediction model based on the principle of maximum information entropy, and its basic working process is: (1) use various single prediction models to carry out wind farm power prediction; (2) use the actual value of wind farm power as the wind farm to be predicted The central point of the power, the central moment of each order of the wind farm power is obtained. Since the distribution of wind speed and power does not satisfy the normal distribution, in addition to the second-order central moment as the statistical characteristic quantity, the information of the third-order central moment and the fourth-order central moment also needs to be added; (3) the wind farm obtained by various prediction models Statistical characteristics of power are used as constraint information, and the principle of maximum information entropy is used to solve the problem.
 Step 204: Considering the influence of the number of start-up and shutdown of wind turbines, predict the total power of the multi-model wind turbines in operation in the wind farm; the total power of the wind turbines in operation is the short-term predicted power.
 The above step 203 obtains the predicted power of a single fan of a certain type. In order to obtain the total predicted power of the wind farm, the influence of the number of start-up and shutdown of the wind must be considered. The information of the operating units is manually entered into the system by the staff through human-computer interaction.
 According to the fan startup and shutdown plan entered manually, use the following formula to obtain the total predicted power of a certain type of fan:
 P i =N i P i '
 In the formula: N i is the number of fans of type i in operation, P i ′ is the predicted power of the i-th type fan unit, P i is the total predicted power of the i-th type fan. Then the total predicted power of the wind field is:
 Step 3: The weighted optimization of ultra-short-term forecast power and short-term forecast power is obtained from the calculation of step 1 and step 2, and the final short-term forecast power is calculated by using the combined forecast method based on BP neural network.
 exist figure 1In the combined prediction based on BP neural network, the results of "ultra-short-term (0-4h) power prediction" and the first 4h of "short-term (0-72h) power prediction" are input as input, and processed by the combined optimization method of BP neural network to obtain The final short-term predicted power of the wind field within 0-72 hours. Its BP neural network structure is shown in Figure 4 , the BP neural network adopts a three-layer network structure including only one hidden layer, the number of neurons in the input layer is 4, and the number of neurons in the output layer is 1, which is the predicted power value. The algorithm is determined with the minimum root mean square error of the training samples as the goal, and there are 6. The transfer function of hidden layer neurons adopts S-type tangent function, the transfer function of output layer neurons adopts linear function purelin, and the training algorithm adopts LM algorithm.
 Figure 4 Each neuron in the middle input layer is x, y, ε x , ε y , the specific meaning is:
 x: ultra-short-term (0-4h) power prediction results;
 y: short-term (0-72h) power prediction result;
 ε x : Average error of short-term ultra-short-term (0-4h) power forecast;
 ε y : Average error of short-term (0-72h) power forecast in the near future.
 The importance of different input parameters for the prediction results is different. Therefore, each input parameter is multiplied by a different weight coefficient to improve the prediction accuracy, and the weight coefficient is manually set. Since within 0-1h, the accuracy of ultra-short-term power prediction results is higher than that of short-term power prediction results, so within 0-1h, x and ε x The weight value is larger, and within 1-4h, y and ε y The weight value is greater.
 The present invention takes ultra-short-term power prediction results, average error and short-term power prediction results, and average error as input, and processes them through a combined prediction method based on BP neural network to obtain the final short-term predicted power of the wind farm. This method improves the future of the wind farm by 72 The accuracy of hourly predicted power provides a reliable basis for the power grid to formulate a reasonable dispatch plan.
 The above is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of changes or modifications within the technical scope disclosed in the present invention. Replacement should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the protection scope of the claims.