Combined model-based fan output short-term prediction method and system

Abstract

The invention discloses a fan output short-term prediction method and system based on a combined model, and the method comprises the steps: carrying out the stabilization of obtained fan output time series data, and obtaining first time series data; inputting the first time sequence data into a preset differential autoregressive moving average model for rule prediction to obtain first prediction data; inputting the first prediction data which does not meet the preset regularity test condition into a preset long-short memory network for regularization prediction to obtain second prediction data; and according to the first prediction data and the second prediction data meeting the regularity test condition, performing prediction through a preset radial basis function neural network to obtain a prediction value of fan output. According to the method, stable and continuous output time sequence data are obtained through prediction of the combined model, so that a more accurate fan output prediction value is predicted.

Description

Technical Field

[0001] This application belongs to the field of short-term power output prediction of power grids, and particularly relates to a short-term prediction method and system for wind turbine power output based on a combined model. Background Technique

[0002] Under the guidance of the "dual carbon" goal, the development of new energy has entered the fast lane, especially wind power. The incorporation of a large number of wind turbines into the power grid has made the radiation area of wind power grow day by day, forming a huge wind power grid. However, due to the characteristics of strong irregularity and difficulty in prediction of wind power, the randomness of the output of wind turbines is caused, making it impossible to support the voltage stability of the power grid. At the same time, affected by factors such as wind speed changes and turbulence, the output of wind turbines will fluctuate and intermittently start and stop, and even suddenly automatically shed load from the full-load operation state. Because the randomness and unpredictability of wind power will bring problems to the arrangement and implementation of traditional dispatching and power generation plans. Therefore, the prediction of the output of wind turbines in the power system becomes more important.

[0003] Traditional wind turbine power output prediction algorithms always appear in the form of single algorithms. However, when facing the complex current situation of wind turbine power output prediction, due to the great difficulty in collecting data for wind power prediction and the frequent occurrence of missing and abnormal data, a single such model is difficult to predict and analyze the wind turbine power output. Summary of the Invention

[0004] This application proposes a short-term prediction method and system for wind turbine power output based on a combined model, which identifies the regularity of the wind turbine power output time series through the combined model to predict a more accurate wind turbine power output prediction value.

[0005] The first aspect of this application provides a short-term prediction method for wind turbine power output based on a combined model, and the method includes:

[0006] Perform stationary processing on the obtained wind turbine power output time series data to obtain the first time series data;

[0007] Input the first time series data into a preset autoregressive integrated moving average model for regularity prediction to obtain the first prediction data;

[0008] Input the first prediction data that does not meet the preset regularity test condition into a preset long short-term memory network for regularization prediction to obtain the second prediction data;

[0009] According to the first prediction data and the second prediction data that meet the regularity test condition, perform prediction through a preset radial basis neural network to obtain the prediction value of the wind turbine power output.

[0010] The above solution first performs a stationary processing on the obtained original fan output time series data to obtain first time series data with a certain degree of stationarity, which is convenient for subsequent time series prediction; then, through the autoregressive integrated moving average model, it predicts and identifies the time series data with a certain periodicity in the first time series data to obtain first prediction data containing high-precision regularity; then, the irregular data in the first prediction data is input into the long short-term memory network for regularized prediction, and second prediction data with a certain degree of regularity is predicted from the irregular time series data, providing accurate data support for subsequent output prediction; finally, the first prediction data and the second prediction data that contain a certain degree of regularity and have passed the residual white noise detection are used to perform output prediction through a radial basis neural network in combination with the regularity of the data, and a predicted value of the fan output with high precision is obtained.

[0011] In a possible implementation method of the first aspect, the obtained fan output time series data is subjected to stationary processing to obtain first time series data, specifically:

[0012] Preprocess the obtained fan output time series data;

[0013] Perform a stationarity test on the preprocessed fan output time series data, and then perform differencing on the fan output time series data that fails the stationarity test to obtain the stationary fan output time series data, denoted as the first time series data.

[0014] The above solution first preprocesses the fan output time series data, corrects the missing data and abnormal data therein, and then performs stationary processing on the preprocessed data to obtain stationary time series data. Since the number of parameters to be estimated in the stationary data is small, the data operation efficiency can be greatly improved.

[0015] In a possible implementation method of the first aspect, the first time series data is input into a preset autoregressive integrated moving average model for regular prediction to obtain first prediction data, specifically:

[0016] Perform autocorrelation analysis on the first time series data to obtain autocorrelation parameters;

[0017] According to the set model prediction accuracy, combined with the autocorrelation parameters, determine the first model parameters of the autoregressive integrated moving average model;

[0018] According to the first model parameters, use the autoregressive integrated moving average model to predict the regularity of the first time series data to obtain first prediction data.

[0019] The above solution first performs autocorrelation analysis on the first time-series data to obtain autocorrelation parameters that can reflect the internal variables of the data, providing data support for obtaining the regularity of the first time-series data. Then, by combining the model prediction accuracy and the autocorrelation parameters, the first model parameters of the autoregressive integrated moving average (ARIMA) model are restricted. While reducing the complexity of the model by reducing the input model parameters, the model prediction effect is ensured. Finally, based on the first model parameters, the regularity of the first time-series data is predicted to obtain regular first prediction data.

[0020] In a possible implementation method of the first aspect, the first prediction data that does not meet the preset regularity test condition is input into a preset long short-term memory (LSTM) network for regularity prediction to obtain second prediction data. Specifically:

[0021] According to the preset regularity test condition, the first prediction data is tested, and the irregular time-series data is identified from the first prediction data that does not meet the regularity test condition.

[0022] According to the irregular time-series data, the second model parameters of the LSTM network are set.

[0023] According to the second model parameters, the irregular time-series data is regularly predicted to obtain second prediction data.

[0024] The above solution screens out the irregular time-series data with large randomness and no certain periodicity from the first prediction data through the regularity test condition, and then constructs the regularity of the irregular time-series data through the LSTM network. By selectively retaining and deleting the data, certain regularity is identified from the data with weak regularity, providing accurate data support for subsequent output prediction.

[0025] In a possible implementation method of the first aspect, the regularity test condition is specifically:

[0026] The specific expression of the regularity test condition is:

[0027]

[0028] In the formula, E(x 1 ) is the mean of x 1 in the first prediction data, E(x 2 ) is the mean of x 2 in the first prediction data, V(x 1 ) is the variance of x 1 in the first prediction data, V(x 2 ) is the variance of x 2 in the first prediction data, σ is the average degree of fluctuation of the first prediction data around its mean, Cov(xt , x t-s ) is the covariance between x in the first prediction data t and x t-s .

[0029] In a possible implementation method of the first aspect, according to the second model parameter, the irregular time series data is regularly predicted to obtain second prediction data. Specifically:

[0030] According to the second model parameter at the current time point, update the unit state at the current time point in the irregular time series data;

[0031] According to the unit state at the current time point, determine the output value at the current time point in the irregular time series data, and update the second model parameter at the next time point;

[0032] When the output values at all time points in the irregular time series data are determined, the second prediction data is obtained.

[0033] The above solution updates the unit state at the current time point in the irregular time series data through the second model parameter at each time point, and then updates the corresponding output value through the unit state, realizing selective retention and deletion of data, highlighting the regularity of the overall time series data. Then, according to the output value at the current time point, the second model parameter at the next time point is updated, making the time series data at the upper and lower time points have a certain continuity.

[0034] In a possible implementation method of the first aspect, it further includes:

[0035] Integrate the first prediction data and the second prediction data that meet the regularity test conditions and input them into the radial basis neural network to construct a fan output power prediction model;

[0036] Obtain the predicted value of the fan output power through the fan output power prediction model.

[0037] The second aspect of the present application provides a short-term prediction system for fan output power based on a combined model. The system includes: a stationary processing module, a regular data prediction module, an irregular data prediction module, and an output data prediction module;

[0038] Among them, the stationary processing module is used to perform stationary processing on the obtained fan output power time series data to obtain first time series data;

[0039] The regular data prediction module is used to input the first time series data into a preset autoregressive integrated moving average model for regular prediction to obtain first prediction data;

[0040] The irregular data prediction module is used to input the first prediction data that does not meet the preset regularity test conditions into a preset long short-term memory network for regularized prediction to obtain the second prediction data;

[0041] The output data prediction module is used to perform prediction through a preset radial basis neural network based on the first prediction data and the second prediction data that meet the regularity test conditions to obtain the predicted value of the fan output.

[0042] A third aspect of the present application provides a terminal device, including: a processor and a memory, where the memory stores a computer program, and when the processor executes the computer program, the steps of a short-term prediction method for fan output based on a combined model according to any one of the embodiments of the present application are implemented.

[0043] A fourth aspect of the present application provides a storage medium that stores computer-readable program codes, and when the computer-readable program codes are executed, the steps of a short-term prediction method for fan output based on a combined model according to any one of the embodiments of the present application are implemented. Description of the Drawings

[0044] To more clearly illustrate the technical solutions of the present application, the drawings required for implementation will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0045] Figure 1 is a specific flowchart of a short-term prediction method for fan output based on a combined model provided by an embodiment of the present application;

[0046] Figure 2 is a structural diagram of an LSTM model of a short-term prediction method for fan output based on a combined model provided by an embodiment of the present application;

[0047] Figure 3 is a structural diagram of an RBF neural network of a short-term prediction method for fan output based on a combined model provided by an embodiment of the present application;

[0048] Figure 4 is a structural diagram of a short-term prediction system for fan output based on a combined model provided by an embodiment of the present application;

[0049] Figure 5 is a structural diagram of a terminal device provided by an embodiment of the present application. Detailed Embodiments

[0050] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0051] It should be understood that the step numbers used in the text are only for convenient description and are not intended to limit the order of execution of the steps.

[0052] First Embodiment

[0053] The current power grid dispatching and power generation plans are formulated and implemented based on conditions such as the safety and reliability of the power grid and predicted load. Among them, the assessment of whether the output of wind turbines can support the voltage stability of the power grid is particularly important. Due to the randomness and unpredictability of wind power, it is difficult to cope with the complex wind turbine output situations by relying on a single model for data prediction. Multiple models need to be used to accurately predict different characteristics of the output data in order to obtain better prediction of data with strong periodicity and volatility.

[0054] As Figure 1 shown, Figure 1 This is a schematic flow chart of a specific method for short-term prediction of wind turbine output of a combined model provided by an embodiment of the present application. The method for short-term prediction of wind turbine output of the combined model in this embodiment includes steps S1 to S4, which are described in detail as follows:

[0055] Step S1, perform stationary processing on the obtained time series data of wind turbine output to obtain the first time series data.

[0056] In the embodiments of the present application, since the obtained original time series data of wind turbine output often has missing values and data anomalies, making it inconvenient for time series data analysis. In order to make the original time series data of wind turbine output have a certain periodicity or stationarity before prediction analysis and reduce the complexity of data processing, it is necessary to preprocess the collected original time series data of wind turbine output before prediction calculation to complete the compensation and replacement of missing values and abnormal values.

[0057] The compensation and replacement of missing values and abnormal values in the data are often related to their positions in the time series data. When the missing values and abnormal values exist at the beginning or end of the entire time series data, they can be directly deleted. However, if the missing values and abnormal values exist in the middle position of the time series data, they cannot be directly deleted. There are generally 5 processing methods:

[0058] 1. Sequence average replacement: Use the average value of the entire time series data for compensation and replacement;

[0059] 2. Mean replacement of adjacent points: Use the mean of several adjacent points of the missing value or outlier for compensation and replacement;

[0060] 3. Median replacement of adjacent points: Use the median of several adjacent points of the missing value or outlier for compensation and replacement;

[0061] 4. Linear interpolation replacement: Use multiple adjacent points of the missing value or outlier for linear fitting to obtain the intermediate point value for compensation and replacement;

[0062] 5. Linear trend replacement of adjacent points: Use the number of adjacent periods as the independent variable and the sequence value as the dependent variable for regression to obtain the missing value or outlier.

[0063] Optionally, in the embodiments of the present application, because the wind turbine output time series data has strong seasonality and volatility, in order to ensure data accuracy, the linear interpolation method and the linear trend replacement method of adjacent points are used for data compensation and replacement.

[0064] Then, perform a stationarity test on the preprocessed wind turbine output time series data. Generally, the stationarity of the wind turbine output time series data can be roughly estimated directly by observing the time series diagram, or the stationarity of the wind turbine output time series data can be tested by using the hypothesis testing method. The most common ones are ADF (augmented dickey fuller) unit root, autocorrelation coefficient function (ACF), and partial autocorrelation coefficient function (PACF) tests, etc. In the embodiments of the present application, the wind turbine output time series data needs to meet the following 3 conditions to pass the stationarity test. The specific expression is:

[0065]

[0066] In the formula, x 1 is the value at time 1 in the wind turbine output time series data, x 2 is the value at time 2 in the wind turbine output time series data, E(x 1 ) is the mean of x 1 in the wind turbine output time series data, E(x 2 ) is the mean of x 2 in the wind turbine output time series data, V(x 1 ) is the variance of x 1 in the wind turbine output time series data, V(x 2 ) is the variance of x 2 in the wind turbine output time series data, σ is the average degree of fluctuation of the wind turbine output time series data around its mean value, Cov(xt , x t-s ) is the covariance between x in the fan output time series data t and x t-s . u is the average value of the fan output time series data at each moment, and γ is the direct correlation between the fan output time series data at times t and s.

[0067] Because stationary time series data has the characteristics that the mean is a constant, the variance is the mean, and the autocovariance function and autocorrelation coefficient only depend on the translation length of time and are independent of the starting point of time, all the data of the fan output time series data can be used in this application to estimate the overall mean and variance.

[0068] Then, perform multiple differencing processes on the fan output time series data that fails the stationarity test, and finally obtain the stationary fan output time series data, denoted as the first time series data.

[0069] Step S2: Input the first time series data into a preset autoregressive integrated moving average model for regular prediction to obtain the first prediction data.

[0070] In the embodiments of this application, because the stationarity of the first time series data after differencing is strong, the autoregressive integrated moving average model that is sensitive to stationary data is used as the preferred prediction model. In the embodiments of this application, the ARIMA model is used as the autoregressive integrated moving average model. Among them, the ARIMA model is fully called the autoregressive integrated moving average model, which is a time series prediction method. The basic idea is to regard the data sequence formed by the prediction object over time as a random sequence and approximate it with a certain mathematical model. Once this model is identified, the future value can be predicted from the past and current values of the time series. The ARIMA model has the characteristics of high accuracy and fast speed when dealing with data with strong periodicity and time regularity. Also, because most wind power generation power data has strong periodicity, the ARIMA model is selected as the preferred model in the embodiments of this application.

[0071] The ARIMA model contains three very important internal variables, which are called autocorrelation parameters in the embodiments of this application, namely parameters p, d, and q. p represents the lag number of the time series data itself used in the model, also called the autoregressive term; d represents the number of orders of differencing required for the time series data to meet stationarity, also called the differencing term; q represents the lag number of the prediction error used in the model, also called the moving average term.

[0072] Among them, the expressions of the p autoregressive term, q moving average term, and d differencing term are respectively:

[0073]

[0074] In the formula, y t is the value of the time series data at time t, p is the autoregressive term, μ and μ' are the constant terms, and α i , β i are the model correlation coefficients, q is the moving average term, and ε t , ε' t are the error values. L is the maximum likelihood function of the model, which can also be called the sequence width and reflects the degree of interpretation of the model for the data; d is the differencing term.

[0075] Because the ARIMA model is simple and does not require the aid of other external variables, it can complete predictions only relying on its own internal variables. However, this model is highly dependent on the stationarity of the input sequence, which has a certain impact on its prediction accuracy. Therefore, differencing stationarity processing must be performed.

[0076] Generally speaking, the more parameters added to a time series model, the better the model fitting effect, but this comes at the cost of increasing the model complexity. Therefore, the established model needs to seek a balance between the model complexity and the data interpretation ability. In the embodiments of the present application, the minimum information criterion AIC and the Bayesian information criterion BIC are used as the basis to select parameters, that is, the first model parameters of the ARIMA model. The specific expression is:

[0077] A IC = 2T - 2ln(L);

[0078] B IC = 2ln(T) - 2ln(L);

[0079] In the formula, T is the number of the first model parameters, representing the complexity of the ARIMA model; L is the maximum likelihood function of the model, which can also be called the sequence width and reflects the degree of interpretation of the simulation for the data.

[0080] In addition, since AIC and BIC are more suitable for small models, to achieve the best balance effect between the model complexity and the data interpretation ability, the embodiments of the present application need to obtain the minimum values of A IC and B IC . In the ARIMA model, when p and q increase, T increases accordingly, and when the sequence width increases, L increases. In this way, the minimum values of AIC and BIC can be obtained, thereby determining p and q.

[0081] Exemplarily, in the embodiments of the present application, let p be 1 and q be 1. Therefore, the time series prediction model established based on the ARIMA model is the ARIMA(1,1,1) model.

[0082] Finally, according to the constructed ARIMA model, the sequence law of the first time series data is identified and predicted to obtain the first predicted data.

[0083] Step S3: Input the first prediction data that does not meet the preset regularity test conditions into a preset long short-term memory network for regular prediction to obtain second prediction data.

[0084] In the embodiments of the present application, when encountering a sequence with high randomness, the ARIMA model cannot effectively identify the sequence pattern. At this time, a long short-term memory network with stronger learning ability is used to predict the sequence. Before using the long short-term memory network for prediction, the first prediction data is first subjected to a residual white noise test through a preset regularity test condition to identify the time series data that does not meet the regularity test condition and has low regularity, obtaining irregular time series data.

[0085] The residual white noise test is also called the pure random number test. When a sequence is a pure random number sequence, it is meaningless to analyze it using a time series model. Therefore, after the sequence has completed the stationarity analysis or processing, it is necessary to perform a white noise test on the sequence. In the embodiments of the present application, the residual of the first prediction data is subjected to a white noise test. If the residual white noise test fails, it is considered that it is difficult to completely identify the time pattern of the first prediction data by simply using the ARIMA model, and it is necessary to use the LSTM model as the long short-term memory network for prediction.

[0086] Optionally, the residual white noise test in the embodiments of the present application is processed using the MATLAB built-in function Autocorrelation[·], and this function specifically uses Bartlett's theorem as the theoretical basis. Moreover, the residual white noise test is also used as the combined node of the ARIMA model and the LSTM model, so that these two models form a combined model.

[0087] Among them, the specific expression of the regularity test condition is:

[0088]

[0089] In the formula, E(x 1 ) is the mean of x 1 in the first prediction data, E(x 2 ) is the mean of x 2 in the first prediction data, V(x 1 ) is the variance of x 1 in the first prediction data, V(x 2 ) is the variance of x 2 in the first prediction data, σ is the average degree of fluctuation of the first prediction data around its mean at each moment, Cov(x t ,x t-s ) is the covariance between x t and x t-s in the first prediction data.

[0090] It should be emphasized that the time series satisfying the above formula is actually a special stationary sequence.

[0091] The LSTM model is essentially an improved RNN that can solve the long-distance dependence problem that the RNN cannot solve. Because when the time series is too long, the RNN is very likely to forget the information in the relatively distant past time periods. The closer the time point is, the greater the impact on the current input. To overcome this shortcoming, the LSTM model was born.

[0092] The hidden layer of the original RNN has only one state quantity, denoted as the C state, which is used to store short-term states. In the LSTM model, an additional state is added, denoted as h, which is used to store long-term memories. The LSTM model mimics the memory method of cell neurons to build the model, stores the information and memories of past sequences in the form of a cell state vector, and combines the sequence state of the previous time, the sequence state of the current time, and the hidden state of the previous time to build the long-term historical memory of the long-term neuron.

[0093] The LSTM model mainly uses three gates to achieve prediction. The three gates are the forget gate, the input gate, and the output gate. The forget gate determines how much of the cell state C at the previous moment t-1 will be saved to the current state of C t ; the input gate determines how much of the input X t will be saved to C t ; the output gate determines how much of the current state C of the control unit t will be output to the current output value h t .

[0094] In the embodiments of the present application, the forget gate, the input gate, and the output gate of the LSTM model are set through the irregular time series data.

[0095] To better show the specific structure of the above LSTM model, Figure 2 a structural diagram of the LSTM model is provided. As shown in the figure, a horizontal line at the top of the figure from C t-1 to C t is the main structure for time series prediction. Wf, Wi, Wc, and Wo are the weight matrices of the forget gate, the input layer forget gate, the input gate, and the computational unit state respectively, and e is the sigmoid function. The input quantities of the forget gate and the input gate are f t , i t respectively, and the specific expressions are:

[0096] f t = e(W f [h, x t + u);

[0097] i t = e(W i [h, x t + u);

[0098] In the formula, x t is the input data at time t, and u is a constant.

[0099] Based on Figure 2 the LSTM model in

[0100] perform regular prediction on the irregular time series data to obtain the second prediction data.

[0101] Specifically, according to the forget gate, input gate, and output gate of the LSTM model at the current time point, update the cell state and cell memory state at the current time point in the irregular time series data.

[0102] According to the cell state at the current time point, determine the output value at the current time point in the irregular time series data, and update the forget gate, input gate, and output gate of the LSTM model at the next time point.

[0103] Among them, the cell memory state C' at the current time point t , the cell state C at the current time point t and the output value h t C' t = tanh(W c [h t-1 , x t + u);

[0104] C t = f t · C t + i t · C' t ;

[0105] h t = O t · tanh(C t );

[0106] In the formula, O t is the control signal.

[0107] When the output values at all time points in the irregular time series data are determined, the second prediction data is obtained.

[0108] Step S4, according to the first prediction data and the second prediction data that meet the regularity test conditions, perform prediction through a preset radial basis neural network to obtain the predicted value of the fan output.

[0109] In the embodiment of the present application, the first prediction data and the second prediction data that meet the regular inspection conditions are integrated and input into a radial basis function neural network for training and learning to construct a fan output prediction model.

[0110] Among them, the radial basis function neural network is also called the RBF neural network. The RBF neural network is a single-hidden-layer forward feedback neural network. Each node in the hidden layer has a parameter vector called the center, which is used to compare with the network input vector to generate a radially symmetric response; the response of the hidden layer is scaled by the connection weights of the output layer, and then combined to generate the network output. The neurons in the hidden layer are neurons with a Gaussian function as the kernel function. The kernel function is a local activation function with the largest response at the center point, and the response decreases exponentially away from the center point. Compared with other neural network models, the RBF neural network has stronger learning, classification, and approximation capabilities.

[0111] Figure 3 The structure diagram of the RBF neural network is provided. As shown in the figure, data enters the RBF neural network from the input layer, and data prediction is performed through the RBF function of the hidden layer, and then linear weighting is performed, and the predicted value of the fan output is mapped and output from the output layer.

[0112] Among them, the specific expression of the RBF function is:

[0113]

[0114] r = ‖X - c i ‖ 2 , i = 1, 2, …, q;

[0115] In the formula, φ is the radial basis function, r is the distance from X to the center point, σ is the expansion constant of the radial basis function, which determines the expansion width of φ, X is the input data of the RBF neural network, c i is the center point of the i-th neuron of the RBF neural network, and q is the number of nodes in the hidden layer.

[0116] After linear weighting, the output layer maps and outputs:

[0117]

[0118] In the formula, y m is the m-th output value of the RBF neural network, w jk is the weight when the j-th neuron in the hidden layer is connected to the output layer, and k is the k-th hidden layer.

[0119] Implementing the embodiment of the present application has the following beneficial effects:

[0120] In the embodiment of the present application, the original fan output time series data obtained is first subjected to a stationary processing to obtain first time series data with a certain degree of stationarity, which is convenient for subsequent time series prediction; then, the time series data with a certain periodicity in the first time series data is predicted and identified through an ARIMA model to obtain first prediction data containing high-precision regularity; then, the irregular data in the first prediction data is input into an LSTM model for regularized prediction, and second prediction data with a certain degree of regularity is predicted from the irregular time series data, providing accurate data support for subsequent output prediction; finally, the first prediction data and the second prediction data containing a certain degree of regularity are combined through an RBF neural network according to the regularity of the data for output prediction, and a predicted value of the fan output with high precision is obtained.

[0121] Second Embodiment

[0122] Furthermore, in order to implement the short-term fan output prediction system based on the combined model corresponding to the above method embodiment to achieve the corresponding functions and technical effects, Figure 4 A structural diagram of a short-term fan output prediction system based on a combined model is provided. For the sake of illustration, only the parts related to this embodiment are shown. The short-term fan output prediction system based on the combined model provided by the embodiment of the present application includes:

[0123] A stationary processing module 201, configured to perform stationary processing on the obtained fan output time series data to obtain first time series data.

[0124] In the embodiment of the present application, the obtained fan output time series data is preprocessed; the stationary test is performed on the preprocessed fan output time series data, and then the difference processing is performed on the fan output time series data that fails to pass the stationary test to obtain the stationary fan output time series data, denoted as the first time series data.

[0125] A regular data prediction module 202, configured to input the first time series data into a preset autoregressive integrated moving average model for regular prediction to obtain first prediction data.

[0126] In the embodiment of the present application, the autocorrelation analysis is performed on the first time series data to obtain autocorrelation parameters; according to the set model prediction accuracy, combined with the autocorrelation parameters, the first model parameters of the autoregressive integrated moving average model are determined; according to the first model parameters, the autoregressive integrated moving average model is used to predict the regularity of the first time series data to obtain first prediction data.

[0127] An irregular data prediction module 203, configured to input the first prediction data that does not meet the preset regularity test condition into a preset long short-term memory network for regularized prediction to obtain second prediction data.

[0128] In an embodiment of the present application, according to a preset regular inspection condition, the first prediction data is inspected, and the irregular time series data is identified from the first prediction data that does not meet the regular inspection condition; according to the irregular time series data, the second model parameter of the long short-term memory network is set; according to the second model parameter, the irregular time series data is regularly predicted to obtain second prediction data.

[0129] The output data prediction module 204 is configured to predict through a preset radial basis neural network according to the first prediction data and the second prediction data that meet the regular inspection condition, so as to obtain a predicted value of the fan output.

[0130] In an embodiment of the present application, the first prediction data and the second prediction data that meet the regular inspection condition are integrated and input into the radial basis neural network for training and learning to construct a fan output prediction model.

[0131] Wherein, the radial basis neural network is also called an RBF neural network. The RBF neural network is a single-hidden-layer forward feedback neural network. Each node in the hidden layer has a parameter vector, called the center, which is used to compare with the network input vector to generate a radially symmetric response; the response of the hidden layer is scaled through the connection weights of the output layer, and then combined to generate the network output. The neurons in the hidden layer are neurons with a Gaussian function as the kernel function. The kernel function is a local activation function with the maximum response at the center point, and the response decreases exponentially away from the center point. Compared with other neural network models, the RBF neural network has stronger learning, classification, and approximation capabilities.

[0132] In the RBF neural network, data enters the RBF neural network from the input layer, is predicted through the RBF function of the hidden layer, and then linearly weighted, and the predicted value of the fan output is mapped and output from the output layer.

[0133] Wherein, the specific expression of the RBF function is:

[0134]

[0135] r = ‖X - c i ‖ 2 , i = 1, 2,..., q;

[0136] In the formula, φ is the radial basis function, r is the distance from X to the center point, σ is the expansion constant of the radial basis function, which determines the expansion width of φ, X is the input data of the RBF neural network, c i is the center point of the i-th neuron of the RBF neural network, and q is the number of nodes in the hidden layer.

[0137] After linear weighting, the output layer maps and outputs as follows:

[0138]

[0139] In the formula, y m is the m-th output value of the RBF neural network, w jk is the weight when the j-th neuron in the hidden layer is connected to the output layer, and k is the k-th hidden layer.

[0140] In some embodiments, the smoothing processing module 201 is specifically:

[0141] In the embodiments of the present application, since the original fan output time series data obtained often has missing values and data anomalies, making it inconvenient for time series data analysis. In order to make the original fan output time series data have a certain periodicity or stationarity before prediction analysis and reduce the complexity of data processing, it is necessary to preprocess the collected original fan output time series data before prediction calculation to complete the compensation and replacement of missing values and outliers.

[0142] The compensation and replacement of missing values and outliers in the data are often related to their positions in the time series data. When missing values and outliers exist at the beginning or end of the entire time series data, the direct deletion method can be used. However, if missing values and outliers exist in the middle of the time series data, they cannot be directly deleted, and there are generally 5 processing methods:

[0143] 6. Sequence average replacement: Use the average value of the entire time series data for compensation and replacement;

[0144] 7. Average value replacement of adjacent points: Use the average value of several adjacent points of the missing value or outlier for compensation and replacement;

[0145] 8. Median replacement of adjacent points: Use the median of several adjacent points of the missing value or outlier for compensation and replacement;

[0146] 9. Linear interpolation replacement: Use multiple adjacent points of the missing value or outlier for linear fitting to obtain the intermediate point value for compensation and replacement;

[0147] 10. Linear trend replacement of adjacent points: Use the adjacent time periods as independent variables and the sequence values as dependent variables for regression to obtain the missing value or outlier.

[0148] Optionally, in the embodiments of the present application, because the fan output time series data has strong seasonality and volatility, in order to ensure data accuracy, the linear interpolation method and the linear trend replacement method of adjacent points are used for data compensation and replacement.

[0149] Then, a stationarity test is performed on the preprocessed fan output time series data. Generally, the stationarity of the fan output time series data can be roughly estimated by directly observing the time series diagram, or the stationarity of the fan output time series data can be tested using hypothesis testing methods. The most common ones are ADF (augmented dickey fuller) unit root, autocorrelation coefficient function (ACF), and partial autocorrelation coefficient function (PACF) tests, etc. In the embodiments of the present application, the fan output time series data needs to meet the following three conditions to pass the stationarity test. The specific expression is:

[0150]

[0151] In the formula, x 1 is the value at time 1 in the fan output time series data, x 2 is the value at time 2 in the fan output time series data, E(x 1 ) is the mean value of x 1 in the fan output time series data, E(x 2 ) is the mean value of x 2 in the fan output time series data, V(x 1 ) is the variance of x 1 in the fan output time series data, V(x 2 ) is the variance of x 2 in the fan output time series data, σ is the average degree of fluctuation of the fan output time series data around its mean value at each moment, Cov(x t ,x t-s ) is the covariance of x t and x t-s in the fan output time series data, u is the average value of the fan output time series data at each moment, and γ is the direct correlation of the fan output time series data at times t and s.

[0152] Because stationary time series data has the characteristics that its mean is a constant, its variance is the mean, and its autocovariance function and autocorrelation coefficient only depend on the translation length of time and are independent of the starting point of time, in this application, the overall mean and variance can be estimated using all the data of the fan output time series data.

[0153] Then, multiple differencing processes are performed on the fan output time series data that fails the stationarity test, and finally the stationary fan output time series data is obtained, denoted as the first time series data.

[0154] In some embodiments, the regular data prediction module 202 is specifically:

[0155] In the embodiments of the present application, since the first time series data after differential processing has strong stationarity, the autoregressive integrated moving average model (ARIMA model), which is sensitive to stationary data, is used as the preferred prediction model. In the embodiments of the present application, the ARIMA model is adopted as the autoregressive integrated moving average model. Among them, the ARIMA model is fully called the autoregressive integrated moving average model, which is a time series prediction method. The basic idea is to regard the data sequence formed by the prediction object over time as a random sequence and approximate it with a certain mathematical model. Once this model is identified, future values can be predicted from the past and current values of the time series. The ARIMA model has the characteristics of high accuracy and fast speed when dealing with data with strong periodicity and time regularity. Since most wind power generation power data has strong periodicity, the ARIMA model is selected as the preferred model in the embodiments of the present application.

[0156] The ARIMA model contains three very important internal variables, which are called autocorrelation parameters in the embodiments of the present application, namely parameters p, d, and q. p represents the lag number of the time series data itself used in the model, also called the autoregressive term; d represents the number of orders of differencing required for the time series data to satisfy stationarity, also called the differencing term; q represents the lag number of the prediction error used in the model, also called the moving average term.

[0157] Among them, the expressions of the p autoregressive term, the q moving average term, and the d differencing term are as follows:

[0158]

[0159] In the formula, y t is the value of the time series data at time t, p is the autoregressive term, μ and μ' are constant terms, α i , β i are model correlation coefficients, q is the moving average term, ε t , ε' t are error values, L is the maximum likelihood function of the model, which can also be called the sequence width and reflects the degree of interpretation of the model for the data; d is the differencing term.

[0160] Because the ARIMA model is simple and does not need to rely on other external variables, it can complete predictions only relying on its own internal variables. However, this model is very dependent on the stationarity of the input sequence, which has a certain impact on its prediction accuracy. Therefore, differential stationarity processing must be carried out.

[0161] Generally speaking, the more parameters are added to a time series model, the better the model fitting effect, but this comes at the cost of increasing the model complexity. Therefore, the established model needs to seek a balance between the model complexity and the data interpretation ability. In the embodiments of the present application, the minimum information criterion AIC and the Bayesian information criterion BIC are used as the basis to select parameters, that is, the first model parameters of the ARIMA model. The specific expression is:

[0162] A IC = 2T - 2ln(L);

[0163] B IC = 2ln(T) - 2ln(L);

[0164] In the formula, T is the number of the first model parameters, representing the complexity of the ARIMA model; L is the maximum likelihood function of the model, which can also be called the sequence width, reflecting the degree of data interpretation by the simulation.

[0165] In addition, since AIC and BIC are more suitable for small models, to achieve the best balance effect between the model complexity and the data interpretation ability, the embodiments of the present application need to obtain the minimum values of A IC and B IC . In the ARIMA model, when p and q increase, T increases accordingly, and when the sequence width increases, L increases. In this way, the minimum values of AIC and BIC can be obtained, so as to determine p and q.

[0166] Exemplarily, in the embodiments of the present application, let p be 1 and q be 1. Therefore, the time series prediction model established on the basis of the ARIMA model is the ARIMA(1,1,1) model.

[0167] Finally, according to the constructed ARIMA model, the sequence law of the first time series data is identified and predicted to obtain the first predicted data.

[0168] In some embodiments, the irregular data prediction module 203 is specifically:

[0169] In the embodiments of the present application, when encountering a sequence with large randomness, the ARIMA model cannot effectively identify the sequence law. At this time, a long short-term memory network with stronger learning ability is used to predict the sequence. Before using the long short-term memory network for prediction, the residual white noise test is first performed on the first predicted data through a preset regularity test condition to identify the time series data that does not meet the regularity test condition and has low regularity, so as to obtain the irregular time series data.

[0170] The residual white noise test is also known as the pure random number test. When a sequence is a pure random number sequence, it is meaningless to analyze it using a time series model. Therefore, after the stationarity analysis or processing of the sequence, a white noise test needs to be performed on the sequence. In the embodiment of the present application, a white noise test is performed on the residual of the first predicted data. If the residual white noise test fails, it is considered that it is difficult to completely identify the time law of the first predicted data simply using the ARIMA model, and the LSTM model needs to be used as the long short-term memory network for prediction.

[0171] Optionally, the residual white noise test in the embodiment of the present application is processed using the MATLAB built-in function Autocorrelation[·], and this function specifically uses Bartlett's theorem as the theoretical basis. Moreover, the residual white noise test is also used as a combined node of the ARIMA model and the LSTM model, so that these two models form a combined model.

[0172] Among them, the regularity test condition, the specific expression is:

[0173]

[0174] In the formula, E(x 1 ) is the mean value of x 1 in the first predicted data, E(x 2 ) is the mean value of x 2 in the first predicted data, V(x 1 ) is the variance of x 1 in the first predicted data, V(x 2 ) is the variance of x 2 in the first predicted data, σ is the average degree of fluctuation of the first predicted data around its mean value, Cov(x t ,x t-s ) is the covariance of x t and x t-s in the first predicted data.

[0175] It should be emphasized that the time series that satisfies the above formula is actually a special stationary sequence.

[0176] The LSTM model is essentially an improved RNN, which can solve the long-distance dependence problem that the RNN cannot solve. Because when the time series is too long, the RNN is very easy to forget the information in the relatively long time period in the front, and the closer the time point is, the greater the impact on the current input. To overcome this shortcoming, the LSTM model was born.

[0177] The hidden layer of the original RNN has only one state variable, denoted as the C state, which is used to store short-term states. In the LSTM model, an additional state, denoted as h, is added to store long-term memories. The LSTM model mimics the memory method of cell neurons to build the model, stores the information and memories of past sequences in the form of a cell state vector, and combines the sequence state at the previous time, the sequence state at the current time, and the hidden state at the previous time to construct the long-term historical memory of the long-term neuron.

[0178] The LSTM model mainly uses three gates to achieve prediction. The three gates are the forget gate, the input gate, and the output gate. The forget gate determines how much of the cell state C at the previous moment t-1 will be saved to the current state of C t ; The input gate determines how much of the input X t will be saved to C t ; The output gate determines how much of the current state C of the control unit t will be output to the current output value h t .

[0179] In the embodiments of the present application, through the irregular time series data, the forget gate, the input gate, and the output gate of the LSTM model are set.

[0180] Based on the LSTM model, the irregular time series data is regularly predicted to obtain second prediction data.

[0181] Specifically, according to the forget gate, the input gate, and the output gate of the LSTM model at the current time point, the cell state and the cell memory state at the current time point in the irregular time series data are updated.

[0182] According to the cell state at the current time point, the output value at the current time point in the irregular time series data is determined, and the forget gate, the input gate, and the output gate of the LSTM model at the next time point are updated.

[0183] Among them, the cell memory state C' at the current time point t , the cell state C at the current time point t and the output value h t , the specific expressions are:

[0184] C' t =tanh(W c [h t-1 ,x t +u);

[0185] C t =f t ·C t +i t ·C' t ;

[0186] h t = O t ·tanh(C t );

[0187] In the formula, O t is the control signal.

[0188] After determining the output values of all time points in the irregular timing data, second prediction data is obtained.

[0189] Implementing the embodiments of the present application has the following beneficial effects:

[0190] In the embodiments of the present application, first, the obtained original fan output timing data is subjected to a stationary processing to obtain first timing data with a certain degree of stationarity, which is convenient for subsequent timing prediction; then, the ARIMA model is used to predict and identify the timing data with a certain periodicity in the first timing data to obtain first prediction data containing high-precision regularity; then, the irregular data in the first prediction data is input into the LSTM model for regularity prediction, and second prediction data with a certain degree of regularity is predicted from the irregular timing data, providing accurate data support for subsequent output prediction; finally, the first prediction data and the second prediction data with a certain degree of regularity are combined through the RBF neural network according to the regularity of the data for output prediction, and a prediction value of the fan output with high precision is obtained.

[0191] Furthermore, Figure 5 is a structural diagram of a terminal device provided by an embodiment of the present application. As Figure 5 shown, the terminal device 3 of this embodiment includes: at least one processor 30 (only one is shown in Figure 5 ), a memory 31, and a computer program 32 stored in the memory 31 and executable on the at least one processor. When the processor 30 executes the computer program 32, the steps of a method for short-term prediction of fan output based on a combined model according to any one of the embodiments of the present application can be implemented.

[0192] The terminal device 3 can be a computing device such as a desktop computer, a cloud server, and a laptop computer. The computing device may include, but is not limited to, the processor 30 and the memory 31. Figure 5 Merely examples of the terminal device 3 are given, which do not constitute a limitation on the terminal device 3 and may include more or fewer components than those shown in the figure.

[0193] The embodiments of the present application provide a storage medium that stores computer-readable program code, and when the computer-readable program code is executed, the steps of the above-mentioned method for short-term prediction of fan output based on a combined model are implemented.

[0194] The specific embodiments described above further elaborate on the purpose, technical solutions, and beneficial effects of this application. It should be understood that the above are only specific embodiments of this application and are not used to limit the protection scope of this application. It is particularly pointed out that for those skilled in the art, any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of this application shall be included within the protection scope of this application.

Claims

1. A short-term prediction method for wind turbine output based on a combined model, characterized in that: include: Stabilizing the acquired wind turbine output time series data to obtain first time series data; Inputting the first time series data into a preset differential autoregressive moving average model for regular prediction to obtain first prediction data; Inputting the first prediction data that does not meet the preset regularity test condition into the preset long short-term memory network for regularization prediction to obtain second prediction data; According to the first prediction data and the second prediction data satisfying the regularity test condition, prediction is performed through a preset radial basis function neural network to obtain a predicted value of the wind turbine output.

2. The method for short-term prediction of wind turbine output based on a combined model according to claim 1 is characterized in that: The obtained wind turbine output time series data is stabilized to obtain first time series data, which is specifically: Preprocess the acquired wind turbine output time series data; The preprocessed wind turbine output time series data is subjected to a stationarity check, and then the wind turbine output time series data that fails the stationarity check is subjected to a differential process to obtain the stationarized wind turbine output time series data, which is recorded as the first time series data.

3. The method for short-term prediction of wind turbine output based on a combined model according to claim 1, characterized in that: The first time series data is input into a preset differential autoregressive moving average model for regular prediction to obtain first prediction data, specifically: Performing autocorrelation analysis on the first time series data to obtain autocorrelation parameters; Determine the first model parameter of the differential autoregressive moving average model according to the set model prediction accuracy and in combination with the autocorrelation parameter; According to the first model parameters, the law of the first time series data is predicted by the differential autoregressive moving average model to obtain first predicted data.

4. The method for short-term prediction of wind turbine output based on a combined model according to claim 1, characterized in that: The first prediction data that does not meet the preset regularity test condition is input into the preset long short-term memory network for regularization prediction to obtain the second prediction data, specifically: According to a preset regularity inspection condition, the first prediction data is inspected, and irregular time series data is identified from the first prediction data that does not meet the regularity inspection condition; According to the irregular time series data, setting the second model parameters of the long short-term memory network; According to the second model parameters, the irregular time series data is regularized and predicted to obtain second prediction data.

5. The method for short-term prediction of wind turbine output based on a combined model according to claim 4 is characterized in that: The regularity test conditions are specifically: The regularity test condition is specifically expressed as follows: Where E(x1) is the mean of x1 in the first prediction data, E(x2) is the mean of x2 in the first prediction data, V(x1) is the variance of x1 in the first prediction data, V(x2) is the variance of x2 in the first prediction data, σ is the average degree of fluctuation of the first prediction data around its mean at each moment, Cov(x t ,x t-s ) is the first prediction data x t With x t-s The covariance of .

6. The method for short-term prediction of wind turbine output based on combined model according to claim 4 is characterized in that: According to the second model parameter, the irregular time series data is subjected to regularized prediction to obtain second prediction data, specifically: updating the unit state at the current time point in the irregular time series data according to the second model parameter at the current time point; Determine the output value of the irregular time series data at the current time point according to the unit state at the current time point, and update the second model parameter at the next time point; After the output values ​​of all time points in the irregular time series data are determined, the second prediction data is obtained.

7. The method for short-term prediction of wind turbine output based on a combined model according to any one of claims 1 to 6, characterized in that: Also includes: Integrate the first prediction data and the second prediction data that meet the regularity test condition and input them into the radial basis function neural network to construct a wind turbine output prediction model; The predicted value of the fan output is obtained through the fan output prediction model.

8. A short-term prediction system for wind turbine output based on a combined model, characterized in that: include: Stabilization processing module, regularity data prediction module, irregularity data prediction module and output data prediction module; The stabilization processing module is used to perform stabilization processing on the acquired wind turbine output time series data to obtain first time series data; The regularity data prediction module is used to input the first time series data into a preset differential autoregressive moving average model to perform regularity prediction to obtain first predicted data; The irregularity data prediction module is used to input the first prediction data that does not meet the preset regularity test condition into the preset long short-term memory network for regularity prediction to obtain the second prediction data; The output data prediction module is used to perform prediction through a preset radial basis function neural network according to the first prediction data and the second prediction data that meet the regularity test condition, so as to obtain a predicted value of the wind turbine output.

9. A terminal device, characterized in that: The method comprises a processor and a memory, wherein the memory stores a computer program, and when the processor executes the computer program, the method implements the steps of a method for short-term prediction of wind turbine output based on a combined model as described in any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium stores computer-readable program codes, which, when executed, implement the steps of a method for short-term prediction of wind turbine output based on a combined model according to any one of claims 1 to 7.

Images