A method for locating and predicting low-frequency forced oscillation disturbance sources in power systems

Abstract

The present invention relates to a method for locating and predicting a disturbance source of low-frequency forced oscillation in a power system, comprising the following steps: obtaining the active power daytime plan of conventional units and the predicted daytime power of wind farms and updating the data at preset time intervals; based on the active power daytime plan of conventional units and the predicted daytime power of wind farms, using a sliding time window to perform dynamic low-frequency forced oscillation detection on the active power daytime plan of conventional units and the predicted daytime power of wind farms within a future preset time interval, obtaining the active power data of conventional units and wind farms within the possible oscillation time window to construct a numerical matrix; inputting the numerical matrix into a trained CNN-SVM positioning model, predicting the potential conventional units and wind farms that may cause low-frequency forced oscillations within the future preset time interval and the corresponding window occurrence time. Compared with the prior art, the present invention has the advantages of being able to achieve accurate positioning results in a small sample case.

Description

Technical Field

[0001] The present invention relates to the technical field of low-frequency forced oscillation disturbance source identification, and in particular to a method for locating and predicting a low-frequency forced oscillation disturbance source in an electric power system. Background Art

[0002] As a clean energy with broad development prospects, the application of wind power technology in the power grid has received widespread attention. With the continuous increase in the proportion of renewable energy such as wind power, the structure of the system has become more complex, and the safety and stability issues it faces have become more prominent. Low-frequency oscillation, as an important issue affecting the safety and stability of the power system, can generally be divided into negatively damped low-frequency oscillation and low-frequency forced oscillation. So far, safety accidents caused by low-frequency oscillation have occurred many times around the world. Continuous low-frequency oscillation will pose a serious threat to the power grid, and may even cause the collapse of the entire power grid, resulting in immeasurable economic losses. Part of the above-mentioned low-frequency oscillation phenomenon is negatively damped low-frequency oscillation caused by insufficient system damping, which can be suppressed by installing power system stabilizers and other measures. The other part is low-frequency forced oscillation caused by external periodic disturbance sources, which has a wider impact range, greater harm, and a greater probability of occurrence. The oscillation cannot be suppressed by the measures taken for negatively damped low-frequency oscillation, and it is necessary to locate the disturbance source and remove it to suppress the oscillation.

[0003] However, with the emergence of new power systems with a high proportion of renewable energy and a high proportion of electronic equipment, the frequency of low-frequency forced oscillation phenomena has increased and the location of disturbance sources has become more diverse. The traditional physical positioning method of locating the disturbance source by deriving criteria from the system linearization model has become increasingly limited. Therefore, how to design a positioning algorithm suitable for the actual new power system to accurately identify the location of the low-frequency forced oscillation disturbance source has become an urgent problem to be solved. In recent years, with the gradual popularization of PMU in the power grid and the rapid development of artificial intelligence technology, a new idea has been provided to solve the problem of locating the low-frequency forced oscillation disturbance source. Among them, artificial intelligence technology represented by deep learning has achieved remarkable results in natural language processing, image recognition and other fields by establishing deep neural networks to analyze and process complex problems. Its research in the fields of transient stability assessment of power systems and load forecasting is also common. Therefore, it is worth studying the use of artificial intelligence technology to solve the problem of locating the low-frequency forced oscillation disturbance source in the power grid. In addition, with the continuous increase in wind power penetration, the risk of low-frequency forced oscillation caused by wind speed fluctuations has also increased. It is also important to use wind farm power prediction to broaden the scope of application of disturbance source positioning so as to help dispatchers predict wind farms that may cause low-frequency forced oscillations in the future. The problem of locating low-frequency forced oscillation disturbance sources is essentially to calculate the collected power grid data information according to a specific algorithm and give a judgment result, which is similar to a classification problem. At present, intelligent algorithms represented by convolutional neural networks (CNNs) can effectively extract feature vectors that represent data information with their powerful network self-learning ability, and use Softmax classifiers to classify the extracted feature vectors, overcoming the dependence of traditional classification methods on manual extraction links, and can have good applicability in complex nonlinear systems. However, the Softmax classifier used by CNN can only make the output data of the fully connected layer conform to the probability distribution, and cannot play a role in improving the classification effect. In addition, the model training time is long. In addition, the accuracy of the model training results depends on a large number of training samples, which leads to a decrease in the classification accuracy in the case of small samples where there is a lack of low-frequency forced oscillation data in the actual power grid. Summary of the invention

[0004] The purpose of the present invention is to provide a method for locating and predicting the disturbance source of low-frequency forced oscillation in an electric power system in order to address the scarcity of low-frequency forced oscillation data in the power grid and the accuracy and rapidity required for the disturbance source positioning result. Firstly, a CNN-SVM positioning model is constructed by replacing the Softmax classifier of the output layer in the CNN model with an SVM classifier that is suitable for small samples and has a better classification effect. The active power data collected from the export side of conventional units and wind farms in the power grid within the low-frequency forced oscillation time window is processed and introduced into the CNN-SVM positioning model. The model is trained with a certain number of samples, so that real-time low-frequency forced oscillation detection can be performed on the active power of conventional units and wind farms in the power grid, and the online positioning and prediction of the disturbance source can be completed using the CNN-SVM positioning model.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A method for locating a low-frequency forced oscillation disturbance source in a power system comprises the following steps:

[0007] Perform real-time low-frequency forced oscillation detection on the active power of conventional units and wind farms in the current time window of the power grid, obtain the active power data of conventional units and wind farms in the oscillation time window and construct a numerical matrix;

[0008] The numerical matrix is ​​input into the trained CNN-SVM positioning model to locate the conventional units or wind farms that cause low-frequency forced oscillations in the power grid online.

[0009] The method for obtaining the data matrix is:

[0010] The dynamic approximate entropy method is used to detect low-frequency forced oscillations in the active power data collected from the outlet side of conventional units and wind farms in the power grid. If low-frequency forced oscillations are detected in a certain time window, a numerical matrix is ​​formed based on the active power data of conventional units and wind farms in the window, which can be expressed as:

[0011] ,

[0012] In the formula, is the normalized active power data of conventional units and wind farms in the power grid within the low-frequency forced oscillation time window, that is, the active power data contained in one sample; For the M Conventional units / wind farms in the T Active power at a moment; M is the total number of conventional units and wind farms in the power grid; T is the total number of sampling points in the time window.

[0013] The training process of the CNN-SVM positioning model is:

[0014] Obtain the numerical matrix formed by the active power data of conventional units and wind farms in different low-frequency forced oscillation time windows and the corresponding disturbance source labels to form a low-frequency forced oscillation disturbance source training sample set ,in, is a set of sample numerical matrices, For the i The numerical matrix of samples, n is the sample size; is the sample label data set, For the i The labels of samples, , is a 0 / 1 variable, 1 represents the m Conventional units / wind farms are disturbance sources, and 0 represents non-disturbance sources, that is, the label data is one-hot encoded;

[0015] The disturbance source sample set Divide into training sample sets according to preset ratios and test sample set ,in p and q are the number of samples in the training sample set and the test sample set respectively;

[0016] The training set Input the CNN-SVM positioning model and obtain a CNN-SVM positioning model that meets the requirements through training and learning:

[0017] ,

[0018] In the formula, For the sample to be tested The output result obtained as the input of the CNN-SVM localization model is the disturbance source position label, and the label data is in numerical form; For the training set Learn the trained CNN-SVM positioning model.

[0019] The CNN-SVM positioning model includes an input layer, a feature extraction module, a feature integration module and an SVM classifier connected in sequence, wherein the feature extraction module extracts features from the active power numerical matrix of all conventional units and wind farms in the input layer within a low-frequency forced oscillation time window, the feature integration module integrates the features extracted by the feature extraction module, and the SVM classifier classifies the feature vectors obtained by integrating the fully connected layer to obtain a low-frequency forced oscillation disturbance source positioning label.

[0020] The feature extraction module includes a convolution layer, a pooling layer and a Dropout layer connected in sequence, and the feature integration module adopts a fully connected layer, wherein the activation functions of the convolution layer and the fully connected layer adopt the ReLU function, and a Dropout layer is connected after the fully connected layer to prevent overfitting.

[0021] The SVM classifier uses linear SVM to classify data;

[0022] For a linearly separable training data set, the hyperplane obtained by maximizing the interval or solving the corresponding convex quadratic programming problem is expressed as:

[0023] ,

[0024] In the formula, is the normal vector of the hyperplane; is the intercept of the hyperplane; the hyperplane is express;

[0025] Assume that the noisy training data set is ,because Linearly inseparable, that is, there are unique sample points Cannot satisfy Constraints of

[0026] Introducing slack variables , eliminate the unique sample points in the training data set, so that each remaining sample point Linearly separable:

[0027] ,

[0028] In addition, for each slack variable Pay a price and rewrite the objective function of the linear SVM classifier as:

[0029] ,

[0030] In the formula, C is the penalty parameter, C The size of determines the importance of the loss caused by the specific sample point. C The bigger the value, the higher the importance;

[0031] By introducing slack variables , modify the interval maximization to soft interval maximization, linear SVM for linearly inseparable The learning problem is transformed into the following convex quadratic programming problem:

[0032] ,

[0033] For a linearly inseparable training dataset, the solution to the convex quadratic programming problem is and , we get the optimal hyperplane and classification decision function, that is, we get the linear SVM model:

[0034] ,

[0035] ,

[0036] In the formula, is the normal vector of the optimal hyperplane of the linear SVM; is the intercept of the optimal hyperplane of the linear SVM; Represents the optimal hyperplane of linear SVM; is the classification decision function of the linear SVM.

[0037] The solution process of the convex quadratic programming problem is as follows:

[0038] The Lagrangian function corresponding to the convex quadratic programming problem of linear SVM is:

[0039] ,

[0040] In the formula, and is a Lagrange multiplier, and and ;

[0041] By solving the minima and maxima problem of the Lagrangian function step by step, the dual problem of the convex quadratic programming problem of the linear SVM is obtained as follows:

[0042] ,

[0043] The optimal solution is obtained by solving the dual problem of the convex quadratic programming problem , .

[0044] A method for predicting a low-frequency forced oscillation disturbance source in a power system comprises the following steps:

[0045] Obtain the daily active power plan of conventional units and the predicted daily power of wind farms and update the data at preset time intervals;

[0046] Based on the daytime active power plan of conventional units and the predicted daytime power of wind farms, a sliding time window is used to perform dynamic low-frequency forced oscillation detection on the daytime active power plan of conventional units and the daytime predicted power of wind farms within a preset time interval in the future. A numerical matrix is ​​constructed using the above method, and the numerical matrix is ​​input into the trained CNN-SVM positioning model to predict the potential conventional units and wind farms that may cause low-frequency forced oscillations within the preset time interval in the future and the occurrence time of the corresponding window.

[0047] The daytime power of the wind farm is predicted by the meteorological extrapolation method. The theoretical power calculation model of each wind turbine is established by using the relationship between wind speed and power data. The theoretical power of the wind farm is obtained by superimposing the theoretical power of all wind turbines in the field. The calculation method is as follows:

[0048] ,

[0049] In the formula, For wind farms t Theoretical power at the moment; M is the number of wind turbines in the entire site; For wind farm m Typhoon generator set t Theoretical power at the moment.

[0050] The active power daytime plan of the conventional units is calculated by the daytime plan rolling scheduling model, specifically:

[0051] The wind farm power and load forecasts are re-performed for the remaining time period of the day at preset time intervals, and a daytime plan rolling dispatch model for the active power of conventional units is established based on the new forecast values ​​with the goal of minimizing the power generation cost of conventional units:

[0052]

[0053] In the formula, For the i The conventional unit is t Rolling plan value for the time period; , and Respectively i The coefficients of the quadratic, linear and conventional terms of the power generation cost of each conventional unit; N is the number of conventional units in the system; is the current period;

[0054] The daytime plan rolling dispatch model of the active power of conventional units meets the following constraints:

[0055] 1) Upper and lower bounds of active power of conventional units:

[0056] ,

[0057] In the formula, , Respectively i The conventional unit is t The upper and lower limits of active power in a time period;

[0058] 2) Conventional unit ramp rate constraints

[0059] ,

[0060] In the formula, , Respectively i Conventional units from t -1 period to t The maximum values ​​of the lowering force and the lifting force allowed in the time period;

[0061] 3) Load balancing constraints

[0062] ,

[0063] ,

[0064] ,

[0065] In the formula, For the h The load is t Average forecast value for the time period; For the h The load in t Time period s The predicted value at the moment; For the j The wind farm is t Average forecast value for the time period; For the j Wind farms in t Time period s The theoretical power forecast value at the time, that is, the active power forecast value; H , M are the number of loads and wind farms respectively; T for t The total number of sampling points of the time period data is related to the signal sampling frequency;

[0066] According to the daytime rolling optimization scheduling model, the continuous particle swarm algorithm is used to solve it, and the wind farm and load forecast power data are used to obtain the active power daytime plan of each conventional unit, and the data is refreshed and calculated at preset time intervals.

[0067] Compared with the prior art, the present invention has the following beneficial effects:

[0068] (1) The present invention introduces slack variables to eliminate unique sample points in the training data set, making each remaining sample point linearly separable, thus avoiding the introduction of complex physical models.

[0069] (2) The present invention uses a convolutional neural network to extract features from active power signals of conventional units and wind farms, thereby enhancing the characteristics of the original signals and reducing the interference of noise on the signals, so that the CNN-SVM positioning model still has a high recognition accuracy under noise interference.

[0070] (3) The present invention dynamically collects the active power daytime plan of conventional units and the daytime predicted power data of wind farms within the oscillation occurrence window through a sliding time window, so that dispatchers can predict the potential wind farms that may cause low-frequency forced oscillations within the next 15 minutes, so as to take corresponding measures in advance to reduce unnecessary economic losses. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 is a flow chart of the positioning method of the present invention;

[0072] Figure 2 It is a schematic diagram of the structure of the CNN-SVM positioning model of the present invention;

[0073] Figure 3 is a flow chart of the prediction method of the present invention;

[0074] Figure 4 This is a flow chart of the rolling scheduling of the active power daytime plan of the conventional generator set of the present invention;

[0075] Figure 5 It is a schematic diagram of the daytime active power plan of conventional units and the daytime predicted power results of wind farms in an embodiment. DETAILED DESCRIPTION

[0076] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0077] Example 1

[0078] This embodiment provides a method for locating a low-frequency forced oscillation disturbance source in a power system. Figure 1 As shown, the following steps are included:

[0079] S101, performing real-time low-frequency forced oscillation detection on the active power of conventional units and wind farms in the power grid within the current time window, obtaining active power data of conventional units and wind farms within the oscillation time window and constructing a numerical matrix;

[0080] S102, inputting the numerical matrix into the trained CNN-SVM positioning model, and online locating the conventional units or wind farms that induce low-frequency forced oscillations in the power grid.

[0081] When low-frequency forced oscillation occurs in the system, the active power, frequency and other electrical quantities of the generator sets in the power grid will show oscillation information reflecting the characteristics of the disturbance source, among which the oscillation information contained in the active power is particularly obvious. Therefore, the dynamic approximate entropy method with engineering practicality is first used to detect low-frequency forced oscillations on the active power data collected from the outlet side of conventional units and wind farms in the power grid. If low-frequency forced oscillations are detected in a certain time window, the numerical matrix formed by the active power data of conventional units and wind farms in the window is used as the input of the CNN-SVM initialization positioning model. The numerical matrix representation method is as follows: The dynamic approximate entropy method is used to detect low-frequency forced oscillations on the active power data collected from the outlet side of conventional units and wind farms in the power grid. If low-frequency forced oscillations are detected in a certain time window, a numerical matrix is ​​formed based on the active power data of conventional units and wind farms in the window, which is expressed as:

[0082] (1)

[0083] In the formula, is the normalized active power data of conventional units and wind farms in the power grid within the low-frequency forced oscillation time window, that is, the active power data contained in one sample; For the M Conventional units / wind farms in the T Active power at a moment; M is the total number of conventional units and wind farms in the power grid; T is the total number of sampling points in the time window.

[0084] According to the above method, the numerical matrix formed by the active power data of conventional units and wind farms in different low-frequency forced oscillation time windows and the corresponding disturbance source labels are obtained to form a low-frequency forced oscillation disturbance source training sample set. ,in, is a set of sample numerical matrices, For the i The numerical matrix of samples, n is the sample size; is the sample label data set, For the i The labels of samples, , is a 0 / 1 variable, 1 represents the m Conventional units / wind farms are disturbance sources, and 0 represents non-disturbance sources, that is, the label data is one-hot encoded.

[0085] The CNN-SVM positioning model includes an input layer, a feature extraction module, a feature integration module and an SVM classifier connected in sequence, wherein the feature extraction module extracts features from the active power numerical matrix of all conventional units and wind farms in the input layer within a low-frequency forced oscillation time window, the feature integration module integrates the features extracted by the feature extraction module, and the SVM classifier classifies the feature vectors obtained by integrating the fully connected layer to obtain the low-frequency forced oscillation disturbance source positioning label. The feature extraction module includes a convolutional layer, a pooling layer and a Dropout layer connected in sequence, and the feature integration module uses a fully connected layer, wherein the parameter optimization of the CNN part adopts the Adam optimization training algorithm, and the activation function of the convolutional layer and the fully connected layer adopts the ReLU function. To prevent overfitting, a Dropout layer is connected after the fully connected layer.

[0086] In order to intuitively display the CNN-SVM positioning model constructed by this method, Figure 2 The composition structure of the model and the role of each part in the low-frequency forced oscillation disturbance source location process are introduced. Among them, the convolution layer and the pooling layer are responsible for feature extraction of the active power numerical matrix of all conventional units and wind farms in the input layer within a low-frequency forced oscillation time window. The Dropout layer is to prevent overfitting in the training process. The fully connected layer is responsible for integrating the extracted features. The SVM is responsible for classifying the feature vectors integrated by the fully connected layer, and finally obtaining the low-frequency forced oscillation disturbance source location label.

[0087] Considering the training dataset There may be a certain number of noise points in the training set that make the data linearly inseparable, so linear SVM is used for data classification.

[0088] Assume that the noisy training data set is ,because Linearly inseparable, that is, there are unique sample points Cannot satisfy Constraints, so it is necessary to eliminate these singular points in the training data set to make the remaining sample points linearly separable. Among them, for linearly separable training data sets, the hyperplane obtained by maximizing the interval or solving the corresponding convex quadratic programming problem is expressed as: , where is the normal vector of the hyperplane; is the intercept of the hyperplane; the hyperplane is express.

[0089] Introducing slack variables , so that each of the remaining sample points Linearly separable:

[0090] (2)

[0091] In addition, for each slack variable Pay a price and rewrite the objective function of the linear SVM classifier as:

[0092] (3)

[0093] In the formula, C is the penalty parameter, C The size of determines the importance of the loss caused by the specific sample point. C The bigger it is, the more importance it receives.

[0094] By introducing slack variables , modify the interval maximization to soft interval maximization, linear SVM for linearly inseparable The learning problem is transformed into the following convex quadratic programming problem:

[0095] (4)

[0096] For a linearly inseparable training dataset, the solution of the convex quadratic programming problem (4) is and , we can get the optimal hyperplane and classification decision function, that is, the linear SVM model:

[0097] (5)

[0098] (6)

[0099] In the formula, is the normal vector of the optimal hyperplane of the linear SVM; is the intercept of the optimal hyperplane of the linear SVM; Represents the optimal hyperplane of linear SVM; is the classification decision function of the linear SVM.

[0100] The solution process of the convex quadratic programming problem (4) is as follows:

[0101] The Lagrangian function corresponding to the convex quadratic programming problem of linear SVM is:

[0102] (7)

[0103] In the formula, and is a Lagrange multiplier, and and .

[0104] By solving the minimax problem of Lagrangian function (5) step by step, the dual problem of the convex quadratic programming problem of linear SVM is obtained as follows:

[0105] (8)

[0106] The optimal solution is obtained by solving the dual problem of the convex quadratic programming problem , , and finally the optimal hyperplane of linear SVM is obtained and classification decision function .

[0107] The disturbance source sample set Divide into training sample sets according to preset ratios and test sample set ,in p and q are the number of samples in the training sample set and the test sample set respectively.

[0108] The training set Input the CNN-SVM positioning model and obtain a CNN-SVM positioning model that meets the requirements through training and learning:

[0109] (9)

[0110] In the formula, For the sample to be tested The output result obtained as the input of the CNN-SVM localization model is the disturbance source position label, and the label data is in numerical form; For the training set Learn the trained CNN-SVM positioning model.

[0111] By performing real-time low-frequency forced oscillation detection on the active power of conventional units and wind farms in the current time window in the power grid, the active power of conventional units and wind farms in the oscillation time window is substituted into the trained CNN-SVM positioning model in the form of a numerical matrix, and the conventional units or wind farms that cause low-frequency forced oscillations in the power grid can be located online.

[0112] Example 2

[0113] In order to help power grid dispatchers predict in advance the potential disturbance sources that may cause low-frequency forced oscillations in the future, this embodiment provides a method for predicting the disturbance sources of low-frequency forced oscillations in a power system, such as Figure 3 As shown, the following steps are included:

[0114] S201, obtaining the daytime active power plan of conventional units and the predicted daytime power of the wind farm and updating the data at preset time intervals;

[0115] S202, based on the active power daytime plan of the conventional units and the predicted daytime power of the wind farm, using a sliding time window to perform dynamic low-frequency forced oscillation detection on the active power daytime plan of the conventional units and the predicted daytime power of the wind farm at a preset time interval in the future, and constructing a numerical matrix using the method described in Example 1;

[0116] S203, inputting the numerical matrix into the trained CNN-SVM positioning model, predicting the potential conventional units and wind farms that may cause low-frequency forced oscillations in the future preset time intervals and the corresponding window occurrence time.

[0117] The daytime power of the wind farm is predicted by the meteorological extrapolation method. The theoretical power calculation model of each wind turbine is established by using the relationship between wind speed and power data. The theoretical power of the wind farm is obtained by superimposing the theoretical power of all wind turbines in the field. The calculation method is as follows:

[0118] (10)

[0119] In the formula, For wind farms t Theoretical power at the moment; M is the number of wind turbines in the entire site; For wind farm m Typhoon generator set t Theoretical power at the moment.

[0120] The active power daytime plan of conventional units is calculated by the daytime plan rolling scheduling model. Since the active power daytime plan of conventional units cannot take into account the disturbances on the conventional unit side, and the daytime predicted power of the wind farm calculated by wind speed can take into account the disturbances caused by wind fluctuations during this period, the CNN-SVM positioning model can ultimately locate and predict the wind farms that may cause low-frequency forced oscillations in the future. The specific method is as follows.

[0121] First, in order to obtain a more accurate daytime active power plan of conventional units (this embodiment uses a 15-minute period, with a total of 96 periods a day), taking into account the fact that the accuracy of wind farm power forecasting and load forecasting decreases with the increase of time scale, it is proposed to re-forecast the wind farm power and load for the remaining periods of the day every 15 minutes, and establish a daytime active power plan rolling scheduling model for conventional units based on the new forecast values ​​with the goal of minimizing the power generation cost of conventional units:

[0122] (11)

[0123] In the formula, For the i The conventional unit is t Rolling plan value for the time period; , and Respectively i The coefficients of the quadratic, linear and conventional terms of the power generation cost of each conventional unit; N is the number of conventional units in the system; For the current period.

[0124] The daily plan rolling dispatch model of active power of conventional units meets the following constraints:

[0125] 1) Upper and lower bounds of active power of conventional units:

[0126] (12)

[0127] In the formula, , Respectively i The conventional unit is t The upper and lower limits of active power in a time period;

[0128] 2) Conventional unit ramp rate constraints

[0129] (13)

[0130] In the formula, , Respectively i Conventional units from t -1 period to t The maximum values ​​of the lowering force and the lifting force allowed in the time period;

[0131] 3) Load balancing constraints

[0132] (14)

[0133] (15)

[0134] (16)

[0135] In the formula, For the h The load is t Average forecast value for the time period; For the h The load in t Time period s The predicted value at the moment; For the j The wind farm is tAverage forecast value for the time period; For the j Wind farms in t Time period s The theoretical power forecast value at the time, that is, the active power forecast value; H , M are the number of loads and wind farms respectively; T for t The total number of data sampling points in a time period is related to the signal sampling frequency.

[0136] According to the daytime rolling optimization scheduling model, the wind farm and load forecast power data are used to obtain the active power daytime plan of each conventional unit, and the above three types of data are refreshed and calculated every 15 minutes. Since the daytime plan is a continuous optimization problem, the continuous particle swarm algorithm can be used to solve it. The calculation process of the rolling scheduling model for the active power daytime plan of conventional units in this embodiment is as follows: Figure 4 As shown in the figure t 0 is the current period.

[0137] By setting a sliding time window of a certain width to collect the active power daytime plan of conventional units and the predicted power data of wind farms, the dynamic approximate entropy method is used to perform dynamic low-frequency forced oscillation detection for the next 15 minutes with equal width steps. Once the occurrence of low-frequency forced oscillation is detected, the active power data of conventional units and wind farms in the oscillation time window are imported into the trained CNN-SVM positioning model in the form of a numerical matrix, and finally the potential wind farms and corresponding window occurrence times that may cause low-frequency forced oscillations in the next 15 minutes are predicted. When the active power daytime plan of conventional units and the predicted power of wind farms are updated every 15 minutes, the disturbance source is predicted for the next period according to the above method, and the window time corresponding to the occurrence of low-frequency forced oscillation and the disturbance source prediction results are sent to the dispatcher.

[0138] The following introduces the prediction method of low-frequency forced oscillation disturbance source based on CNN-SVM, taking the power grid system with 4 conventional units G1~G4 and 2 wind farms W1 and W2 as an example. Assume that the active power daytime plan of G1~G4 and the daytime predicted power of W1 and W2 are as follows: Figure 5 As shown, the daytime forecast power of W1 and W2 is obtained by the aforementioned wind farm power forecasting method, the active power daytime plan of G1~G4 is obtained by the daytime rolling scheduling model of conventional units, and the active power daytime plan data of G1~G4 follows the daytime forecast power of W1, W2 and the load and is refreshed every 15 minutes.

[0139] Set the sliding time window width to w (like w = 20s), collect the first t 0~t The active power signals of G1~G4 and W1, W2 in each time window of the 0+1 period are collected, and the low-frequency forced oscillation detection of the active power signals in the window is performed using the dynamic approximate entropy method. If a low-frequency forced oscillation is detected in a window, the active power signals of G1~G4 and W1, W2 in the window are imported into the trained CNN-SVM positioning model in the form of a numerical matrix, and finally the wind farm that may cause low-frequency forced oscillation in the window is located and predicted; if no low-frequency forced oscillation is detected, the disturbance source positioning prediction is not performed. When the G1~G4 active power daytime plan is updated every 15 minutes, the disturbance source is predicted for the next period according to the above method, and the window time of low-frequency forced oscillation and the disturbance source prediction result are sent to the dispatcher, so that the corresponding suppression measures can be taken in advance to prevent the occurrence of oscillation.

[0140] By setting a sliding time window, dynamic low-frequency forced oscillation detection is performed on the active power daytime plan of conventional units and the predicted power of wind farms in the power grid. The active power daytime plan of conventional units and the predicted power of wind farms within the oscillation time window are substituted into the trained CNN-SVM positioning model in the form of numerical matrices to predict the potential wind farms that may cause low-frequency forced oscillations in the future, thereby providing early warning information for power grid dispatchers to formulate corresponding plans to prevent the occurrence of oscillations.

Claims

1. A method for locating and predicting a low-frequency forced oscillation disturbance source in a power system, characterized in that: The following steps are involved: Obtain the daily active power plan of conventional units and the predicted daily power of wind farms and update the data at preset time intervals; Based on the active power daytime plan of conventional units and the predicted daytime power of wind farms, a sliding time window is used to perform dynamic low-frequency forced oscillation detection on the active power daytime plan of conventional units and the predicted daytime power of wind farms within a preset time interval in the future, and the active power data of conventional units and wind farms within the time window where oscillation may occur are obtained to construct a numerical matrix; The numerical matrix is ​​input into the trained CNN-SVM positioning model to predict the potential conventional units and wind farms that may cause low-frequency forced oscillations and the corresponding window occurrence time within the preset time interval in the future; The active power daytime plan of the conventional units is calculated by the daytime plan rolling scheduling model, specifically: The wind farm power and load forecasts are re-performed for the remaining time periods of the day at preset time intervals, and a daytime rolling scheduling model for the active power of conventional units is established based on the new forecast values ​​with the goal of minimizing the power generation cost of conventional units: ， In the formula, For the i The conventional unit is t Rolling plan value for the time period; , and Respectively i The coefficients of the quadratic, linear and conventional terms of the power generation cost of each conventional unit; N is the number of conventional units in the system; For the current period.

2. A method for locating and predicting a low-frequency forced oscillation disturbance source in a power system according to claim 1, characterized in that: The predicted daytime power of the wind farm is obtained by meteorological extrapolation. The theoretical power calculation model of each wind turbine is established by using the relationship between wind speed and power data. The theoretical power of the wind farm is obtained by superimposing the theoretical power of all wind turbines in the field. The calculation method is as follows: ， In the formula, For wind farms t Theoretical power at the moment; M is the number of wind turbines in the entire site; For wind farm m Typhoon generator set t Theoretical power at the moment.

3. The method for locating and predicting a low-frequency forced oscillation disturbance source in a power system according to claim 1, characterized in that: The daytime plan rolling dispatch model of the active power of conventional units meets the following constraints: 1) Upper and lower bounds of active power of conventional units: ， In the formula, , Respectively i The conventional unit is t The upper and lower limits of active power in a time period; 2) Conventional unit ramp rate constraints ， In the formula, , Respectively i Conventional units from t -1 period to t The maximum values ​​of the lowering force and the lifting force allowed in the time period; 3) Load balancing constraints ， ， ， In the formula, For the h The load is t Average forecast value for the time period; For the h The load in t Time period s The predicted value at the moment; For the j The wind farm is t Average forecast value for the time period; For the j Wind farms in t Time period s The theoretical power forecast value at the time, that is, the active power forecast value; H , M are the number of loads and wind farms respectively; T for t The total number of sampling points of the time period data is related to the signal sampling frequency; According to the daytime rolling optimization scheduling model, the continuous particle swarm algorithm is used to solve it, and the wind farm and load forecast power data are used to obtain the active power daytime plan of each conventional unit, and the data is refreshed and calculated at preset time intervals.

4. The method for locating and predicting a low-frequency forced oscillation disturbance source in a power system according to claim 1, characterized in that: The method for obtaining the numerical matrix is: The dynamic approximate entropy method is used to detect low-frequency forced oscillations in the active power data collected from the outlet side of conventional units and wind farms in the power grid. If low-frequency forced oscillations are detected in a certain time window, a numerical matrix is ​​formed based on the active power data of conventional units and wind farms in the window, which can be expressed as: ， In the formula, is the normalized active power data of conventional units and wind farms in the power grid within the low-frequency forced oscillation time window, that is, the active power data contained in one sample; For the M Conventional units / wind farms in the T Active power at a moment; M is the total number of conventional units and wind farms in the power grid; T is the total number of sampling points in the time window.

5. The method for locating and predicting a low-frequency forced oscillation disturbance source in a power system according to claim 1, characterized in that: The training process of the CNN-SVM positioning model is: Obtain the numerical matrix formed by the active power data of conventional units and wind farms in different low-frequency forced oscillation time windows and the corresponding disturbance source labels to form a low-frequency forced oscillation disturbance source training sample set ,in, is a set of sample numerical matrices, For the i The numerical matrix of samples, n is the sample size; is the sample label data set, For the i The labels of samples, , is a 0 / 1 variable, 1 represents the m Conventional units / wind farms are disturbance sources, and 0 represents non-disturbance sources, that is, the label data is one-hot encoded; The disturbance source sample set Divide into training sample sets according to preset ratios and test sample set ,in p and q are the number of samples in the training sample set and the test sample set respectively; The training set Input the CNN-SVM positioning model and obtain a CNN-SVM positioning model that meets the requirements through training and learning: ， In the formula, For the sample to be tested The output result obtained as the input of the CNN-SVM localization model is the disturbance source position label, and the label data is in numerical form; For the training set Learn the trained CNN-SVM positioning model.

6. The method for locating and predicting a low-frequency forced oscillation disturbance source in a power system according to claim 1, characterized in that: The CNN-SVM positioning model includes an input layer, a feature extraction module, a feature integration module and an SVM classifier connected in sequence, wherein the feature extraction module extracts features from the active power numerical matrix of all conventional units and wind farms in the input layer within a low-frequency forced oscillation time window, the feature integration module integrates the features extracted by the feature extraction module, and the SVM classifier classifies the feature vectors obtained by integrating the fully connected layer to obtain a low-frequency forced oscillation disturbance source positioning label.

7. A method for locating and predicting a low-frequency forced oscillation disturbance source in a power system according to claim 6, characterized in that: The feature extraction module includes a convolution layer, a pooling layer and a Dropout layer connected in sequence, and the feature integration module adopts a fully connected layer, wherein the activation functions of the convolution layer and the fully connected layer adopt the ReLU function, and a Dropout layer is connected after the fully connected layer to prevent overfitting.

8. A method for locating and predicting a low-frequency forced oscillation disturbance source in a power system according to claim 6, characterized in that: The SVM classifier uses linear SVM to classify data; For a linearly separable training data set, the hyperplane obtained by maximizing the interval or solving the corresponding convex quadratic programming problem is expressed as: ， In the formula, is the normal vector of the hyperplane; is the intercept of the hyperplane; the hyperplane is express; Assume that the noisy training data set is , n is the sample size, For the i The numerical matrix of samples, For the i The labels of samples; Linearly inseparable, that is, there are unique sample points Cannot satisfy Constraints of Introducing slack variables , eliminate the unique sample points in the training data set, so that each remaining sample point Linearly separable: ， In addition, for each slack variable Pay a price and rewrite the objective function of the linear SVM classifier as: ， In the formula, C is the penalty parameter, C The size of determines the importance of the loss caused by the specific sample point. C The bigger the value, the higher the importance; By introducing slack variables , modify the interval maximization to soft interval maximization, linear SVM for linearly inseparable The learning problem is transformed into the following convex quadratic programming problem: ， For a linearly inseparable training dataset, the solution to the convex quadratic programming problem is and , we get the optimal hyperplane and classification decision function, that is, we get the linear SVM model: ， ， In the formula, is the normal vector of the optimal hyperplane of the linear SVM; is the intercept of the optimal hyperplane of the linear SVM; Represents the optimal hyperplane of linear SVM; is the classification decision function of the linear SVM.

9. A method for locating and predicting a low-frequency forced oscillation disturbance source in a power system according to claim 8, characterized in that: The solution process of the convex quadratic programming problem is as follows: The Lagrangian function corresponding to the convex quadratic programming problem of linear SVM is: ， In the formula, and is a Lagrange multiplier, and and ; By solving the minima and maxima problem of the Lagrangian function step by step, the dual problem of the convex quadratic programming problem of the linear SVM is obtained as follows: ， The optimal solution is obtained by solving the dual problem of the convex quadratic programming problem , .

Images