SVG parameter optimization and identification method based on convolutional neural network

Abstract

The SVG controller parameter identification method based on convolutional neural network of the present invention obtains different combinations of various SVG parameters and the reactive power, voltage and current curve data corresponding to the output through BPA simulation software, and screens them to obtain the final data set; then, a neural convolutional network model that can reversely infer SVG parameters based on reactive power dynamic curve data is built based on the acquired data set; finally, RTDS measured curve data is used as model input to quickly identify SVG parameter values. This method combines convolutional neural network with SVG controller parameter identification, provides a new method for SVG controller parameter identification, ensures the accuracy of parameter prediction results, and improves identification efficiency.

Description

Technical Field

[0001] The invention relates to the field of electric power information technology, and in particular to an SVG parameter optimization identification method based on a convolutional neural network. Background Art

[0002] The emergence of Flexible AC Transmission Technology (FACTS) provides a new technical means to improve the reliability and economy of power grids. Static VAR generator (SVG), as an important member of the FACTS family, is widely used in improving the voltage quality of power systems and improving the stability of system operation. Accurate SVG controller model parameters are particularly important for the correctness of power system simulation analysis. However, many manufacturers do not provide corresponding SVG control parameters due to technical confidentiality, so SVG controller parameter identification is very necessary.

[0003] In the literature, a parameter identification method based on controller hardware in-the-loop testing was proposed in the SVG controller parameter identification based on RTDS hardware in-the-loop testing. The adopted particle swarm algorithm is simple but easy to fall into the local optimal solution. In the study of intelligent identification method of SVG model parameters considering the random characteristics of wind farms, a parameter identification method based on RTDS hardware in-the-loop testing was proposed. The SVG response data obtained by the test was used as the measured data. For different controller parameter combinations, BPA software was used for transient simulation. The parameters were identified according to the least squares index of the transient simulation results and the measured data. The SVG controller parameters can be accurately identified. However, there are many SVG controller parameters, and it takes a long time to identify each parameter.

[0004] In recent years, the research on deep learning in power system control has been recognized by people from all walks of life. The new generation of artificial intelligence technology represented by deep learning has great advantages in feature self-learning driven by big data, strong nonlinear fitting, neural network modeling, etc. At present, deep learning does not require the establishment of complex mathematical models and can quickly obtain results. It has been applied to power system fault analysis and location. However, there are few studies on the combination of deep learning and SVG controller parameter identification. Summary of the invention

[0005] In order to solve the problems of slow SVG parameter identification speed and poor accuracy, the present invention combines deep learning with SVG controller parameter identification to significantly improve the efficiency and accuracy of parameter identification.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The SVG parameter optimization and identification method based on convolutional neural network includes:

[0008] Step S1: establishing an equivalent mathematical model of the SVG system in the same operating environment as the SVG measured curve;

[0009] Step S2: Calculate the reactive power trajectory sensitivity, voltage trajectory sensitivity and current trajectory sensitivity of each parameter using the perturbation method, explore the trajectory sensitivity values ​​under different observation quantities, set a threshold, and select parameters whose trajectory sensitivity is greater than the set threshold;

[0010] Step S3: establishing a data set based on the parameters obtained above and performing standardization on the data set;

[0011] Step S4: Build a convolutional neural network model and use the data set to complete model training;

[0012] Step S5: Use the model to identify SVG parameters and obtain the final identification result.

[0013] Further: in step S1, the SVG measured curve is the RTDS measured curve, and the simulation tool used is PSD-BPA.

[0014] Furthermore, in step S2, the reactive power trajectory sensitivity, voltage trajectory sensitivity and current trajectory sensitivity of each parameter are calculated using the perturbation method to explore the trajectory sensitivity values ​​under different observation quantities, specifically:

[0015] Set a short-circuit fault, perform transient calculation, and record the reactive power curve, current curve, and voltage curve at the location where the SVG bus is connected; increase the selected parameter by 5% on the basis of the initial value, perform transient calculation again, obtain the output curve, and then calculate the reactive power trajectory sensitivity, current trajectory sensitivity, and voltage trajectory sensitivity of the selected parameter. The calculation formula is as follows:

[0016]

[0017] In the formula, are the reactive trajectory sensitivity, current trajectory sensitivity, and voltage trajectory sensitivity, respectively. N is the number of sampling points. Q0, I0, and U0 are the reactive value, current value, and voltage value obtained when the parameters take the initial values, respectively. Q1, I1, and U1 are the reactive value, current value, and voltage value obtained when the parameters increase by 5% on the basis of the initial values, respectively.

[0018] Furthermore, the specific process of step S3 is as follows:

[0019] SVG parameters [T1, T2, T3, T4, T5, T S ,K I ,K D ,I CMAX ] as the parameters to be identified, the ranges of each parameter are: T1 is 0.00585-0.0715, T2-T5 is 0.9-1.1, TS 0.00522-0.00638, K I 702-858, K D 0.0252-0.0308, I CMAX 1.1-1.2;

[0020] After determining the identification parameters and parameter range, starting from the lower limit, each time the SVG parameters are randomly selected within their respective ranges, each time the parameters change, the data of the reactive power curve, current curve and voltage curve and the corresponding SVG parameters are recorded to construct a data set; each time the values ​​are taken, there will be a certain error between the BPA simulation and RTDS measured reactive data, and the objective function is used to filter out the data with an error of not less than 2000 to obtain the final data set; the objective function is represented by the square of the error between the RTDS measured reactive power and the BPA simulated reactive power, and its calculation formula is as follows:

[0021]

[0022] In the formula, E Z is the error calculation result, t0 and t1 are the start and end time of the error calculation, Q base is the reactive power data measured by RTDS, Q test Reactive power data for BPA simulation.

[0023] The data of the data set is standardized, and the calculation formula is as follows:

[0024]

[0025] In the formula, X is the value of each parameter before processing, X MIN To its minimum value, X MAX is its maximum value, and Y is the processed label.

[0026] Furthermore, the specific process of step S4 is as follows:

[0027] The model includes an input layer, a convolution layer, a pooling layer, three fully connected layers and an output layer. The corresponding SVG parameters are inverted based on the reactive dynamic curve data. The input data of the input layer is a two-dimensional matrix composed of reactive, voltage and current dynamic curve data, and the output data of the output layer is the identified SVG parameters. The convolution layer uses a convolution kernel to perform a convolution operation, moves along the time step to extract the motion characteristics of the system in the current time zone, and the moving step is 1. The mapping relationship between the parameters and the curve data is constructed, and the activation function is used to perform a nonlinear activation operation on the result after the convolution calculation. The pooling layer performs a dimensionality reduction operation on the result, flattens the two-dimensional matrix data output by the convolution operation, and converts it into one-dimensional data as the input of the fully connected layer.

[0028] Before model training, the sample data set is randomly shuffled, 80% of the data is taken as the training set, and the remaining 20% ​​is used as the test set; during model training, the loss function is propagated in reverse, and the weights and bias values ​​in each layer are continuously corrected. After reaching the number of training times, testing begins; during model testing, a threshold is set for the loss function. Test results with loss function values ​​exceeding the threshold are unacceptable, while test results below this threshold are acceptable. When the test set accuracy reaches 95%, the model training is considered complete.

[0029] Furthermore, the specific process of step S5 is: using the RTDS measured curve data as the input of the model, and directly obtaining the final SVG parameter identification result through model calculation.

[0030] In summary, the method of the present invention uses the power system simulation software BPA to model and simulate SVG, and then screens out the main SVG parameters that have a greater impact on the SVG reactive dynamic curve according to the trajectory sensitivity, which can reduce the number of parameters involved in the identification and reduce the parameter identification time. Secondly, the range of SVG parameters is set, and the corresponding data set of SVG parameters and result curves is obtained by BPA simulation. The SVG parameters are used as label data, and the label data is standardized. A convolutional neural network model is built with PyTorch, and then the model is trained through the data set. The final model can identify the SVG parameters based on the actual reactive dynamic curve or the reactive dynamic curve obtained by devices such as RTDS, etc., solve the problems of poor stability and difficulty in convergence that often occur in traditional algorithms, reduce the complexity of parameter identification, improve the identification accuracy of parameters that have a greater impact on the controller, and improve the identification efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0032] Figure 2 This is a model diagram of an SVG controller in an embodiment of the present invention;

[0033] Figure 3 This is a model diagram of a single-machine infinite system connected to SVG in an embodiment of the present invention;

[0034] Figure 4 This is a CNN structure diagram in an embodiment of the present invention;

[0035] Figure 5 Schematic diagram of a convolution kernel in an embodiment of the present invention;

[0036] Figure 6 This is a structural diagram of a fully connected layer in an embodiment of the present invention;

[0037] Figure 7 This is a training flow chart in an embodiment of the present invention;

[0038] Figure 8 This is the identification result diagram using the particle swarm optimization algorithm;

[0039] Fig. 9 This is a diagram of the identification results using the method of the present invention. DETAILED DESCRIPTION

[0040] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0041] like Figure 1 The SVG parameter optimization identification method shown is shown in the figure. The detailed process is as follows.

[0042] S1. Establish an equivalent mathematical model with the same operating environment as the SVG measured curve.

[0043] The measured curve of SVG adopts the measured curve of RTDS, and the equivalent mathematical model is established using the simulation tool PSD-BPA (BPA for short). BPA has functions such as power flow calculation, transient stability simulation calculation, short circuit calculation, and small interference stability calculation. The power flow data file is a DAT text file defined by BPA that provides data and instructions for power flow calculation, and can be directly edited and modified. The transient calculation file is similar to the power flow data file. The dynamic parameters of the components, fault operations, calculations, and output controls are all input in the form of cards, and stability calculations under different interference modes can be performed by setting the parameters of the relevant cards.

[0044] like Figure 2 The SVG controller model shown in the figure, in which: V is the SVG output voltage, V REF is the reference voltage, V SCS is the auxiliary signal, V T is the system side voltage, I S is the SVG output current, T1 is the time constant of the filter and the measurement circuit, T2 and T3 are the first-order leading time constant and the second-order lagging time constant, T4 and T5 are the second-order leading time constant and the third-order lagging time constant, T P is the time constant of the proportional link, T S is the SVG response delay, K P is the magnification of the proportional link, K I is the magnification of the integral link, K D is the slope of the VI characteristic curve of SVG, X T is the equivalent reactance between SVG and the system. There are six limiting links in the figure: V MAX is the upper limit of the voltage limit link, V MIN is the lower limit of the voltage limiting link, I CMAX is the maximum capacitive current, I LMAX is the maximum inductive current, VSMAX and V SMIN Output limit for the proportional and integral links, V SMAX and V SMIN The calculation formula is as follows:

[0045]

[0046] Figure 3 It is a classic single-machine infinite model, into which SVG is connected for subsequent simulation calculations.

[0047] S2. Calculate the trajectory sensitivity of each SVG parameter and select the main parameters.

[0048] The perturbation method is used to calculate the reactive power trajectory sensitivity, voltage trajectory sensitivity, and current trajectory sensitivity of each parameter, explore the trajectory sensitivity value under different observations, set the threshold, and screen out parameters whose trajectory sensitivity is greater than the set threshold. The specific calculation process is as follows:

[0049] 1) According to the SVG controller model in BPA, the initial parameters to be identified are [T1, T2, T3, T4, T5, T S ,T P ,K P ,K I ,K D ,V MAX ,V MIN ,I CMAX ,I LMAX ](X T = is the system equivalent reactance, which is determined according to the manufacturer's given value without identification). The parameter values ​​in the VG / VG+ card in the transient data file are used as initial values. The initial values ​​are T1 = 0.005, T2 = T3 = T4 = T5 = 1, T P =0.5, T S =0.003, K P =0.05, K I =600,K D =0.02, V MAX =1, V MIN =-1,I CMAX =I LMAX =1.1. When t = 0.072, Figure 2 A three-phase short-circuit fault is set at the SVG busbar. The fault is cleared after 0.1 second. Transient calculation is performed and the output reactive power value, current value and voltage value of the SVG device are recorded.

[0050] 2) Increase T1 by 5%, keep the other parameters unchanged, and repeat the same calculation as in the previous step. Increase each of the other parameters by 5% (when one parameter changes, the other parameters remain unchanged), and obtain the simulated reactive power curve, current curve, and voltage curve after each parameter changes.

[0051] 3) Calculate the trajectory sensitivity of each parameter, the calculation formula is as follows:

[0052]

[0053] In the formula, are the reactive trajectory sensitivity, current trajectory sensitivity, and voltage trajectory sensitivity, respectively. N is the number of sampling points. Q0, I0, and U0 are the reactive value, current value, and voltage value obtained when the parameters take the initial values, respectively. Q1, I1, and U1 are the reactive value, current value, and voltage value obtained when the parameters increase by 5% on the basis of the initial values, respectively.

[0054] According to the above steps, the trajectory sensitivity is calculated by an automated program, and the calculation results are shown in Table 1.

[0055] Table 1 Parameter trajectory sensitivity

[0056]

[0057] As shown in Table 1, the voltage sensitivity of each parameter is close and difficult to compare, and the voltage sensitivity is smaller than the current sensitivity and reactive sensitivity, so the voltage sensitivity is not used as a judgment indicator. The sensitivity of each parameter obtained by the current sensitivity and reactive sensitivity is the same, so in the subsequent calculation, only the reactive power output by the SVG is selected as the indicator for judging the sensitivity.

[0058] According to the trajectory sensitivity calculation results, the parameter V MAX , V MIN The trajectory sensitivity is 0, and the parameter K P , T P The trajectory sensitivity is close to 0. In the subsequent identification, these four parameters are used as initial values ​​without identification. The required identification parameters after screening according to trajectory sensitivity are: [T1, T2, T3, T4, T5, T S , K I , K D , I CMAX ]. For these parameters, analyze the impact of each parameter on the fault characteristics.

[0059] S3. Acquisition and processing of datasets.

[0060] SVG parameters [T1, T2, T3, T4, T5, T S , K I , KD , I CMAX ] as the parameters to be identified, the ranges of each parameter are: T1 is 0.00585-0.0715, T2-T5 is 0.9-1.1, T S 0.00522-0.00638, K I 702-858, K D 0.0252-0.0308, I CMAX It is 1.1-1.2.

[0061] After determining the identification parameters and parameter range, starting from the lower limit, each time the SVG parameters are randomly selected within their respective ranges, each time the parameters change, the data of the reactive power curve, current curve, and voltage curve and the corresponding SVG parameters are recorded to construct a data set. Each time the values ​​are taken, since there will be a certain error in the reactive data of BPA simulation and RTDS measurement, it is necessary to use the objective function to filter out the data with an error of not less than 2000 to obtain the final data set. The objective function is represented by the square of the error between the reactive power measured by RTDS and the reactive power simulated by BPA, and its expression is as follows:

[0062]

[0063] In the formula, E Z is the error calculation result, t0 and t1 are the start and end time of the error calculation, Q base is the reactive power data measured by RTDS, Q test Reactive power data for BPA simulation.

[0064] Because the values ​​of the parameters to be identified vary greatly, different values ​​have different features. When training the model, the model will be more guided by certain features with large values ​​during the gradient descent process, while those with small values ​​will be ignored. Therefore, it is necessary to make different features have the same scale, so that when training the neural network, the influence of different features on the parameters can be consistent. This is achieved by standardizing the input and output data of the data set through formula (3).

[0065]

[0066] In the formula, X is the value of each parameter before processing, X MIN To its minimum value, X MAX is its maximum value, and Y is the processed label.

[0067] S4. Build a CNN model and train it to obtain the trained model.

[0068] The parameter identification model inputs time series data and outputs the result of parameter estimation in the form of a one-dimensional vector. Since the input data format is similar to a single-channel image, each unit position can be regarded as a pixel point of the image, so the CNN model, which is often used for image feature extraction, is selected. Figure 4 As shown in the figure, the CNN model of this scheme includes an input layer, a convolutional layer, a pooling layer, three fully connected layers and an output layer.

[0069] The input layer input data is a two-dimensional matrix consisting of reactive power, voltage and current dynamic curve data.

[0070] The convolution layer extracts hidden features from the original data. The dimension size of the convolution kernel is the same as the horizontal size of the data, which is a 3×3 matrix. Among them, the circled area in the input feature is the local receptive field, that is, the input data matrix is ​​locally perceived, and then the local information is integrated at a higher level to obtain global information. Therefore, the convolution only needs to move down in the original input data matrix with a step size of 1 to extract more local information. The subsequent results are then nonlinearly activated by the activation function to enhance the expression ability of the neural network. The activation function uses the Rule function, and the Relu function will make the output of some neurons 0, which causes the sparsity of the network, reduces the amount of calculation, and reduces the interdependence of parameters, alleviates the occurrence of overfitting problems, and increases the nonlinear relationship between the layers of the neural network.

[0071]

[0072] Taking a convolution kernel as an example, the input feature has only one channel, and a convolution kernel of size 3×3 is used for convolution operation, such as Figure 5 As shown, the eigenvalue T1 of the first position of the output is as follows,

[0073]

[0074] Where p is the position of the current local receptive field, X i，j is the value of the input data (i, j) at the position corresponding to the current local receptive field, b is the bias of the current convolutional layer, and w ij is the weight at the position of the convolution kernel (i, j). The current convolution layer has multiple convolution kernels. The weights of each convolution kernel are not shared, but the bias is shared. That is, after the convolution operation is performed on multiple convolution kernels, the bias of the current convolution layer is added to generate multiple feature maps, which are input to the next layer. Each feature map represents a channel of the graph in the next layer. When performing the convolution operation, the results on each channel need to be added, and finally the common bias b is added.

[0075] The pooling layer follows the convolution layer and is used to compress data and parameters and reduce the amount of calculation. The input time step of the present invention is relatively large (greater than 100), so it is chosen to add a pooling layer after the convolution layer for dimensionality reduction. The pooling layer selects MaxPooling, which is a 2×2 matrix. The calculation process is similar to the convolution calculation, but it only obtains the maximum feature as the output during the movement. The Padding value is 2, that is, the feature matrix after the convolution surrounds 2 layers of data with a value of 0, so that no feature data will be missed during the subsequent pooling operation, and all the features extracted by the convolution can be fully utilized. The obtained data is then expanded into one-dimensional data as the input of the fully connected layer.

[0076] Each neuron in the fully connected layer is fully connected to all neurons in the previous layer, which can integrate the local information with category distinction in the convolutional layer or pooling layer. The structure of the fully connected layer is as follows: Figure 6 As shown, three fully connected layers (including three hidden layers) are selected to integrate the features extracted previously.

[0077] The output layer is the parameter prediction result output after mapping of multiple fully connected layers, that is, the identified SVG parameters.

[0078] After the model is built, the model is trained.

[0079] Before training, in order to make the target output data conform to the normal distribution, the labels need to be standardized according to formula (3). Then the sample data set is randomly shuffled to avoid the influence of time series, and then 80% of the data is taken as the training set and the remaining 20% ​​as the test set.

[0080] Network training. The experiment was conducted on an NVIDIA GeForce RTX 3060Laptop GPU. Pytorch was used to build and train the neural network, and the loss function MSELoss was used to calculate the difference between the predicted value and the true value. The training flow chart is as follows Figure 7 As shown in the figure, the loss function is transferred backwards during the training process, and the weights and bias values ​​in each layer are continuously corrected by the gradient descent method (solving the minimum value of each layer along the direction of gradient descent). The loss function is the mean square error (MSE):

[0081]

[0082] Where T is the label (target output result), Y is the output result of the neural network, n is the number of SVG parameters, and t i is the SVG parameter of T at position i, y i is the SVG parameter for Y at position i.

[0083] During the training process, you can choose to use an optimizer to update and calculate the network parameters that affect model training and model output, so that they approach or reach the optimal value, thereby minimizing the loss function. The optimizer selects the Adam optimizer and sets the learning rate to 0.0001.

[0084] After reaching the number of training times, testing is performed and a threshold ε is set for the loss function. The loss function value of the test result below ε is an acceptable result, and the loss function value above ε is an unacceptable test result. Finally, when the acceptable result of the test set reaches 95%, the training is considered successful, and the final model is obtained.

[0085] S5. Use the model to identify SVG parameters and obtain the final identification result;

[0086] After obtaining the trained model, the measured curve data of RTDS can be used as the input of the model, and the final SVG parameter identification results can be directly obtained through model calculation.

[0087] The main purpose of using neural networks to estimate system parameters is to reduce the amount of calculation for parameter identification and shorten the calculation time, but the premise is to ensure the accuracy of parameter prediction results. It is necessary to verify the accuracy and feasibility of this identification method before further discussing its improvement in efficiency. Therefore, the experimental results mainly focus on two aspects: accuracy and time.

[0088] Figure 8 , 9 The RTDS curve in the middle is the measured curve. Figure 8 The optimization curve in the figure is the identification curve obtained using the particle swarm optimization method. Fig. 9 The BPA curve in the figure is the identification result curve obtained by using the convolutional neural network method. From the simulation curve, it can be seen that both methods can obtain good fitting curves, and relatively speaking, this method has higher identification accuracy. From the perspective of time, compared with the particle swarm optimization method, although this method takes a longer training time, it has a faster identification speed and higher identification accuracy.

[0089] The comparison of identification results and accuracy is shown in Table 2 and Table 3. In terms of time, the parameter identification process of this method only took 3.97 minutes, including 3.96 minutes of training time and 0.6 seconds of identification process after training, which is better than the particle swarm optimization method. From this point of view, the parameter identification using neural network can greatly shorten the time of the identification process and ensure high identification accuracy.

[0090] Table 2 Comparison of identification results (a)

[0091]

[0092] Table 3 Comparison of identification results (b)

[0093]

[0094]

[0095] The above are specific embodiments of the present invention, but the protection scope of the present invention should not be limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention, so the protection scope of the present invention should be based on the protection scope defined in the claims.

Claims

1. The SVG parameter optimization identification method based on convolutional neural network is characterized by: include: Step S1: establishing an equivalent mathematical model of the SVG system in the same operating environment as the SVG measured curve; Step S2: Calculate the reactive power trajectory sensitivity, voltage trajectory sensitivity and current trajectory sensitivity of each parameter using the perturbation method, explore the trajectory sensitivity values ​​under different observation quantities, set a threshold, and select parameters whose trajectory sensitivity is greater than the set threshold; Step S3: establishing a data set based on the parameters obtained above and performing standardization on the data set; Step S4: Build a convolutional neural network model and use the data set to complete model training; Step S5: using the model to identify SVG parameters and obtain the final identification result; The specific process of step S3 is as follows: SVG Parameters As the parameters to be identified, the range of each parameter is: is 0.00585-0.0715, - 0.9-1.1, is 0.00522-0.00638, is 702-858, is 0.0252-0.0308, 1.1-1.2; After determining the identification parameters and parameter range, starting from the lower limit, each time the SVG parameters are randomly selected within their respective ranges, each time the parameters change, the data of the reactive power curve, current curve and voltage curve and the corresponding SVG parameters are recorded to construct a data set; each time the values ​​are taken, there will be a certain error between the BPA simulation and RTDS measured reactive data, and the objective function is used to filter out the data with an error of not less than 2000 to obtain the final data set; the objective function is represented by the square of the error between the RTDS measured reactive power and the BPA simulated reactive power, and its calculation formula is as follows: ； In the formula, is the error calculation result, and is the start time and end time of calculating the error, is the reactive power data measured by RTDS, Reactive power data for BPA simulation; The data of the data set is standardized, and the calculation formula is as follows: ； In the formula, are the parameter values ​​before processing, is its minimum value, For its maximum value, is the processed label; The specific process of step S4 is as follows: The model includes an input layer, a convolution layer, a pooling layer, three fully connected layers and an output layer. The corresponding SVG parameters are inverted based on the reactive dynamic curve data. The input data of the input layer is a two-dimensional matrix composed of reactive, voltage and current dynamic curve data, and the output data of the output layer is the identified SVG parameters. The convolution layer uses a convolution kernel to perform a convolution operation, moves along the time step to extract the motion characteristics of the system in the current time zone, and the moving step is 1. The mapping relationship between the parameters and the curve data is constructed, and the activation function is used to perform a nonlinear activation operation on the result after the convolution calculation. The pooling layer performs a dimensionality reduction operation on the result, flattens the two-dimensional matrix data output by the convolution operation, and converts it into one-dimensional data as the input of the fully connected layer. Before model training, the sample data set is randomly shuffled, 80% of the data is used as the training set, and the remaining 20% ​​is used as the test set; During model training, the loss function is propagated in reverse, and the weights and bias values ​​in each layer are continuously corrected. After reaching the number of training times, testing begins. During model testing, a threshold is set for the loss function. Test results with loss function values ​​exceeding the threshold are unacceptable, while test results below this threshold are acceptable. When the test set accuracy reaches 95%, the model training is considered complete.

2. The SVG parameter optimization and identification method based on convolutional neural network according to claim 1, characterized in that: In step S1, the SVG measured curve is the RTDS measured curve, and the simulation tool used is PSD-BPA.

3. The SVG parameter optimization and identification method based on convolutional neural network according to claim 1, characterized in that: In step S2, the reactive power trajectory sensitivity, voltage trajectory sensitivity and current trajectory sensitivity of each parameter are calculated by using the perturbation method to explore the trajectory sensitivity values ​​under different observation values, specifically: Set a short-circuit fault, perform transient calculation, and record the reactive power curve, current curve, and voltage curve at the location where the SVG bus is connected; increase the selected parameter by 5% on the basis of the initial value, perform transient calculation again, obtain the output curve, and then calculate the reactive power trajectory sensitivity, current trajectory sensitivity, and voltage trajectory sensitivity of the selected parameter. The calculation formula is as follows: （1） In the formula, They are reactive trajectory sensitivity, current trajectory sensitivity, and voltage trajectory sensitivity. is the number of sampling points, They are the reactive power value, current value, and voltage value obtained when the parameters take their initial values. They are the reactive power value, current value and voltage value obtained by increasing the parameters by 5% on the basis of the initial values.

4. The SVG parameter optimization and identification method based on convolutional neural network according to claim 1, characterized in that: The specific process of step S5 is: using the RTDS measured curve data as the input of the model, and directly obtaining the final SVG parameter identification result through model calculation.

Images