A power system transient stability evaluation method based on multi-source information

Abstract

The application discloses a kind of belonging to power system transient stability evaluation based on multi-source information's transient stability evaluation method.Firstly, by changing system operating state and fault condition, using power system analysis program for batch time domain simulation to construct massive data set for IEEE39 node system.Then, considering the influencing factors affecting system transient stability, the operating information and fault information are used as the input features of the convolutional neural network model.In order to better realize the comprehensive utilization of operating information and fault information, and avoid the influence of feature dimension difference on accuracy, two different feature fusion schemes are adopted for phased fusion of features, and different transient stability evaluation models are formed.Then, the transient stability evaluation model established by training is used to evaluate the transient stability of the test set, and the accuracy, precision, recall and index of the confusion matrix tool are used to evaluate the performance of the model.Finally, the t-SNE algorithm is used to visualize the established transient stability evaluation model.

Description

Technical Field

[0001] This invention belongs to the field of transient stability assessment of power systems, and particularly relates to a transient stability assessment method based on multi-source information and deep learning. Background Technology

[0002] Transient stability assessment is one of the main issues in power system security and stability analysis. Accurate transient stability assessment is crucial for formulating power system risk prevention and control strategies. Therefore, transient stability assessment methods play an important role in the safe and stable operation of power systems. Current transient stability assessment methods can be mainly divided into time-domain simulation methods and direct methods, each with its own adaptive time scale and data types.

[0003] Time-domain simulation (TDS), also known as the indirect method, is based on the idea of ​​using numerical integration algorithms to solve the time-varying differential-algebraic equations describing the system before and after the disturbance, and then judging the system's stability based on the rotor oscillation curves of each generator. Due to its good model adaptability and reliable results, this algorithm has been widely used in academic research and practical production. However, the time-domain simulation algorithm also has the following limitations: large computational load, long computation time, and low computational efficiency; although it can obtain the time-varying differential-algebraic equations of the disturbed system through time-domain simulation, there is still no unified and clear stability criterion available. It can only make a rough judgment based on engineering experience methods such as the oscillation trend of the power angle curve or the maximum relative power angle difference within a certain time window, and the accuracy of the judgment needs to be improved. The direct method, utilizing energy functions and Lyapunov theory, analyzes power system stability. Compared to time-domain simulation, this algorithm eliminates the need for time-consuming simulations of the system after a fault and provides a measure of system instability. Besides the energy function method, the Extended Equal Area Criteria (EEAC) is another important branch of the direct method. This method is based on the "equal area rule" of a single-machine infinite bus system, dividing the multiple generators in the system into two coherent groups, thus representing an equivalent single-machine infinite bus system. The algorithm is quite effective when the system transient process exhibits a two-group swing mode, but it is difficult to apply to analyzing multiple swing modes. Both time-domain simulation and the direct method have certain disadvantages in terms of computational efficiency, model adaptability, and algorithm stability, requiring further research.

[0004] Deep learning, a branch of machine learning, aims to build deeper structures to enhance the model's ability to capture hidden features in massive amounts of data. Compared with traditional transient stability assessment methods, deep learning has a series of hidden layers capable of nonlinear transformations, thus enabling it to tackle more complex environments and problems. However, the "black box" nature of deep learning limits the effectiveness of algorithm improvement methods. To address these issues, this invention proposes a transient stability assessment method based on multi-source information, comprehensively utilizing multi-source information in the power system to improve prediction accuracy. Therefore, applying deep learning theory to the power system field is a crucial support for the transformation and development of energy and power, and an inevitable choice for power grid development. Summary of the Invention

[0005] To achieve the above objectives, this invention provides a power system transient stability assessment method based on multi-source information and deep learning, the transient stability assessment method comprising the following steps:

[0006] Step 1: By changing the system operating state and fault conditions, a massive dataset is constructed by performing batch time-domain simulations on the IEEE 39-bus system using Power System Analysis and Synthesis Program (PSASP) 7.x.

[0007] Step 2: Taking into account the factors affecting the transient stability of the system, the operation information and fault information are used as input features of the Convolutional Neural Network (CNN) model. Two different feature fusion schemes are used to fuse the features in stages to form different transient stability evaluation models.

[0008] Step 3: Use the transient stability evaluation model established through training to evaluate the transient stability of the test set and obtain the results;

[0009] Step 4: Use the accuracy (Acc), precision (Prec), recall (Recall), and F1-score metrics from the confusion matrix tool to comprehensively evaluate the model performance;

[0010] Step 5: Visualize the established transient stability assessment model.

[0011] Optionally, the transient stability assessment model can be expressed as the following formula:

[0012]

[0013] in, t represents the model's output, i.e., the transient stability assessment result at time t. y, as the model's input, represents the power system's operational information (such as voltage, load, and generator output). t, L, l, p lThese are also the inputs to the model, namely, fault time, fault location, line, and load level.

[0014] This invention is the first to use fault information and operational information, after feature extraction, as input features for a model, forming a power system transient stability assessment method based on multi-source data fusion. Compared with current transient stability assessment methods, this invention significantly improves assessment accuracy and generalization ability, providing a basis for rational power grid dispatch and possessing industrial application value.

[0015] Optionally, in step 1, a massive dataset is constructed by performing batch time-domain simulations on the IEEE 39-bus system using Power System Analysis and Synthesis Program (PSASP) 7.x. The fault parameters are set as follows: six load conditions with line load levels ranging from 70% to 120% increasing in 10% increments; all faults are three-phase short-circuit faults; the durations are set to 0.1, 0.15, and 0.2 seconds, respectively; the fault locations are 1%, 50%, and 99%; the simulation duration is 5 seconds; and a total of 1836 initial sample data are generated.

[0016] Optionally, in step 2, the operating information and fault information are used as input features of the Convolutional Neural Network (CNN) model, and two different feature fusion schemes are used to fuse the features in stages. This invention selects four features as fault features: fault time, fault location, line and load level; and selects features such as voltage, load and generator output in the power system as operating features. The different feature fusion schemes are as follows: (1) Parallel input in the input stage, keeping the network structure and network parameters of the model unchanged, fusing the system operating features and fault features in the input stage of the model, and inputting them into the model in parallel to complete the training; (2) Serial input in the output stage, keeping the network structure and network parameters of the model unchanged, fusing the system operating features and fault features in the output stage of the model, and inputting them into the model in serial manner to complete the training; (3) As a control group, fault features are not considered, and only the operating features are used as input samples to complete the model training.

[0017] Optionally, in step 3, 70% of the samples are randomly selected by stratified sampling as the training set, while the remaining 30% are selected as the test set.

[0018] Optionally, the Convolutional Neural Network (CNN) optimizes the algorithm by setting parameters and iteratively calculating the accuracy predictions for different convolutional kernels and learning rates. Then, after fixing the parameters, the impact of different optimization methods (RMSprop, SGD, and Adam) on the model is compared, and finally the Adam algorithm is selected as the optimization algorithm.

[0019] Optionally, the comprehensive evaluation indicators Acc, Precision, Recall, and F1-score in step 4 can be expressed by the following formula:

[0020]

[0021]

[0022]

[0023]

[0024] TP, TN, FP, and FN represent the classification results of the confusion matrix.

[0025] Optionally, in step 5, the t-SNE algorithm is used to map the features in different schemes to a low-dimensional space to visualize the evaluation model.

[0026] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0027] The transient stability assessment method based on multi-source information and deep learning described in this invention, on the one hand, improves the model's ability to extract features and reduce dimensionality of the original data by using a Convolutional Neural Network (CNN) to perform deep feature mining on the input data, and reduces overfitting during training by introducing dropout technology (derived from the Adam approach), thereby improving prediction accuracy. On the other hand, by comprehensively considering the factors affecting the transient stability of the system, it uses operational information and fault information as input features of the CNN model, which better realizes the comprehensive utilization of operational information and fault information, avoids the impact of feature dimensionality differences on accuracy, and thus further improves prediction accuracy. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the overall process of the transient stability assessment method based on multi-source information and deep learning described in this invention.

[0029] Figure 2 This is a schematic diagram of the Convolutional Neural Network (CNN) model described in this invention.

[0030] Figure 3 This is a schematic diagram of a feature fusion scheme.

[0031] Figure 4 The learning curves of the CNN model are shown when different optimization algorithms (RMSprop, SGD, Adam) are used.

[0032] Figure 5 The diagram shows the confusion matrices for models using different feature fusion schemes. In the figure, (a) represents the control group, (b) represents Scheme 1 (parallel input in the input stage), and (c) represents Scheme 2 (serial input in the output stage).

[0033] Figure 6 This is a visualization of the results of mapping features from the control group scheme to a low-dimensional space using the T-SNE algorithm.

[0034] Figure 7 This is a visualization of the results of mapping the features in Scheme 1 to a low-dimensional space using the T-SNE algorithm.

[0035] Figure 8 This is a visualization of the results of mapping the features in Scheme 2 to a low-dimensional space using the T-SNE algorithm. Detailed Implementation

[0036] The present invention will be further described below with reference to specific embodiments.

[0037] This invention addresses the problems of high computational cost and low accuracy in current transient stability assessments. In this embodiment, power system operation information and fault information are used as input features to a Convolutional Neural Network (CNN) model, and transient stability assessment is performed using deep learning methods. The implementation scheme is described in detail below.

[0038] Reference Figure 1 As shown, Figure 1 This is a flowchart illustrating the overall process of the power system transient stability assessment method based on multi-source information and deep learning according to the present invention. The transient stability assessment method based on multi-source information and deep learning includes the following steps:

[0039] Step 1: Taking the New England 10-unit 39-bus system as an example, the generator model adopts a second-order classical model, and the load adopts a constant impedance mode. By changing the system operating state and fault conditions, a massive dataset is constructed by batch time-domain simulation using Power System Analysis and Synthesis Program (PSASP) 7.x.

[0040] Step 2: Taking into account all factors affecting the transient stability of the system, operational information and fault information are used as input features of the CNN model. Three different feature fusion schemes are employed to fuse the features in stages, resulting in different transient stability evaluation models. (Refer to...) Figure 2 As shown, Figure 2This diagram illustrates a Convolutional Neural Network (CNN) model. A complete CNN network consists of three layers: convolutional layers, pooling layers, and fully connected layers. The convolutional layers primarily extract features from the input layer and enhance the model's ability to process non-linear data through activation functions. The computation process of the convolutional layer is represented as follows:

[0041]

[0042] Among them, H i W represents the features of the i-th layer of a convolutional neural network. i This represents the weight vector of the i-th layer; This represents the convolution operation between the convolution kernel and the features of the previous layer, where b is the offset vector of the i-th layer. i The features of the next layer are obtained by summing them and then applying the nonlinear activation function f(x).

[0043] Pooling layers primarily extract and reduce the dimensionality of data from convolutional layers, not only reducing information redundancy but also preventing overfitting to some extent. The calculation process of a pooling layer is represented as follows:

[0044] H i =subsampling(H i-1 )

[0045] By fully connecting to all neurons in the previous layer, the fully connected layer integrates class-discriminating local information from the convolutional or pooling layers. The training objective of a Convolutional Neural Network (CNN) is to minimize a function of the network's loss, typically using the cross-entropy function.

[0046]

[0047] During training, Convolutional Neural Networks (CNNs) typically employ gradient descent for optimization, updating the trainable parameter weights (W) of each layer of the CNN layer by layer. i ) and offset (b i The learning rate parameter (η) is used to adjust the backpropagation speed.

[0048]

[0049]

[0050] Three different feature fusion schemes are referenced. Figure 3As shown in the control group, fault characteristics are not considered; only operational characteristics are used as input samples, and a convolutional neural network model is trained using the training set. As shown in Scheme 1, the network structure and parameters of the model remain unchanged, and the system operational characteristics and fault characteristics are fused at the input stage of the model and input into the model in parallel. A new convolutional neural network model is then retrained using a new training set. As shown in Scheme 2, the network structure and parameters of the model remain unchanged, and the system operational characteristics and fault characteristics are fused at the output stage of the model and input into the model serially to complete the training.

[0051] Reference Figure 4 As shown, Figure 4 The curve represents the loss function value after fixing the convolution kernel and learning rate. Figure 4 The left side of the graph shows the loss function value curves on the training set. Figure 4 The right-hand graph shows the loss function curves on the test set. The dashed line represents RMSprop, the dotted line represents SGD, and the solid line represents Adam. It can be seen that although the blue line converges faster than the dotted and solid lines in the training set, the dashed line shows an increasing trend in the loss function in the test set, exceeding the dotted and solid lines in the later stages of training, indicating overfitting. Therefore, the Adam algorithm is chosen as the optimization algorithm.

[0052] Step 3: Use the transient stability evaluation model established through training to evaluate the transient stability of the test set and obtain the results.

[0053] Step 4: To visually demonstrate the impact of different feature fusion methods on model performance, this invention first uses a confusion matrix to directly observe the model's performance across various categories. The confusion matrix is ​​referenced... Figure 5 As shown in the figure. Then, the accuracy (Acc), precision (Prec), recall, and F1-score metrics from the confusion matrix tool are used to comprehensively evaluate the model performance.

[0054] Step 5: Use the t-SNE algorithm to map the features from different schemes to a low-dimensional space, and then visualize the evaluation model. (Refer to...) Figure 6 , 7 As shown in Figure 8, Figure 6 , 7 Figures 8 and 9 are visualizations of the phased feature fusion model using three different feature fusion schemes.

[0055] The power system transient stability assessment method based on multi-source information and deep learning described in this invention employs the adaptive moment estimation algorithm (Adam) as its gradient optimization algorithm. The model uses the 'dropout' method to reduce overfitting during training. During training, dropout randomly discards hidden neurons in the network with a certain probability, essentially setting the input and output of those neurons to zero. This effectively reduces the number of internal parameters in the model and also indirectly increases the diversity of the model's input data, thus mitigating overfitting to some extent.

[0056] This invention addresses the problem of high computational complexity and low accuracy in current transient stability assessments. In its embodiments, power system operation information and fault information are used as input features of a convolutional neural network (CNN) model. Transient stability assessment is performed using deep learning methods, resulting in higher accuracy and stronger generalization ability, and thus has engineering application value.

[0057] The above-described embodiments are merely preferred embodiments of this invention and are not intended to limit the scope of this invention. Any equivalent structural or functional transformations made using the description and drawings of this invention, or any direct or indirect applications in other related technical fields, are similarly included within the scope of protection of this invention.

Claims

1. A method for evaluating the transient stability of a power system based on multi-source information and deep learning, characterized in that, The transient stability assessment method includes the following steps: Step 1: By changing the system operating state and fault conditions, a massive dataset is constructed by performing batch time-domain simulations on the IEEE 39-bus system using the Power System Analysis and Synthesis Program 7.x. Step 2: Taking into account the factors affecting the transient stability of the system, the operation information and fault information are used as input features of the convolutional neural network model. Two different feature fusion schemes are used to fuse the features in stages to form different transient stability evaluation models. Step 3: Use the transient stability evaluation model established through training to evaluate the transient stability of the test set and obtain the results; Step 4: Use the accuracy (Acc), precision (Prec), recall (Recall), and F1-score metrics from the confusion matrix tool to comprehensively evaluate the model performance; Step 5: Visualize the established transient stability assessment model using the t-SNE algorithm; In step 2, the operation information and fault information are used as input features of the convolutional neural network model, and the features are fused in stages using two feature fusion schemes: parallel and serial. The operation information and fault information are selected as features, specifically, four features in the fault information, namely fault time, fault location, line and load level, are selected as fault features. The voltage, load and generator output features in the power system are selected as operation features.

2. The power system transient stability assessment method based on multi-source information and deep learning as described in claim 1, characterized in that, The transient stability assessment model can be expressed as the following formula: in, t represents the output of the model, i.e., the transient stability assessment result at time t; y represents the input of the model, which is the operating information of the power system; t, L, l, p l These are also the inputs to the model, namely, fault time, fault location, line, and load level.

3. The power system transient stability assessment method based on multi-source information and deep learning as described in claim 1, characterized in that, In step 1, the fault parameters of the dataset were set as follows: six load conditions with line load levels ranging from 70% to 120% and increasing by 10% increments. All faults were three-phase short-circuit faults. The durations of the three-phase short-circuit faults were set to 0.1, 0.15, and 0.2 seconds, respectively. The fault locations were 1%, 50%, and 99%, respectively. The simulation duration was 5 seconds, and a total of 1836 initial sample data were generated.

4. The power system transient stability assessment method based on multi-source information and deep learning as described in claim 1, characterized in that, The parallel feature fusion scheme is as follows: during the input stage, input is performed in parallel, keeping the network structure and network parameters of the model unchanged, and the system operation features and fault features are fused during the input stage of the model, and input into the model in a parallel manner to complete the training. The serial feature fusion scheme is as follows: the output stage is input serially, keeping the network structure and network parameters of the model unchanged, and the system operation features and fault features are fused in the output stage of the model and input into the model serially to complete the training; a model that only considers the operation features is set as a control group.

5. A power system transient stability assessment method based on multi-source information and deep learning as described in any one of claims 1-3, characterized in that, In step 3, 70% of the samples are randomly selected by stratified sampling as the training set, while the remaining 30% are used as the test set.

6. A power system transient stability assessment method based on multi-source information and deep learning as described in any one of claims 1-3, characterized in that, In step 2, parameters are set and the algorithm is optimized in a loop. The accuracy prediction is calculated for different convolution kernels and learning rates. Then, after fixing the parameters, the impact of different optimization methods such as RMSprop, SGD and Adam on the model is compared. Finally, the Adam algorithm is selected as the optimization algorithm.

7. A power system transient stability assessment method based on multi-source information and deep learning as described in any one of claims 1-3, characterized in that, In step 4, the comprehensive evaluation indicators Accuracy (Acc), Precision (Prec), Recall (Recall), and F1-score can be expressed by the following formula: TP, TN, FP, and FN represent the classification results of the confusion matrix.

8. A power system transient stability assessment method based on multi-source information and deep learning as described in any one of claims 1-3, characterized in that, In step 5, the t-SNE algorithm is used to map the high-dimensional nonlinear features in different schemes to a low-dimensional space, and to visualize the evaluation model.

Images