New energy grid-connected system transient voltage stability evaluation method based on bidirectional transfer learning

Abstract

A transient voltage stability evaluation method for a new energy grid-connected system based on bidirectional transfer learning comprises the following steps: establishing a database containing a large amount of historical simulation result information, and fitting an operation scene of a current system through the database; calculating transient voltage stability indexes TVSI of the samples in the database, and dividing the TVSI into different transient voltage stability categories; based on the data set in the database, a long short-term memory (LSTM) network model is trained offline, and an LSTM-based power system transient voltage stability evaluation model is established; in consideration of the situation that the operation condition of the power system changes, the data set is finely adjusted through bidirectional transfer learning, so that the transient voltage stability evaluation model of the power system can adapt to new working conditions. When the change amplitude of the transient voltage of the power system is large, the transient voltage stability evaluation method for the new energy grid-connected system based on bidirectional transfer learning is adopted to perform evaluation in time, and if the system is unstable, an alarm is immediately given to a worker, so that the transient voltage stable operation of the power system is ensured.

Description

Technical Field

[0001] The present invention relates to the field of transient voltage stability assessment of power systems, and in particular to a transient voltage stability assessment method for a new energy grid-connected system based on bidirectional transfer learning. Background Art

[0002] In recent years, the capacity of renewable energy grid-connected power generation has grown rapidly, and electricity generated by renewable energy has accounted for a considerable proportion. However, the output power of renewable energy generators often fluctuates rapidly and violently, which will have a significant impact on the transient voltage stability of the power system. In recent years, a number of accidents caused by renewable energy generation have occurred around the world. Therefore, rapid and reliable transient voltage stability assessment of power systems combined with renewable energy generation should be paid attention to.

[0003] The patent document with the authorization announcement number CN115935264B discloses a method for evaluating the transient voltage stability of a power grid based on synchronous learning of spatiotemporal information. Through various transient faults, time domain simulation is performed to obtain a transient sample set, and then the spatial adjacency matrix and the time series adjacency matrix are constructed, and the transient voltage stability classification evaluation is performed using the graph convolutional neural network algorithm. However, this method does not take into account the unpredictability of the system operating conditions, which leads to the lack of robustness of the model. The patent document with the authorization announcement number CN116031879A discloses a hybrid intelligent feature selection method adapted to the transient voltage stability evaluation of the power system. A high-dimensional time series sample data set is obtained by analyzing stability-related factors, constructing the original feature set, and simulating transient time domains in multiple scenarios. The feature validity measurement and preliminary screening are performed based on the T-Relief algorithm, and then feature selection and stability evaluation based on the improved swarm intelligence algorithm are performed. However, this method does not take into account the problem of grid connection of new energy generators, which may lead to a decrease in the accuracy of transient voltage stability evaluation results, and even unreliable evaluation results.

[0004] The biggest disadvantage of traditional time-domain simulation is that it is time-consuming, especially with the grid connection of new energy power generation equipment, the scale of power systems is getting larger and larger, and the integration of power electronic equipment is increasing. However, due to the large output fluctuations of renewable energy power generation equipment such as wind power and photovoltaic power generation equipment, the requirements for the rapidity of transient voltage stability assessment have increased. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a transient voltage stability assessment method for a new energy grid-connected system based on bidirectional transfer learning. When the transient voltage of the power system changes greatly, an assessment is performed in a timely manner. If the system becomes unstable, an alarm is immediately sent to the staff to ensure the stable operation of the transient voltage of the power system.

[0006] The technical solution adopted by the present invention is:

[0007] The transient voltage stability assessment method of a new energy grid-connected system based on bidirectional transfer learning includes the following steps:

[0008] Step 1: Establish a database containing a large amount of historical simulation result information, and fit the operation scenario of the current system through the database;

[0009] Step 2: Calculate the transient voltage stability index TVSI of the samples in the database and divide them into different transient voltage stability categories;

[0010] Step 3: Based on the data set in the database, the long short-term memory network LSTM model is trained offline to establish a transient voltage stability assessment model for the power system based on LSTM;

[0011] Step 4: Considering the situation where the operating conditions of the power system change, the data set is fine-tuned through bidirectional transfer learning so that the power system transient voltage stability assessment model can adapt to the new operating conditions.

[0012] In the step 1, an initial database containing a large amount of historical simulation result information is established for the situation of grid-connected renewable energy power generation, wherein the database includes data such as generator active power, generator reactive power, load active power, load reactive power, voltage amplitude and phase angle; thereafter, the operation scenario of the current system is fitted through the database, wherein the established operation scenario includes parameters such as fault line, fault location, fault duration, fault type, etc., and multiple scenarios are constructed for testing to verify the advantages of the method of the present invention.

[0013] In step 1, in order to adapt to specific operation scenarios, the distance between operation scenarios needs to be defined accordingly:

[0014] Define the i-th operation scenario s i As shown below:

[0015] s i = {V i ,θ i ,P Gi ,Q Gi ,P Li ,Q Li} (1);

[0016] Where V i is the voltage amplitude of node i; θ i is the voltage phase angle of node i; P Gi is the active power of the generator at node i; Q Gi is the reactive power of the generator at node i; P Li is the load active power of node i; Q Li is the load reactive power of node i.

[0017]

[0018] Where V i 1 is the voltage amplitude of the first scene node i; V i n is the voltage amplitude of the nth scene node i; is the voltage phase angle of the first scenario node i; is the voltage phase angle of the nth scene node i; is the active power of the generator of the first scenario node i; is the active power of the generator at the nth scenario node i; is the reactive power of the generator at the first scenario node i; is the reactive power of the generator at the nth scenario node i; is the load active power of the first scenario node i; is the load active power of the nth scene node i; is the reactive power of the load at the first scenario node i; is the reactive power of the load at the nth scenario node i.

[0019] Taking the distance between scene s0 and scene s1 as an example, the definition is as follows:

[0020]

[0021] Where D(s0,s1) is the distance between scene s0 and scene s1; V i,0 is the voltage amplitude of node i in scenario s0; V i,1 is the voltage amplitude of node i in scenario s1; θ i,0 is the voltage phase angle of node i in scenario s0; θ i,1 is the voltage phase angle of node i in scenario s1; P Gi,0 is the active power of the generator at node i in scenario s0; P Gi,1 is the active power of the generator at node i in scenario s1; Q Gi,0 is the reactive power of the generator at node i in scenario s0; Q Gi,1 is the reactive power of the generator at node i in scenario s1; P Li,0 is the load active power of node i in scenario s0; P Li,1 is the active power of the load at point i in scenario s1; Q Li,0 is the reactive power of the load at node i in scenario s0; Q Li,1 is the reactive power of the load at node i in scenario s1.

[0022] If the distance between two scenes is very small, then the two scenes are considered to have the highest similarity. Take the reciprocal of D(·) -1 As the weight of s1, to fit the scene s0.

[0023] In the step 2, by calculating the transient voltage stability index TVSI, multiple samples are divided into unstable, stable and stable with fault-induced voltage delayed recovery phenomenon, and are marked with numbers 0, 1 and 2 respectively.

[0024] The samples are classified by transient voltage stability index TVSI as follows:

[0025]

[0026] Where n is the number of system nodes; T is the voltage recovery time; t cr is the fault removal time; TVDI j,t is the transient voltage deviation index of node i at time t.

[0027] The transient voltage deviation index (TVDI) in the transient voltage stability index TVSI can be expressed as:

[0028]

[0029] Where V j,t is the voltage amplitude measured at node j at time t; V j,0 is the voltage amplitude before the fault of node j; α is taken as 10%, which is the threshold of unacceptable voltage deviation amplitude. The transient voltage stability index TVSI is used to evaluate the transient voltage stability of the samples collected in the entire system, and three labels of 0, 1 and 2 are assigned respectively.

[0030] During the simulation, all samples whose bus node voltage change trajectories remain no lower than the pre-fault voltage within a stable time of more than 3 seconds are marked as stable; further, by evaluating the transient voltage stability index TVSI of the stable sample, the stable samples are reclassified according to whether they are prone to fault-induced voltage delayed recovery phenomenon. The stable samples with TVSI>4 are marked as stable samples with fault-induced voltage delayed recovery phenomenon, and the others are stable samples; the simulation process lasts for 5 seconds, and the samples whose bus voltage change trajectory is lower than the pre-fault voltage are classified as unstable.

[0031] In the step 3, a long short-term memory network LSTM model is trained offline, and a transient voltage stability assessment model for a power system based on LSTM is established by learning the time dependence of the system dynamics after the disturbance. The specific steps are as follows: first, data preparation is performed, data is collected and preprocessed, and the data is divided into a training set and a test set; second, an LSTM model is constructed, an input layer, one or more LSTM layers, and a fully connected layer are defined, each LSTM layer consists of multiple LSTM units and is trained using a training data set; then, a model evaluation is performed, and the model performance is evaluated on an independent training set or test set; finally, the model structure or parameters are adjusted according to the evaluation results.

[0032] Among them, the specific working principles of the forget gate, input gate and output gate of the LSTM model are as follows:

[0033] f t =σ(W f ·[h t-1 ,x t ]+b f ) (6);

[0034] i t =σ(W i ·[h t-1 ,x t ]+b i ) (7);

[0035] o t =σ(W o ·[h t-1 ,x t ]+b o ) (8);

[0036] In the formula, f t For the forget gate; i t is the input gate; o t is the output gate; W f , W i , W o is the weight matrix; b f , b i , b o is the bias vector; σ(·) is the activation function; h t-1 is the transient voltage timing characteristic information output at the previous moment; x t It is the transient voltage timing characteristic information input at the current moment.

[0037] The transformation process of the memory unit is described as follows:

[0038]

[0039] In the formula, is the candidate memory gate; tanh is the activation function; W c is the weight matrix; b c is the bias vector; c t is the memory unit of the LSTM model, indicating the state of the memory unit at the current time step t; is the element product; c t-1 is the state of the memory unit at the previous time step t-1; W c ·[h t-1 ,x t ] is the product of the time series feature information and the weight matrix.

[0040] Hidden layer state h t The representation is as follows:

[0041]

[0042] In the formula, tanh(·) represents a nonlinear activation function. The unique structure of using memory units instead of hidden layer nodes in the LSTM model enables it to extract time-domain related features for a long time.

[0043] In step 3, the LSTM-based power system transient voltage stability assessment model mainly consists of five parts: input layer, LSTM layer, fully connected layer, softmax layer and output layer, as shown in Figure 2 In this evaluation model, the input layer is the system dynamic measurement collected after the failure occurs; the LSTM layer is used to capture the potential time dependency between time series data, and uses the internal memory of the LSTM model and a unique gating mechanism to process the input in chronological order; the fully connected layer plays the role of dimensionality reduction in the whole process, and data is discarded at this layer to prevent the model from overfitting.

[0044] In the softmax layer, the softmax activation function is used to map hidden features to the final evaluation results, realizing multi-classification of the evaluation model. The process can be expressed as:

[0045] y t =softmax(W S ·h t +b S ) (12);

[0046] In the formula, y t is the predicted label of the sample; h t is a hidden feature; W S and b S They are the weight matrix and bias vector that need to be updated during the offline training of the model; softmax(·) is the activation function.

[0047] In step 4, in the actual operation of the power system, the operating conditions may undergo some changes that have not been considered. Bi-TL is used to fine-tune the data set, and the existing data set and model are used to help the model adapt to the new operating conditions, avoiding time-consuming and laborious data labeling work. The learning of bidirectional transfer learning Bi-TL includes the following steps:

[0048] S4.1: Use the information in the target domain data for forward learning, called query (+), to obtain the true label of the sample whose label the current evaluation model is most uncertain about, which helps to eliminate uncertainty and bring the most effective information to the model; this is achieved through marginal sampling:

[0049]

[0050] In the formula, x is the input sample variable; x + is the marginal value of variable x; and Model Posterior probability predictions for the first and second most likely classes; and are the labels of the first and second most likely categories of variable x; D T is the target domain; U is the unlabeled sample in the target domain.

[0051] The smaller the margin, the higher the sample information content. This can also be achieved through entropy sampling, expressed as:

[0052]

[0053] In the formula, y i Contains all possible classes. Samples with high entropy contain relatively rich information. The result of forward learning is B+ target samples with the largest uncertainty, forming a set Query the stable label of the sample through the TVSI indicator After that, it is moved to the labeled training set. The added labeled target domain samples are represented as: in, is the marginal value set of variable x; Y t + for The corresponding tag set; is the target domain sample; x i is the input sample variable; i For x i The hyperparameter B+ is chosen to strike a balance between accuracy and adaptation speed as well as the convenience of sample labeling.

[0054] S4.2: Use the information in the initial domain data for backward learning, called query (-), analyze the existing initial domain data, intelligently remove interference information, and expand the number of labeled samples for fine-tuning by leveraging the information in the initial database.

[0055] Specifically, it starts from the initial domain D S The labeled samples that deviate from the target domain distribution are deleted. Therefore, the initial domain samples that conform to the target domain distribution will be retained. In this way, the information in the initial domain data is used to expand the number of labeled samples used for fine-tuning.

[0056] Backward learning removes the initial domain samples of the maximum posterior probability change. The specific function implementation process is as follows:

[0057]

[0058] In the formula, x - is the query value of variable x; and The models are and The posterior probability vectors of different stability categories are calculated; the L1 norm ||·||1 is the integral of the changes in each dimension of the posterior probability vector; x is the input sample variable; y is the label of variable x.

[0059] Backward learning is repeated B times to remove B initial domain samples that no longer represent the target domain. The removed sample labels are expressed as: in, is the set of selected samples; Y t - for The corresponding label; is the target domain sample.

[0060] The combination of forward learning and backward learning can comprehensively learn the knowledge of target domain data and initial domain data, reduce the labeling process of new costs in the target domain, and the effect of the generated hybrid training data set is better than the randomly selected target domain data. The present invention provides a transient voltage stability assessment method for a new energy grid-connected system based on bidirectional transfer learning, and the technical effects are as follows: 1) The present invention designs a method for fitting the current system operation scenario through an initial database for the grid-connected situation of new energy power generation, and uses the information stored in the database to quickly analyze the transient voltage stability of the target operation scenario through linear data fitting.

[0061] 2) The present invention uses an indicator TVSI for measuring the voltage stability of samples. By calculating TVSI, the stability of samples can be divided more specifically compared with the binary classification method used in the traditional evaluation process.

[0062] 3) In view of the fact that some unconsidered changes may occur in the system operation status, the present invention fine-tunes the data set through Bi-TL, using the existing data set and model to help the model adapt to the new operating conditions, avoiding time-consuming and labor-intensive data labeling and model retraining. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a flow chart of the method for evaluating transient voltage stability of a power system proposed by the present invention.

[0064] Figure 2 It is a diagram of the LSTM structure used in the present invention.

[0065] Figure 3 It is a schematic diagram of the fine-tuning process of transfer learning proposed in the present invention.

[0066] Figure 4 It is a schematic diagram of the IEEE 39-node system in an example of the present invention. DETAILED DESCRIPTION

[0067] The purpose of the present invention is to construct a bidirectional transfer learning power system transient voltage stability assessment model that can adapt to the access of new energy sources. It establishes an initial database containing the grid-connected scenarios of new energy power generation, and multi-classifies the sample state by calculating the transient voltage stability index (TVSI). On the basis of the long short-term memory network (LSTM) model, it further combines bidirectional transfer learning (Bi-TL) to fine-tune the data set, so that the LSTM model can adapt to new working conditions without retraining a new model. Thus, a transient voltage stability assessment method for a new energy grid-connected system based on bidirectional transfer learning is provided.

[0068] A transient voltage stability assessment method for renewable energy grid-connected systems based on bidirectional transfer learning, such as Figure 1 As shown, the following steps are included:

[0069] Step 1: For the situation of renewable energy power generation grid connection, an initial database containing a large amount of historical simulation result information is established, and the operation scenario of the current system is fitted through the database;

[0070] Step 2: For the samples in the database, classify them into different transient voltage stability categories by calculating the transient voltage stability index (TVSI);

[0071] Step 3: Offline training of the Long Short-Term Memory (LSTM) model based on the data set in the database to establish a transient voltage stability assessment model for the power system based on LSTM;

[0072] Step 4: Considering the situation where the operating conditions of the power system change, the data set is fine-tuned through bidirectional transfer learning (Bi-TL) to enable the LSTM model to adapt to the new operating conditions and build a Bi-TL-LSTM transient voltage stability assessment framework.

[0073] In step 1, specifically, for the situation of renewable energy power generation grid connection, an initial database containing a large amount of historical simulation result information is established, and the operation scenario of the current system is fitted through the database. In order to adapt to specific operation scenarios, the distance between operation scenarios needs to be defined accordingly.

[0074] The biggest disadvantage of traditional time-domain simulation is that it is time-consuming, especially with the grid connection of new energy power generation equipment, the scale of the power system is getting larger and larger, and the integration of power electronic equipment is increasing, and the time consumption is becoming more and more significant. However, due to the large output fluctuations of renewable energy power generation equipment such as wind power and photovoltaic power generation, the rapidity of transient voltage stability assessment is required to be increased. Therefore, a database containing a large amount of historical simulation result information is established. By fitting the current operating scenario through the database, the model can quickly analyze the transient voltage stability of the new energy power generation grid-connected system at this time.

[0075] Define the i-th operation scenario s i As shown below:

[0076] s i = {V i ,θ i ,P Gi ,Q Gi ,P Li ,Q Li} (1)

[0077] Where V i is the voltage amplitude of node i; θ i is the voltage phase angle of node i; P Gi is the active power of the generator at node i; Q Gi is the reactive power of the generator at node i; P Li is the load active power of node i; Q Li is the load reactive power of node i.

[0078]

[0079] Where Vi 1 is the voltage amplitude of the first scene node i; V i n is the voltage amplitude of the nth scene node i; is the voltage phase angle of the first scenario node i; is the voltage phase angle of the nth scene node i; is the active power of the generator of the first scenario node i; is the active power of the generator at the nth scenario node i; is the reactive power of the generator at the first scenario node i; is the reactive power of the generator at the nth scenario node i; is the load active power of the first scenario node i; is the load active power of the nth scene node i; is the reactive power of the load at the first scenario node i; is the reactive power of the load at the nth scenario node i.

[0080] Taking the distance between scene s0 and scene s1 as an example, the definition is as follows:

[0081]

[0082] Where D(s0,s1) is the distance between scene s0 and scene s1; V i,0 is the voltage amplitude of node i in scenario s0; V i,1 is the voltage amplitude of node i in scenario s1; θ i,0 is the voltage phase angle of node i in scenario s0; θ i,1 is the voltage phase angle of node i in scenario s1; P Gi,0 is the active power of the generator at node i in scenario s0; P Gi,1 is the active power of the generator at node i in scenario s1; Q Gi,0 is the reactive power of the generator at node i in scenario s0; Q Gi,1 is the reactive power of the generator at node i in scenario s1; P Li,0 is the load active power of node i in scenario s0; P Li,1 is the active power of the load at point i in scenario s1; Q Li,0 is the reactive power of the load at node i in scenario s0; Q Li,1 is the reactive power of the load at node i in scenario s1.

[0083] If the distance between two scenes is very small, the two scenes are considered to have the highest similarity. The inverse of D(i) is taken as the weight of s1 to fit scene s0.

[0084] In step 2, specifically, by calculating TVSI, many samples are divided into unstable, stable, and stable with fault-induced voltage delayed recovery phenomenon, and are marked with numbers 0, 1, and 2, respectively.

[0085] The principle of classifying samples by TVSI is as follows:

[0086]

[0087] Where n is the number of system nodes; T is the voltage recovery time; t cr is the fault removal time; TVDI j,t is the transient voltage deviation index of node i at time t.

[0088] The transient voltage deviation index (TVDI) in TVSI can be expressed as:

[0089]

[0090] Where V j,t is the voltage amplitude measured at node j at time t; V j,0 is the voltage amplitude before the fault of node j; α is taken as 10%, which is the threshold of unacceptable voltage deviation amplitude. The transient voltage stability index TVSI is used to evaluate the transient voltage stability of the samples collected in the entire system, and three labels of 0, 1 and 2 are assigned respectively.

[0091] In step three, specifically, the LSTM model is trained offline, and a transient voltage stability assessment model for the power system based on LSTM is established by learning the time dependence of the system dynamics after the disturbance. The specific working principles of the forget gate, input gate, and output gate of LSTM are as follows:

[0092] f t =σ(W f ·[h t-1 ,x t ]+b f ) (6)

[0093] i t =σ(W i ·[h t-1 ,x t ]+b i ) (7)

[0094] o t =σ(W o ·[h t-1 ,x t ]+b o ) (8)

[0095] In the formula, f t For the forget gate; i t is the input gate; o t is the output gate; W f , W i , W o is the weight matrix; b f , b i , b o is the bias vector; σ(·) is the activation function; h t-1 is the transient voltage timing characteristic information output at the previous moment; x t It is the transient voltage timing characteristic information input at the current moment.

[0096] The transformation process of the memory unit is described as follows:

[0097]

[0098] In the formula, is the candidate memory gate; tanh is the activation function; W c is the weight matrix; b c is the bias vector; c t is the memory unit of the LSTM model, indicating the state of the memory unit at the current time step t; is the element product; c t-1 is the state of the memory unit at the previous time step t-1; W c ·[h t-1 ,x t ] is the product of the time series feature information and the weight matrix.

[0099] Hidden layer state h t The representation is as follows:

[0100]

[0101] In the formula, tanh(·) represents a nonlinear activation function. The unique structure of using memory units instead of hidden layer nodes in the LSTM model enables it to extract time-domain related features for a long time.

[0102] The transient voltage stability assessment model based on LSTM mainly consists of five parts: input layer, LSTM layer, fully connected layer, softmax layer and output layer. Figure 2 As shown in Figure 2. In this evaluation model, the input is the system dynamic measurements collected after the failure. The LSTM layer is used to capture the potential time dependencies between time series data. The internal memory of the LSTM network is used to process the input in chronological order using a unique gating mechanism.

[0103] In the whole process, the fully connected layer plays the role of dimensionality reduction and implements data discarding at this layer to prevent the model from overfitting. In the softmax layer, the softmax activation function is used to map hidden features to the final evaluation results, realizing the multi-classification of the evaluation model. The process can be expressed as:

[0104] y t =softmax(W S ·h t +b S ) (12)

[0105] In the formula, y t is the predicted label of the sample; h t is a hidden feature; W S and b S They are the weight matrix and bias vector that need to be updated during the offline training of the model; softmax(·) is the activation function.

[0106] In step 4, specifically, in the actual operation of the power system, the operating conditions may undergo some changes that have not been considered. Through Bi-TL, the data set is fine-tuned, and the existing data set and model are used to help the model adapt to the new operating conditions, avoiding time-consuming and laborious data labeling work. Among them, Bi-TL learning includes forward learning using the information in the target domain data and backward learning using the information in the initial domain data. The combination of forward learning and backward learning can comprehensively learn the knowledge of the target domain data and the initial domain data, reducing the labeling process of new costs in the target domain.

[0107] Fine-tuning is one of the most popular transfer learning methods when using deep learning. This method is simple and effective. The diagram of fine-tuning is as follows Figure 3 As shown in the figure, the initial data set of the training model is defined as the initial domain data, and the new data set generated after some unconsidered changes in the operating conditions of the actual power system are defined as the target domain data. The fine-tuning process starts with the evaluation model network trained with the initial domain data. As the system undergoes unexpected changes, the parameters of the trained model are fine-tuned with reference to the data of the target domain and optimized with a small learning rate. The parameters of the evaluation model trained based on the initial domain data reflect the knowledge learned from the initial domain data. Fine-tuning is a technique that allows knowledge transfer by initializing the target model using the parameters of the source model. With the foundation of the transferred knowledge, fine-tuning can not only provide a good starting point for the target model, but also significantly reduce the sample labeling and iterative training work required compared to training a new model from scratch. In addition, fine-tuning is also easier to implement than most other transfer learning techniques.

[0108] Bi-TL combines forward and backward learning to form a mixed sample set, including new knowledge actively introduced from the target domain and existing information actively inherited from the initial domain. The goal of forward learning is to select the most worthy target samples for annotation, so as to obtain better performance than selecting other samples. Backward learning is to find suitable samples in the initial domain while reducing the labeling of target domain samples to invisibly expand the training set. Then, the pre-trained model is fine-tuned using the constructed mixed sample set so that it can adapt to the new operating conditions. Starting from the initial domain data, forward learning gradually introduces "good" target domain samples into the training set, while backward learning gradually removes "bad" initial domain samples in each cycle.

[0109] (1) Forward learning uses information in the target domain data, also known as the query (+). Its purpose is to obtain the true labels of the samples whose labels the current model is most uncertain about based on uncertainty sampling, which helps to eliminate uncertainty and bring the most information to the model. This is achieved through marginal sampling:

[0110]

[0111] In the formula, x is the input sample variable; x + is the marginal value of variable x; and Model Posterior probability predictions for the first and second most likely classes; and are the labels of the first and second most likely categories of variable x; D T is the target domain; U is the unlabeled sample in the target domain. The smaller the margin, the higher the sample information content. It can also be achieved through entropy sampling, expressed as:

[0112]

[0113] In the formula, y i Contains all possible classes. Samples with high entropy contain relatively rich information. The result of forward learning is B+ target samples with the largest uncertainty, forming a set Query the stable label of the sample through the TVSI indicator After that, it is moved to the labeled training set. The added labeled target domain samples are represented as: in, is the marginal value set of variable x; Y t + for The corresponding tag set; is the target domain sample; x i is the input sample variable; i For x iThe hyperparameter B+ is chosen to strike a balance between accuracy and adaptation speed as well as the convenience of sample labeling.

[0114] (2) Backward learning utilizes information in the initial domain data, also known as query (-). The initial domain data contains valuable information, but also interference information. The purpose of backward learning is to explore the existing initial domain data and intelligently remove interference information while retaining transferable data.

[0115] Specifically, it starts from the initial domain D S The labeled samples that deviate from the target domain distribution are removed. Therefore, the initial domain samples that conform to the target domain distribution will be retained. In this way, the information in the initial domain data is utilized to actually expand the number of labeled samples used for fine-tuning.

[0116] Backward learning removes the initial domain samples of the maximum posterior probability change. The specific function implementation process is as follows:

[0117]

[0118] In the formula, x - is the query value of variable x; and The models are and The posterior probability vectors of different stability categories are calculated; the L1 norm ||·||1 is the integral of the changes in each dimension of the posterior probability vector; x is the input sample variable; y is the label of variable x.

[0119] Backward learning is repeated B times to remove B initial domain samples that no longer represent the target domain. The removed sample labels are expressed as: in, is the set of selected samples; Y t - for The corresponding label; is the target domain sample.

[0120] (3) Combination of forward learning and backward learning. In Bi-TL, the combination of forward learning and backward learning can comprehensively learn the knowledge of the target domain data and the initial domain data, reduce the labeling process of new costs in the target domain, and the generated mixed training data set is better than the randomly selected target domain data.

[0121] Example:

[0122] The embodiments used in the present invention are based on Figure 4The IEEE 39-node system shown in the figure includes 30 nodes, 10 generators and 39 buses. The generator at the 37th node is replaced by a double-fed wind turbine to simulate the grid-connected scenario of renewable energy power generation. This test includes all the steps described in the method of the present invention. The test was conducted on a computer equipped with an Intel Core i5 CPU and 4GB RAM, and the test results were obtained. Both the LSTM and Bi-TL models were implemented based on the MATLAB platform.

[0123] The simulation process was set to the three-phase short-circuit fault condition that had the most serious impact on the system. The fault duration was 0.2 seconds, and the simulation process lasted for 5 seconds. Then, different fault locations were set according to a certain rule to simulate the topological changes caused by different faults. Finally, the initial sample library obtained 4170 stable samples, 2162 unstable samples, and 300 stable samples with fault-induced voltage delay recovery phenomenon. 80% of the sample set was randomly selected for offline training, and the remaining 20% ​​was used for performance testing. After 5-fold cross-validation, stable test results were obtained. Taking into account the different contingencies under different operating conditions, the total load demand was set to 80%, 100%, and 120% of the baseline level, respectively. The three-phase short-circuit fault was set from the end of a transmission line, and the fault location was changed multiple times. The parameter settings of the evaluation model LSTM are shown in Table 1. The false detection rate Mis, false alarm rate Fal, and accuracy rate A are used. CC To measure the classification performance of the transient voltage stability assessment model, the calculation formula is as follows:

[0124] Mis=N fs / N×100% (16);

[0125] Fal=N fu / N×100% (17);

[0126] A CC =(NN fs -N fu ) / N×100% (18);

[0127] Where: N is the total number of samples, N fs is the number of samples that are actually unstable but evaluated as stable, N fu The number of samples that are evaluated as unstable for true stability.

[0128] Table 1 Parameter settings of the evaluation model LSTM

[0129] Hyperparameters value Learning Rate 0.0001 Discard rate 0.25 LSTM cell size 256 Batch size 64 LSTM epoch 200

[0130] The evaluation performance of the model obtained through testing is: Mis = 0.31%, Fal = 0.22%, A CC=98.91%. In order to verify the superiority of the transient voltage stability assessment model of the present invention, it is compared with other traditional methods, such as support vector machines (SVM), recurrent neural network (RNN), extreme learning machine (ELM) and convolutional neural network (CNN). The test performance of each model is shown in Table 2. The test results show that the performance of the assessment model of the present invention is superior as a whole, which meets the purpose of the present invention.

[0131] Table 2 Test comparison of each model

[0132]

[0133] The LSTM evaluation model with Bi-TL processing and the conventional LSTM evaluation model were tested under the same conditions. As shown in Table 3, it can be seen from the test results that the addition of Bi-TL improves the evaluation effect of the evaluation model, making it more accurate, which meets the purpose of the present invention.

[0134] Table 3 Performance comparison of the LSTM evaluation model with Bi-TL processing and the conventional LSTM evaluation model

[0135] method Mis / % Fal / % <![CDATA[A CC / %]]> LSTM(Bi-TL) 0.31 0.22 98.91 LSTM 0.42 0.25 98.78

[0136] In order to verify the adaptability of the model system to unknown operating conditions, the fault settings of the test system are changed, and new samples are generated for testing the model. The final test results are shown in Table 4. It can be seen from the test results that the model can accurately measure the transient voltage stability of the system under different operating conditions, which meets the purpose of the present invention.

[0137] Table 4 Test results of new samples on the test model

[0138] Type of accident Specific topology changes <![CDATA[A CC / %]]> N-1 50% of line 14-15 is disconnected 98.89 N-1 Line 33-19 is disconnected at the 33 head end 98.83 N-1 Line 8-9 is disconnected at the end of 9 98.88

[0139] In order to verify whether the processing speed of the model can meet the real-time online transient voltage safety assessment, a processing speed test was conducted on the IEEE39 node system, and the final processing speed test results are shown in Table 5. According to the data acquisition speed of the actual synchronized phasor measurement unit, the processing speed of a system snapshot should be less than 0.033 seconds. It can be seen from the test results that the model meets the actual needs and meets the purpose to be achieved by the present invention.

[0140] Table 5 IEEE 39-node system processing speed test results

[0141] Test system Offline training time Test suite processing time IEEE 39 Node 80.63 seconds (5305 samples) 2.21 seconds (1327 samples)

Claims

1. A transient voltage stability assessment method for renewable energy grid-connected systems based on bidirectional transfer learning, characterized in that The following steps are involved: Step 1: Establish a database containing a large amount of historical simulation result information, and fit the operation scenario of the current system through the database; Step 2: Calculate the transient voltage stability index TVSI of the samples in the database and divide them into different transient voltage stability categories; Step 3: Based on the data set in the database, the long short-term memory network LSTM model is trained offline to establish a transient voltage stability assessment model for the power system based on LSTM; Step 4: Considering the situation where the operating conditions of the power system change, the data set is fine-tuned through bidirectional transfer learning so that the power system transient voltage stability assessment model can adapt to the new operating conditions.

2. The transient voltage stability assessment method for a new energy grid-connected system based on bidirectional transfer learning according to claim 1 is characterized in that: In the step 1, an initial database containing a large amount of historical simulation result information is established for the situation of new energy power generation being connected to the grid, wherein the database includes generator active power, generator reactive power, load active power, load reactive power, voltage amplitude and phase angle data; thereafter, the operation scenario of the current system is fitted through the database, wherein the established operation scenario includes fault line, fault location, fault duration, fault type parameters, and multiple scenarios are constructed for testing.

3. The transient voltage stability assessment method for a new energy grid-connected system based on bidirectional transfer learning according to claim 2 is characterized in that: In step 1, in order to adapt to specific operation scenarios, the distance between operation scenarios is defined accordingly: Define the i-th operation scenario s i As shown below: s i ={V i ,θ i ,P Gi ,Q Gi ,P Li ,Q Li } (1); Where V i is the voltage amplitude of node i; θ i is the voltage phase angle of node i; P Gi is the active power of the generator at node i; Q Gi is the reactive power of the generator at node i; P Li is the load active power of node i; Q Li is the load reactive power of node i; Where V i 1 is the voltage amplitude of the first scene node i; V i n is the voltage amplitude of the nth scene node i; is the voltage phase angle of the first scenario node i; is the voltage phase angle of the nth scene node i; is the active power of the generator of the first scenario node i; is the active power of the generator at the nth scenario node i; is the reactive power of the generator at the first scenario node i; is the reactive power of the generator at the nth scenario node i; is the load active power of the first scenario node i; is the load active power of the nth scene node i; is the reactive power of the load at the first scenario node i; is the reactive power of the load at the nth scene node i; The distance between scene s0 and scene s1 is defined as follows: Where D(s0,s1) is the distance between scene s0 and scene s1; V i,0 is the voltage amplitude of node i in scenario s0; V i,1 is the voltage amplitude of node i in scenario s1; θ i,0 is the voltage phase angle of node i in scenario s0; θ i,1 is the voltage phase angle of node i in scenario s1; P Gi,0 is the active power of the generator at node i in scenario s0; P Gi,1 is the active power of the generator at node i in scenario s1; Q Gi,0 is the reactive power of the generator at node i in scenario s0; Q Gi,1 is the reactive power of the generator at node i in scenario s1; P Li,0 is the load active power of node i in scenario s0; P Li,1 is the active power of the load at point i in scenario s1; Q Li,0 is the reactive power of the load at node i in scenario s0; Q Li,1 is the reactive power of the load at node i in scenario s1; If the distance between two scenes is very small, the two scenes are considered to have the highest similarity; take the reciprocal of D(·) D(·) -1 As the weight of s1, to fit the scene s0.

4. The transient voltage stability assessment method for a new energy grid-connected system based on bidirectional transfer learning according to claim 1 is characterized in that: In the step 2, by calculating the transient voltage stability index TVSI, multiple samples are divided into unstable, stable and stable with fault-induced voltage delayed recovery phenomenon, and are marked with numbers 0, 1 and 2 respectively; The samples are classified by transient voltage stability index TVSI as follows: Where n is the number of system nodes; T is the voltage recovery time; t cr is the fault clearing time; TVDI j,t is the transient voltage deviation index of node i at time t; The transient voltage deviation index in the transient voltage stability index TVSI is expressed as: Where V j,t is the voltage amplitude measured at node j at time t; V j,0 is the voltage amplitude before the fault of node j; α is the threshold of unacceptable voltage deviation amplitude; the transient voltage stability index TVSI is used to evaluate the transient voltage stability of the samples collected from the entire system, and three labels of 0, 1 and 2 are assigned respectively.

5. The transient voltage stability assessment method for a new energy grid-connected system based on bidirectional transfer learning according to claim 4 is characterized in that: During the simulation, all samples whose bus node voltage change trajectories remain no lower than the pre-fault voltage within a stable time of more than 3 seconds are marked as stable; further, by evaluating the transient voltage stability index TVSI of the stable sample, the stable samples are reclassified according to whether they are prone to fault-induced voltage delayed recovery phenomenon. The stable samples with TVSI>4 are marked as stable samples with fault-induced voltage delayed recovery phenomenon, and the others are stable samples; the simulation process lasts for 5 seconds, and the samples whose bus voltage change trajectory is lower than the pre-fault voltage are classified as unstable.

6. The transient voltage stability assessment method for a new energy grid-connected system based on bidirectional transfer learning according to claim 1 is characterized in that: In the step 3, a long short-term memory network LSTM model is trained offline, and a transient voltage stability assessment model for a power system based on LSTM is established by learning the time dependence of the system dynamics after the disturbance. The specific steps are as follows: first, data preparation is performed, data is collected and preprocessed, and the data is divided into a training set and a test set; second, an LSTM model is constructed, an input layer, one or more LSTM layers, and a fully connected layer are defined, each LSTM layer consists of multiple LSTM units and is trained using a training data set; then, a model evaluation is performed, and the model performance is evaluated on an independent training set or test set; finally, the model structure or parameters are adjusted according to the evaluation results.

7. The transient voltage stability assessment method for a new energy grid-connected system based on bidirectional transfer learning according to claim 6 is characterized in that: The specific working principles of the forget gate, input gate, and output gate of the LSTM model are as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ) (6); i t =σ(W i ·[h t-1 ,x t ]+b i ) (7); the t =σ(W o ·[h t-1 ,x t ]+b o ) (8); In the formula, f t For the forget gate; t is the input gate; o t is the output gate; W f , W i , W o is the weight matrix; b f 、b i 、b o is the bias vector; σ(·) is the activation function; h t-1 is the transient voltage timing characteristic information output at the previous moment; x t The transient voltage timing characteristic information input at the current moment; The transformation process of the memory unit is described as follows: In the formula, is the candidate memory gate; tanh is the activation function; W c is the weight matrix; b c is the bias vector; c t is the memory unit of the LSTM model, indicating the state of the memory unit at the current time step t; is the element product; c t-1 is the state of the memory unit at the previous time step t-1; W c ·[h t-1 ,x t ] is the product of the time series feature information and the weight matrix; Hidden layer state h t The representation is as follows: Where tanh(·) represents a nonlinear activation function. The unique structure of using memory units instead of hidden layer nodes in the LSTM model enables it to extract time-domain related features for a long time.

8. The transient voltage stability assessment method for a new energy grid-connected system based on bidirectional transfer learning according to claim 7 is characterized in that: The LSTM-based power system transient voltage stability assessment model consists of an input layer, an LSTM layer, a fully connected layer, a softmax layer, and an output layer. In this assessment model, the input layer is the system dynamic measurement quantity collected after the fault occurs; the LSTM layer is used to capture the potential time dependency between time series data, and the internal memory of the LSTM model is used to process the input in chronological order through a gating mechanism; in the whole process, the fully connected layer plays the role of dimensionality reduction, and data is discarded on this layer to prevent the model from overfitting. In the softmax layer, the softmax activation function is used to map hidden features to the final evaluation results, realizing multi-classification of the evaluation model. The process can be expressed as: y t =softmax(W S ·h t +b S ) (12)? In the formula, y t is the predicted label of the sample; h t is a hidden feature; W S and b S They are the weight matrix and bias vector that need to be updated during the offline training of the model; softmax(·) is the activation function.

9. The transient voltage stability assessment method for a new energy grid-connected system based on bidirectional transfer learning according to claim 1 is characterized in that: In step 4, the learning of bidirectional transfer learning Bi-TL includes the following steps: S4.1: Use the information in the target domain data for forward learning, called query (+), to obtain the true labels of the samples whose labels the current evaluation model is least sure about; this is achieved through marginal sampling: In the formula, x is the input sample variable; x + is the marginal value of variable x; and Model Posterior probability predictions for the first and second most likely classes; and are the labels of the first and second most likely categories of variable x, respectively; D T is the target domain; U is the unlabeled sample in the target domain; The smaller the margin, the higher the information content of the sample; this can also be achieved through entropy sampling, expressed as: In the formula, y i Contains all possible classes; the result of forward learning is B+ target samples with the largest uncertainty, forming a set Query the stable label Y of the sample through the TVSI indicator t + After that, it is moved to the labeled training set; the added labeled target domain sample is represented as: in, is the marginal value set of variable x; Y t + for The corresponding tag set; is the target domain sample; x i is the input sample variable; i For x i The corresponding labels; the hyperparameter B+ is selected to strike a balance between accuracy and adaptation speed and the convenience of sample labeling; S4.2: Use the information in the initial domain data for backward learning, called query (-), analyze the existing initial domain data, remove interference information, and expand the number of labeled samples for fine-tuning by utilizing the information in the initial database; Specifically, it starts from the initial domain D S The labeled samples that deviate from the target domain distribution are deleted; therefore, the initial domain samples that conform to the target domain distribution will be retained; in this way, the information in the initial domain data is utilized, which actually expands the number of labeled samples used for fine-tuning; Backward learning removes the initial domain samples of the maximum posterior probability change. The specific function implementation process is as follows: In the formula, x - is the query value of variable x; and The models are and The calculated posterior probability vectors of different stability categories; L1 norm ||·||1 is the integral of the changes in each dimension of the posterior probability vector; x is the input sample variable; y is the label of variable x; Backward learning is repeated B-times to remove B-initial domain samples that no longer represent the target domain; the removed sample labels are expressed as: in, is the set of selected samples; Y t - for The corresponding label; is the target domain sample.

Images