A power system stability evaluation method for high proportion of new energy access

Abstract

The present application relates to the high proportion of new energy access-oriented power system stability evaluation method, including step one: based on the high proportion of new energy access power system operation characteristics, the construction transient stability evaluation characteristic set;Step two: propose based on the STA-GLN power system transient stability evaluation model of space-time double attention mechanism;Step three: introduce integrated learning and transfer learning improve STA-GLN power system transient stability evaluation model;Based on the integrated STA-GLN multi-task transient stability evaluation method of AdaBoost and transfer learning, the space-time double attention mechanism is introduced to deeply mine the transient characteristics, solve the problem of the preceding power system transient stability evaluation research that the evaluation accuracy decreases due to the change of topological structure;The integrated learning method combined with AdaBoost reduces the imbalance of samples and reduces the misjudgment of instability;The transfer learning is introduced to solve the problem of long training time caused by model complexity, and the rapidity of evaluation is ensured.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a power system stability assessment method for high-proportion renewable energy integration. Background Technology

[0002] Currently, the production model relying on traditional fossil fuels is changing, and energy supply is developing towards low-carbon and clean energy. Clean energy becoming the main energy source is an inevitable trend. Against this backdrop, my country's wind power installed capacity has shown a year-on-year upward trend. In 2019, the installed capacity of wind power reached 7.1 times that of 2010. According to the State Grid Energy Research Institute, by 2050, more than 75% of power generation will come from clean energy, of which new energy power generation represented by wind power and photovoltaics will become the largest power source, accounting for about 40% of power generation. With the continuous integration of high proportions of renewable energy and the improvement of the degree of power electronics, the power system structure is becoming increasingly complex, which threatens the stability of the power system.

[0003] Transient stability assessment of power systems is one of the key issues in maintaining the safe and stable operation of power systems. Traditional transient stability assessment and analysis methods are mostly driven by mathematical models, such as time-domain simulation and direct methods. Time-domain simulation has the disadvantage of excessive calculation time when applied online, and the increased calculation cost weakens its practical value in the engineering field. Direct methods have smaller requirements and shorter time consumption, but it is more difficult to construct energy functions, and the model has poor adaptability, making it difficult to be widely used. With the continuous development of digital technology in power systems, data-driven transient stability assessment methods for power systems have gradually emerged.

[0004] Currently, most data mining methods used for transient stability assessment are shallow learning models, such as support vector machines, decision trees, and extreme learning machines. Their limitations lie in their limited ability to represent complex functions, high computational complexity, and poor generalization ability. Deep learning, as the latest branch of artificial intelligence, has brought new ideas for obtaining characteristic information of power grid data and modeling and analyzing stability problems with its powerful data mining and self-learning capabilities. Deep learning uses multi-layer nonlinear network structures to approximate complex functions and learns the distributed feature representation of input data, and has the ability to extract essential features from a small sample set. However, it still has problems such as poor adaptability to power system topology changes, sample imbalance, and excessively long assessment time.

[0005] In summary, there is an urgent need for a power system transient stability assessment method that can address the problems of poor topology adaptability, difficulty in learning unstable samples, and long model training time that still exist in artificial intelligence-based transient stability assessment methods, and has excellent performance. Summary of the Invention

[0006] This invention provides a power system stability assessment method for high-proportion renewable energy integration. It is an integrated STA-GLN multi-task transient stability assessment method based on AdaBoost and transfer learning. It introduces a spatiotemporal dual attention mechanism to deeply mine transient characteristics, addressing the problem of decreased assessment accuracy due to topology changes in previous power system transient stability assessment studies. The ensemble learning method combined with AdaBoost reduces sample imbalance and instability misjudgments. Transfer learning is introduced to address the problem of long training times caused by model complexity, ensuring rapid assessment.

[0007] To achieve the above objectives, the present invention employs the following technical solution:

[0008] A power system stability assessment method for high-proportion renewable energy integration includes the following steps:

[0009] Step 1: Based on the operating characteristics of power systems with a high proportion of renewable energy access, construct a transient stability assessment feature set;

[0010] Step 2: Propose a STA-GLN power system transient stability assessment model based on a spatiotemporal dual attention mechanism;

[0011] Step 3: Introduce ensemble learning and transfer learning to improve the STA-GLN power system transient stability assessment model.

[0012] Furthermore, step one is specifically as follows:

[0013] 1) Generate a sample set;

[0014] 2) Perform electromechanical transient simulation by setting up a three-phase ground fault on 34 parallel transmission lines without transformers to generate a system transient instability sample set;

[0015] 3) Construct a transient stability feature set, taking the transient power angle stability and transient voltage stability of the power system as the research objects, and select feature quantities from two dimensions: time and space;

[0016] 4) Select transient stability assessment indicators.

[0017] Furthermore, step two is detailed as follows:

[0018] 1) From a spatial perspective, spatial attention mechanism is introduced into graph convolutional neural network to construct SA-GCN sub-model and extract input features that integrate the topological connectivity characteristics of power system;

[0019] 2) From a temporal perspective, the temporal attention mechanism is introduced into the LSTM encoder-decoder to construct a TA-LSTM sub-model. The importance of temporal features at each time step is calculated, and the learning intensity of the evaluation model for important time step input features is adjusted by weighting. The transient characteristics in the input features are then explored to improve the model performance.

[0020] 3) By adopting a cascaded structure and combining the SA-GCN sub-model and the TA-LSTM sub-model, a power system transient stability assessment model based on STA-GLN is obtained.

[0021] Furthermore, step three is detailed as follows:

[0022] 1) Introduce ensemble learning by using the STA-GLN power system transient stability assessment model as a sub-model, integrating it through AdaBoost, and using the SAMME algorithm to dynamically adjust the input sample weights of each STA-GLN sub-classifier;

[0023] 2) Introduce transfer learning to transfer model parameters of the STA-GLN sub-classifier;

[0024] Furthermore, the transfer learning transfers the model parameters of the first STA-GLN sub-classifier to the next sub-classifier, and then fine-tunes the parameters of the sub-classifier using the training set with updated weights to obtain a new STA-GLN sub-classifier. This process is then repeated.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] 1) Introduce a spatiotemporal dual attention mechanism to deeply explore transient characteristics and solve the problem of decreased assessment accuracy due to topological changes in the aforementioned power system transient stability assessment research;

[0027] 2) Combining AdaBoost with ensemble learning methods reduces the risk of imbalanced samples and thus reduces instability and misjudgments;

[0028] 3) Introduce transfer learning to solve the problem of long training time caused by model complexity, and ensure rapid evaluation. Attached Figure Description

[0029] Figure 1 This is a diagram of the New England 39-node system described in this invention.

[0030] Figure 2 This is a structural diagram of the TSA model based on STA-GLN as described in this invention.

[0031] Figure 3 This is a generalization capability diagram of Scheme 1 in the embodiment of the present invention.

[0032] Figure 4This is a generalization capability diagram of Scheme 2 in the present invention. Detailed Implementation

[0033] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0034] This invention provides a power system stability assessment method for high-proportion renewable energy integration, comprising the following steps:

[0035] Step 1: Based on the operating characteristics of power systems with a high proportion of renewable energy access, construct a transient stability assessment feature set;

[0036] Step 2: Propose a STA-GLN power system transient stability assessment model based on a spatiotemporal dual attention mechanism;

[0037] Step 3: Introduce ensemble learning and transfer learning to improve the STA-GLN power system transient stability assessment model.

[0038] See Figure 1 Furthermore, step one is specifically as follows:

[0039] 1) Generate a sample set, using the England 10-machine 39-node system as the test system, with a reference frequency of 60Hz;

[0040] 2) Perform electromechanical transient simulation, set up three-phase grounding short-circuit faults on 34 parallel transmission lines without transformers to generate a system transient instability sample set, and add sample generation for N-1 disconnection and N-2 disconnection systems on the basis of the fully connected system;

[0041] 3) Construct a transient stability feature set, taking the transient power angle stability and transient voltage stability of the power system as the research objects, and select feature quantities from two dimensions: time and space. The time series feature input feature formula of the overall model is as follows:

[0042] X = (x1, ..., x) k )=(U T ,θ T ) T (1)

[0043] Where U and Θ represent the bus voltage magnitude and phase angle, respectively, as shown in the following formula:

[0044]

[0045]

[0046] Where k is the number of sampling points, u n,t θ n,tThe voltage amplitude and phase angle of bus n at time t are respectively. Through evaluation experiments, the number of sampling points is determined to be k=5, that is, the selected time-series electrical characteristics are sampled 5 times within the selected sampling window range.

[0047] 4) Select a transient stability assessment index. The Transient Stability Index (TSI) is selected as the transient power angle stability criterion for the sample data. The TSI formula is as follows:

[0048]

[0049] Where, Δδ max To determine the maximum power angle difference between any two generators in the system 4 seconds after troubleshooting, n TSI For transient stability margin, n TSI When the value is greater than 0, the system is considered to be in a stable state, and the sample label is marked as 1; n TSI When the value is less than 0, the system is considered to be in an unstable state, and the sample label is marked as 0. For the voltage stability category, the practical criterion for transient voltage stability is adopted, that is, the bus node voltage is less than or equal to 1 second after the fault. Nodes that meet this condition are identified as stable nodes and marked as 1, and nodes that do not meet the condition are identified as unstable nodes and marked as 0. The sample label has 39 dimensions.

[0050] See Figure 2 Furthermore, step two is specifically as follows:

[0051] 1) From a spatial perspective, the electrical features X and the adjacency matrix A are input into the graph convolutional neural network layer i, which incorporates a spatial attention mechanism. The adjacency matrix element a ij The definition is as follows:

[0052]

[0053] Based on the bus node, the SA-GCN network integrates topological connectivity characteristics with node input features, and calculates the cosine similarity of the input features of each node through a spatial attention mechanism, as shown in the following formula:

[0054]

[0055] Attention weights are assigned to each node, and the spatial attention coefficients of system node i to node j are as follows:

[0056]

[0057] Among them, H l For the inter-layer propagation rules of the l-th layer of a graph convolutional neural network (GCN), γ (l)As an attention-guided propagation parameter, COS is used to calculate the cosine similarity of hidden layer features. The greater the similarity, the greater its influence on the central node.

[0058] 2) From a temporal perspective, the output fusion feature XG of the SA-GCN network is used as the input of the TA-LSTM network for secondary feature extraction. The cosine similarity between the input fusion feature at each sampling time i and the total input fusion features is calculated to obtain the temporal attention coefficient for the input features at important times. The specific formula is as follows:

[0059]

[0060] Update the hidden state of TA-LSTM using time attention coefficients:

[0061] h1` = contact(h i ,a i H) (9)

[0062] The influence of input features at different times on the evaluation results is quantified into weights for updating the hidden state of the LSTM network through attention calculation. The new hidden state is obtained and decoded, then processed by Softmax to obtain a 2D output representing the probability of transient power angle stability and instability of the system. The prediction result of the transient power angle stability of the power system is obtained by comparing it with the set classification threshold. The model is as follows:

[0063] y ST =softmax(f c (H`)) (10)

[0064]

[0065] 3) By adopting a cascaded structure and combining the SA-GCN sub-model and the TA-LSTM sub-model, a power system transient stability assessment model based on STA-GLN is obtained.

[0066] Furthermore, step three is detailed as follows:

[0067] 1) Introduce ensemble learning by using the STA-GLN power system transient stability assessment model as a sub-model, integrating it through AdaBoost, and using the SAMME algorithm to dynamically adjust the input sample weights of each STA-GLN sub-classifier;

[0068] The formula for the transient stable training sample set is as follows:

[0069] X train ={(x1,y1),(x2,y2),…,(x n ,y n (12)

[0070] Where x is the feature vector of the input sample, y is the class of the input sample, and the initial weights of the training set samples are assigned as follows:

[0071]

[0072] 2) Introduce transfer learning to transfer model parameters of the STA-GLN sub-classifier;

[0073] Based on the accuracy of the weak learners generated after each iteration in classifying all samples in the training set and the overall classification, the sample weights of the training set for the weak learners entering the next iteration are adjusted. The weak learners generated in each iteration are then integrated to obtain the final strongly supervised classification model. The SAMME algorithm is used to classify the STA-GLN model, and the model is updated using a weighted probability estimation method, calculating weighted class probability estimates. For the m-th sub-classifier, sample x i The probability of being classified into the Kth class is given by the following formula:

[0074]

[0075] Then, the new sub-classifier h(k) is calculated based on the weighted probability estimate, as shown in the following formula:

[0076]

[0077] Normalize and update sample weights w m+1,i :

[0078]

[0079] Finally, the strong classifier is obtained by weighting:

[0080]

[0081] in, The largest category is the classification category of the model, Z. m w will be used as a normalization factor mi The value is restricted to [0, 1], guaranteeing

[0082] First, a first STA-GLN sub-classifier is trained using the training set. Then, in the next iteration of the AdaBoost ensemble, the model parameters of the first STA-GLN sub-classifier are transferred through transfer learning and assigned to the next sub-classifier. The sub-classifier is then fine-tuned using the training set with updated weights to obtain a new STA-GLN sub-classifier. This process is repeated, with the model parameters of the STA-GLN sub-classifier generated in each iteration being transferred from the STA-GLN sub-classifier generated in the previous iteration. The sub-classifier model parameters are transferred through transfer learning, and the transferred STA-GLN sub-classifier is fine-tuned using training sets with different sample weights.

[0083] The following embodiments are implemented based on the technical solution of the present invention, providing detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments. Unless otherwise specified, the methods used in the following embodiments are conventional methods.

[0084]

Example

[0085] See Figure 3 and Figure 4 Taking the New England 10-machine 39-node system as an example, the voltage amplitude and phase angle of 39 buses are selected as time-series input features. The original sample set is re-divided into the form shown in Table 1, which shows the stable and unstable sample situations of the four datasets. Some samples are extracted from the three datasets in a hierarchical manner to maintain the consistency with the category ratio of the original sample set and form a mixed dataset D in a ratio of 2:1:1.

[0086] Table 1

[0087] Dataset Stablize Instability total A (Fully wired) 2436 1235 3671 B(N-1 disconnected line) 1517 958 2475 C(N-2 broken wire) 845 755 1600 D (Stratified sampling mixture) 1249 757 2006

[0088] The different combinations of network layers in SA-GCN and TA-LSTM affect the evaluation performance of the STA-GLN model. We designed STA-GLN combined networks with different layer combinations and extracted one-third of the sample data from dataset A to test the transient stability evaluation performance of the models, thus determining the optimal structure of the STA-GLN transient stability evaluation model. Table 2 shows the performance of STA-GLN models with different layer combinations; when both SA-GCN and TA-LSTM networks have 2 layers, the transient stability evaluation performance of STA-GLN is better.

[0089] Table 2

[0090]

[0091] To verify the evaluation performance of the proposed STA-GLN-based transient stability evaluation model without considering noise and missing data, dataset D was selected. Networks used to construct the STA-GLN model were chosen to build corresponding transient stability evaluation models, including GCN, SA-GCN, LSTM, TA-LSTM, and GCN-LSTM, and performance comparison analysis was performed.

[0092] Table 3

[0093] Evaluation methods <![CDATA[P ACC / %]]> <![CDATA[P REC / %]]> <![CDATA[R EC / %]]> <![CDATA[F1 / %]]> GCN 90.41 94.08 96.62 95.33 SA-GCN 91.10 94.70 96.62 95.65 LSTM 92.47 95.36 97.30 96.32 TA-LSTM 93.15 95.39 97.97 96.67 GCN-LSTM 94.52 96.67 97.97 97.32 STA-GLN 96.58 98.64 97.97 98.31

[0094] As shown in Table 3, the STA-GLN network achieved an accuracy of 96.58% and an F1 score of 98.31% in the comparative experiment, demonstrating a good balance between precision and recall, with all four metrics being the highest.

[0095] Different algorithms were selected and their performance compared with the multi-task transient stability evaluation model proposed in this invention. The transient power angle stability evaluation of Task 1 is a binary classification problem, and the evaluation metric used is the accuracy P calculated based on the transient stability confusion matrix. ACC Precision P rec Recall rate R ec Given the F1 score, the transient voltage stability assessment in Task 2 is a multi-dimensional label classification problem. The Jaccard similarity coefficient is chosen as the evaluation metric. A sample set E is constructed from all samples to verify the performance superiority of the integrated STA-GLN multi-task transient stability assessment model proposed in this invention.

[0096] Table 4

[0097]

[0098] Compared with other models, the model proposed in this invention performs best in the entire sample set. The AdaBoost-STA-GLN multi-task evaluation model has an accuracy of 98.76% and a recall of 99.58%, indicating that the model can effectively mine the power angle characteristics and voltage characteristics of the system transient process.

[0099] To verify the anti-interference capability of the proposed AdaBoost-STA-GLN multi-task transient stability evaluation model, Gaussian white noise with a signal-to-noise ratio of 30dB and 50dB was added to the original sample set E to form noise interference sample sets G and H, respectively. Simulation results showed that the sample feature amplitudes oscillated after adding both types of Gaussian white noise, negatively impacting the extracted transient characteristics. The AdaBoost-STA-GLN multi-task model was trained and tested with other models on the noise sample set. The anti-interference capability test results of each model are shown in Table 5, which illustrates the anti-interference capabilities of different ensemble models.

[0100] Table 5

[0101]

[0102] Comparing the performance of each model under external disturbances, the AdaBoost-STA-GLN multi-task transient stability assessment model proposed in this invention showed the smallest performance decline among all ensemble models. The accuracy of Task 1 decreased by 0.16% and 0.69%, respectively, while the Jaccard coefficient for Task 2 decreased by 0.20% and 0.75%, respectively. The proposed AdaBoost-STA-GLN multi-task model exhibits better robustness than other ensemble models. To verify the adaptability of the AdaBoost-STA-GLN multi-task assessment model to power system topology changes, training and testing sets composed of different sample sets were used. The generalization performance of the multi-task model was tested using two schemes: Scheme 1 involved training the model on a fully connected sample set A and testing it on an N-1 disconnected sample set B; Scheme 2 involved training the model on a mixed sample set consisting of the fully connected sample set A and the N-1 disconnected sample set B and testing it on an N-2 disconnected sample set C. The performance indicators of the AdaBoost-STA-GLN multi-task evaluation model and the comparison model were obtained under the two schemes. The AdaBoost-STA-GLN multi-task evaluation model proposed in this invention has good generalization performance, strong adaptability to power system topology changes, and can perform accurate transient stability evaluation in systems with frequent changes in operating conditions.

Claims

1. A power system stability assessment method for high-proportion renewable energy integration, characterized in that, Includes the following steps: Step 1: Based on the operating characteristics of power systems with a high proportion of renewable energy access, construct a transient stability assessment feature set; Step 2: Propose a STA-GLN power system transient stability assessment model based on a spatiotemporal dual attention mechanism, as follows: 1) From a spatial perspective, spatial attention mechanism is introduced into graph convolutional neural network to construct SA-GCN sub-model and extract input features that integrate the topological connectivity characteristics of power system; 2) From a temporal perspective, the temporal attention mechanism is introduced into the LSTM encoder-decoder to construct a TA-LSTM sub-model, calculate the importance of temporal features at each time step, adjust the learning intensity of the evaluation model for important time step input features in a weighted manner, and explore the transient characteristics in the input features to improve model performance. 3) By adopting a cascaded structure and combining the SA-GCN sub-model and the TA-LSTM sub-model, a power system transient stability assessment model based on STA-GLN is obtained; Step 3: Introduce ensemble learning and transfer learning to improve the STA-GLN power system transient stability assessment model, as detailed below: 1) Introducing ensemble learning: The STA-GLN power system transient stability assessment model is used as a sub-model and integrated through AdaBoost. The SAMME algorithm is used to dynamically adjust the input sample weights of each STA-GLN sub-classifier. 2) Introduce transfer learning to transfer the model parameters of the STA-GLN subclassifier.

2. The power system stability assessment method for high-proportion renewable energy access according to claim 1, characterized in that, Step one is described in detail as follows: 1) Generate a sample set; 2) Perform electromechanical transient simulation by setting up a three-phase ground fault on 34 parallel transmission lines without transformers to generate a system transient instability sample set; 3) Construct a transient stability feature set, taking the transient power angle stability and transient voltage stability of the power system as the research objects, and select feature quantities from two dimensions: time and space; 4) Select transient stability assessment indicators.

3. The power system stability assessment method for high-proportion renewable energy access according to claim 1, characterized in that, The transfer learning transfers the model parameters of the first STA-GLN sub-classifier and assigns them to the next sub-classifier. The sub-classifier is then fine-tuned using the training set with updated weights to obtain a new STA-GLN sub-classifier. This process is then repeated.

Images