Transient stability evaluation method for photovoltaic energy storage grid-connected system

Abstract

The application relates to a photovoltaic energy storage grid-connected system transient stability evaluation method, which comprises the following steps: obtaining operation data of a photovoltaic energy storage grid-connected system under different disturbance scenes, calculating a transient power angle stability index and a stability margin change rate based on the operation data; constructing a basic comprehensive stability index model, the input being the transient power angle stability index, and the output being a comprehensive stability index; constructing an online comprehensive stability index correction model, the input being the stability margin change rate, the comprehensive stability index and real-time state of charge data of energy storage, and the output being a final online corrected comprehensive stability index; and identifying the transient state of the system under different fault scenes based on the final online corrected comprehensive stability index. The application realizes accurate, efficient and real-time evaluation of the transient stability of the photovoltaic energy storage grid-connected system, and has the core advantages of high evaluation precision, strong anti-interference capability, wide scene adaptability, efficient real-time early warning and strong practicability, and can effectively support the safe and stable operation of a power grid.

Description

Technical Field

[0001] This application relates to the field of power system operation and control technology, specifically to a transient stability assessment method for photovoltaic energy storage grid-connected systems. Background Technology

[0002] Photovoltaic power generation, as a clean and renewable energy source, exhibits significant intermittency and volatility. Its power output is affected by various factors such as solar radiation intensity, weather changes, and day-night cycles, leading to voltage fluctuations and flicker when connected to the grid. With the large-scale grid integration of new energy sources, the connection of photovoltaic energy storage systems to the main grid has a significant impact on system transient stability. Their bidirectional power flow characteristics can optimize cluster regulation and control, ensuring stable system operation and improving power quality and reliability. Energy storage systems smooth power output through peak and valley regulation, enabling coordinated interaction between photovoltaic power generation and the grid, reducing the impact of intermittent fluctuations on the grid, enhancing grid voltage stability, alleviating grid peak-shaving pressure, optimizing power resource allocation, and providing strong support for the economical and stable operation of the grid.

[0003] However, existing technologies have significant drawbacks: 1. Low assessment accuracy: Existing technologies have not constructed a feature system that fits the transient characteristics of photovoltaic energy storage. Key fault features are not prominent, there is significant redundancy and interference, and most use a single model, which is prone to overfitting and insufficient generalization ability. At the same time, the judgment criteria are fixed and cannot be adapted to different scenarios, resulting in low accuracy and unreliable results in transient stability identification. 2. Poor scenario adaptability: Existing technologies cannot adapt to different grid topologies and different photovoltaic energy storage installed capacities. When facing atypical scenarios, a complete model needs to be retrained, resulting in low adaptation efficiency and weak engineering applicability, making it difficult to meet the diverse operational needs of actual power grids. 3. Insufficient real-time performance and early warning: Existing technologies have slow data acquisition and model inference speeds, making it impossible to achieve online real-time assessment of transient stability; they lack accurate stability margin prediction capabilities, cannot predict instability risks in advance, and lack a sound hierarchical early warning and active intervention mechanism, making it difficult to prevent fault escalation.

[0004] Therefore, this invention proposes a transient stability assessment method for photovoltaic energy storage grid-connected systems. Summary of the Invention

[0005] To overcome the shortcomings of the existing technology, this application provides a transient stability assessment method for photovoltaic energy storage grid-connected systems, which solves the technical problems of poor adaptability, insufficient accuracy, and difficulty in engineering implementation of existing assessment methods. The specific technical solution is as follows.

[0006] A transient stability assessment method for a photovoltaic energy storage grid-connected system includes the following steps.

[0007] S1. Data Acquisition and Preprocessing: Acquire operational data of the photovoltaic energy storage grid-connected system under different disturbance scenarios, and eliminate data errors and dimensional differences through preprocessing.

[0008] S2. Input Feature Construction and Optimization: Construct a multi-dimensional input feature set that fits the transient characteristics of the photovoltaic energy storage grid-connected system, and improve the model's accuracy in identifying transient faults through feature optimization.

[0009] S3. Transient stability determination and label construction: Based on the physical operating characteristics of the power system, a two-dimensional transient stability determination standard of power angle and voltage is established to label the training samples as stable / instable, ensuring that the labels are accurate and reasonable.

[0010] S4. Construct a transient stability evaluation model based on gcForest-AdaBoost. Through hybrid algorithm optimization, improve the problems of overfitting and insufficient generalization ability of the single gcForest model, and achieve accurate identification of the transient state of the system under different fault scenarios.

[0011] S5. Model Optimization: Through anti-interference training and cross-scenario adaptation optimization, the robustness and generalization ability of the model under complex working conditions and changing power grid scenarios are further improved.

[0012] S6. Online Transient Stability Assessment and Early Warning: Performs real-time transient stability assessment on photovoltaic energy storage grid-connected systems, outputs assessment results and early warning information, supports the safe and stable operation of the power grid, and prevents the fault from escalating further.

[0013] As an optional solution of this invention, data acquisition and preprocessing provide high-quality basic data for the entire evaluation method. A PSASP simulation model that fits the actual engineering situation is built based on IEEE 39 nodes, which can be flexibly adapted to photovoltaic installed capacity of 10MW-100MW. Various disturbance scenarios are fully simulated and repeated to ensure data representativeness. Eighteen types of core electrical quantities and energy storage parameters are collected. After cleaning and normalization preprocessing, multiple types of dynamic noise are injected to simulate the actual PMU measurement environment. This overcomes the shortcomings of traditional data acquisition scenarios that are single and out of touch with reality, and lays a reliable foundation for subsequent feature construction and model training.

[0014] As an optional solution of the present invention, a high-quality feature set adapted to transient assessment is constructed by input feature construction and optimization. Multi-dimensional basic features covering buses, power sources, generators, lines and loads are screened to construct transient derivative features and energy storage status correlation features, and establish the correlation between energy storage regulation capability and grid stability. Feature weights are dynamically allocated through a time-series attention enhancement module to highlight key fault features. Then, random forest screening and PCA dual dimensionality reduction are used to solve the problems of traditional feature redundancy and lack of key information highlighting, while taking into account both feature comprehensiveness and model training efficiency.

[0015] As an optional solution of this invention, transient stability determination and label construction achieve accurate labeling of sample labels, breaking through the limitations of traditional fixed thresholds. Based on the scene feature library, the power angle and voltage determination thresholds are dynamically adjusted to adapt to different photovoltaic installed capacities and grid topologies. A comprehensive stability index model is constructed by adopting a dual-dimensional determination logic of power angle and voltage, dynamically allocating weight coefficients, and adjusting the power angle and voltage weights according to the photovoltaic installed capacity to avoid the bias of single-dimensional determination and provide high-quality labeled samples for model training.

[0016] As an optional solution of the present invention, a gcForest-AdaBoost hybrid model is constructed to solve the problem of insufficient evaluation accuracy of a single model. The gcForest multi-granularity scanning is used to enhance feature representation, and the sample weights are optimized iteratively by combining AdaBoost. Dropout layers and early stopping strategies are added to the cascaded forest to effectively suppress overfitting, achieve accurate identification of transient faults, ensure that the model training process is controllable and the results are reliable, and break through the limitations of the generalization ability of traditional single models.

[0017] As an optional solution of this invention, model optimization and adaptation improve the model's anti-interference ability and scene adaptability. The invention points are: using wavelet decomposition to separate multi-source noise and adopting adaptive suppression strategies for different noise types; achieving sample equalization through the ADASYN algorithm to solve the problem of unstable sample identification bias; constructing a scene feature library and combining transfer learning to achieve rapid adaptation to multiple power grid topologies and multiple installed capacity scenarios without retraining the complete model, which greatly improves the model's engineering applicability.

[0018] As an optional solution of this invention, online real-time evaluation is achieved through high-frequency acquisition by PMU and rapid model inference; an LSTM model is used to predict the stability margin trend in the next 3-5 seconds, and an online correction model linking energy storage SOC and stability margin is constructed; a three-level early warning and active intervention mechanism is designed, coupled with an incremental model update strategy, to adapt to engineering operation and maintenance processes, and can be directly implemented in typical scenarios such as IEEE 39 / 118 nodes, breaking through the problems of traditional evaluation only making judgments, not providing early warnings, and being difficult to implement.

[0019] The technical solution of this application has achieved the following beneficial effects.

[0020] This application improves assessment accuracy by constructing a basic comprehensive stability index model and an online comprehensive stability index correction model, combined with the gcForest-AdaBoost hybrid model and normalization calculation; enhances anti-interference capability by relying on error closed-loop control and multi-scenario optimization coefficients; achieves wide-scenario adaptation by dynamically adjusting thresholds and multi-granularity scanning through scenario library; and introduces stability margin change rate, time series prediction and fast inference to ensure real-time and efficient early warning. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the transient stability assessment method for a photovoltaic energy storage grid-connected system in this application.

[0022] Figure 2 This is a flowchart of the gcForest-AdaBoost strong classifier for the transient stability assessment method of the photovoltaic energy storage grid-connected system in the embodiments of this application. Detailed Implementation

[0023] The present application will now be further described with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present application and should not be construed as limiting the scope of protection of the present application. It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present application.

[0024] like Figure 1 and Figure 2 As shown, this invention discloses a transient stability assessment method for a photovoltaic energy storage grid-connected system, comprising the following steps.

[0025] S1. Data Acquisition and Preprocessing: Acquire operational data of the photovoltaic energy storage grid-connected system under different disturbance scenarios, and eliminate data errors and dimensional differences through preprocessing.

[0026] S2. Input Feature Construction and Optimization: Construct a multi-dimensional input feature set that fits the transient characteristics of the photovoltaic energy storage grid-connected system, and improve the model's accuracy in identifying transient faults through feature optimization.

[0027] S3. Transient stability determination and label construction: Based on the physical operating characteristics of the power system, a two-dimensional transient stability determination standard of power angle and voltage is established to label the training samples as stable / instable, ensuring that the labels are accurate and reasonable.

[0028] S4. Construct a gcForest-AdaBoost hybrid evaluation model: Construct a transient stability evaluation model based on gcForest-AdaBoost. Through hybrid algorithm optimization, improve the problems of easy overfitting and insufficient generalization ability of the single gcForest model, and achieve accurate identification of the transient state of the system under different fault scenarios.

[0029] S5. Model Optimization: Through anti-interference training and cross-scenario adaptation optimization, the robustness and generalization ability of the model under complex working conditions and changing power grid scenarios are further improved.

[0030] S6. Online Transient Stability Assessment and Early Warning: Performs real-time transient stability assessment on photovoltaic energy storage grid-connected systems, outputs assessment results and early warning information, supports the safe and stable operation of the power grid, and prevents the fault from escalating further.

[0031] Example 1.

[0032] This embodiment 1 discloses a transient stability assessment method for a photovoltaic energy storage grid-connected system, wherein the specific detailed steps of step S1 are as follows.

[0033] 1. Simulation Model Construction: Using the power system simulation software PSASP, a power grid simulation model including photovoltaic energy storage units is constructed. Based on the IEEE 39-bus system, 10 traditional generator sets, 39 buses, and 46 transmission lines are configured. The system base power is set to 1000MVA, and the rated frequency is 50Hz. The photovoltaic energy storage units are connected to bus 39 of the IEEE 39-bus system. This bus is the system load center bus and has strong power regulation capabilities. The photovoltaic installed capacity can be configured from 10MW to 100MW according to actual project needs. The energy storage system uses lithium battery energy storage, with the initial SOC value set to 60%-80%, and the charging and discharging power limited to 30%-50% of the photovoltaic installed capacity.

[0034] 2. Simulated Disturbance Scenarios: Simulate various large disturbance scenarios that may occur in the actual operation of the photovoltaic energy storage grid-connected system, including: three-phase short circuit fault, line tripping fault, photovoltaic output sudden drop fault, large capacity load switching fault, photovoltaic output sudden change caused by extreme weather, and multiple fault superposition, etc. Each disturbance scenario is simulated 50 to 100 times to ensure the diversity and representativeness of the data.

[0035] 3. Data Acquisition: The simulation software's data acquisition module collects real-time system operation data for 0.5 seconds before, during, and 3 seconds after a disturbance. The acquisition frequency is set to 100Hz. The acquired data includes: bus voltage amplitude, bus voltage phase angle, generator power angle, generator angular velocity, generator terminal voltage, generator terminal current, generator electromagnetic power, generator active / reactive power, photovoltaic power station active / reactive power, energy storage power station active / reactive power, energy storage SOC, AC line active / reactive power, AC line voltage, AC line phase angle, and bus load active / reactive power, totaling 18 categories of core electrical quantities and energy storage status parameters.

[0036] 4. Data Preprocessing: First, the collected data is cleaned to remove outliers and missing values. Then, the cleaned data is normalized using the min-max normalization method to map all data to the [0,1] interval, eliminating the influence of different units. Finally, the original transient stable dataset is constructed, categorized and stored according to the type of perturbation, for subsequent feature construction and model training.

[0037] 5. Simulation noise injection: During the simulation dataset construction phase, weak Gaussian white noise, impulse noise, and harmonic noise are injected, and a dynamic noise intensity change scenario is set to simulate the actual PMU measurement environment, providing a data foundation for subsequent model anti-interference training.

[0038] In this technical solution, by building a realistic power grid simulation model, simulating comprehensive and representative disturbance scenarios, and collecting accurate and comprehensive operational data, and through scientific preprocessing and noise injection, high-quality and highly adaptable basic data is provided for the entire evaluation method, laying a reliable foundation for subsequent model training and evaluation.

[0039] Example 2.

[0040] This embodiment 2 discloses a transient stability assessment method for a photovoltaic energy storage grid-connected system, wherein the specific detailed steps of step S2 are as follows.

[0041] 1. Basic Feature Selection and Determination: Select raw electrical quantities that are easy to measure and have clear physical meaning during system operation as basic features, specifically including...

[0042] (1) Bus dimension: voltage amplitude and voltage phase angle of 39 buses, totaling 78 basic characteristics.

[0043] (2) Power supply dimension: grid connection voltage, active power, and reactive power of photovoltaic power stations, grid connection voltage, active power, and reactive power of energy storage power stations, a total of 6 basic characteristics.

[0044] (3) Generator dimension: 70 basic characteristics of 10 generators, including power angle, angular velocity, terminal voltage, terminal current, electromagnetic power, active power and reactive power.

[0045] (4) Line and load dimensions: active power, reactive power, voltage and phase angle of 46 AC lines, active power and reactive power of 39 busbars, totaling 282 basic characteristics.

[0046] 2. Constructing Derivative Features: Based on the basic features, construct derivative features that reflect the dynamic characteristics of the transient process, specifically including...

[0047] (1) Transient process derivative characteristics: The first-order difference method is used to calculate the bus voltage change rate, generator power angle change rate, and photovoltaic / energy storage active power ramp rate. Each basic characteristic corresponds to one derivative characteristic, with a total of 436 derived characteristics, which are used to quantify the dynamic response speed of the system after disturbance.

[0048] (2) Energy storage system state correlation characteristics: The SOC, SOC change rate and reserve capacity ratio of energy storage power station are introduced. The SOC change rate is calculated by the SOC difference between adjacent times, and the reserve capacity ratio is calculated based on the current SOC of energy storage and the charging and discharging limit. A total of 3 derived characteristics are established to establish a direct correlation mapping between energy storage regulation capability and grid transient stability.

[0049] 3. Temporal Feature Attention Enhancement Processing: Construct a temporal feature attention enhancement module to dynamically focus on and enhance temporal features in transient processes. The specific operations are as follows.

[0050] 1) Time-series feature division: All features are divided into millisecond-level features and second-level features according to the time scale; millisecond-level features include instantaneous changes in bus voltage, current and power angle; second-level features include slowly changing features such as energy storage SOC and load changes, which are calculated by averaging the data collected at 100Hz using a 1s sliding window.

[0051] 2) Attention Weight Calculation: An attention mechanism is used to calculate the attention weight of each time-series feature. The weight calculation is positively correlated with the amplitude and duration of the feature's sudden change. The weight threshold is dynamically adjusted based on the scenario feature library: for key features with a bus voltage drop ≥10% and a duration ≥50ms, the attention weight is increased to 1.8-2.0; for features with a sudden drop in photovoltaic energy storage output ≥20%, the attention weight is increased to 1.5-1.7; for routine features under stable operating conditions, such as bus voltage and load power under normal operating conditions, the attention weight is reduced to 0.5-0.7. The above weight ranges were all optimized and determined through multi-scenario simulation comparison experiments.

[0052] 3) Feature fusion: The enhanced time-series features are fused with the basic features and derived features to form the final input feature set, which contains a total of 879 features. This ensures the comprehensiveness of the features, highlights the role of key fault features, suppresses the interference of redundant features, and improves the accuracy of subsequent model evaluation.

[0053] 4. Feature Dimensionality Reduction and Optimization: First, random forest feature importance analysis is used to retain the top 60% of key features with clear physical meaning. Then, principal component analysis (PCA) algorithm is used to reduce the dimensionality of the fused feature set, retaining principal components with a cumulative variance contribution rate of ≥95%, eliminating redundant features, and compressing the feature dimension to 100-150 dimensions. This reduces the computational cost of model training, avoids the curse of dimensionality, and ensures the efficiency and accuracy of model training.

[0054] This technical solution provides high-quality input features for the model by constructing comprehensive basic features and targeted derived features, combined with attention enhancement to highlight key features and dual dimensionality reduction optimization to balance accuracy and efficiency, thereby improving the accuracy and efficiency of subsequent evaluation.

[0055] Example 3.

[0056] This embodiment 3 discloses a transient stability assessment method for a photovoltaic energy storage grid-connected system, wherein the specific detailed steps of step S3 are as follows.

[0057] 1. Setting the judgment criteria parameters.

[0058] (1) Transient power angle stability judgment parameters: Set the threshold for calculating the transient power angle stability index, and set the simulation time to 3s after the disturbance occurs. Combining offline simulation and online data, the threshold for judging the maximum relative power angle difference is dynamically adjusted by the threshold mapping algorithm based on the scene feature library: if the photovoltaic installed capacity of the current scene in the scene feature library is ≥50MW, the threshold is adjusted to 28° (0.489rad); if the photovoltaic installed capacity is <50MW, the threshold is adjusted to 32° (0.559rad).

[0059] (2) Transient voltage stability judgment parameters: Combining offline simulation and online data, the threshold mapping algorithm based on the scene feature library is used to dynamically adjust the bus node voltage threshold and duration safety threshold: If the current scene in the scene feature library is the IEEE 118-node system, the voltage threshold is adjusted to 0.87 pu and the duration threshold is adjusted to 0.15 s; if it is the IEEE 39-node system, the voltage threshold is adjusted to 0.83 pu and the duration threshold is adjusted to 0.25 s.

[0060] 2. Transient work angle stability determination and labeling.

[0061] (1) Calculate the transient power angle stability index: Use the transient power angle stability index calculation formula to calculate the maximum relative power angle difference between any two generators during the simulation time.

[0062] (2) Labeling rules: The system is judged based on the scene adaptive threshold. When the maximum relative power angle is less than the corresponding scene threshold and greater than 0.3, the system is judged to be transiently stable in power angle, and the power angle stability label of the corresponding sample is recorded as 1. When the maximum relative power angle is greater than or equal to the corresponding scene threshold or less than or equal to 0.3, the system is judged to be transiently unstable in power angle, and the power angle stability label of the corresponding sample is recorded as 0.

[0063] 3. Transient voltage stability assessment and labeling.

[0064] (1) Voltage data statistics: After the disturbance occurs, the duration t of the voltage of each bus node being lower than the corresponding scenario threshold; the duration t of being lower than 0.85pu; and the transient voltage stability index is calculated at the same time.

[0065] (2) Labeling rules: If t≤0.2s and transient voltage stability index>0.3, the voltage of the bus node is determined to be stable, and the voltage stability label of the corresponding sample is recorded as 1; if t>0.2s or T transient voltage stability index≤0.3, the voltage of the bus node is determined to be unstable, and the voltage stability label of the corresponding sample is recorded as 0; each sample corresponds to the voltage stability labels of 39 bus nodes, forming a multidimensional label vector.

[0066] 4. Final Sample Label Determination: A basic comprehensive stability index model is constructed by combining power angle stability and voltage stability labels. Weighting coefficients are dynamically adjusted based on the scenario feature library. If the photovoltaic installed capacity of the current scenario in the scenario feature library is ≥50MW, the power angle stability weight is adjusted accordingly. Adjusted to 0.5, voltage stability weight. Adjust to 0.5; if the photovoltaic installed capacity is <50MW, Adjusted to 0.7. Adjusted to 0.3; baseline value =0.6, =0.4. The basic comprehensive stability index model is as follows.

[0067] .

[0068] ; . .

[0069] in, The weighting coefficient for the power angle stability index; This is the weighting coefficient for the voltage stability index; is the transient power angle stability index, used to quantify the power angle stability of the system, with a value range of [0,1]; n is the total number of bus nodes in the system, in this scheme n=39, corresponding to the IEEE 39-bus system. Let be the transient voltage stability index of the k-th bus node, used to quantify the voltage stability of that node, with a value range of [0,1]. The calculation is based on the normalized calculation of the bus node voltage over-limit duration and the voltage stability limit duration. It is a basic comprehensive stability index. It is the maximum relative power angle difference between any two generators within 3 seconds after the disturbance occurs, in degrees, corresponding to a radian value of 0.523 rad (30°); t is a time variable, with a value range of [0, 3s], that is, the time period from 0 to 3 seconds after the disturbance occurs; It is the power angle of the i-th generator at time t, in degrees, reflecting the operating phase state of the generator rotor; It is the power angle of the j-th generator at time t, in degrees, i≠j (i.e., any two different generators). It is the power angle stability limit value, which is taken as 30° (0.523 rad) in this scheme. It is the core threshold for determining whether the power angle of the system is stable. is the voltage stability limit duration, which is the upper limit of the safe duration when the bus voltage is lower than the threshold; -t is the actual duration for which the voltage of the k-th bus node is lower than 0.85pu-per-unit value after the disturbance occurs, in seconds; k is the bus node number, ranging from 1 to n; 0.85pu is the bus node voltage judgment threshold, pu is the per-unit value, based on the system rated voltage.

[0070] like If the value is greater than 0.3, the sample is considered a stable sample, and the final label is recorded as 1; if... If the value is ≤0.3, the sample is determined to be an unstable sample, and the final label is recorded as 0, ensuring that the label can fully reflect the transient stable state of the system.

[0071] This technical solution employs a threshold mapping algorithm based on a scenario feature library. It dynamically adjusts the power angle and voltage thresholds according to photovoltaic installed capacity and grid topology, avoiding the poor adaptability of fixed thresholds and adapting to different operating scenarios. A basic comprehensive stability index model is constructed to accurately integrate the dual-dimensional judgment results, ensuring that the final sample label truly reflects the system's transient state and avoiding label bias.

[0072] Example 4.

[0073] This embodiment 4 discloses a transient stability assessment method for a photovoltaic energy storage grid-connected system, wherein the specific detailed steps of step S4 are as follows.

[0074] 1. Initialize the weights for each data point in the dataset, with the sample set as follows: The sample size is The sample initial weight distribution is as follows: Sample initialization weights Ensure that all samples have the same weights during the initial training phase.

[0075] The core advantage of the gcForest model lies in its enhanced input feature representation through multi-granularity scanning technology, adapting to the multi-dimensional and multi-timescale coupling characteristics of transient features in photovoltaic energy storage grid-connected systems. The specific multi-granularity scanning process is as follows: A sliding window is introduced to scan the input features, using multiple sliding windows of different sizes simultaneously to include more temporal and spatial correlation features in the input feature vector. Parallel computing is employed to optimize the multi-granularity scanning process. The specific multi-granularity scanning process is as follows.

[0076] Suppose we input a model The input feature samples of dimension L are sampled using a sliding window of length L to obtain... indivual First, a feature subsample vector is generated; second, after the subsample vector is used in the cascaded forest model, each layer of the forest will yield a feature subsample vector of length 1. Different probability vectors (of the number of categories), The number of categories; that is, the length that can be obtained is The representation vector. If the dimension is The sliding window has a total of The number of dimensions can be used to obtain the output results. .

[0077] .

[0078] In the formula, For the first A sliding window; For each layer Number of forests.

[0079] Multi-granularity scanning selects sub-samples of different lengths, which is beneficial for associating the dimensionality and feature attributes of the samples, thereby strengthening the cascaded forest.

[0080] 2. Set the iteration round to Second-rate, ,based on Generate weak classifiers for: In the formula, It is a weak classifier; This represents the training mapping function using gcForest, a multi-granularity cascaded forest, as the base learner. The input is the feature set X and the sample weight distribution. The output is the weak classifier obtained through training. T represents the total number of iterations or the total number of weak classifiers.

[0081] 3. The weak classifier generated in this iteration In the present The weighted error rate is [missing value].

[0082] .

[0083] In the formula, It is the weight of the i-th sample in the n-th iteration; It is the predicted value of the i-th sample during the nth weak learning period. It is the true value of the i-th sample; It is the weighted error rate.

[0084] 4. Calculation Weighting coefficients in the set of strong classifiers : .

[0085] 5. Update the sample weight distribution based on the classification results to prepare for the next iteration.

[0086] .

[0087] In the formula, It is the sample weight distribution for the next iteration. It is the sample weight in round t.

[0088] 6. After the iteration is completed, the final strong classifier is output, which is the ensemble deep random forest model gcForest-AdaBoost.

[0089] .

[0090] In the formula, To ultimately integrate a strong classifier, the gcForest-AdaBoost model outputs: sign( ) is the sign function, used to output the final binary classification result; T is the total number of iterations in AdaBoost, the total number of weak classifiers; is the weighting coefficient of the t-th weak classifier; the larger the coefficient, the higher the reliability of the weak classifier. is the weak classifier obtained in the t-th round of training, with gcForest as the base learner; x is the input sample feature vector, which is the vector after feature optimization in step S2.

[0091] Add a dropout layer to the cascaded forest of gcForest, set the dropout ratio to 0.2~0.5, and adopt an early stopping strategy. Stop training when the accuracy on the validation set has not improved for 5~10 consecutive rounds to prevent the model from overfitting.

[0092] In this technical solution, by constructing a gcForest-AdaBoost hybrid model, combining multi-granularity scanning to enhance feature representation and adding overfitting control strategies, the shortcomings of a single model are compensated for, achieving accurate identification of transient faults and ensuring that the model training process is controllable and the results are reliable.

[0093] Example 5.

[0094] This embodiment discloses a transient stability assessment method for a photovoltaic energy storage grid-connected system, wherein step S5 is detailed as follows.

[0095] 1. Anti-interference training.

[0096] (1) Multi-source noise injection: Three types of typical PMU measurement noise are injected into the training dataset, including Gaussian white noise, impulse noise and harmonic noise.

[0097] (2) Adaptive suppression of multi-source noise: Configure the adaptive suppression module for multi-source noise and perform adaptive suppression of multi-source noise. The specific operation process is as follows.

[0098] ① Noise identification and separation: The training data after noise injection is decomposed into Gaussian white noise, impulse noise and harmonic noise by wavelet decomposition algorithm. The wavelet basis is selected as db4 wavelet and the decomposition level is 5. After decomposition, 5 levels of low frequency approximation components and 5 levels of high frequency detail components can be obtained. This method can effectively separate noise from clean signals.

[0099] ② Targeted Noise Suppression: For Gaussian white noise, an adaptive Kalman filter algorithm is used for suppression. The filter coefficient is dynamically adjusted according to the noise intensity; the higher the noise intensity, the larger the filter coefficient, with a value range of 0.1-0.5. Through dynamic adjustments via state and gain updates, precise suppression of Gaussian white noise is achieved. For impulse noise, a combination of median filtering and threshold filtering is used. A dynamic noise threshold is set, which is adjusted according to the amplitude of the measured data, and is set to 1.2 times the mean of this type of data, accurately eliminating impulse noise. For harmonic noise, a notch filter is introduced, and the notch parameters are automatically adjusted according to the harmonic frequency. The damping coefficient is set to 0.707, achieving adaptive suppression of harmonic noise.

[0100] (3) Sample Balancing: The Adaptive Synthetic Sampling (ADASYN) algorithm is used to oversample the unstable minority class samples. Specifically, the number of synthetic samples is adjusted according to the sample density, generating fewer synthetic samples for unstable samples with higher density and more synthetic samples for unstable samples with lower density. The k-nearest neighbors (k=5) of the minority class samples are selected. New synthetic samples are generated between the minority class samples and their nearest neighbors using linear interpolation. The oversampling ratio is set to 1:1 to ensure that the number of stable and unstable samples in the training set is equal. This strategy effectively avoids model bias caused by sample imbalance and improves the accuracy of identifying unstable samples.

[0101] 2. Model Performance Validation: Input the anti-interference training and test sets into the gcForest-AdaBoost model. Evaluate model performance using confusion matrix metrics, including accuracy, recall, precision, F1 score, and area under the AUC curve. The required accuracy is ≥95%, recall ≥94%, precision ≥94%, F1 score ≥94%, and AUC ≥0.95. Each metric is calculated using corresponding logic. Accuracy is the proportion of correctly identified samples to the total number of samples; recall is the proportion of correctly identified unstable samples to all unstable samples; precision is the proportion of correctly identified unstable samples to all samples judged as unstable; F1 score is the harmonic mean of recall and precision; and AUC is the area under the receiver operating characteristic curve. In these formulas, TP represents true positives, TN represents true negatives, FP represents false positives, and FN represents false negatives. If the metric requirements are not met, return to step S4 to adjust the model parameters and retrain until the requirements are met.

[0102] 3. Cross-scene adaptation optimization.

[0103] (1) Constructing a scenario feature library: Pre-construct scenario feature libraries for different power grid topologies and different photovoltaic energy storage installed capacities, specifically including.

[0104] ① Power Grid Topology Scenarios: Covers two typical power grid topologies: IEEE 39-node and IEEE 118-node, storing node parameters, line parameters, and generator parameters for each topology.

[0105] ② Photovoltaic energy storage installed capacity scenarios: covering four typical photovoltaic installed capacities of 10MW, 30MW, 50MW and 100MW, with corresponding energy storage installed capacities of 3MW, 9MW, 15MW and 30MW, storing the photovoltaic output characteristics and energy storage charging and discharging parameters for each capacity.

[0106] ③ Adaptation parameter storage: Store the optimal hyperparameter combination and feature weight allocation standard of the gcForest-AdaBoost model for each scene to form a scene feature library, which facilitates rapid adaptation of the model in the future.

[0107] (2) Cross-scenario migration and adaptation: When the topology and installed capacity of the photovoltaic energy storage grid-connected system change, the model automatically calls the corresponding parameters in the scenario feature library. The specific adaptation process is as follows.

[0108] ① Scene Recognition: By collecting real-time system parameters, the current system automatically identifies the corresponding scene. The real-time parameters include the number of nodes, the number of lines, and the installed capacity of photovoltaic energy storage. A scene similarity calculation model is introduced.

[0109] .

[0110] In the formula, S represents the scene similarity; the smaller S is, the higher the similarity. Q represents the number of scene features. Let m be the weight of the m-th scene feature. These are the feature values ​​of the current system scene. These are the feature values ​​of the reference scene in the scene feature library.

[0111] Setting a similarity threshold S < 0.1 defines a highly similar scene, allowing direct reuse of reference scene parameters.

[0112] The scenario feature library not only includes two typical topologies (IEEE 39 / 118 nodes) and four typical photovoltaic (PV) installed capacities (10 / 30 / 50 / 100MW), but also supports extensions to scenarios with arbitrary grid topologies and installed capacities. When the system operates in atypical scenarios not predefined in the feature library (such as 20MW PV capacity or custom grid topologies), the model automatically matches the most similar reference scenario through scenario similarity calculation and calls the corresponding optimal hyperparameters. Subsequently, through supplementary training with a small number of new scenario samples and fine-tuning via transfer learning, rapid adaptation to new scenarios is achieved without retraining the complete model, significantly improving the system's generalization ability and engineering applicability.

[0113] ② Transfer Learning Fine-tuning: The optimal hyperparameters in the scene feature library are called, and the cascaded forest structure and weak classifier parameters of the gcForest-AdaBoost model are quickly fine-tuned through the transfer learning algorithm without retraining the entire model; the parameter fine-tuning model of transfer learning is as follows.

[0114] .

[0115] In the formula For the fine-tuned model parameters, For reference scene model parameters, The learning rate, ranging from 0.01 to 0.05. The gradient of the loss function. This is a sample set for new scenarios.

[0116] ③ Supplemental training with new samples: Supplement the model with a small number of new scene samples to ensure that the model is adapted to the new scene.

[0117] ④ Adaptation Validation: Verify the model adaptation effect through the test set of the new scenario. The accuracy is required to be ≥93%. If the requirement is not met, adjust the transfer learning parameters and readjust the training until the requirement is met. This mechanism can shorten the model adaptation time by more than 70%.

[0118] In this technical solution, various PMU measurement noises are accurately separated and suppressed through multi-source noise injection and targeted suppression strategies, thereby improving the model's anti-interference ability. The ADASYN algorithm is used to achieve sample balance, avoid model bias, and improve the accuracy of unstable sample identification.

[0119] Example 6.

[0120] This embodiment discloses a transient stability assessment method for a photovoltaic energy storage grid-connected system, wherein the specific detailed steps of step S6 are as follows.

[0121] 1. Real-time data acquisition: The system operation data is acquired in real time through the PMU phasor measurement unit. The acquisition frequency is consistent with that during model training, such as 100Hz. The acquired data includes all feature parameters in the input feature set, including basic features and derived features, and is transmitted to the model inference terminal in real time to ensure the real-time performance of the data.

[0122] 2. Real-time assessment and stability margin prediction.

[0123] (1) Real-time evaluation: The real-time data collected by the PMU is preprocessed, including normalization and feature enhancement, and then input into the trained gcForest-AdaBoost model. The model quickly outputs the transient stability / instability determination result of the system, and the inference time is controlled within 50ms to ensure the real-time performance of the evaluation. At the same time, the model evaluation confidence is output, and the model evaluation confidence model is.

[0124] .

[0125] In the formula, C∈[0,1], C>0.5 is considered stable, C≤0.5 is considered unstable, and the closer C is to 1 or 0, the higher the confidence level. T is the total number of weak classifiers or the number of iterations in the gcForest-AdaBoost ensemble model. Represents the t-th weak classifier Weighting coefficients in the entire ensemble model. This represents the classification prediction result of the t-th weak classifier for the input feature vector X. Typically, the output value is +1, indicating a stable classification, or -1, indicating an unstable classification. X is the input feature vector, representing the multi-dimensional feature vector input to the gcForest-AdaBoost model for inference. It includes key features such as electrical quantities and state quantities that have been preprocessed and feature-enhanced from real-time acquisition by the PMU.

[0126] (2) Dynamic prediction of transient stability margin: The transient stability margin dynamic prediction module is used to perform dynamic prediction of transient stability margin. The specific operation is as follows.

[0127] ① Prediction Model Construction: An LSTM time series prediction model is adopted. Real-time data-feature vectors of nearly 100ms are input to construct a transient stability margin prediction model. The LSTM model has 3 hidden layers, each with 64 neurons, and 100 iterations. The Adam optimizer is used. The output of the LSTM model is mapped to the cell state after weight matrix and bias processing through the sigmoid activation function to obtain the transient stability margin in the next 3-5s.

[0128] ② Stability Margin Prediction: Based on current PMU measurement data, energy storage SOC status, and grid topology parameters, the transient stability margin (TSI) trend of the system within the next 3-5 seconds is predicted using an LSTM time series prediction model. The TSI value ranges from 0 to 1, with a larger value indicating a more stable system. The stability margin change rate is introduced and calculated as the ratio of the difference in stability margin between two adjacent moments to the time interval (0.1s). When the change rate is negative, it indicates a decrease in stability margin, and the faster the rate of decrease, the higher the risk of system instability.

[0129] ③ Prediction accuracy verification: The prediction error is controlled within 5%. The prediction error is calculated by dividing the absolute difference between the predicted value and the actual value by the actual value and then multiplying by 100%. If the prediction error exceeds 5%, the LSTM model parameters are adjusted and the prediction model is retrained.

[0130] 3. Comprehensive Transient Stability Margin Determination: To avoid the limitations of single-dimensional evaluation and to address the problem that the traditional comprehensive stability index has fixed weights and cannot adapt to changes in online real-time operating conditions, this invention constructs an online comprehensive stability index correction model to determine the comprehensive transient stability margin. The online comprehensive stability index correction model is as follows.

[0131] .

[0132] In the formula, It is the final comprehensive stability index after online correction, with a value range of [0,1]. It takes into account the dynamic changes of real-time operating conditions and is used for the final system-level transient stability determination. The larger the value, the stronger the system transient stability, which can effectively improve the accuracy of determination. It is a basic comprehensive stability index with a value range of [0,1], which serves as the basis for online correction; It is the stability margin change rate correction coefficient, with a value range of [0.3, 0.5]. It is a fixed coefficient optimized by experience, used to dynamically adapt to the changing trend of stability margin and amplify or weaken the impact of stability margin changes on the final judgment result. The rate of change of stability margin is calculated as the ratio of the difference in stability margin between two adjacent moments to a 0.1s time interval; it is dimensionless. A value less than 0 indicates a decrease in stability margin. >0 indicates an increase in stability margin; is the energy storage SOC state correction coefficient, with a value range of [0.2, 0.4]. It is a fixed coefficient optimized empirically and used to correlate the impact of energy storage regulation capability on system stability, reflecting the coupling relationship between energy storage state and system transient stability. The energy storage SOC state correction coefficient is determined through multi-scenario simulation optimization. SOC is the real-time state of charge of energy storage, with a value range of [60%, 80%]. It is consistent with the initial setting of the energy storage system mentioned above and is the core state parameter collected in real time by the PMU. It is the reference state of charge of energy storage, with a fixed value of 70%, which is the optimal operating state of charge of the energy storage system and serves as the reference value for SOC correction; It is the upper limit of the state of charge of energy storage, with a fixed value of 80%, which is consistent with the initial setting of the energy storage system前文 and is the safety upper limit for energy storage charging and discharging; It is the lower limit of the state of charge of energy storage, with a fixed value of 60%, which is consistent with the initial setting of the energy storage system前文 and is the safety lower limit for energy storage charging and discharging.

[0133] By performing a clipping operation on the minimum value of 1 and the maximum value of 0, it is ensured that the online comprehensive stability index always satisfies 0 ≤ ≤ 1.

[0134] The online comprehensive stability index is constrained within the interval [0, 1] through a clipping function to ensure the numerical rationality.

[0135] 4. Hierarchical warning and active intervention.

[0136] (1) Three-level warning mechanism: Based on the TSI index - that is, the final comprehensive stability index after online correction Adopt a three-level warning mechanism, clarify the warning thresholds and corresponding intervention strategies, and optimize the warning trigger logic in combination with the change rate of the stability margin.

[0137] ① Green warning TSI > 0.7: The system is operating normally, there is no need for intervention, only the operating state of the system is monitored in real time, and the evaluation result is updated every 10s; TSI refers to the final comprehensive stability index after online correction .

[0138] ② Yellow warning 0.3 < TSI ≤ 0.7: There are potential instability risks in the system. A smooth power regulation command is automatically sent to the energy storage system to adjust the active / reactive power output of the energy storage system. The active power regulation range is 5% - 10%, and the reactive power regulation range is 3% - 5%. This range is optimized through multi-scenario simulations of the IEEE 39-node system to improve the system damping.

[0139] ③ Red warning TSI ≤ 0.3: The system is on the verge of instability. Immediately trigger an emergency control strategy, quickly cut off non-critical loads, and the cut-off capacity is 5% - 10% of the total system load. At the same time, quickly adjust the energy storage system for emergency support and increase the energy storage charging and discharging power to 80% - 100% of the rated power. Non-critical loads are: loads that are not essential for residents' daily life, not for power supply to core grid equipment, and have a proportion of ≤ 5% in the impact on system stability.

[0140] (2) Warning signal output: The warning signal is real-time fed back to the operation and maintenance terminal, including the warning level, the current TSI value, the change trend of the stability margin, and the recommended intervention measures, which is convenient for operation and maintenance personnel to grasp the system state in real time and assist in decision-making.

[0141] 5. The model is dynamically updated.

[0142] (1) Update cycle setting: The model is updated regularly according to the changes in the system's operating status. Under normal operating conditions, it is updated once every 3 months. When the system experiences a major failure, topology change, or adjustment of photovoltaic energy storage capacity, the model is updated immediately.

[0143] (2) Update process: Collect system operation data and fault data within the update cycle, preprocess the data and supplement it to the training set, and use incremental training to fine-tune the parameters of the gcForest-AdaBoost model and the LSTM prediction model. There is no need to retrain the entire model, which ensures that the model evaluation accuracy is stable in the long term.

[0144] (3) Update and verification: After the update is completed, the model performance is verified through the test set to ensure that the accuracy, recall and other indicators still meet the requirements. If they do not meet the requirements, incremental training is performed again.

[0145] Real-time online assessment is achieved through PMU real-time data acquisition and rapid inference. Combined with LSTM model, the stability margin trend is accurately predicted. With a three-level early warning and active intervention mechanism with dynamic correction and convenient model update strategy, it is adapted to the deployment of PMU equipment, control logic of photovoltaic energy storage units and operation and maintenance process in engineering. It can be directly implemented in typical power grid scenarios such as IEEE 39 / 118 nodes to achieve real-time monitoring of transient stability, early warning and long-term reliable operation.

[0146] In this technical solution, online real-time evaluation is achieved through PMU real-time data acquisition and rapid inference. Combined with LSTM model, the stability margin trend is accurately predicted. With a three-level early warning and active intervention mechanism with dynamic correction, as well as a convenient model update strategy, real-time monitoring of transient stability, early warning, and long-term reliable operation are realized.

[0147] Furthermore, the stability margin change rate correction coefficient The objective function was determined through multi-scenario orthogonal simulation experiments and objective function optimization. The specific steps are as follows.

[0148] 1. Define the objective function: Using the weighted sum of the unstable sample recall and the overall evaluation accuracy as the evaluation index, construct an objective function to solve for the optimal stability margin change rate correction coefficient. The formula is as follows.

[0149] .

[0150] In the formula: for The corresponding comprehensive evaluation score, the highest score corresponds to This is the optimal value; ; As weight; To assess overall accuracy, The recall rate for unstable samples must be ≥98%.

[0151] 2. Orthogonal simulation experiment.

[0152] Based on the IEEE 39 / 118 node system, covering various disturbance scenarios, the stability margin change rate correction coefficient is... Eleven candidate values ​​were set in the interval [0.3, 0.5] with a step size of 0.02. Each candidate value corresponds to multiple scenario simulations, and the collected evaluation data is used for function calculation.

[0153] 3. Optimization calculation and stability margin change rate correction coefficient Sure.

[0154] Correction coefficients for the rate of change of stability margin in each group Substituting the corresponding simulation data into the objective function, and considering the constraints: unstable sample recall ≥ 98%, overall accuracy ≥ 96%, and warning response time ≤ 50ms, the stability margin change rate correction coefficient with the highest score is selected. As the optimal value.

[0155] The optimization process must meet the requirement of achieving the highest comprehensive evaluation score, while ensuring that the recall rate of unstable samples is ≥98%, the overall evaluation accuracy is ≥96%, and the early warning response time is ≤50ms.

[0156] The stability margin change rate correction coefficient was calculated. =0.40 provides the best overall performance, and the optimal value range for adapting to different scenarios is [0.38, 0.42]. In engineering, 0.40 is preferred.

[0157] 4. Verification Description: Verification shows that... <0.38 is prone to being missed. The false alarm rate is too high above 0.42; only the range [0.38, 0.42] can achieve a balance between accurate identification and low false alarms.

[0158] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for transient stability assessment of a photovoltaic energy storage grid-connected system, characterized in that, Includes the following steps: A basic comprehensive stability index model is constructed. The input of the basic comprehensive stability index model includes the transient power angle stability index, and the output is the basic comprehensive stability index. An online comprehensive stability index correction model is constructed. The input of the online comprehensive stability index correction model is the basic comprehensive stability index, the stability margin change rate and the real-time state of charge data of energy storage, and the output is the final comprehensive stability index after online correction. Based on the final comprehensive stability index after online correction, the transient state of the system under different fault scenarios is identified.

2. The transient stability assessment method for photovoltaic energy storage grid-connected systems according to claim 1, characterized in that... The transient power angle stability index is obtained by calculating the maximum relative power angle difference between any two generators after the disturbance.

3. The transient stability assessment method for photovoltaic energy storage grid-connected systems according to claim 2, characterized in that... The stability margin change rate is the ratio of the difference in transient stability margin between two adjacent moments to the time difference between those two moments.

4. The transient stability assessment method for photovoltaic energy storage grid-connected systems according to claim 1, characterized in that... The online comprehensive stability index correction model is as follows: ； In the formula, It is the final composite stability index after online correction; It is a basic comprehensive stability index; It is the correction coefficient for the rate of change of stability margin; rate of change of stability margin; It is the energy storage SOC state correction factor; SOC It is the real-time state of charge of the energy storage; It is the energy storage reference state of charge; This is the upper limit of the energy storage state of charge; It is the limit of the energy storage state of charge.

5. The transient stability assessment method for photovoltaic energy storage grid-connected systems according to claim 4, characterized in that... The stability margin change rate correction coefficient is determined by combining multi-scenario orthogonal simulation experiments with objective function optimization.

6. The transient stability assessment method for photovoltaic energy storage grid-connected systems according to claim 1, characterized in that... The basic comprehensive stability index model is as follows: ； ； ； ； in, The weighting coefficient for the power angle stability index; This is the weighting coefficient for the voltage stability index; The transient power angle stability index is given by n; n is the total number of bus nodes in the system. The transient voltage stability index of the k-th bus node; It is a basic comprehensive stability index; It is the maximum relative power angle difference between any two generators within 3 seconds after the disturbance occurs; t is the time variable; It is the power angle of the i-th generator at time t; It is the power angle of the j-th generator at time t; It is the stability limit value of the work angle; is the voltage stability limit duration; k is the bus node number.

7. The transient stability assessment method for photovoltaic energy storage grid-connected systems according to claim 6, characterized in that... The transient voltage stability index The calculation is based on the normalized calculation of the bus node voltage over-limit duration and the voltage stability limit duration.

8. The transient stability assessment method for photovoltaic energy storage grid-connected systems according to claim 1, characterized in that... It also includes the step of constructing a transient stability evaluation model based on the ensemble deep random forest algorithm, specifically including the following steps: Initialize the weights of all samples in the dataset to keep the weights consistent during the initial training phase. A multi-granularity scanning technique is adopted, which introduces multi-size sliding windows to scan input features and enhanced feature representations in parallel, and optimizes the scanning process through parallel computing. Set the iteration rounds, construct weak classifiers based on multi-granularity cascaded forests, and calculate the weighted error rate and weighting coefficient of each weak classifier; Update the sample weight distribution based on the classification results and complete multiple rounds of iterative training; After the iteration is completed, a hybrid strong classifier is output, which is the ensemble deep random forest algorithm model after training.

9. The transient stability assessment method for photovoltaic energy storage grid-connected systems according to claim 1, characterized in that... Before constructing the basic integrated stability index model, the process also includes establishing transient stability criteria in two dimensions: power angle and voltage. This involves the following steps: A threshold mapping algorithm based on a scene feature library is adopted to dynamically adjust the maximum relative power angle difference judgment threshold, the bus node voltage threshold and the duration safety threshold for transient power angle stability. Calculate the maximum relative power angle difference between any two generators during the simulation time, compare it with the dynamically adjusted transient power angle stability judgment threshold, determine whether the power angle is stable or not, and mark the corresponding label. The duration for which the voltage at each bus node is lower than the corresponding threshold after dynamic adjustment is statistically analyzed after a disturbance occurs. The transient voltage stability index is calculated to determine whether the voltage is stable or not and to label it accordingly, thus forming a multi-dimensional label vector. By combining power angle stability labels and voltage stability labels, and dynamically adjusting the weight coefficients based on the scenario feature library, a system stability index model is constructed and a comprehensive stability index is calculated. The stability of a sample is then determined based on the system stability index, and the final label is assigned.

10. The transient stability assessment method for a photovoltaic energy storage grid-connected system according to claim 1, characterized in that... It also includes a step for dynamic prediction of transient stability margin, specifically including the following steps: A time-series prediction model is adopted, and real-time operating data features are input to construct a stability margin prediction model. The input of the stability margin prediction model is the real-time operating data features, and the output is the transient stability margin of the system in the future period and its changing trend. Based on real-time measurement data, energy storage status and grid parameters, predict the trend of system stability margin changes; Control the prediction error; if it exceeds the error range, adjust the model parameters and retrain. Calculate the rate of change of stability margin to assess the system stability risk.

Images