Method for transient voltage prediction and stability evaluation of dc receiving end system based on mcnn-bilstm-am

Abstract

This invention discloses a method for transient voltage prediction and stability assessment of DC receiving-end systems based on MCNN-BiLSTM-AM, belonging to the field of transient stability analysis of new energy power systems. The method includes: generating transient voltage stability analysis samples of DC receiving-end systems containing new energy sources; constructing a transient voltage prediction model based on MCNN-BiLSTM-AM; learning and training the transient voltage prediction model based on sample feature acquisition; assessing the transient voltage stability margin of the DC receiving-end system based on the predicted trajectory; and simulation verification of the transient voltage prediction and stability assessment model of the DC receiving-end system. This invention utilizes short-term operational information at the initial stage of a fault to quickly predict the system voltage change trend, and further assesses the system's transient voltage stability margin, thereby achieving rapid and stable analysis of transient voltage prediction and stability of DC receiving-end power systems containing new energy sources.

Description

Technical Field

[0001] This invention belongs to the field of smart grid technology, and in particular to power system stability analysis technology. It relates to a method for predicting transient voltage and assessing stability of DC receiving-end systems by integrating a multi-scale convolutional-bidirectional long short-term memory neural network with an attention mechanism (MCNN-BiLSTM-AM). Background Technology

[0002] Driven by the "dual carbon" goals and the energy structure transformation strategy, the scale of new energy power generation has continued to grow. According to relevant statistics, wind power and photovoltaic power generation capacity together account for approximately 47.3% of the total installed power generation capacity in China, exceeding that of thermal power. With the large-scale grid connection of new energy sources and the introduction of numerous power electronic devices, the structure and dynamic characteristics of the power system have become more complex. Furthermore, the output of new energy sources is random and volatile, making it more susceptible to voltage fluctuations and even voltage instability under fault disturbances. Therefore, transient voltage stability has become one of the important factors affecting the safe and stable operation of new power systems, making rapid prediction of system voltage changes and stability margin assessment crucial.

[0003] Existing power system transient voltage stability analysis primarily relies on time-domain simulation methods. This method establishes a detailed dynamic model of the power system, performs time-domain integration calculations on the dynamic process of the system under fault disturbances, thereby obtaining the system voltage change trajectory. Based on this, voltage stability indices or stability criteria are constructed to evaluate the system's transient voltage stability and stability margin. While time-domain simulation methods can accurately reflect the system's dynamic response process, they involve significant computational loads, long simulation times, and require high accuracy from the system model.

[0004] With the large-scale grid connection of new energy sources and the increasing complexity of power grid structures, the integration of numerous power electronic devices further complicates the dynamic characteristics of the system, increasing the computational burden of simulation analysis. Simultaneously, since stability criteria typically rely on simulation results, relevant indicators struggle to fully reflect the dynamic voltage changes in complex power grids, thus affecting the accuracy of assessment results. In actual power grid operation, transient stability analysis is usually required to be completed within a short time after a fault occurs to provide a basis for power grid transient control and safety decisions. However, traditional time-domain simulation-based transient voltage stability analysis methods struggle to simultaneously meet the requirements of speed and accuracy. Summary of the Invention

[0005] In view of the deficiencies and shortcomings of existing technologies, the purpose of this invention is to provide a method for transient voltage prediction and stability assessment of DC receiving-end systems that integrates MCNN-BiLSTM-AM. This method utilizes short-term operational information at the initial stage of a fault to rapidly predict system voltage change trends and achieves rapid assessment of the transient voltage stability margin of the DC receiving-end system. Based on this invention, a data-driven approach based on artificial intelligence is implemented. By establishing a mapping relationship between system operating characteristics and transient voltage change trajectories, rapid prediction of system voltage change trends is achieved, and further, the system transient voltage stability margin is assessed. This enables rapid and stable analysis of transient voltage prediction and stability in DC receiving-end power systems containing renewable energy sources.

[0006] According to a first aspect of the present invention, a method for transient voltage prediction and stability assessment of a DC receiving-end system based on MCNN-BiLSTM-AM is proposed, comprising the following steps:

[0007] Step 1: Construct a receiving-end power system model including new energy generating units and DC transmission system, and generate system transient voltage stability analysis samples through time-domain simulation;

[0008] Step 2: Construct a multi-timescale convolutional bidirectional long short-term memory neural network transient voltage prediction model that incorporates an attention mechanism. The multi-timescale convolutional neural network is used to extract feature information at different time scales, the bidirectional long short-term memory network is used to learn the temporal dependencies in the voltage change process, and the key feature information is weighted by introducing an attention mechanism to predict the system voltage change trajectory.

[0009] Step 3: Preprocess and extract features from the generated sample data, divide it into training set, validation set and test set, train the transient voltage prediction model by defining a loss function, optimize the network parameters during training, and evaluate the model prediction performance by evaluation index to verify the model's ability to predict the transient voltage change trajectory of the system.

[0010] Step 4: Based on the trained transient voltage prediction model, predict the voltage change trajectory of key nodes in the system, and construct a transient voltage stability criterion table in conjunction with the power system safety guidelines and new energy grid connection technical specifications. By analyzing the system voltage recovery characteristics and voltage deviation, the transient voltage stability and stability margin of the DC receiving end system can be evaluated.

[0011] In a further embodiment, in step 1, time series samples are obtained by collecting system operation characteristics during the fault occurrence period and within a short time trajectory after the fault is cleared. Each sample consists of feature vectors from multiple sampling times. Specifically, one sampling point is selected at the steady state time before the fault, six sampling points are selected evenly during the fault duration, and five sampling points are selected evenly within a 0.1 s time range after the fault is cleared, thereby forming a time series sample containing 12 sampling times.

[0012] In a further embodiment, in step 1, when constructing the transient simulation model of the receiving-end power system containing new energy generating units and DC transmission system, the system operation mode is set as follows: the system load level varies within the range of 90% to 110% of the rated load, the output of new energy generating units varies within the range of 60% to 100%, and the system power balance is ensured by adjusting the active power output of synchronous generators, so that the system node voltage is maintained within the normal operating range of 0.95 to 1.05 pu;

[0013] The system disturbance settings are as follows: the fault type is a three-phase short-circuit fault of the transmission line, the fault location is set at 20%, 50% and 80% of the total line length, the fault duration is set at 0.15 s, 0.20 s and 0.25 s, respectively, and the system simulation time is set at 5 s; in order to highlight the system voltage stability characteristics and reduce the input dimension, the synchronous generator node, the new energy grid connection node, the DC feed-in node and the voltage weak node are selected as key monitoring nodes.

[0014] In a further embodiment, in step 2, the transient voltage prediction model consists of a multi-timescale convolutional feature extraction module, a bidirectional long short-term memory network module, an attention mechanism module, and a fully connected output layer. The predicted transient voltage change trajectory is output through the fully connected layer, wherein:

[0015] The multi-timescale convolutional feature extraction module uses a parallel multi-convolutional kernel structure to extract features from the input time series. The convolutional kernel sizes are 1, 3 and 5, respectively, to extract dynamic change features at different time scales. The features extracted by each convolutional branch are fused to form a unified multi-scale feature representation.

[0016] The bidirectional long short-term memory network module adopts a two-layer BiLSTM structure, with 128 neurons in the first BiLSTM hidden layer and 256 neurons in the second BiLSTM hidden layer, in order to extract temporal dependencies in time series data.

[0017] The attention mechanism module adopts a multi-head attention mechanism structure, with 4 attention heads set to adaptively allocate weights to the feature importance at different time steps.

[0018] Finally, the feature vector weighted by the attention mechanism is input into the fully connected layer, and the transient voltage prediction values ​​of the key nodes of the system within the future time window are output, thereby obtaining the prediction results of the transient voltage change trajectory of the system.

[0019] In a further embodiment, in step 3, the generated sample data is preprocessed and features are extracted, and the data is divided into training, validation, and test sets. The transient voltage prediction model is trained by defining a loss function, the network parameters are optimized during training, and the model's prediction performance is evaluated using evaluation metrics. This includes the following steps:

[0020] The original sample dataset is preprocessed, and the input feature vector is normalized to eliminate differences in the units of different features.

[0021] The sample dataset is divided into a training set, a validation set, and a test set, with the training set, validation set, and test set each having a ratio of 70%, 15%, and 15%, respectively.

[0022] During the model training phase, mean squared error (MSE) is used as the loss function. The overall prediction accuracy of the model is measured by calculating the error between the model's predicted value and the actual voltage trajectory. The model parameters are continuously optimized by minimizing the loss function.

[0023] During model training, the validation set is used to monitor model performance, and training is stopped when the validation set error no longer decreases. After model training is completed, the test set is used to evaluate the model's prediction performance.

[0024] In a further embodiment, in step 4, the voltage change trajectory of key system nodes is predicted based on the trained transient voltage prediction model. A transient voltage stability criterion table is constructed in conjunction with the power system safety guidelines and new energy grid connection technical specifications. By analyzing the system voltage recovery characteristics and voltage deviation, the transient voltage stability and stability margin of the DC receiving-end system are assessed. This includes the following processes:

[0025] Using a trained MCNN-BiLSTM-AM-based transient voltage prediction model, the voltage change process of key nodes in the system under fault disturbance conditions is predicted, and the voltage change trajectory of the system within a certain time range after the fault occurs is obtained. Then, feature extraction is performed on the predicted voltage trajectory to obtain the minimum voltage value of key nodes during the transient process and the key voltage features of the voltage recovery process.

[0026] After obtaining the voltage change trajectory of key system nodes, a binary stability criterion table based on voltage-duration is constructed according to the requirements for transient voltage stability operation of the power system. This binary stability criterion table uses voltage amplitude and corresponding permissible duration as two evaluation dimensions to describe the allowable voltage drop range and duration constraints under fault disturbances. The criterion table is constructed based on the new energy grid connection technology standards and the power system safety operation specifications. According to the requirements of the new energy grid connection technology standards and the constructed voltage-time boundary, during the voltage drop process, the voltage of key system nodes satisfies the following constraints: when the node voltage gradually decreases from 0.9 pu to 0.2 pu, its corresponding permissible duration gradually shortens as the voltage amplitude decreases; that is, the lower the node voltage, the shorter its permissible duration. Specifically, when the node voltage is 0.9 pu, the corresponding permissible duration does not exceed 2.0 s; when the node voltage is 0.2 pu, the corresponding permissible duration does not exceed 0.625 s. When the node voltage recovers to 0.9 pu or above, the system voltage should gradually recover to the rated voltage operating range.

[0027] Based on the aforementioned voltage-duration relationship, a binary criterion region for system transient voltage stability is constructed, forming the system transient voltage stability boundary curve. After completing the voltage stability criterion table construction, the predicted voltage trajectories of key system nodes are compared with the constructed voltage-time stability boundary. When the predicted voltage trajectory of the system lies above the stability boundary curve throughout the entire transient process, i.e., the system voltage does not fall below the allowable lower voltage limit under the corresponding time condition at any given time, the system is determined to be in a transient voltage stable state. If the predicted voltage trajectory of the system falls below the stability boundary curve for a certain period of time, the system is determined to be in a transient voltage unstable state.

[0028] In a further embodiment, after determining the system's stable state, a transient voltage stability margin index is introduced to evaluate the system's operating state:

[0029] Starting from the fault clearing moment, the difference between the predicted system voltage trajectory and the binary criterion stability boundary is integrated over the time interval from fault clearing to the end of the simulation to obtain the area difference between the system voltage trajectory and the stability boundary; this area difference reflects the safe distance of the system voltage trajectory relative to the stability boundary.

[0030] When the calculated area difference is positive, it is determined that the overall voltage trajectory of the system is above the stability boundary and the system has a transient voltage stability margin.

[0031] When the area difference is close to zero, the system is determined to be close to the transient voltage stability limit.

[0032] When the area difference is negative, it is determined that the system voltage trajectory has exceeded the stability boundary and the system is in a transient voltage instability state.

[0033] In conjunction with the above embodiments, the method proposed in this invention employs a binary stability criterion table to determine the transient voltage stability state of the system. The binary stability criterion table uses node voltage dip amplitude and voltage duration as two evaluation dimensions to describe the voltage dip characteristics and allowable duration range of the system under fault disturbances. Specifically, the voltage dip amplitude constraint is set according to the low-voltage ride-through requirements in the grid connection technical specifications for new energy generating units, ensuring that the new energy generating units can maintain grid connection operation during voltage dips. Simultaneously, a corresponding maximum allowable duration is set for different voltage levels to form a voltage-time stability constraint relationship. When the voltage trajectory of key system nodes satisfies the voltage-time constraint relationship in the binary stability criterion table during fault disturbances—that is, the minimum node voltage is not lower than the minimum allowable voltage value for the corresponding voltage level, and the voltage dip duration does not exceed the maximum allowable duration for the corresponding voltage level—then the system is considered to meet the transient voltage stability condition.

[0034] Furthermore, based on the steady-state determination, to further quantify the transient voltage stability of the system, this invention proposes a multi-binary table transient voltage safety margin assessment method that considers the low-voltage ride-through characteristics of new energy sources. Specifically, taking the fault clearing moment as the starting point, within the time range after fault clearing, the predicted voltage trajectory of key system nodes is scanned layer by layer using binary tables, and the voltage trajectory is compared with the stability boundary corresponding to each layer of binary stability criterion table to determine the distance relationship between the system voltage operating state and the stability boundary. When the system voltage trajectory is above the current binary stability boundary, it indicates that the system meets the stability constraint conditions of that layer, and the next layer of binary stability criteria is then evaluated. The system scans a criterion table; when the voltage trajectory reaches or falls below a certain stability boundary, subsequent scanning stops, and the system transient voltage safety margin is calculated based on the corresponding time interval. Through the above multi-binary table scanning calculation method, the system transient voltage safety margin index can be obtained. This index comprehensively reflects the impact of factors such as the system voltage drop depth, voltage recovery process, and low-voltage ride-through capability of new energy units on system stability. A large transient voltage safety margin indicates that the system has strong anti-disturbance capability and voltage support capability. When the safety margin gradually decreases, it indicates that the system operating state is gradually approaching the stability boundary, which can be used to quickly assess the system transient voltage stability.

[0035] Compared with existing technologies, the transient voltage prediction and stability assessment method for DC receiving-end systems integrating MCNN-BiLSTM-AM proposed in this invention has the following advantages and beneficial effects:

[0036] (1) This invention utilizes short-term operational data from the initial stage of a system fault to construct time-series samples, and employs a deep learning model combining multi-time-scale convolutional neural networks and bidirectional long short-term memory networks to predict the trajectory of transient voltage changes in the system. Compared with traditional transient stability assessment methods that rely on complete time-domain simulation calculations, this invention can rapidly predict the subsequent voltage change trend of the system using only short-term information from the initial stage of a fault, thereby significantly improving the computational efficiency of transient voltage stability assessment and meeting the speed requirements of online safety assessment of power systems;

[0037] (2) Based on the transient voltage prediction results, this invention, combined with the power system safety guidelines and the technical specifications for new energy grid connection, constructs a multi-binary meter transient voltage stability margin assessment method that takes into account the low-voltage ride-through characteristics of new energy, thereby achieving a quantitative assessment of the transient voltage stability level of the DC receiving-end system. Compared with existing methods that rely solely on a single voltage stability criterion, this invention comprehensively considers the low-voltage ride-through operation characteristics of new energy units and the system's safe operation requirements, improving the accuracy and engineering applicability of the transient voltage stability assessment results. Attached Figure Description

[0038] Figure 1 This is a flowchart of a method for predicting transient voltage and evaluating stability of a DC receiving-end system by fusing MCNN-BiLSTM-AM according to an embodiment of the present invention.

[0039] Figure 2 This is a schematic diagram of the predictive model structure according to an embodiment of the present invention.

[0040] Figure 3 This is a diagram of a BiLSTM structure according to an embodiment of the present invention.

[0041] Figure 4 This is a diagram showing the predicted transient voltage trajectory of a load node according to an embodiment of the present invention.

[0042] Figure 5 This is a graph showing the prediction results of rapid recovery of DC inverter-side bus voltage according to an embodiment of the present invention.

[0043] Figure 6 This is a graph showing the predicted results of the slow recovery of the DC inverter-side bus voltage according to an embodiment of the present invention.

[0044] Figure 7 This is a diagram showing the predicted transient voltage trajectory of a wind turbine according to an embodiment of the present invention. Detailed Implementation

[0045] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0046] Referring to the accompanying drawings, the method for transient voltage prediction and stability assessment of a DC receiving-end system based on MCNN-BiLSTM-AM according to an embodiment of the present invention includes the following steps:

[0047] A method for transient voltage prediction and stability assessment of DC receiving-end systems based on MCNN-BiLSTM-AM is characterized by the following steps:

[0048] Step 1: Construct a receiving-end power system model including new energy generating units and DC transmission system, and generate system transient voltage stability analysis samples through time-domain simulation;

[0049] Step 2: Construct a multi-timescale convolutional bidirectional long short-term memory neural network transient voltage prediction model that incorporates an attention mechanism. The multi-timescale convolutional neural network is used to extract feature information at different time scales, the bidirectional long short-term memory network is used to learn the temporal dependencies in the voltage change process, and the key feature information is weighted by introducing an attention mechanism to predict the system voltage change trajectory.

[0050] Step 3: Preprocess and extract features from the generated sample data, divide it into training set, validation set and test set, train the transient voltage prediction model by defining a loss function, optimize the network parameters during training, and evaluate the model prediction performance by evaluation index to verify the model's ability to predict the transient voltage change trajectory of the system.

[0051] Step 4: Based on the trained transient voltage prediction model, predict the voltage change trajectory of key nodes in the system, and construct a transient voltage stability criterion table in conjunction with the power system safety guidelines and new energy grid connection technical specifications. By analyzing the system voltage recovery characteristics and voltage deviation, the transient voltage stability and stability margin of the DC receiving end system can be evaluated.

[0052] As an optional implementation, in step 1, a receiving-end power system model including new energy generating units and DC transmission systems is constructed, and transient voltage stability analysis samples of the system are generated through time-domain simulation. Specifically, the following process is included:

[0053] Construct a receiving-end power system model that includes new energy generating units and DC transmission system, obtain system topology, line parameters, load model, induction motor parameters, DC transmission parameters and wind turbine parameters, and set different new energy output levels as well as different fault types and fault locations;

[0054] Transient operating data of the system is generated by time-domain simulation, and the system operating characteristics during the fault occurrence and short time trajectory after fault clearance are collected. Transient voltage stability analysis samples are constructed by extracting key electrical quantities. Considering that new energy power sources and DC transmission systems have a significant impact on the transient voltage stability of the system, operating variables that can reflect the key dynamic characteristics of the system are selected as input features in the process of sample feature construction. The model input data and output data are defined to form a sample dataset.

[0055] In the embodiments of the present invention, a transient simulation model of the receiving-end power system containing new energy generating units and DC transmission systems is constructed. Time-domain simulation calculations are carried out by combining different operating modes and typical fault scenarios, and key operating features of the system are extracted to construct a transient voltage stability sample dataset.

[0056] It should be understood that the power system model includes a typical AC network topology model, a synchronous generator model, a load model, a new energy generator model, and a DC transmission system model. The AC network topology model is used to describe the network topology structure and electrical parameters of the system bus, transmission lines, and transformers. The synchronous generator model is used to reflect the dynamic response characteristics of conventional power sources during transient processes. The load model adopts a comprehensive load model that includes static ZIP loads and dynamic loads of induction motors. The new energy generator model is used to describe the operating characteristics of wind turbines and take into account their low voltage ride-through capability. The DC transmission system model is used to describe the coupling relationship between converter stations and DC lines and the AC system.

[0057] After system modeling is completed, transient simulation calculations are conducted by setting different load levels, renewable energy output levels, and typical transmission line fault scenarios. Dynamic operating data of key system nodes and equipment are extracted from the simulation results. To highlight system voltage stability characteristics and reduce input dimensionality, in this embodiment, synchronous generator nodes, renewable energy grid-connected nodes, DC feed-in nodes, and voltage-weak nodes are selected as key monitoring nodes. Furthermore, multiple sampling times are selected within a short timeframe before the fault occurs, during the fault duration, and after the fault is cleared to construct a time series sample of system operating characteristics.

[0058] As an optional embodiment, in step 1, the model input data is represented by the input feature variable X:

[0059] X=[X g X w X DC X L X IM ];

[0060] Among them, X g Indicates the operating state characteristics of a synchronous generator; X wIndicates the operating characteristics of new energy generator sets; X DC Indicates the operating characteristics of a DC transmission system; X L Indicates static load characteristics; X IM Indicates the load characteristics of the induction motor;

[0061] The synchronous generator operating status characteristic X g Represented as: X g =[X g1 ,X g2 ,X g3 , ..., X gn ]; where X is the eigenvector of the i-th synchronous generator. gi Represented as: X gi =[ δ i , ω i , P Gi U Gi In the formula, δ i ω represents the power angle of the i-th synchronous generator; i P represents the rotor speed of the i-th synchronous generator; Gi U represents the active power output of the i-th synchronous generator; Gi represents the terminal voltage of the i-th synchronous generator; n represents the number of synchronous generators in the system.

[0062] The operating status characteristic X of the new energy generator set w Represented as: X w = [X w1 , X w2 , X w3 , ..., X wm ]; where X is the feature vector of the i-th new energy unit. wi For: X wi =[P WFi Q WFi U WFi In the formula, P WFi Q represents the active power output of the i-th renewable energy unit; WFi Indicates the first Each new energy unit outputs reactive power; U WFi This indicates the grid connection voltage of the new energy generating unit; m indicates the number of new energy generating units in the system.

[0063] Among them, the new energy unit is modeled using a doubly-fed induction generator model, and its key electrical parameters include stator resistance R. s Rotor resistance R r Stator reactance X s Rotor reactance X r and excitation reactance X m The rated power of the wind turbine is PN The rated voltage is U N The operating characteristics X of the DC transmission system DC Represented as: X DC =[ U dci U dcj , P dci ,P dcj Q dci Q dcj ,α i , γ j ]; Among them, U dci U represents the DC voltage on the rectifier side. dcj P represents the DC voltage on the inverter side. dci P represents the DC transmission power on the rectifier side. dcj Indicates the DC transmission power on the inverter side; Q dci This indicates the reactive power demand of the rectifier-side converter station; Q dcj Indicates the reactive power demand of the inverter-side converter station; α i , indicates the firing angle on the rectifier side; γ j Indicates the inverter-side arc extinction angle;

[0064] The system load is described using a comprehensive load model, including a static ZIP load model and an induction motor dynamic load model. The ZIP model consists of three parts: constant impedance, constant current, and constant power, used to describe the static characteristics X of the load. L The induction motor model is used to reflect the dynamic behavior of the load under voltage disturbances and to describe the load characteristics X of the induction motor. IM .

[0065] Specifically, in step 1, time series samples are obtained by collecting system operation characteristics during the fault occurrence period and within a short time trajectory after fault clearance. Each sample consists of feature vectors from multiple sampling times (sampling points). In this example, the number of sampling points T is chosen to be 12. Specifically, one sampling point is selected at the steady-state moment before the fault, six sampling points are evenly selected during the fault duration, and five sampling points are evenly selected within a 0.1 s time range after fault clearance, thereby forming a time series sample containing 12 sampling times.

[0066] In step 1, when constructing the transient simulation model of the receiving-end power system containing new energy generating units and DC transmission system, the system operation mode is set as follows: the system load level varies within the range of 90% to 110% of the rated load, the output of new energy generating units varies within the range of 60% to 100%, and the system power balance is ensured by adjusting the active power output of synchronous generators, so that the system node voltage is maintained within the normal operating range of 0.95 to 1.05 pu.

[0067] The system disturbance settings are as follows: the fault type is a three-phase short-circuit fault of the transmission line, the fault location is set at 20%, 50% and 80% of the total line length, the fault duration is set at 0.15 s, 0.20 s and 0.25 s, respectively, and the system simulation time is set at 5 s; in order to highlight the system voltage stability characteristics and reduce the input dimension, the synchronous generator node, the new energy grid connection node, the DC feed-in node and the voltage weak node are selected as key monitoring nodes.

[0068] Based on this, a system operation feature time series composed of multiple sampling times is constructed as the model input sample, and the transient voltage change trajectory of key system nodes is used as the model prediction target; at the same time, in terms of sample label determination, transient stability criteria are used to classify the samples.

[0069] As an optional example, the process of classifying samples includes:

[0070] Define the voltage at the critical node of the system as V. i (t), where i is the critical node number and t is the simulation time;

[0071] Define the minimum system voltage as V min =min( V i (t));

[0072] When a system fault or disturbance occurs, for non-new energy nodes, if V is satisfied... min If the voltage is less than 0.75 pu and the duration of this state is greater than 1 s, the system is determined to have experienced transient voltage instability; otherwise, the system is determined to be in a transient voltage stable state.

[0073] This allows us to obtain a transient stable sample dataset containing input features and stable labels, providing a data foundation for training subsequent transient voltage prediction models.

[0074] As an optional implementation, in step 2, the transient voltage prediction model consists of a multi-timescale convolutional feature extraction module, a bidirectional long short-term memory network module, an attention mechanism module, and a fully connected output layer. The predicted transient voltage change trajectory is output through the fully connected layer, wherein:

[0075] The multi-timescale convolutional feature extraction module uses a parallel multi-convolutional kernel structure to extract features from the input time series. The convolutional kernel sizes are 1, 3 and 5, respectively, to extract dynamic change features at different time scales. The features extracted by each convolutional branch are fused to form a unified multi-scale feature representation.

[0076] The bidirectional long short-term memory network module adopts a two-layer BiLSTM structure, with 128 neurons in the first BiLSTM hidden layer and 256 neurons in the second BiLSTM hidden layer, in order to extract temporal dependencies in time series data.

[0077] The attention mechanism module adopts a multi-head attention mechanism structure, with 4 attention heads set to adaptively allocate weights to the feature importance at different time steps.

[0078] Finally, the feature vector weighted by the attention mechanism is input into the fully connected layer, and the transient voltage prediction values ​​of the key nodes of the system within the future time window are output, thereby obtaining the prediction results of the transient voltage change trajectory of the system.

[0079] Specifically, as an optional example, combining Figure 2 The transient voltage prediction model structure based on MCNN-BiLSTM-AM is shown. This model integrates MCNN, BiLSTM, and AM to establish the mapping relationship between system operating characteristics and the transient voltage change trajectory after a fault. The transient voltage prediction model mainly consists of an MCNN module, a BiLSTM module, and an AM module, and outputs the predicted transient voltage change trajectory through a fully connected layer. The one-dimensional convolutional layer (Conv1D) is mainly used for time series feature extraction, capturing local change features in the sequence through convolution operations. The kernel size is represented by k, and different kernel sizes can extract features at different scales. A batch normalization (BN) layer is typically added after the convolutional layer. This layer standardizes the output of each layer, making the data distribution more stable, thereby accelerating model training and improving model convergence performance. Finally, a rectified linear activation function (Rectified Linear Activation Function) is used. The ReLU unit performs nonlinear mapping on the features; in the multi-scale convolutional structure, the features extracted by different convolutional layers need to be fused. Therefore, the feature concatenation operation (Concatenate, Concat) is used to concatenate multiple feature vectors along a specific dimension to form a richer feature representation, providing more comprehensive input information for subsequent network layers; to further illustrate the network composition and parameter settings of the constructed MCNN-BiLSTM-AM transient voltage prediction model, Table 1 gives the type, specific parameters and output feature map size of each network layer.

[0080] Table 1. Parameters of MCNN-BiLSTM-AM Network

[0081] Network layer type Specific parameters Output feature map size Input layer The input is a multivariate transient feature sequence within a short time window before, during, and after the fault is cleared. 12×D One-dimensional convolutional layer branch 1 The number of convolution kernels is 64, the kernel size k=1, and the stride is 1. 12×64 One-dimensional convolutional layer branch 2 The number of convolution kernels is 64, the kernel size k=3, and the stride is 1. 12×64 One-dimensional convolutional layer branch 3 The number of convolution kernels is 64, the kernel size k=5, and the stride is 1. 12×64 Batch Normalization Layer Batch normalization is performed on the output of the convolutional layer. 12×64 ReLU activation layer Nonlinear mapping using a linear rectified activation function 12×64 Feature splicing layer The outputs of the three convolutional branches are concatenated along the feature dimension. 12×192 BiLSTM Layer 1 A bidirectional long short-term memory network with 128 hidden units, using tanh as the internal activation function, returns a complete temporal output. 12×256 BiLSTM Layer 2 A bidirectional long short-term memory network with 256 hidden units, using tanh as the internal activation function, returns a complete temporal output. 12×512 Multi-head attention layer The attention heads are 4, and weighted allocation is performed on the bidirectional temporal features. 12×512 Fully connected layer 128 neurons, ReLU activation function 1×128 Output layer Output 200 transient voltage sampling points, using linear output. 1×200

[0082] In the table, D represents the dimension of the input features at each time step; since the number of final input feature variables can vary with the selection of nodes, it is retained as symbol D here; the size of the output feature map is expressed as "time step × feature dimension".

[0083] The input layer receives multivariate transient feature sequences within a short time window before and after the fault. The MCNN module extracts multi-scale local temporal features through three one-dimensional convolutional branches with different kernel sizes. After feature concatenation, the features are input into two layers of BiLSTM network to model bidirectional temporal dependencies. Then, the key features are highlighted through the attention mechanism. Finally, the predicted transient voltage change trajectory is output by the fully connected layer.

[0084] When the time series samples obtained in step (1) are input into the MCNN module, since the transient process of the power system has obvious multi-time-scale dynamic characteristics and the change law of system variables under different time scales are different, a parallel multi-convolutional kernel structure is used to extract features from the input time series. In this embodiment, three different sizes of one-dimensional convolutional kernels are set, with kernel sizes of 1, 3 and 5 respectively. Among them, the convolutional kernel with size 1 is used to extract the local instantaneous features of the input sequence; the convolutional kernel with size 3 is used to capture the dynamic change features of the medium time scale; and the convolutional kernel with size 5 is used to extract the dynamic change law over a longer time range. After each convolutional branch performs parallel convolution calculation on the input feature sequence, it obtains the feature representation under different receptive fields, and forms a unified multi-scale feature representation through feature splicing or fusion, thereby realizing multi-time-scale modeling of the transient dynamic features of the system.

[0085] The features output by the multi-scale convolution module are input into the bidirectional long short-term memory network module. Figure 3 (See BiLSTM structure diagram) to further extract temporal dependencies from time series data; the BiLSTM network consists of a forward long short-term memory network and a backward long short-term memory network, which can simultaneously utilize historical and future information from the time series to model the transient dynamic evolution process of the power system; in this embodiment, a two-layer BiLSTM network structure is set, with 128 neurons in the first layer and 256 neurons in the second layer, thereby improving the model's ability to express complex transient dynamic characteristics by extracting higher-level temporal features layer by layer.

[0086] An attention mechanism module is introduced after the BiLSTM module to adaptively assign weights to the feature importance at different time steps. By calculating the attention weights of features at each time step, the model can automatically focus on key features that have a significant impact on the transient voltage changes of the system, thereby improving the model's ability to extract key dynamic information. In this embodiment, a multi-head attention mechanism structure is adopted, with 4 attention heads. Multiple attention heads are used to calculate the feature weights of different subspaces in parallel to enhance the model's ability to express complex feature relationships.

[0087] The attention-weighted feature vector is input into the fully connected layer, and the predicted transient voltage values ​​of key system nodes in the future time period are output, thereby obtaining the prediction results of the transient voltage change trajectory of the system. Through the synergistic effect of multi-time-scale convolutional feature extraction, bidirectional temporal feature learning, and attention-weighted feature, the model's ability to learn the dynamic change law of transient voltage in the power system can be effectively improved, enabling rapid prediction of the transient voltage trajectory of the system after a fault, and providing a reliable data foundation for subsequent transient voltage stability margin assessment.

[0088] In a further embodiment, in step 3, the generated sample data is preprocessed and features are extracted, and the training set, validation set, and test set are divided. The model parameters are iteratively updated by defining a loss function and combining it with an optimization algorithm. The network parameters are optimized during training, and the model's prediction performance is evaluated using evaluation metrics, thereby obtaining a transient voltage prediction model with good prediction performance. This includes the following steps:

[0089] The original sample dataset is preprocessed, and the input feature vector is normalized to eliminate differences in the dimensions of different features. Since the sample data consists of the input feature sequence and the corresponding transient voltage change trajectory of the system, where the input features are time series feature vectors composed of the operating parameters of the key equipment of the system, and the output is the transient voltage change trajectory of the key nodes of the system after the fault; in order to avoid the impact of differences in the dimensions of different features on model training, the input feature data is normalized to map each feature variable to a uniform numerical range, thereby improving the stability and convergence speed of the model training process.

[0090] The sample dataset is divided into a training set, a validation set, and a test set, with the ratios of 70%, 15%, and 15% for the training set, validation set, and test set, respectively. The training set is used for learning model parameters, the validation set is used for parameter tuning and overfitting monitoring during model training, and the test set is used for evaluating the predictive performance after model training is completed.

[0091] During the model training phase, the training samples are input into the MCNN-BiLSTM-AM transient voltage prediction model constructed in step 2. The model extracts multi-scale feature information from the input data through a multi-time-scale convolutional neural network, learns the temporal dependencies in the time series data using a bidirectional long short-term memory network, and then assigns weights to features at different time steps through an attention mechanism, thereby outputting the transient voltage prediction values ​​of key system nodes within the future time window.

[0092] To measure the difference between the model's predicted results and the actual voltage trajectory, this embodiment uses the Mean Squared Error (MSE) as the loss function for model training. The average of the squared errors between the predicted and actual voltage values ​​is calculated to measure the overall prediction error of the model. The calculation form is as follows:

[0093] ;

[0094] Where N is the number of samples, V i This represents the true voltage value of the i-th sample. The voltage value is predicted by the model. By minimizing this loss function, the model parameters are continuously optimized, allowing the model to gradually learn the transient voltage change law of the system.

[0095] During the model parameter optimization process, the backpropagation algorithm based on gradient descent is used to update the network parameters. In this embodiment, the Adam optimization algorithm is used as the parameter optimization method for model training. By adaptively adjusting the learning rate, the model training efficiency is improved and the convergence speed is accelerated. During the training process, the model continuously updates the network weight parameters through multiple rounds of iterative learning, so that the loss function gradually decreases and eventually converges.

[0096] During model training, validation set samples are used to verify the model's performance. When the validation set error no longer decreases, model training can be stopped to avoid overfitting. After model training is completed, test set samples are used to evaluate the model's predictive performance. By comparing the transient voltage change trajectory of key nodes predicted by the model with the actual voltage trajectory calculated by simulation, and calculating the mean square error of the prediction results, the model's ability to predict the dynamic change process of transient voltage in the system is evaluated. The smaller the mean square error, the higher the accuracy of the model's prediction of the transient voltage change trend.

[0097] Through the above sample training and model evaluation process, a transient voltage prediction model with good predictive performance can be obtained, providing a reliable model for subsequent evaluation of the transient voltage stability margin of DC receiving-end systems based on predicted voltage trajectories.

[0098] Furthermore, in step 4, based on the trained transient voltage prediction model, the voltage change trajectory of key system nodes is predicted. A transient voltage stability criterion table is constructed in conjunction with the power system safety guidelines and new energy grid connection technical specifications. By analyzing the system voltage recovery characteristics and voltage deviation, the transient voltage stability and stability margin of the DC receiving-end system are assessed. This includes the following processes:

[0099] Using a trained MCNN-BiLSTM-AM-based transient voltage prediction model, the voltage change process of key nodes in the system under fault disturbance conditions is predicted, and the voltage change trajectory of the system within a certain time range after the fault occurs is obtained. Subsequently, feature extraction is performed on the predicted voltage trajectory to obtain key voltage features such as the minimum voltage value and voltage recovery time during the transient process of the system. Based on the guidelines for safe operation of power systems and the technical specifications for grid connection of new energy units, a transient voltage stability criterion for the system is constructed.

[0100] Specifically, using the transient voltage prediction model trained in step 3, the voltage dynamics of the system after a fault disturbance are predicted, and the transient voltage change trajectory of the key node within the prediction time window is obtained: V(t), t∈[t0, t... e ], t0, t e These represent the time of fault occurrence and the time of prediction end, respectively. V(t) represents the trajectory of the critical node voltage over time, which is used to describe the dynamic voltage response of the system under fault disturbance conditions.

[0101] To assess the transient voltage stability of the system, it is necessary to construct the permissible voltage dip range. This embodiment constructs a system voltage-duration stability boundary based on the grid connection technical specifications for new energy generating units and power system operation standards. The stability boundary mainly refers to GB / T 19963.1-2021; the above standard sets forth clear requirements for the low-voltage ride-through capability of new energy generating units, namely, when a voltage dip occurs in the system, the new energy generating units should maintain grid-connected operation within the specified voltage level and duration range. According to the above standard, a correspondence between system voltage amplitude and permissible duration can be established.

[0102] Table 2. Binary Table of Transient Voltage Constraints

[0103] Voltage amplitude (pu) Allowed duration (seconds) Voltage amplitude (pu) Allowed duration (seconds) 0.2 pu 0.6250s 0.6 pu 1.4107s 0.3 pu 0.8214s 0.7 pu 1.6071s 0.4 pu 1.0179s 0.8 pu 1.8036s 0.5 pu 1.2143s 0.9 pu 2.000s

[0104] Based on the voltage-time correspondence described above, the system voltage-time stability boundary curve V can be constructed. b (t), this boundary curve describes the minimum voltage level allowed by the system during the transient process; the predicted voltage trajectory V(t) and the stability boundary curve V bAfter V(t), the transient voltage stability of the system can be determined; when the predicted voltage trajectory of the system satisfies: V(t) ≥ V b When V(t) is within the allowable range, the voltage drop amplitude and duration are within acceptable limits, and the system is in a transient voltage stable state; when V(t) < V(t) occurs for a certain period of time... b When (t), it indicates that the voltage drop or duration of the system exceeds the allowable range, and the system is in a transient voltage instability state.

[0105] Based on the determination of the stable state, to further quantify the transient voltage stability of the system, this invention uses the area difference between the voltage trajectory and the stability boundary curve as the transient voltage stability margin index; the specific calculation method is as follows; taking the fault clearing time t... c Starting from [t], within the time interval [t] c , t e The area difference M between the predicted voltage trajectory and the voltage-time stability boundary curve is calculated internally.

[0106] ;

[0107] Where M is the system transient voltage stability margin; V(t) is the predicted voltage trajectory at the system's critical nodes; V b (t) represents the voltage-time stability boundary curve; t c Indicates the fault clearing time; t e To predict the end time.

[0108] When the calculated stability margin M is greater than 0, it indicates that the system voltage trajectory is located above the stability boundary curve, and the system has a certain transient voltage stability margin. When the stability margin M is close to zero, it indicates that the system is operating close to the stability boundary. When M < 0, it indicates that the system voltage trajectory is below the stability boundary curve, and the system may experience transient voltage instability.

[0109] The transient voltage stability margin index of the system can be obtained through the above calculation method. This index can comprehensively reflect the impact of factors such as voltage drop depth, voltage duration, voltage recovery process, and low voltage ride-through capability of new energy units on the system stability under fault disturbance conditions. When the transient voltage stability margin is large, it indicates that the system has strong anti-disturbance capability and voltage support capability. When the stability margin gradually decreases, it indicates that the system operating state is gradually approaching the stability boundary, and corresponding control measures should be taken to improve the system operating stability.

[0110] Therefore, a binary stability criterion table is constructed using node voltage dip amplitude and voltage duration as two evaluation dimensions to determine the transient voltage stability of the system. The voltage dip amplitude constraint is set according to the low voltage ride-through requirements in the grid connection technical specifications for new energy units to ensure that new energy units can maintain grid connection during voltage dips. At the same time, a corresponding maximum allowable duration is set for different voltage levels to form a voltage-time stability constraint relationship. When the voltage trajectory of the critical nodes of the system satisfies the voltage-time constraint relationship in the binary stability criterion table during fault disturbances, that is, when the minimum node voltage is not lower than the minimum allowable voltage value of the corresponding voltage level and the voltage dip duration does not exceed the maximum allowable duration of the corresponding voltage level, the system is considered to meet the transient voltage stability condition.

[0111] Specifically, after obtaining the voltage change trajectory of key system nodes, a voltage-duration binary stability criterion table is constructed according to the requirements for transient voltage stability operation of the power system. This binary stability criterion table uses voltage amplitude and corresponding permissible duration as two evaluation dimensions to describe the allowable voltage drop range and duration constraints under fault disturbances. The criterion table is constructed based on the new energy grid connection technology standards and the power system safety operation specifications. According to the requirements of the new energy grid connection technology standards and the constructed voltage-time boundary, specifically referring to the relevant requirements for voltage transient operation range in GB / T 19963.1—2021, during the voltage drop process, the voltage of key system nodes satisfies the following constraint relationship: when the node voltage gradually decreases from 0.9 pu to 0.2 pu, its corresponding permissible duration gradually shortens as the voltage amplitude decreases; that is, the lower the node voltage, the shorter its permissible duration. Specifically, when the node voltage is 0.9 pu, the corresponding permissible duration does not exceed 2.0 s; when the node voltage is 0.2 pu, the corresponding permissible duration does not exceed 0.625 s. When the node voltage recovers to 0.9 pu or above, the system voltage should gradually recover to the rated voltage operating range.

[0112] Furthermore, based on the aforementioned voltage-duration relationship, a binary criterion region for system transient voltage stability is constructed, forming a system transient voltage stability boundary curve. After completing the construction of the voltage stability criterion table, the predicted voltage trajectories of key system nodes are compared with the constructed voltage-time stability boundary. When the predicted voltage trajectory of the system lies above the stability boundary curve throughout the entire transient process, i.e., the system voltage does not fall below the allowable lower voltage limit under the corresponding time condition at any given time, the system is determined to be in a transient voltage stable state. If the predicted voltage trajectory of the system falls below the stability boundary curve for a certain period of time, the system is determined to be in a transient voltage unstable state.

[0113] In a further optional embodiment, after determining the system's stable state, a transient voltage stability margin index is introduced to evaluate the system's operating state in order to further quantify the degree of system transient voltage stability.

[0114] Starting from the fault clearing moment, the difference between the predicted system voltage trajectory and the binary criterion stability boundary is integrated over the time interval from fault clearing to the end of the simulation to obtain the area difference between the system voltage trajectory and the stability boundary; this area difference reflects the safe distance of the system voltage trajectory relative to the stability boundary.

[0115] When the calculated area difference is positive, it is determined that the overall voltage trajectory of the system is above the stability boundary and the system has a transient voltage stability margin.

[0116] When the area difference is close to zero, the system is determined to be close to the transient voltage stability limit.

[0117] When the area difference is negative, it is determined that the system voltage trajectory has exceeded the stability boundary and the system is in a transient voltage instability state.

[0118] Therefore, combining the above assessment of transient voltage safety margin using a multi-binary table that takes into account the low-voltage ride-through characteristics of new energy sources, taking the fault clearing moment as the starting point, within the time range after fault clearing, the predicted voltage trajectories of key system nodes are scanned layer by layer using binary tables. The voltage trajectories are then compared with the stability boundaries corresponding to each layer of binary stability criterion tables to determine the distance relationship between the system voltage operating state and the stability boundary. When the system voltage trajectories are above the current binary stability boundary, it indicates that the system meets the stability constraints of that layer, and the scanning of the next layer of binary stability criterion tables continues. When the voltage trajectories reach or fall below the current binary stability boundary, the system continues scanning. When a certain stability boundary is reached, subsequent scanning is stopped, and the system transient voltage safety margin is calculated based on the corresponding time interval. Through the above multi-binary table scanning calculation method, the system transient voltage safety margin index can be obtained. This index can comprehensively reflect the impact of factors such as the system voltage drop depth, voltage recovery process, and low voltage ride-through capability of new energy units on system stability. When the transient voltage safety margin is large, it indicates that the system has strong anti-disturbance capability and voltage support capability. When the safety margin gradually decreases, it indicates that the system operating state is gradually approaching the stability boundary, which can be used to achieve rapid assessment of the system transient voltage stability.

[0119] Below, we will use a specific example to test and verify the effectiveness of the proposed transient voltage prediction and stability margin assessment method for DC receiving-end systems by establishing a DC receiving-end power system simulation model on the PSASP simulation platform, taking the IEEE 39-bus system containing wind power as an example.

[0120] Step 1: Generate transient voltage stability analysis samples for DC receiving-end systems containing new energy sources.

[0121] This DC receiving-end power system simulation model is an extension and modification based on the IEEE 10-machine 39-bus standard example. The system includes 39 AC nodes, 10 synchronous generators, 46 transmission lines, and 12 transformers. The system's rated frequency is f=50Hz, and the rated voltage base is V_base=345kV. In this embodiment, a comprehensive load model is used, which includes static ZIP loads and dynamic load components from induction motors. Typical parameters of the induction motors are set as follows: rated power P GN =2MW, rated voltage U N =6kV, stator resistance R s =0.01pu, rotor resistance R r =0.015pu, stator reactance X s =0.15pu, rotor reactance X r =0.15pu, magnetizing reactance X m =3.0 pu.

[0122] To reflect the impact of different load dynamic characteristics on the system's transient voltage stability, different proportions of induction motors were set in the integrated load model, with proportions of 10%, 30%, and 50%, respectively. The wind turbine was modeled using a doubly-fed induction generator model, which can effectively reflect the dynamic response and reactive power regulation capability of the wind turbine during voltage disturbances. The rated capacity of the wind turbine is S. WF =200MW, rated voltage U WF =690V, connected to the grid bus via a step-up transformer; the wind turbine control system has low-voltage ride-through capability, maintaining grid-connected operation and providing some reactive power support when the grid voltage drops; the system includes a high-voltage direct current (HVDC) transmission line to transmit power to the receiving-end system; the DC system consists of a rectifier-side converter station, an inverter-side converter station, and a DC line; the rated transmission capacity of the DC transmission system is P. dc =1000MW, rated DC voltage U dc =±500kV; the rectifier-side converter station adopts constant current control mode, and the inverter-side converter station adopts constant arc extinction angle control mode to ensure the stable operation of the DC transmission system; the main control parameters are set as follows: rectifier-side firing angle α i =15°, inverter-side arc extinction angle γ j =18°; Simultaneously, a reactive power compensation device is configured on the AC side of the converter station, with a compensation capacity Q. c It accounts for approximately 40% of the DC transmission power to meet the reactive power requirements during converter station operation.

[0123] After completing the system simulation model construction, in order to generate data samples for model training and testing, it is necessary to perform a large number of time-domain simulation calculations on the system under different operating scenarios and fault disturbance conditions. During the sample generation process, transient operating samples of the system are constructed by setting different system operating modes and fault disturbance scenarios. Specifically, the system load level varies within the range of 90% to 110% of the rated load and is adjusted in steps of 5%; the wind turbine output varies within the range of 60% to 100% of the rated output and is adjusted in steps of 20%; under different operating modes, the output of the synchronous generator is adjusted accordingly to maintain the system power balance and ensure that the voltage of each node is kept within the normal operating range of 0.95 pu to 1.05 pu.

[0124] In terms of fault scenario setting, in order to simulate the impact of different fault disturbances on the transient voltage stability of the system, 33 transmission lines in the system were selected as fault lines in sequence, and fault points were set at 20%, 50% and 80% of the length of each line. The fault type was three-phase short circuit fault, and different fault durations were set, namely 0.1s, 0.15s and 0.2s.

[0125] By combining different operating modes with different fault scenarios, various transient disturbance operating conditions of the system were constructed, generating a total of 13,365 simulation samples. Time-domain simulation calculations were performed on each sample, with the simulation duration set to 5 seconds, and the voltage change trajectories of key system nodes during the transient process were recorded.

[0126] During the sample data acquisition process, the voltage change characteristics during the fault occurrence period and within a short time range after the fault is cleared are extracted as model input features to reflect the dynamic response behavior of the system in the early stage of the disturbance. At the same time, the complete voltage change trajectory of the key nodes of the system within a certain time range after the fault disturbance is used as model output data to describe the transient voltage recovery process of the system.

[0127] During the sample generation and data acquisition process, automated simulation scripts are written in MATLAB to call and control the PSASP software, enabling batch simulation calculations under multiple operating modes and multiple fault scenarios. Specifically, the MATLAB program automatically generates PSASP simulation input data files and calls PSASP to complete time-domain simulation calculations. At the same time, it automatically extracts and organizes simulation result data, thereby forming a large-scale transient voltage sample dataset for model training and testing.

[0128] Combination Figure 2Taking the constructed transient voltage prediction model structure as an example, the time series samples constructed in step 1 are input into the multi-timescale convolutional neural network module. A parallel multi-convolutional kernel structure is used to extract features from the input sequence. As mentioned above, the kernel sizes of the three different sizes of one-dimensional convolutional kernels are k1=1, k2=3, and k3=5, respectively, which are used to extract the instantaneous local features of the input sequence. This captures dynamic change features at medium time scales and extracts dynamic change patterns over longer time ranges. Each convolutional branch performs one-dimensional convolution operations on the input feature sequence to obtain feature maps. After the convolutional layers, BN and ReLU are added sequentially to standardize and nonlinearly map the features to improve the model training stability. Subsequently, feature concatenation is performed on the features output by each convolutional branch, thus forming a unified multi-scale feature representation and realizing multi-timescale modeling of the system's transient dynamic features.

[0129] The features output by the multi-timescale convolutional module are input into the BiLSTM module to further extract the temporal dependencies in the time series data. The BiLSTM network consists of a forward long short-term memory network and a backward long short-term memory network, which can simultaneously utilize historical and future information of the time series to model the transient dynamic evolution process of the power system. In this embodiment, a two-layer BiLSTM network structure is adopted, in which the number of neurons in the first layer of BiLSTM hidden layer is set to 128 and the number of neurons in the second layer of BiLSTM hidden layer is set to 256. By extracting higher-level temporal features layer by layer, the model's ability to express the complex transient dynamic characteristics of the system is enhanced.

[0130] An attention mechanism module is introduced after the BiLSTM module to adaptively assign weights to the importance of features at different time steps. By weighting the time series features output by BiLSTM through the attention mechanism, the model can automatically focus on key features that have a significant impact on the transient voltage changes of the system, thereby improving the model's ability to extract important dynamic information. In this embodiment, a multi-head attention mechanism structure is adopted, with the number of attention heads set to 4. The weight distribution of different feature subspaces is calculated in parallel by multiple attention heads to enhance the model's ability to express complex feature relationships.

[0131] The feature vectors weighted by the attention mechanism are input into the fully connected layer for mapping, and the transient voltage prediction values ​​of the key nodes of the system in the future time period are output, thereby obtaining the predicted results of the transient voltage change trajectory of the key nodes of the system.

[0132] During model training, the 13,365 sample data generated in step 1 are divided into training and test sets according to a certain ratio. 70% of the samples are used for model training, and 30% of the samples are used for model performance verification. The MCNN-BiLSTM-AM transient voltage prediction model is iteratively trained using the training set data, so that the model gradually learns the mapping relationship between system operating characteristics and transient voltage change trajectory.

[0133] During model training, mean squared error was used as the model loss function, and the Adam optimization algorithm was employed to optimize and update the network parameters. The training process was set with a learning rate of 0.001, a batch size of 64, and a maximum training epoch of 100. The network weight parameters were continuously adjusted using backpropagation to gradually reduce the error between the predicted voltage trajectory and the simulated voltage trajectory. Once the model training was complete, the trained MCNN-BiLSTM-AM transient voltage prediction model was obtained. This model can predict the voltage changes of key system nodes after a fault based on system operating state characteristics and fault disturbance information, providing a predictive voltage data foundation for subsequent transient voltage stability assessment and safety margin calculation.

[0134] Finally, after obtaining the transient voltage prediction model trained in step 3, the voltage change trajectory of key nodes in the system is predicted based on the system operating state input samples, and the transient voltage stability state and stability margin of the DC receiving-end system are evaluated based on the predicted voltage trajectory. Using the MCNN-BiLSTM-AM transient voltage prediction model constructed and trained in steps 2 and 3, the voltage change process of key nodes in the system under fault disturbance conditions is predicted, and the voltage change trajectory of the system within a certain time range after the fault occurs is obtained. Subsequently, feature extraction is performed on the predicted voltage trajectory to obtain the minimum voltage value of key nodes in the transient process of the system and key voltage features such as the voltage recovery process.

[0135] After obtaining the voltage change trajectories of key system nodes, a voltage-duration binary stability criterion table is constructed based on the requirements for transient voltage stability operation of the power system. The binary criterion table uses voltage amplitude and corresponding permissible duration as two evaluation dimensions to describe the allowable voltage drop range and duration constraints under fault disturbances. The criterion table is constructed based on new energy grid connection technology standards and power system safety operation specifications, specifically referring to the relevant requirements for voltage transient operation range in GB / T 19963.1—2021. Based on the aforementioned standards and the voltage-time boundary curve adopted in this paper, the system's critical node voltages should meet the following constraints during voltage dips: when the node voltage drops to 0.2 pu, its duration should not exceed 0.625 s; thereafter, the node voltage should gradually rise along the recovery boundary, and at 0.3 pu, 0.4 pu, 0.5 pu, 0.6 pu, 0.7 pu, 0.8 pu, and 0.9 pu, the corresponding allowable durations should not exceed 0.8214 s, 1.0179 s, 1.2143 s, 1.4107 s, 1.6071 s, 1.8036 s, and 2.0 s, respectively; when the node voltage recovers to 0.9 pu or above, the system voltage should gradually recover to the rated voltage operating range.

[0136] Based on the above analysis and stability margin assessment, in a DC receiving-end system, the inverter-side bus voltage is significantly affected by the fault location and severity. When a near-end fault or a minor fault with a short duration occurs, the inverter-side bus voltage can usually recover quickly after the fault is cleared. However, when a far-end fault or a more severe fault occurs, the DC system recovery process requires a large amount of reactive power. If reactive power support is insufficient during the recovery process, the voltage recovery process will be slow, or even the voltage may fail to recover. Furthermore, in AC-DC coupled systems, voltage instability is often accompanied by power angle stability issues, and the system may experience voltage oscillations or even oscillatory instability.

[0137] Therefore, when considering different recovery modes of transient voltage in a DC receiving-end system, the samples are divided into two categories based on whether there is obvious oscillation during the voltage recovery process: voltage recovery samples without oscillation and voltage recovery samples with oscillation. Among them, voltage recovery samples without oscillation include voltage recovery samples with rapid voltage recovery and voltage recovery samples with slow voltage recovery. The above classification is mainly based on the voltage recovery mode for preliminary classification, while the final stable state of the system is determined based on whether the duration of the system bus voltage being lower than 0.75 pu exceeds 1 second. When the duration is greater than 1 second, the system corresponding to the sample is considered to be in a voltage instability state.

[0138] For samples exhibiting significant oscillation characteristics during voltage recovery, a first-stage transient power angle stability assessment model is used to screen for power angle stability and instability scenarios. Based on this, samples determined to be power angle stable are further subjected to transient voltage prediction and stability assessment. For stable samples with rapid voltage recovery, the predicted voltage curve highly coincides with the actual voltage trajectory obtained from PSASP time-domain simulation, indicating that the constructed MCNN-BiLSTM-AM transient voltage prediction model can accurately predict relatively smooth voltage time series. When the node with the smallest transient voltage stability margin is a load node, the predicted transient voltage trajectory is as follows: Figure 4 As shown;

[0139] Further analysis was conducted using the DC inverter-side bus as a key observation node. When a fault occurred on the far end of the DC inverter-side line, the inverter-side bus voltage recovered rapidly after the fault was cleared. The predicted transient voltage trajectory is shown below. Figure 5 As shown; when a three-phase short-circuit fault occurs on lines 23-24, i.e., the lines near the DC inverter side, the inverter bus voltage exhibits a decreasing and then slowly recovering process after the fault is cleared. The prediction results for its instability samples are as follows. Figure 6 As shown in the figure; simulation results show that for this type of voltage time series where the voltage first decreases and then gradually recovers, the proposed model can also track its changing trend well and make effective predictions.

[0140] Besides the voltage at critical load nodes and the DC inverter side bus, renewable energy units typically operate in a low-voltage ride-through state when the system voltage is low during or after a fault. Therefore, the voltage at renewable energy nodes also requires close monitoring. Taking the wind turbine terminal voltage as an example, its transient voltage trajectory prediction results are as follows: Figure 7 As shown in the figure; the simulation results show that the voltage change trend predicted by the model is basically consistent with the simulation results, and can accurately reflect the voltage change characteristics of the wind turbine during the low voltage ride-through process.

[0141] To further verify the effectiveness of the proposed transient stability assessment method, the assessment performance of the two-stage transient stability assessment model was statistically analyzed. The results show that the accuracy of the first-stage transient power angle stability assessment model is 97.16%, the misclassification rate is 0.52%, and the omission rate is 2.32%. In the second stage, transient voltage prediction and stability assessment were further performed on samples judged to be stable in power angle. The mean square error of the transient voltage trajectory prediction at the node with the most severe voltage drop was 0.0033. The accuracy of transient voltage stability assessment in this stage was 96.6%, the misclassification rate was 1.12%, and the omission rate was 2.32%.

[0142] Based on the combined results of the two-stage evaluation, the overall accuracy of the two-stage transient stability evaluation model reached 98.65%. Specifically, the misclassification rate of power angle stability being judged as power angle instability was 0.07%, and the missed judgment rate of power angle instability being judged as stability was 0.07%. The misclassification rate of voltage stability being judged as instability was 0.29%, and the missed judgment rate of voltage instability being judged as stability was 1.19%. As can be seen from the above results, the two-stage evaluation model improves the evaluation accuracy by approximately 2.05% compared to the single-stage transient voltage stability evaluation model, verifying the effectiveness and accuracy of the method of the present invention in transient voltage prediction and stability margin evaluation.

[0143] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A method for transient voltage prediction and stability assessment of a DC receiving-end system based on MCNN-BiLSTM-AM, characterized in that, Includes the following steps: Step 1: Construct a receiving-end power system model including new energy generating units and DC transmission system, and generate system transient voltage stability analysis samples through time-domain simulation; Step 2: Construct a multi-timescale convolutional bidirectional long short-term memory neural network transient voltage prediction model that incorporates an attention mechanism. The multi-timescale convolutional neural network is used to extract feature information at different time scales, the bidirectional long short-term memory network is used to learn the temporal dependencies in the voltage change process, and the key feature information is weighted by introducing an attention mechanism to predict the system voltage change trajectory. Step 3: Preprocess and extract features from the generated sample data, divide it into training set, validation set and test set, train the transient voltage prediction model by defining a loss function, optimize the network parameters during training, and evaluate the model prediction performance by evaluation index to verify the model's ability to predict the transient voltage change trajectory of the system. Step 4: Based on the trained transient voltage prediction model, predict the voltage change trajectory of key nodes in the system, and construct a transient voltage stability criterion table in conjunction with the power system safety guidelines and new energy grid connection technical specifications. By analyzing the system voltage recovery characteristics and voltage deviation, the transient voltage stability and stability margin of the DC receiving end system can be evaluated.

2. The method for transient voltage prediction and stability assessment of a DC receiving-end system based on MCNN-BiLSTM-AM according to claim 1, characterized in that, In step 1, a receiving-end power system model including new energy generating units and DC transmission systems is constructed. Time-domain simulation is used to generate transient voltage stability analysis samples for the system. This process includes the following steps: Construct a receiving-end power system model that includes new energy generating units and DC transmission system, obtain system topology, line parameters, load model, induction motor parameters, DC transmission parameters and wind turbine parameters, and set different new energy output levels as well as different fault types and fault locations; Transient operating data of the system is generated by time-domain simulation, and the system operating characteristics during the fault occurrence and short time trajectory after the fault is cleared are collected. Transient voltage stability analysis samples are constructed by extracting key electrical quantities, and the model input data and output data are defined to form a sample dataset.

3. The method for transient voltage prediction and stability assessment of DC receiving-end systems based on MCNN-BiLSTM-AM according to claim 2, characterized in that, In step 1, the model input data is represented by the input feature variable X: X=[X g ,X w ,X DC ,X L ,X IM ]; Among them, X g Indicates the operating state characteristics of a synchronous generator; X w Indicates the operating characteristics of new energy generator sets; X DC Indicates the operating characteristics of a DC transmission system; X L Indicates static load characteristics; X IM Indicates the load characteristics of the induction motor; The synchronous generator operating status characteristic X g Represented as: X g =[X g1 ,X g2 ,X g3 , ..., X gn ]; where X is the eigenvector of the i-th synchronous generator. gi Represented as: X gi =[ δ i , ω i , P Gi U Gi In the formula, δ i ω represents the power angle of the i-th synchronous generator; i P represents the rotor speed of the i-th synchronous generator; Gi U represents the active power output of the i-th synchronous generator; Gi represents the terminal voltage of the i-th synchronous generator; n represents the number of synchronous generators in the system. The operating status characteristic X of the new energy generator set w Represented as: X w = [X w1 X w2 X w3 , ..., X wm ]; where X is the feature vector of the i-th new energy unit. wi For: X wi =[P WFi Q WFi U WFi In the formula, P WFi Q represents the active power output of the i-th renewable energy unit; WFi Indicates the first Each new energy unit outputs reactive power; U WFi This indicates the grid connection voltage of the new energy generating unit; m indicates the number of new energy generating units in the system. Among them, the new energy unit is modeled using a doubly-fed induction generator model, and its key electrical parameters include stator resistance R. s Rotor resistance R r Stator reactance X s Rotor reactance X r and excitation reactance X m The rated power of the wind turbine is P N The rated voltage is U N The operating characteristics X of the DC transmission system DC Represented as: X DC =[ U dci U dcj , P dci ,P dcj Q dci Q dcj , α i ,γ j ]; Among them, U dci U represents the DC voltage on the rectifier side. dcj P represents the DC voltage on the inverter side. dci Indicates the DC transmission power on the rectifier side; P dcj Indicates the DC transmission power on the inverter side; Q dci This indicates the reactive power demand of the rectifier-side converter station; Q dcj Indicates the reactive power demand of the inverter-side converter station; α i , indicates the firing angle on the rectifier side; γ j Indicates the inverter-side arc extinction angle; The system load is described using a comprehensive load model, including a static ZIP load model and an induction motor dynamic load model. The ZIP model consists of three parts: constant impedance, constant current, and constant power, used to describe the static characteristics X of the load. L The induction motor model is used to reflect the dynamic behavior of the load under voltage disturbances and to describe the load characteristics X of the induction motor. IM .

4. The method for transient voltage prediction and stability assessment of DC receiving-end system based on MCNN-BiLSTM-AM according to claim 2, characterized in that, In step 1, time series samples are obtained by collecting system operation characteristics during the fault occurrence period and within a short time trajectory after the fault is cleared. Each sample consists of feature vectors from multiple sampling times. Specifically, one sampling point is selected at the steady state before the fault, six sampling points are selected evenly during the fault duration, and five sampling points are selected evenly within a 0.1 s time range after the fault is cleared, thus forming a time series sample containing 12 sampling times.

5. The method for transient voltage prediction and stability assessment of a DC receiving-end system based on MCNN-BiLSTM-AM according to claim 1, characterized in that, In step 1, when constructing the transient simulation model of the receiving-end power system containing new energy generating units and DC transmission system, the system operation mode is set as follows: the system load level varies within the range of 90% to 110% of the rated load, the output of new energy generating units varies within the range of 60% to 100%, and the system power balance is ensured by adjusting the active power output of synchronous generators, so that the system node voltage is maintained within the normal operating range of 0.95 to 1.05 pu. The system disturbance settings are as follows: the fault type is a three-phase short-circuit fault of the transmission line, the fault location is set at 20%, 50% and 80% of the total line length, the fault duration is set at 0.15 s, 0.20 s and 0.25 s, respectively, and the system simulation time is set at 5 s; in order to highlight the system voltage stability characteristics and reduce the input dimension, the synchronous generator node, the new energy grid connection node, the DC feed-in node and the voltage weak node are selected as key monitoring nodes.

6. The method for transient voltage prediction and stability assessment of a DC receiving-end system based on MCNN-BiLSTM-AM according to claim 1, characterized in that, In step 1, the transient stability criterion is used to classify the samples: Define the voltage at the critical node of the system as V. i (t), where i is the critical node number and t is the simulation time; Define the minimum system voltage as V min =min( V i (t)); When a system fault or disturbance occurs, for non-new energy nodes, if V is satisfied... min If the voltage is less than 0.75 pu and the duration of this state is greater than 1 s, the system is determined to have experienced transient voltage instability; otherwise, the system is determined to be in a transient voltage stable state.

7. The method for transient voltage prediction and stability assessment of a DC receiving-end system based on MCNN-BiLSTM-AM according to any one of claims 1-6, characterized in that, In step 2, the transient voltage prediction model consists of a multi-timescale convolutional feature extraction module, a bidirectional long short-term memory network module, an attention mechanism module, and a fully connected output layer. The predicted transient voltage change trajectory is output through the fully connected layer, wherein: The multi-timescale convolutional feature extraction module uses a parallel multi-convolutional kernel structure to extract features from the input time series. The convolutional kernel sizes are 1, 3 and 5, respectively, to extract dynamic change features at different time scales. The features extracted by each convolutional branch are fused to form a unified multi-scale feature representation. The bidirectional long short-term memory network module adopts a two-layer BiLSTM structure, with 128 neurons in the first BiLSTM hidden layer and 256 neurons in the second BiLSTM hidden layer, in order to extract temporal dependencies in time series data. The attention mechanism module adopts a multi-head attention mechanism structure, with 4 attention heads set to adaptively allocate weights to the feature importance at different time steps. Finally, the feature vector weighted by the attention mechanism is input into the fully connected layer, and the transient voltage prediction values ​​of the key nodes of the system within the future time window are output, thereby obtaining the prediction results of the transient voltage change trajectory of the system.

8. The method for transient voltage prediction and stability assessment of a DC receiving-end system based on MCNN-BiLSTM-AM according to claim 1, characterized in that, In step 3, the generated sample data is preprocessed and features are extracted. The dataset is divided into training, validation, and test sets. The transient voltage prediction model is trained using a defined loss function. During training, network parameters are optimized, and the model's prediction performance is evaluated using evaluation metrics. This includes the following steps: The original sample dataset is preprocessed, and the input feature vector is normalized to eliminate differences in the units of different features. The sample dataset is divided into a training set, a validation set, and a test set, with the training set, validation set, and test set each having a ratio of 70%, 15%, and 15%, respectively. During the model training phase, mean squared error (MSE) is used as the loss function. The overall prediction accuracy of the model is measured by calculating the error between the model's predicted value and the actual voltage trajectory. The model parameters are continuously optimized by minimizing the loss function. During model training, the validation set is used to monitor model performance, and training is stopped when the validation set error no longer decreases. After model training is completed, the test set is used to evaluate the model's prediction performance.

9. The method for transient voltage prediction and stability assessment of a DC receiving-end system based on MCNN-BiLSTM-AM according to claim 1, characterized in that, In step 4, based on the trained transient voltage prediction model, the voltage change trajectory of key system nodes is predicted. A transient voltage stability criterion table is constructed in conjunction with power system safety guidelines and new energy grid connection technical specifications. By analyzing the system voltage recovery characteristics and voltage deviation, the transient voltage stability and stability margin of the DC receiving-end system are assessed. This includes the following processes: Using a trained MCNN-BiLSTM-AM-based transient voltage prediction model, the voltage change process of key nodes in the system under fault disturbance conditions is predicted, and the voltage change trajectory of the system within a certain time range after the fault occurs is obtained. Then, feature extraction is performed on the predicted voltage trajectory to obtain the minimum voltage value of key nodes during the transient process and the key voltage features of the voltage recovery process. After obtaining the voltage change trajectory of the key nodes of the system, a voltage-duration binary stability criterion table is constructed according to the requirements of transient voltage stability operation of the power system. The binary stability criterion table uses voltage amplitude and corresponding permissible duration as two evaluation dimensions to describe the allowable voltage drop range and duration constraints of the system under fault disturbances. The criterion table is constructed based on the new energy grid connection technology standards and the power system safety operation specifications. According to the requirements of the new energy grid connection technology standards and the constructed voltage-time boundary, during the voltage drop process, the voltage of key system nodes satisfies the following constraints: when the node voltage gradually decreases from 0.9 pu to 0.2 pu, its corresponding permissible duration gradually shortens as the voltage amplitude decreases; that is, the lower the node voltage, the shorter its permissible duration. Specifically, when the node voltage is 0.9 pu, the corresponding permissible duration does not exceed 2.0 s; when the node voltage is 0.2 pu, the corresponding permissible duration does not exceed 0.625 s. When the node voltage recovers to 0.9 pu or above, the system voltage should gradually recover to the rated voltage operating range. Based on the aforementioned voltage-duration relationship, a binary criterion region for system transient voltage stability is constructed, forming the system transient voltage stability boundary curve. After completing the voltage stability criterion table construction, the predicted voltage trajectories of key system nodes are compared with the constructed voltage-time stability boundary. When the predicted voltage trajectory of the system lies above the stability boundary curve throughout the entire transient process, i.e., the system voltage does not fall below the allowable lower voltage limit under the corresponding time condition at any given time, the system is determined to be in a transient voltage stable state. If the predicted voltage trajectory of the system falls below the stability boundary curve for a certain period of time, the system is determined to be in a transient voltage unstable state.

10. The method for transient voltage prediction and stability assessment of a DC receiving-end system based on MCNN-BiLSTM-AM according to claim 9, characterized in that, After determining the system's stable state, a transient voltage stability margin index is introduced to evaluate the system's operating status. Starting from the fault clearing moment, the difference between the predicted system voltage trajectory and the binary criterion stability boundary is integrated over the time interval from fault clearing to the end of the simulation to obtain the area difference between the system voltage trajectory and the stability boundary; this area difference reflects the safe distance of the system voltage trajectory relative to the stability boundary. When the calculated area difference is positive, it is determined that the overall voltage trajectory of the system is above the stability boundary and the system has a transient voltage stability margin. When the area difference is close to zero, the system is determined to be close to the transient voltage stability limit. When the area difference is negative, it is determined that the system voltage trajectory has exceeded the stability boundary and the system is in a transient voltage instability state.

Images