Multi-dimensional fault jump prediction method and system for high proportion of new energy grid connection

Abstract

The application provides a multi-dimensional fault jump prediction method and system for high-proportion new energy grid connection, wherein the method first collects and processes multi-source heterogeneous data such as new energy characteristics, power grid operation, V2G interaction and fault history to form a core feature set; then a multi-dimensional fault jump prediction model is built to extract the coupling features between new energy fluctuation, V2G two-way interaction and power grid parameters, realize the joint prediction of fault position, time, amplitude and phase; then the model is trained and verified by using historical samples to obtain the final version; then real-time data is input into the model to obtain the initial prediction result, and the final multi-dimensional prediction result is output by combining historical data and current working condition dynamic correction; finally, the fault risk level is divided according to this, the V2G charging and discharging control and the power grid scheduling strategy are linked, and the prevention and control scheme is generated; the application can realize multi-dimensional advance prediction and closed-loop prevention and control of fault jump in the high-proportion new energy grid connection scene containing V2G two-way interaction.

Description

Technical Field

[0001] This invention relates to the field of power system renewable energy grid connection safety technology, and in particular to a multi-dimensional fault jump prediction method and system for high-proportion renewable energy grid connection. Background Technology

[0002] The intermittent and fluctuating power output of new energy sources, coupled with the randomness and bidirectional power surges of electric vehicle charging and discharging in V2G mode, can easily cause sudden changes in electrical parameters such as voltage, current, and phase at the grid connection point, thereby inducing fault point jumps. These fault point jumps include not only changes in fault location but also changes in jump time, amplitude, and phase. In severe cases, they may cause equipment overcurrent, inverter lockout, power angle instability, and cascading disconnection of new energy units from the grid, threatening the safe and stable operation of the power grid and the reliability of power supply.

[0003] Existing technologies for analyzing and predicting grid-connected faults in new energy sources mainly fall into two categories.

[0004] The first category comprises traditional fault analysis methods based on equivalent circuit models. These methods typically rely on linearization assumptions and relatively stable operating conditions, making them more suitable for unidirectional energy flow scenarios. When a high proportion of new energy sources are integrated and bidirectional V2G power interaction exists, the system exhibits significant nonlinearity, strong coupling, and strong time-varying characteristics, easily leading to a decrease in the applicability and computational accuracy of traditional models. Furthermore, these methods usually focus on amplitude or phase analysis after a fault occurs, making it difficult to predict the location and timing of fault transitions in advance, thus failing to meet the proactive prevention and control requirements in complex grid-connected scenarios.

[0005] The second category consists of data-driven anomaly monitoring or fault identification methods. While these methods possess some nonlinear modeling capabilities, they still suffer from the following shortcomings: First, they fail to deeply integrate core interactive data in V2G scenarios, such as the number of connections, state of charge, charging and discharging power, and connection / disconnection timing, making it difficult to accurately depict the coupling relationship between "new energy fluctuations—V2G bidirectional interaction—grid parameter changes—fault jumps." Second, the prediction dimensions are relatively singular, focusing primarily on amplitude and phase analysis after anomaly identification or fault occurrence, lacking joint prediction of fault jump location and time, making it difficult to provide complete support for fault isolation and recovery decisions. Third, existing models typically lack specific optimization for V2G bidirectional power timing characteristics, resulting in high false alarm and false negative rates, and failing to achieve closed-loop linkage between prediction results and V2G control and grid dispatch.

[0006] Therefore, there is an urgent need for a multi-dimensional fault jump prediction method and system for high-proportion renewable energy grid connection, which can realize multi-dimensional early prediction and closed-loop prevention and control of fault jump in high-proportion renewable energy grid connection scenarios with V2G bidirectional interaction. Summary of the Invention

[0007] The purpose of this invention is to provide a multi-dimensional fault jump prediction method and system for high-proportion renewable energy grid connection, aiming to solve the technical problem of difficult accurate prediction and proactive prevention and control of fault jump in high-proportion renewable energy grid connection scenarios with V2G bidirectional interaction.

[0008] To achieve the above objectives, in a first aspect, the present invention provides a multi-dimensional fault jump prediction method for high-proportion renewable energy grid connection, the steps of which include:

[0009] S1. Collect multi-source heterogeneous operation data, which includes at least new energy characteristic data, power grid operation data, V2G interaction data and fault history data, and preprocess the multi-source heterogeneous operation data to obtain a core feature set for fault jump prediction.

[0010] S2. Construct a multi-dimensional fault jump prediction model based on the core feature set. The multi-dimensional fault jump prediction model is used to extract the coupling features between new energy fluctuations, V2G bidirectional power interaction and grid parameter changes, and to jointly predict the location, time, amplitude and phase of the fault jump.

[0011] S3. The multidimensional fault jump prediction model is trained and validated using historical sample data to obtain the final prediction model;

[0012] S4. Input the real-time collected operating data into the final prediction model to obtain the initial prediction result of the fault jump, and dynamically correct the initial prediction result by combining the historical prediction results and the current operating conditions to obtain the final multidimensional prediction result.

[0013] S5. Based on the final multidimensional prediction results, classify the fault risk level, and generate a fault prevention and control scheme by linking the V2G charging and discharging control strategy and the power grid dispatching strategy according to the fault risk level.

[0014] As a further improvement to the above technical solution, in step S1, the new energy characteristic data includes one or more of the following: new energy power station output data, environmental meteorological data, and output fluctuation data.

[0015] The power grid operation data includes one or more of the following: grid connection point voltage, grid connection point current, system frequency, bus voltage, line impedance, harmonic parameters, and three-phase imbalance parameters.

[0016] The V2G interaction data includes one or more of the following: number of electric vehicles connected, battery state of charge, charging and discharging power, charging and discharging mode, and connection or disconnection timing data.

[0017] As a further improvement to the above technical solution, in step S1, the preprocessing includes data cleaning, missing value imputation, data standardization, temporal feature extraction, and relevant feature screening.

[0018] The time-series feature extraction is based on a sliding time window to extract one or more of the following features from multi-source heterogeneous operating data: trend features, periodic features, fluctuation features, and abrupt change features.

[0019] The relevant feature filtering is used to filter strongly correlated features related to fault jump events and construct the core feature set.

[0020] As a further improvement to the above technical solution, step S1 also includes: aligning the access timing changes and charging / discharging power changes in the V2G interaction data with the voltage, current, and frequency changes in the grid operation data in time, and constructing a coupling feature that reflects the propagation correlation between new energy fluctuations, V2G bidirectional power disturbances, and grid state changes.

[0021] As a further improvement to the above technical solution, the multidimensional fault jump prediction model includes a feature extraction layer, a temporal correlation modeling layer, a key feature adaptive focusing layer, and a multi-task joint output layer connected in sequence.

[0022] The feature extraction layer is used to extract local coupling features from multi-source heterogeneous operating data; the temporal correlation modeling layer is used to extract long-term dependency features and bidirectional temporal features in the fault evolution process; the key feature adaptive focusing layer is used to increase the weight of features that significantly affect fault jumps; and the multi-task joint output layer is used to output the prediction results of fault jump location, fault jump time, fault jump amplitude, and fault jump phase.

[0023] As a further improvement to the above technical solution, the feature extraction layer adopts a convolutional neural network layer, the temporal correlation modeling layer includes a self-attention modeling layer and a bidirectional recurrent neural network layer, and the key feature adaptive focusing layer is used to adaptively weight V2G charging and discharging power, battery state of charge, number of electric vehicles connected, new energy power output fluctuations and grid operation parameters to enhance the model's ability to identify key influencing features.

[0024] As a further improvement to the above technical solution, in step S3, a multi-task joint loss function is used to supervise the training of the multi-dimensional fault jump prediction model. The multi-task joint loss function includes at least a classification loss term for fault jump location prediction and a regression loss term for fault jump time, fault jump amplitude and fault jump phase prediction, so that the four prediction tasks share the underlying fault representation and are optimized collaboratively.

[0025] As a further improvement to the above technical solution, step S3 also includes selecting a model for the multidimensional fault jump prediction model through cross-validation and hyperparameter optimization to obtain a final prediction model that is suitable for the current high proportion of new energy grid connection scenario.

[0026] As a further improvement to the above technical solution, in step S4, the dynamic correction includes: weighting and fusing the initial prediction result with the historical prediction result within a preset time window to suppress prediction deviations caused by short-term random fluctuations.

[0027] When changes in V2G access scale, fluctuations in new energy output, or changes in key grid operating parameters exceed preset thresholds are detected, the final prediction model is adjusted online or the initial prediction results are adaptively corrected.

[0028] As a further improvement to the above technical solution, in step S4, different correction rules are adopted for at least two types of predicted quantities among fault jump location, fault jump time, fault jump amplitude and fault jump phase, in order to adapt to the different response of different prediction dimensions to changes in operating conditions.

[0029] As a further improvement to the above technical solution, in step S5, the fault risk level includes at least low risk, medium risk, high risk and extremely high risk, and the fault risk level is determined based on at least two of the following: fault transition time warning window, fault transition amplitude, fault transition phase, fault transition location and prediction confidence.

[0030] As a further improvement to the above technical solution, in step S5,

[0031] When the fault risk level is medium risk or above, a charging and discharging power adjustment command is sent to the V2G control terminal.

[0032] When the fault risk level is high or above, a new energy output adjustment command or power flow adjustment command is sent to the power grid dispatching terminal.

[0033] When the fault risk level is extremely high, a fault isolation plan is generated.

[0034] Secondly, the present invention also provides a prediction system based on the multi-dimensional fault jump prediction method for high-proportion renewable energy grid connection described in the first aspect, comprising:

[0035] The data acquisition module is used to collect data on new energy characteristics, power grid operation, V2G interaction, and historical fault data.

[0036] The data preprocessing module is used to preprocess multi-source heterogeneous operational data and construct a core feature set;

[0037] The predictive modeling module is used to construct a multidimensional fault jump prediction model and jointly predict the location, time, amplitude, and phase of the fault jump based on the core feature set.

[0038] The training and validation module is used to train and validate the multidimensional fault jump prediction model using historical sample data, and generate the final prediction model.

[0039] The dynamic correction module is used to dynamically correct the initial prediction results of fault transitions based on historical prediction results and current operating conditions, generating a final multi-dimensional prediction result; and

[0040] The early warning linkage module is used to classify the fault risk level based on the final multidimensional prediction results, and to generate a fault prevention and control scheme by linking the V2G charging and discharging control strategy and the power grid dispatching strategy.

[0041] As a further improvement to the above technical solution, the prediction modeling module includes a feature extraction unit, a temporal correlation modeling unit, a key feature adaptive focusing unit, and a multi-task joint output unit.

[0042] The feature extraction unit is used to extract the local coupling features between new energy fluctuations, V2G bidirectional power interaction, and changes in grid parameters.

[0043] The temporal correlation modeling unit is used to extract long-term dependency features and bidirectional temporal features of fault evolution;

[0044] The key feature adaptive focusing unit is used to increase the weight of key influencing features;

[0045] The multi-task joint output unit is used to output the prediction results of fault jump location, fault jump time, fault jump amplitude, and fault jump phase.

[0046] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the multi-dimensional fault jump prediction method for high-proportion new energy grid connection described in the first aspect.

[0047] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the multi-dimensional fault jump prediction method for high-proportion renewable energy grid connection described in the first aspect.

[0048] Because the present invention adopts the above technical solutions, the beneficial effects of this application are as follows:

[0049] This invention provides a multi-dimensional fault jump prediction method for high-proportion renewable energy grid connection. By integrating renewable energy characteristic data, grid operation data, V2G interaction data, and historical fault data, and constructing a coupling feature that reflects the propagation correlation between renewable energy fluctuations, V2G bidirectional power disturbances, and grid state changes, the prediction model can characterize the correlation between "renewable energy fluctuations - V2G bidirectional interaction - grid parameter changes - fault jump". Therefore, compared with schemes that rely solely on renewable energy characteristics or single grid operation parameters, this method is more suitable for high-proportion renewable energy grid connection scenarios involving V2G bidirectional power interaction.

[0050] This invention constructs a multi-task joint prediction mechanism for fault jump location, time, amplitude and phase, enabling different prediction dimensions to share the underlying fault representation and optimize collaboratively. Therefore, it can not only determine whether a fault has occurred, but also provide more complete prediction information for fault isolation, recovery control and scheduling decisions, thereby improving the problems of existing technologies with single prediction dimensions and lack of accurate positioning of jump location and advance prediction of jump time.

[0051] This invention employs a hybrid modeling approach that includes CNN feature extraction, self-attention modeling, bidirectional temporal modeling, V2G characteristic adaptive attention, and multi-task output. This approach enables the model to take into account local coupling features, long-term dependency features, and bidirectional temporal features in multi-source data. Therefore, it is beneficial to improve the ability to identify key causes of fault transitions and reduce the interference of weakly correlated features and pseudo-correlated features on the prediction results.

[0052] This invention dynamically corrects the initial prediction results by combining historical prediction results and current operating conditions during the real-time prediction stage. It also performs adaptive correction or online adjustment when the scale of V2G access, the fluctuation of new energy output, or the change of key grid parameters exceeds a preset threshold. Therefore, it can reduce the fluctuation of prediction results caused by short-term random disturbances, improve the continuity and stability of prediction results under complex operating conditions, and thus be more suitable for actual online early warning and linkage control.

[0053] This invention classifies fault risk levels based on the final multidimensional prediction results and generates fault prevention and control schemes by linking V2G charging and discharging control strategies and grid dispatching strategies. It constructs a closed-loop processing mechanism that directly drives control actions based on prediction results. Therefore, it can shift fault prevention and control from passive handling after a fault occurs to proactive early warning and prevention before a fault occurs, which is conducive to reducing the risk of fault expansion and improving the safety and controllability of high-proportion renewable energy grid-connected systems.

[0054] The input data, feature construction method, prediction target, and linkage output of this invention all correspond to the specific application scenario of high-proportion new energy grid connection with V2G bidirectional energy interaction. Therefore, this invention is not an abstract algorithm detached from the actual scenario, but a specific technical solution for new energy grid connection fault early warning and control, and has a clear engineering application basis. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0056] Figure 1 This is a flowchart illustrating a multi-dimensional fault jump prediction method for high-proportion renewable energy grid connection disclosed in this invention.

[0057] Figure 2 This is a schematic diagram illustrating the implementation process of the multidimensional fault jump prediction model disclosed in this invention.

[0058] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] It should be noted that the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0061] Example 1

[0062] See Figures 1-2 This invention provides a multi-dimensional fault jump prediction method for high-proportion renewable energy grid connection, the specific steps of which include:

[0063] S1. Collect multi-source heterogeneous operation data, which includes at least new energy characteristic data, power grid operation data, V2G interaction data and fault history data. Preprocess the multi-source heterogeneous operation data to obtain a core feature set for fault jump prediction.

[0064] Specifically, the new energy characteristic data includes one or more of the following: new energy power plant output data, environmental meteorological data, and output fluctuation data; in this embodiment, the new energy characteristic data includes minute-level light intensity, wind speed, ambient temperature and humidity, photovoltaic module conversion efficiency, wind turbine pitch angle, wind turbine speed, and real-time output data of the new energy power plant.

[0065] The power grid operation data includes one or more of the following: grid connection point voltage, grid connection point current, system frequency, bus voltage, line impedance, harmonic parameters, and three-phase imbalance parameters; in this embodiment, the power grid operation data includes 10ms-level grid connection point three-phase voltage, three-phase current, system frequency, line impedance, bus voltage, harmonic content, and three-phase imbalance data.

[0066] The V2G interaction data includes one or more of the following: number of electric vehicles connected, battery state of charge, charging and discharging power, charging and discharging mode, and connection or disconnection timing data. In this embodiment, the V2G interaction data includes minute-level number of electric vehicles connected, single-vehicle battery SOC status, 10ms-level single-vehicle charging and discharging power, charging and discharging operation mode, and V2G connection / disconnection / power switching timing data.

[0067] Historical fault data includes historical fault types, transition locations, transition times, amplitude / phase transition values, fault causes, and cascading fault propagation path data.

[0068] S2. Construct a multi-dimensional fault jump prediction model based on the core feature set. The multi-dimensional fault jump prediction model is used to extract the coupling features between new energy fluctuations, V2G bidirectional power interaction and grid parameter changes, and to jointly predict the location, time, amplitude and phase of the fault jump.

[0069] S3. Use historical sample data to train and validate the multidimensional fault jump prediction model to obtain the final prediction model.

[0070] Historical datasets containing different V2G access ratios and new energy power output scenarios were used, and the datasets were divided into training, validation, and test sets in an 8:1:1 ratio for supervised training. Specifically, in this embodiment, the hyperparameters for model training were set as follows: the AdamW optimizer was used, the initial learning rate was set to 0.001, the weight decay coefficient was set to 0.0001, the learning rate decay adopted a cosine annealing strategy, T_max was set to 100, and eta_min was set to 0.00001; the training batch size was set to 64, the maximum number of training epochs was set to 100, an early stopping mechanism was adopted, and the patience value was set to 10.

[0071] S4. Input the real-time collected operating data into the final prediction model to obtain the initial prediction result of the fault jump, and dynamically correct the initial prediction result by combining the historical prediction results and the current operating conditions to obtain the final multidimensional prediction result.

[0072] The specific correction process is as follows: Real-time running data is collected and input into the model every 5 minutes to obtain the initial prediction result. The exponential moving average method is used to fuse the previous 5 historical prediction results with the current initial prediction result, with the decay coefficient set to [value missing]. This smooth fusion operation effectively reduces the random fluctuations of individual prediction results. Furthermore, for extreme conditions in grid-connected systems, online model fine-tuning is triggered when the change rate of the number of V2G connected vehicles is ≥50% within 10 minutes, or when the output of new energy vehicles fluctuates by ≥30% within 5 minutes. Rapid incremental updates improve the model's prediction accuracy under extreme conditions.

[0073] S5. Based on the final multidimensional prediction results, classify the fault risk level, and generate a fault prevention and control scheme by linking the V2G charging and discharging control strategy and the power grid dispatching strategy according to the fault risk level.

[0074] Based on the final output's transition time warning window, transition amplitude, and prediction confidence, the system failure risk is divided into four levels:

[0075] Blue Low Risk: Issues early warning prompts and continuously monitors the operational status of the corresponding grid connection point.

[0076] Yellow Medium Risk: Outputs a warning signal and sends a charge and discharge power stabilization command to the V2G aggregator to limit the EV charge and discharge power fluctuation range to ≤10% / min and actively suppress power fluctuation.

[0077] Orange High Risk: Send control commands to the V2G aggregator to reduce charging and discharging power according to a set ratio, suspend new EV access, and switch the charging and discharging mode to constant voltage mode; at the same time, send suggestions for adjusting the output of new energy sources to the grid dispatch system.

[0078] Red Extremely High Risk: Outputs an emergency audible and visual warning signal, sends an emergency shutdown command to the V2G aggregator, and simultaneously sends a fault isolation plan to the power grid dispatch system, executing the target line switching action 3 seconds in advance.

[0079] Through the above-mentioned hierarchical early warning and control linkage strategy, a full-link response from "prediction-early warning" to "control closed loop" is realized, enabling the system to dynamically adjust EV charging and discharging behavior and new energy power output to actively avoid faults, and greatly improving the system's safe and stable operation capability.

[0080] As a preferred embodiment, to ensure the data quality of the input model and fully extract effective information that can characterize the fault evolution trend, the preprocessing specifically includes data cleaning, missing value imputation, data standardization, temporal feature extraction, and strong correlation feature screening. The specific execution process is as follows:

[0081] Data cleaning: An improved isolated forest algorithm is used, with 3σ as the outlier identification threshold, to remove outliers and noisy data from the original multi-source heterogeneous data, so as to prevent noise from interfering with the model's learning of the real operating state.

[0082] Missing value imputation: Considering communication delays or packet loss in real-world industrial scenarios, a WGAN-GP generative adversarial network is used to impute time-series data with a missing value rate of ≤30%. This step effectively solves the problem of missing samples in V2G interaction data and new energy operation data, and also overcomes the deficiency of scarce fault samples in deep learning.

[0083] Data standardization: Since multi-source heterogeneous data includes various types such as power, voltage, and number of vehicles connected, the Z-score standardization method is used to unify the dimensions of all data. This operation eliminates scale differences between different types of data, which helps to accelerate the convergence speed of subsequent hybrid deep learning models and improve prediction accuracy.

[0084] Time-series feature extraction: Deep mining of multi-source heterogeneous operational data is performed based on a sliding time window, specifically setting the sliding window size to 5 minutes and the step size to 1 minute. Within this window, trend features (e.g., linear fitting slope), periodic features (e.g., Fourier transform frequency), volatility features (e.g., the ratio of standard deviation to mean), and abrupt change features (e.g., the peak value of the absolute value of the first difference) are extracted from each time series data. By extracting these four types of features, discrete time series can be transformed into high-order feature expressions reflecting the evolution and abrupt change patterns of the system state.

[0085] Relevant Feature Screening: To reduce data dimensionality and computational redundancy, a Pearson correlation coefficient of 0.3 was used as a threshold to screen features strongly correlated with fault transition events from the massive amount of extracted features. After screening, a 32-dimensional multi-source heterogeneous core feature set was finally constructed, including 12-dimensional V2G interaction features, 10-dimensional new energy characteristic features, and 10-dimensional power grid operation features. Constructing the core feature set not only eliminated weakly correlated and pseudo-correlated features, but also enabled the subsequent prediction model to accurately focus on the core elements that caused the fault.

[0086] As a preferred embodiment, in order to accurately capture the complex nonlinear time-varying correlation in the scenario of high proportion of new energy and V2G access, step S1 further includes: aligning the access timing changes and charging / discharging power changes in the V2G interaction data with the voltage, current, and frequency changes in the grid operation data in time, and constructing a coupling feature that reflects the propagation correlation between new energy fluctuations, V2G bidirectional power disturbances, and grid state changes.

[0087] Specifically, the collected multi-source heterogeneous operational data differ in time scale. For example, grid operation data such as three-phase voltage and current at the grid connection point, and single-vehicle charging and discharging power in V2G interaction data, are high-frequency data at the 10ms level, while data on the number of electric vehicles connected, battery SOC status, and new energy characteristics are mostly low-frequency data at the minute level. In constructing the multi-source heterogeneous core feature set, it is essential to map the time-series data with different sampling frequencies onto the same time reference axis to achieve time alignment. Based on complete time alignment, the timing actions of V2G bidirectional power interaction (such as charging / discharging mode switching and sudden power changes) are further correlated and integrated with the real-time fluctuation trajectory of grid-side electrical parameters (such as grid connection point voltage and frequency), thereby constructing a coupled feature set that can characterize the global evolution of the system. This setup fully aligns with the complex operational characteristics of "bidirectional energy flow, strong coupling, and strong time variation" in a high-proportion V2G new energy grid-connected system. By constructing time-aligned and coupling features, the model can accurately capture and reconstruct the cross-dimensional coupling correlation between "new energy fluctuations - V2G bidirectional interaction - grid parameter changes - fault jumps". This effectively solves the core defect of existing anomaly monitoring methods, which fail to understand the propagation path of bidirectional power disturbances in the grid due to isolated analysis of the new energy source or a single grid parameter. The high-quality coupling features also provide solid data input support for subsequent hybrid deep learning models (such as through CNN layers) to accurately extract spatial coupling features, thus ensuring that the model still has extremely high fault jump prediction accuracy under complex operating conditions such as high-frequency V2G access or large fluctuations in new energy.

[0088] As a preferred embodiment, in order to fully explore the complex coupling relationship between the three ends of "new energy - V2G - power grid" in the grid-connected system, the multi-dimensional fault jump prediction model includes a feature extraction layer, a time-series correlation modeling layer, a key feature adaptive focusing layer, and a multi-task joint output layer connected in sequence.

[0089] The feature extraction layer employs a convolutional neural network (CNN) layer to extract local coupling features from multi-source heterogeneous operating data. In specific implementation, multiple serial convolutional layers combined with max pooling layers can be used to process the input core feature set, which can effectively extract the spatial cross-coupling features of new energy fluctuations, V2G charging and discharging impacts, and grid parameters, overcoming the shortcomings of existing technologies that cannot capture the global spatial correlation of the system due to fragmented analysis of single-sided data.

[0090] The temporal correlation modeling layer comprises a self-attention modeling layer and a bidirectional recurrent neural network layer, used to extract long-term dependency features and bidirectional temporal features during the fault evolution process. Specifically, the self-attention modeling layer is preferably a Transformer self-attention layer with multiple parallel multi-head attention heads, capable of accurately capturing the long- and short-term temporal dependencies of V2G bidirectional power interaction; the bidirectional recurrent neural network layer is preferably a serial bidirectional LSTM layer, used to deeply mine the bidirectional temporal correlation between renewable energy output fluctuations, V2G charging and discharging rhythms, and the fault evolution process. The cascaded combination of these two significantly improves the model's sequence modeling accuracy in dealing with the strong nonlinearity and time-varying characteristics of a high-proportion renewable energy weak grid scenario.

[0091] The key feature adaptive focusing layer (i.e., the V2G characteristic adaptive attention layer) is used to adaptively weight V2G charging and discharging power, battery state of charge (SOC), number of electric vehicles connected, fluctuations in renewable energy output, and grid operating parameters to enhance the model's ability to identify key influencing features. During model training, this layer can dynamically adjust the weight allocation of each input feature, increasing the weight of features significantly affecting fault transitions. This enables the model to adaptively focus on core causal features and eliminate spurious correlation noise in massive and complex grid-connected monitoring data. This not only reduces false alarms and false negatives under complex operating conditions but also significantly improves the interpretability of deep learning black-box models in industrial applications.

[0092] The multi-task joint output layer is used to synchronously output the predicted results of fault transition location, fault transition time, fault transition amplitude, and fault transition phase. Specifically, this is achieved by setting up four parallel network output branches, each using a corresponding activation function to output the preset fault location probability, early warning duration, and voltage / current amplitude and phase transition values. The beneficial effect of this layer design is that it completely breaks through the technical bottlenecks of existing technologies, which suffer from single-dimensional prediction and severe prediction lag. It achieves four-dimensional collaborative early prediction of "location-time-amplitude-phase," providing comprehensive quantitative decision support for dispatchers to formulate accurate fault isolation and recovery strategies in advance, thus preventing the expansion of the fault range from the source.

[0093] In this embodiment, see Figure 2 A hybrid deep learning model of "CNN-Transformer-bidirectional LSTM" adapted to the bidirectional power interaction characteristics of V2G is constructed as the multidimensional fault jump prediction model. The model consists of an input layer, a CNN feature extraction layer, a Transformer self-attention layer, a bidirectional LSTM layer, a V2G characteristic adaptive attention layer, and a multi-task joint output layer.

[0094] Input layer: Receives a 32-dimensional core feature set, sets a separate feature embedding layer for discrete features in V2G interaction features, sets the embedding dimension to 16, and converts discrete features into continuous vectors.

[0095] CNN Feature Extraction Layer: Three sequential convolutional layers are set up with kernel sizes of 3×1, 5×1, and 3×1 respectively, a stride of 1 for each layer, padding mode set to "SAME", and ReLU activation function. Each convolutional layer is followed by a max pooling layer with a kernel size of 2×1 and a stride of 2, used to extract the spatial coupling features of new energy fluctuations, V2G charging and discharging impacts, and grid parameters.

[0096] Transformer self-attention layer: Set 4 parallel multi-head attention heads, set the hidden layer dimension to 64, and set the dropout rate to 0.1 to capture the long and short-term temporal dependencies of V2G bidirectional power interaction.

[0097] Bidirectional LSTM layer: A two-layer serial bidirectional LSTM network is set up, with 128 hidden units per layer, a dropout rate of 0.1, and a recurrent_dropout rate of 0.05. This is used to explore the bidirectional time-series correlation between new energy power output fluctuations, V2G charging and discharging rhythms, and fault evolution processes.

[0098] The V2G feature adaptive attention layer initially assigns weights of 0.3 for V2G charging / discharging power, 0.2 for battery SOC status, 0.2 for the number of electric vehicles connected, 0.2 for new energy output fluctuations, and 0.1 for grid operating parameters. The weights of each feature are adaptively adjusted during model training to focus on core features that significantly impact fault transitions. Initial weights are assigned to each input feature and adaptively adjusted during model training to focus on core features that significantly impact fault transitions and eliminate spurious features to improve model interpretability.

[0099] Multi-task joint output layer: Four parallel output branches are set up: the jump position branch, which uses the Softmax activation function to output the fault occurrence probability of 20 preset fault positions; the jump time branch, which uses the Linear activation function to output the fault early warning duration from 0 to 30 seconds; the jump amplitude branch, which uses the Linear activation function to output the voltage / current amplitude jump value from 0 to 0.5 pu; and the jump phase branch, which uses the Linear activation function to output the voltage phase jump value from 0 to 90°.

[0100] The multi-branch structure enables four-dimensional collaborative early prediction of fault points, effectively avoiding the risk of expanding the scope of faults due to single-dimensional post-event analysis.

[0101] As a preferred embodiment, in order to improve the model's learning efficiency and comprehensive prediction ability for complex fault evolution features, in step S3, a multi-task joint loss function is used to supervise the training of the multi-dimensional fault jump prediction model. The multi-task joint loss function includes at least a classification loss term for fault jump location prediction and a regression loss term for fault jump time, fault jump amplitude and fault jump phase prediction, so that the four prediction tasks share the underlying fault representation and are optimized collaboratively.

[0102] In practical implementation, the calculation formula for the multi-task joint loss function is expressed as follows:

[0103] ;

[0104] in, The cross-entropy loss is the branch at the jump position, which is the classification loss term used for predicting the fault jump position. This represents the mean squared error loss for the jump time branch. This is the mean square error loss of the amplitude branch. This represents the mean square error loss for the phase jump branch. (The above...) , and Together, they constitute the regression loss term used for time, amplitude, and phase prediction. , , , These are the weighting coefficients. By setting the weighting coefficients, the training priority of each prediction task is balanced.

[0105] By placing the classification loss (cross-entropy) and regression loss (mean squared error) under the same joint loss function for backpropagation constraints, the discrete jump-position classification task and the continuous time, amplitude, and phase regression task can fully share the extracted new energy fluctuation and V2G coupling features from the underlying layer. Each branch optimizes collaboratively during training, effectively overcoming the tendency of traditional fragmented single-dimensional prediction models to get trapped in local optima. In joint training, the gradient magnitudes generated by different tasks vary significantly. By introducing and setting reasonable weight coefficients, such as... =0.3、 =0.25、 =0.25、 =0.2, which effectively balances the training priorities of each task and prevents the gradient update process of the entire model from being dominated by an excessively large absolute error in one dimension (such as amplitude or time). This multi-task joint supervised training mechanism ensures that the deep learning model can achieve balanced and high-precision collaborative prediction of the four-dimensional features (position, time, amplitude, and phase) of fault transitions in complex V2G bidirectional interaction scenarios.

[0106] As a preferred embodiment, in order to ensure the generalization ability and prediction accuracy of the model in a complex and ever-changing grid-connected environment, step S3 also includes model selection of the multidimensional fault jump prediction model through cross-validation and hyperparameter optimization to obtain a final prediction model that is suitable for the current high proportion of new energy grid connection scenario.

[0107] Specifically, in this embodiment, the data samples are first scientifically allocated, with the collected historical dataset divided into a training set, a validation set, and a test set in an 8:1:1 ratio. To ensure the sufficiency and comprehensiveness of model learning, the training set is required to contain no fewer than 100,000 valid samples, and the proportion of fault samples is mandatory to be no less than 20%. This setting can fully cover diverse fault transition cases under different V2G access ratios and different renewable energy output scenarios in the real power grid, solving the defect that the model is not sensitive enough to fault characteristics due to the extreme imbalance between normal samples and fault samples.

[0108] During the model training and selection phase, a combination of five-fold cross-validation and Bayesian optimization was used to optimize the hyperparameters of the hybrid deep learning model.

[0109] Specific hyperparameter settings and optimization space include: the AdamW optimizer is used, with an initial learning rate of 0.001 and a weight decay coefficient of 0.0001; the batch size is set to 64, and the maximum number of training epochs is set to 100. To further improve the model's convergence and generalization ability, a cosine annealing strategy is used for learning rate decay, with the period parameter T_max set to 100 and the minimum learning rate eta_min set to 0.00001. Simultaneously, an early stopping mechanism is forcibly introduced during training, with a tolerance value set to 10, meaning training automatically stops when the joint loss on the validation set shows no decreasing trend for 10 consecutive epochs.

[0110] A tuning strategy combining five-fold cross-validation and Bayesian optimization is employed to efficiently and purposefully search for optimal solutions within a massive hyperparameter combination space. This avoids the blindness and inefficiency of traditional manual hyperparameter tuning, enabling the model architecture to accurately adapt to the input 32-dimensional multi-source heterogeneous core feature set. Overfitting risk is effectively mitigated: by introducing a cosine annealing strategy to dynamically adjust the learning rate, coupled with an early stopping mechanism based on validation set performance, overfitting of the model to the training set data can be effectively prevented.

[0111] Through the rigorous cross-validation and parameter fine-tuning process described above, the final prediction model not only possesses extremely high prediction accuracy but also exhibits excellent robustness. Even under extreme grid connection scenarios involving high-frequency V2G access and disconnection, as well as drastic fluctuations in renewable energy output, the final prediction model can still output stable and reliable multi-dimensional early warning signals, fully meeting the safety and control requirements of high-proportion renewable energy grid connection projects.

[0112] As a preferred embodiment, to improve the robustness and dynamic prediction stability of the model in complex grid-connected environments, the implementation process of step S4 is detailed as follows: In step S4, the dynamic correction includes: weighted fusion of the initial prediction result and historical prediction results within a preset time window to suppress prediction deviations caused by short-term random fluctuations. Specifically, the system collects real-time running data into the model every 5 minutes to obtain the initial prediction result. Subsequently, the exponential moving average method is used to fuse the previous 5 historical prediction results with the current initial prediction result.

[0113] The formula for weighted fusion is:

[0114] ;

[0115] In the formula, For the final prediction result, This is the current initial prediction result. The value is the average of the previous five historical predictions. To ensure a reasonable allocation of weights for the current operating condition, the attenuation coefficient is set to 0.9. This setting, through the smoothing effect of the exponential moving average method, effectively reduces the drastic fluctuations in prediction results caused by single measurement errors or short-term random disturbances, ensuring the stability and reliability of the output multidimensional prediction results. Simultaneously, when changes in V2G access scale, fluctuations in renewable energy output, or changes in key grid operating parameters exceed preset thresholds, the final prediction model is adjusted online or the initial prediction results are adaptively corrected.

[0116] Specifically, in this embodiment, the system monitors the real-time operating status. When it detects a change rate of ≥50% in the number of V2G connected vehicles within 10 minutes, or a fluctuation of ≥30% in new energy power output within 5 minutes, the system automatically triggers the online fine-tuning mechanism of the prediction model. The parameters for online fine-tuning are configured as follows: the fine-tuning learning rate is set to 0.0001, and the number of fine-tuning rounds is set to 5. This setting endows the model with adaptive calibration capabilities under extreme and sudden operating conditions (such as a massive number of electric vehicles connecting instantly or a sharp uphill climb in wind and solar power output). Through rapid incremental parameter updates, the model's prediction accuracy under extreme operating conditions is significantly improved.

[0117] Furthermore, in step S4, different correction rules are applied to at least two types of predicted quantities among fault jump location, fault jump time, fault jump amplitude, and fault jump phase to adapt to the different response differences of different prediction dimensions to changes in operating conditions. Since the multi-task joint output layer outputs 20 preset fault location probabilities, fault early warning durations from 0 to 30 seconds, amplitude jump values ​​from 0 to 0.5 pu, and phase jump values ​​from 0 to 90° in parallel, the sensitivity of each prediction dimension to fluctuations in physical conditions is drastically different. By setting differentiated correction rules for different predicted quantities—for example, assigning a higher correction weight to the historical mean for the topology-correlated "jump location," while adopting an online correction strategy that responds more quickly to real-time conditions for the fluctuation-sensitive "jump amplitude" and "jump time"—this setting accurately adapts to the specific laws of the four-dimensional variables "location-time-amplitude-phase" in the electrical physical evolution process, avoiding overfitting or underfitting of the prediction dimensions caused by using a single correction scale, thereby comprehensively improving the overall accuracy of multi-dimensional collaborative prediction.

[0118] As a preferred embodiment, in order to achieve a closed-loop system from multi-dimensional fault prediction to proactive defense and improve the safety management and control capabilities in scenarios with a high proportion of new energy grid connection, the specific implementation method of step S5 is detailed as follows:

[0119] In step S5, the fault risk level includes at least blue (low risk), yellow (medium risk), orange (high risk), and red (extremely high risk). Specifically, the fault risk level is comprehensively classified based on multi-dimensional prediction results, and its evaluation criteria combine at least two of the following: fault transition time warning window, fault transition amplitude (or phase), prediction confidence, and fault transition location.

[0120] This design eliminates the inherent flaw of traditional systems that rely solely on a single electrical parameter exceeding its limit for passive alarms. By integrating time urgency (early warning window), fault destructive force (transition amplitude / phase), and predictive reliability (confidence level), it can accurately quantify the fault evolution stages of the grid-connected system under complex operating conditions, thereby reducing the false alarm rate and missed alarm rate of the system.

[0121] Furthermore, in step S5, based on the dynamically classified risk levels, a differentiated prevention and control linkage strategy is implemented: when the fault risk level is yellow (medium risk) or above, for example, with a jump time warning window of 10s~30s and a prediction confidence level ≥75%, the system outputs a warning signal and sends a charging and discharging power adjustment command to the V2G control terminal (i.e., the V2G aggregator). Specific operations include issuing a charging and discharging power stabilization command to limit the fluctuation range of electric vehicle charging and discharging power to ≤10% / min. In the fault initiation stage, the high-frequency impact caused by V2G bidirectional power interaction is promptly mitigated from the demand side, actively smoothing out fluctuations in grid connection point parameters and suppressing the evolution of the fault.

[0122] When the fault risk level is orange (high risk) or above, for example, when the jump time warning window is 5s~10s, the prediction confidence is ≥85%, and the amplitude / phase jump exceeds the equipment tolerance threshold, the system outputs an audible and visual warning signal. In addition to issuing stricter limiting instructions to the V2G control terminal (such as reducing charging power by 50%, discharging power by 60%, and switching to constant voltage mode), the system also simultaneously sends renewable energy output adjustment instructions or power flow adjustment instructions to the grid dispatching terminal (grid dispatching system). Specifically, it sends a renewable energy output adjustment suggestion with an adjustment range ≤10% of the rated capacity. This coordinated design establishes a cross-end collaborative control link between "source (renewable energy output) - grid (grid dispatching) - load (V2G charging and discharging)". When the system approaches a critical fault state, it forcibly changes the grid-connected power flow distribution through multi-end efforts, avoiding large-scale grid disconnection accidents of renewable energy units caused by limited unilateral adjustment capabilities.

[0123] When the fault risk level is designated as "extremely high risk" (e.g., a transition time warning window ≤ 5s, prediction confidence ≥ 90%), indicating a risk of cascading faults, the system outputs an emergency audible and visual warning signal, generates a fault isolation plan, and sends it to the power grid dispatch system. This plan clarifies the action sequence of the target line switch and sets the action time to 3s before the predicted transition time. Simultaneously, it sends an emergency shutdown command to the V2G aggregator, suspending all unnecessary electric vehicle charging and discharging activities at the corresponding grid connection point. For critical faults with a high probability of triggering cascading propagation, the system fully utilizes the 0-30s advance prediction time difference output by the deep learning model to complete topology isolation before physical equipment damage or grid instability, completely cutting off the fault propagation path and ensuring the absolute safety of the power grid backbone.

[0124] Example 2

[0125] This invention also provides a prediction system based on the multi-dimensional fault jump prediction method for high-proportion renewable energy grid connection described in Embodiment 1. The system specifically includes:

[0126] The data acquisition module is used to collect data on renewable energy characteristics, power grid operation, V2G interaction, and historical fault data. Specifically, this module establishes communication connections with the renewable energy power station SCADA system, the power grid EMS system, the V2G charging and swapping station management platform, and meteorological monitoring stations to obtain multi-source data. This module can comprehensively acquire the core underlying operating status of the "renewable energy-power grid-V2G" three-terminal system, breaking down information silos between systems and providing multi-dimensional underlying data support for complex correlation analysis under high-proportion renewable energy access.

[0127] The data preprocessing module, communicatively connected to the data acquisition module, is used to preprocess multi-source heterogeneous operational data and construct a core feature set. This module is configured to sequentially perform outlier cleaning, data missing value imputation based on generative adversarial networks, dimensional standardization, time-series feature extraction based on sliding windows, and strong correlation feature selection. This module effectively eliminates scale differences and noise interference in heterogeneous data and solves the common problem of missing data samples in real-world industrial scenarios, thereby ensuring that the data input to the prediction model has high quality and strong physical correlation, laying the foundation for improving prediction accuracy.

[0128] The predictive modeling module is communicatively connected to the data preprocessing module and is used to construct a multidimensional fault jump prediction model and jointly predict the location, time, amplitude and phase of the fault jump based on the core feature set.

[0129] The specific network architecture within the predictive modeling module is further subdivided as follows:

[0130] Feature extraction unit: Employing a multi-layer convolutional neural network, this unit extracts local coupling features between renewable energy fluctuations, V2G bidirectional power interaction, and grid parameter variations. This unit effectively captures cross-dimensional spatial coupling features between different electrical quantities using convolutional kernels, overcoming the limitations of single-parameter analysis.

[0131] Temporal correlation modeling unit: Integrating a Transformer self-attention layer and a bidirectional LSTM layer, this unit is used to extract long-term dependency features and bidirectional temporal features of fault evolution. This unit can accurately characterize the complex temporal evolution of bidirectional power fluctuations in high-proportion renewable energy grid-connected systems and deeply mine the sequential information of the preceding and following fault evolution processes.

[0132] Key Feature Adaptive Focusing Unit: This unit is used to increase the weight of key influencing features (such as V2G charging and discharging power, battery SOC state, number of electric vehicles connected, fluctuations in renewable energy output, and grid operating parameters). This unit enables the system to adaptively focus on the core factors that have a decisive impact on fault transition evolution from a massive amount of features, eliminating spurious correlation noise, thereby enhancing the engineering interpretability and anti-interference capability of the deep learning black box model.

[0133] Multi-task joint output unit: This unit has four parallel output branches to synchronously output the predicted results of fault transition location, fault transition time, fault transition amplitude, and fault transition phase. This unit completely breaks through the bottleneck of the single-dimensional prediction of traditional technology, realizing four-dimensional collaborative early prediction of fault characteristics, and providing multi-dimensional data support for the comprehensive formulation of scheduling plans.

[0134] The training and validation module is used to train and validate the multidimensional fault jump prediction model using historical sample data, generating the final prediction model. By constructing a multi-task joint loss function for supervised learning and using cross-validation to complete hyperparameter tuning, this module ensures that multiple prediction tasks can share underlying features and converge collaboratively, guaranteeing the model's strong generalization ability and robustness in dealing with various complex and extreme grid-connected operating conditions.

[0135] The dynamic correction module, communicatively connected to both the prediction modeling module and the data acquisition module, dynamically corrects the initial prediction results of fault transitions based on historical prediction results and current real-time operating conditions, generating the final multi-dimensional prediction results. This module, by introducing an exponential moving average fusion strategy and an online fine-tuning mechanism, effectively mitigates prediction biases caused by short-term random fluctuations and rapidly adapts and calibrates under sudden extreme conditions (such as sudden changes in access scale), ensuring the stability and reliability of the system's online real-time early warning system.

[0136] The early warning linkage module, communicatively connected to the dynamic correction module, is used to classify fault risk levels based on the final multi-dimensional prediction results and generate fault prevention and control schemes in conjunction with V2G charging and discharging control strategies and grid dispatching strategies. This constructs a closed-loop mechanism across the entire "prediction-early warning-control" chain. The system not only outputs tiered early warning signals but also sends instructions to the V2G aggregator and the grid dispatching system. By dynamically adjusting the charging and discharging behavior of electric vehicles and the output of new energy units, it proactively avoids faults, thereby successfully upgrading passive post-fault handling to proactive pre-fault prevention, significantly improving the safe and stable operation capability of the power grid.

[0137] Example 3

[0138] The present invention also provides an apparatus comprising a memory for storing computer program instructions and a processor for executing the computer program instructions, wherein when the computer program instructions are executed by the processor, the apparatus is triggered to perform some or all of the steps in Embodiment 1;

[0139] A processor may include one or more processing units, such as an application processor (AP), a modem processor, a graphics processing unit (GPU), an image signal processor (ISP), a controller, a video codec, a digital signal processor (DSP), a baseband processor, and / or a neural network processing unit (NPU). These different processing units may be independent devices or integrated into one or more processors.

[0140] The controller can serve as the nerve center and command center of an electronic device. Based on the instruction opcode and timing signals, the controller generates operation control signals to control the fetching and execution of instructions.

[0141] The processor may also include memory for storing instructions and data. In some embodiments, the memory in the processor is a cache memory. This memory can store instructions or data that the processor has just used or that are used repeatedly. If the processor needs to use the instruction or data again, it can retrieve it directly from the memory. This avoids repeated accesses, reduces processor waiting time, and thus improves system efficiency.

[0142] Example 4:

[0143] The present invention also provides a computer-readable storage medium storing a computer program, wherein the program, when running, controls the device where the storage medium is located to perform some or all of the steps in Embodiment 1.

[0144] The storage medium may include high-speed RAM memory, and may also include nonvolatile memory, such as at least one disk storage device. It is understood that the storage medium can be any machine-readable medium capable of storing program code, such as random access memory (RAM), magnetic disk, hard disk, solid state disk (SSD), or nonvolatile memory.

[0145] Those skilled in the art will understand that embodiments of the present invention can be provided as methods or storage media. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0146] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. All equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A multi-dimensional fault jump prediction method for high-proportion renewable energy grid connection, characterized in that, The steps include: S1. Collect multi-source heterogeneous operation data, which includes at least new energy characteristic data, power grid operation data, V2G interaction data and fault history data, and preprocess the multi-source heterogeneous operation data to obtain a core feature set for fault jump prediction. S2. Construct a multi-dimensional fault jump prediction model based on the core feature set. The multi-dimensional fault jump prediction model is used to extract the coupling features between new energy fluctuations, V2G bidirectional power interaction and grid parameter changes, and to jointly predict the location, time, amplitude and phase of the fault jump. S3. The multidimensional fault jump prediction model is trained and validated using historical sample data to obtain the final prediction model; S4. Input the real-time collected operating data into the final prediction model to obtain the initial prediction result of the fault jump, and dynamically correct the initial prediction result by combining the historical prediction results and the current operating conditions to obtain the final multidimensional prediction result. S5. Based on the final multidimensional prediction results, classify the fault risk level, and generate a fault prevention and control scheme by linking the V2G charging and discharging control strategy and the power grid dispatching strategy according to the fault risk level.

2. The multi-dimensional fault jump prediction method for high-proportion renewable energy grid connection according to claim 1, characterized in that, In step S1, the new energy characteristic data includes one or more of the following: new energy power station output data, environmental meteorological data, and output fluctuation data; The power grid operation data includes one or more of the following: grid connection point voltage, grid connection point current, system frequency, bus voltage, line impedance, harmonic parameters, and three-phase imbalance parameters. The V2G interaction data includes one or more of the following: number of electric vehicles connected, battery state of charge, charging and discharging power, charging and discharging mode, and connection or disconnection timing data.

3. The multi-dimensional fault jump prediction method for high-proportion renewable energy grid connection according to claim 1, characterized in that, In step S1, the preprocessing includes data cleaning, missing value imputation, data standardization, temporal feature extraction, and relevant feature screening; The time-series feature extraction is based on a sliding time window to extract one or more of the following features from multi-source heterogeneous operating data: trend features, periodic features, fluctuation features, and abrupt change features. The relevant feature filtering is used to filter strongly correlated features related to fault jump events and construct the core feature set.

4. The multi-dimensional fault jump prediction method for high-proportion renewable energy grid connection according to claim 3, characterized in that, Step S1 also includes: aligning the access timing changes and charging / discharging power changes in the V2G interaction data with the voltage, current, and frequency changes in the grid operation data to construct a coupling feature that reflects the propagation correlation between new energy fluctuations, V2G bidirectional power disturbances, and grid state changes.

5. The multi-dimensional fault jump prediction method for high-proportion renewable energy grid connection according to claim 1, characterized in that, The multidimensional fault jump prediction model includes a feature extraction layer, a temporal correlation modeling layer, a key feature adaptive focusing layer, and a multi-task joint output layer connected in sequence. The feature extraction layer is used to extract local coupling features from multi-source heterogeneous operating data; the temporal correlation modeling layer is used to extract long-term dependency features and bidirectional temporal features in the fault evolution process; the key feature adaptive focusing layer is used to increase the weight of features that significantly affect fault jumps; and the multi-task joint output layer is used to output the prediction results of fault jump location, fault jump time, fault jump amplitude, and fault jump phase.

6. The multi-dimensional fault jump prediction method for high-proportion renewable energy grid connection according to claim 5, characterized in that, The feature extraction layer employs a convolutional neural network layer, the temporal correlation modeling layer includes a self-attention modeling layer and a bidirectional recurrent neural network layer, and the key feature adaptive focusing layer is used to adaptively weight V2G charging and discharging power, battery state of charge, number of electric vehicles connected, fluctuations in new energy output, and grid operating parameters to enhance the model's ability to identify key influencing features.

7. The multi-dimensional fault jump prediction method for high-proportion renewable energy grid connection according to claim 1, characterized in that, In step S3, the multi-task joint loss function is used to supervise the training of the multi-dimensional fault jump prediction model. The multi-task joint loss function includes at least a classification loss term for fault jump location prediction and a regression loss term for fault jump time, fault jump amplitude and fault jump phase prediction, so that the four prediction tasks share the underlying fault representation and are optimized collaboratively. Step S3 also includes model selection for the multidimensional fault jump prediction model through cross-validation and hyperparameter optimization to obtain a final prediction model that is suitable for the current high proportion of new energy grid connection scenario.

8. The multi-dimensional fault jump prediction method for high-proportion renewable energy grid connection according to claim 1, characterized in that, In step S4, the dynamic correction includes: weighting and fusing the initial prediction result with the historical prediction results within a preset time window to suppress prediction deviations caused by short-term random fluctuations; When changes in V2G access scale, fluctuations in new energy output, or changes in key grid operating parameters exceed preset thresholds are detected, the final prediction model is adjusted online or the initial prediction results are adaptively corrected.

9. The multi-dimensional fault jump prediction method for high-proportion renewable energy grid connection according to claim 1, characterized in that, In step S5, the fault risk level includes at least low risk, medium risk, high risk and extremely high risk, and the fault risk level is determined based on at least two of the following: fault transition time warning window, fault transition amplitude, fault transition phase, fault transition location and prediction confidence. When the fault risk level is medium risk or above, a charging and discharging power adjustment command is sent to the V2G control terminal. When the fault risk level is high or above, a new energy output adjustment command or power flow adjustment command is sent to the power grid dispatching terminal. When the fault risk level is extremely high, a fault isolation plan is generated.

10. A prediction system based on the multi-dimensional fault jump prediction method for high-proportion renewable energy grid connection as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect data on new energy characteristics, power grid operation, V2G interaction, and historical fault data. The data preprocessing module is used to preprocess multi-source heterogeneous operational data and construct a core feature set; The predictive modeling module is used to construct a multidimensional fault jump prediction model and jointly predict the location, time, amplitude, and phase of the fault jump based on the core feature set. The training and validation module is used to train and validate the multidimensional fault jump prediction model using historical sample data, and generate the final prediction model. The dynamic correction module is used to dynamically correct the initial prediction results of fault transition based on historical prediction results and current operating conditions, and generate the final multidimensional prediction results. The early warning linkage module is used to classify the fault risk level based on the final multidimensional prediction results, and to generate a fault prevention and control scheme by linking the V2G charging and discharging control strategy and the power grid dispatching strategy.

Images