A wind power boosting and collecting station safety state evaluation method and system based on digital twinning

Abstract

The application relates to the technical field of wind power operation and maintenance, and specifically discloses a wind power boosting and collecting station safety state evaluation method and system based on digital twinning, which comprises the following steps: S1, multi-source data acquisition and synchronous preprocessing; S2, digital twinning simulation working condition generation; S3, simulation and measured data fusion and expansion; S4, space-time representation learning and feature extraction; S5, safety state evaluation model training; S6, online estimation, early warning triggering and operation and maintenance decision support; the application can construct a comprehensive sample set which not only follows physical laws but also is consistent with field distribution by performing formalized domain alignment and physical constraint data fusion on high-fidelity digital twinning simulation samples and field measured samples and adopting a generative enhancement strategy based on physical constraints. The application expands the failure working condition space that can be learned by the model, makes up for the shortage of rare or extreme working conditions in field historical samples, and facilitates engineers to rely on and adopt the evaluation results as operation and maintenance judgment basis.

Description

Technical Field

[0001] This invention belongs to the field of wind power operation and maintenance technology, specifically relating to a method and system for assessing the safety status of wind power booster and collection stations based on digital twins. Background Technology

[0002] As the core hub for power transmission from wind farms, the operational safety and stability of wind power booster stations directly determine the reliable grid connection and efficient consumption of new energy power generation systems, making them a crucial link in ensuring the large-scale development of the wind power industry. With a high proportion of wind and solar new energy being connected to the grid, the equipment types in booster stations are becoming increasingly complex, with traditional transformers, circuit breakers, and new power electronic converters coexisting. Furthermore, they face complex operating conditions such as fluctuating wind speeds and frequent load changes, significantly increasing the risk of equipment failure. Condition monitoring and fault diagnosis have become a global challenge for the industry. Against this backdrop, safety condition assessment, as a core technology for predicting equipment degradation trends and preventing major accidents, is crucial for wind farm operation and maintenance decisions due to its accuracy, timeliness, and reliability.

[0003] However, existing methods for assessing the safety status of wind power booster substations still have many technical bottlenecks and are difficult to meet the actual needs under complex operating conditions.

[0004] Traditional assessment methods often employ fixed warning thresholds, which are ill-suited to dynamically changing conditions such as wind speed and load, leading to false alarms or missed alarms. Furthermore, the models fail to quantify their own uncertainties and lack conservative handling mechanisms for cognitive blind spots or data deviations, further reducing the reliability of warnings. In addition, existing methods mostly only output fault determination results, lacking the ability to accurately pinpoint the root cause of the fault, and do not incorporate simulation technology to provide optimized operation and maintenance strategies, making it difficult to support the transition from "passive inspection" to "proactive maintenance."

[0005] In response to this, this application proposes a method and system for assessing the safety status of wind power booster substations based on digital twins, in order to solve the above-mentioned problems. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for assessing the safety status of wind power booster and collection stations based on digital twins, so as to solve the problems mentioned in the background art.

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

[0008] A method for assessing the safety status of wind power booster substations based on digital twins includes the following steps:

[0009] S1. Multi-source data acquisition and synchronous preprocessing: Collect real-time observation data (including voltage, current, temperature, vibration, oil temperature, partial discharge signal and environmental data such as wind speed and direction) from the booster station, and perform timestamp alignment, noise reduction, missing value imputation and outlier detection on the observation data to obtain a time-aligned observation dataset.

[0010] S2. Generation of digital twin simulation conditions: Based on the physical parameters and operating conditions of the booster substation, parameter scanning and fault simulation are performed in the digital twin simulation model according to wind speed, load, topology and typical fault modes to obtain a labeled multi-condition simulation sample dataset.

[0011] S3. Fusion and augmentation of simulation and measured data: The time-aligned observation dataset and the simulation sample dataset are domain aligned and semantically mapped. Generative data augmentation methods are used to augment the scarce fault samples to obtain a fused and augmented comprehensive dataset.

[0012] S4. Spatiotemporal representation learning and feature extraction: Spatiotemporal feature learning is performed on the comprehensive dataset, and high-order feature vectors reflecting equipment operation and hidden fault modes are extracted through deep representation learning methods.

[0013] S5. Safety Status Assessment Model Training: Using the feature vector as input, construct and train a data-driven safety assessment model so that the safety assessment model can output the health index or failure probability distribution of equipment and sites, and obtain the trained safety assessment model.

[0014] S6. Online estimation, early warning triggering, and operation and maintenance decision support: The real-time collected data is processed in the same way as in steps S1 to S4 to obtain the current high-order feature vector. The current high-order feature vector is input into the safety assessment model to calculate the current safety index or failure probability. When the safety index meets the preset or adaptive triggering conditions, an early warning is triggered. Based on the output of the safety assessment model, root cause analysis and maintenance strategy simulation are performed in combination with digital twin simulation to generate operation and maintenance suggestions and optimize the maintenance plan.

[0015] Preferably, the real-time observation data (including voltage, current, temperature, vibration, oil temperature, partial discharge signals, and environmental data such as wind speed and direction) of the collected booster substation are processed by timestamp alignment, noise reduction, missing value imputation, and outlier detection to obtain a cleaned and time-aligned observation dataset, including:

[0016] Wavelet thresholding denoising and robust statistical outlier removal are performed. Missing value imputation adopts a method combining temporal adjacent interpolation and model prediction. Dynamic correction is performed on multi-source clock differences to obtain the observation dataset with higher confidence.

[0017] Preferably, based on the physical parameters and operating conditions of the booster substation, parameter scanning and fault simulation are performed in the digital twin simulation model according to wind speed, load, topology, and typical fault modes to obtain a labeled multi-condition simulation sample dataset, including:

[0018] A multiphysics coupling model was constructed and Monte Carlo parameter sampling was carried out to cover the boundary conditions; for rare fault conditions, a physics-aware GAN was used to augment the samples, resulting in the labeled multi-condition simulation sample dataset.

[0019] Preferably, the step of performing domain alignment and semantic mapping between the observed dataset and the simulated sample dataset, and using generative data augmentation methods to expand the scarce fault samples to obtain a fused and expanded comprehensive dataset includes:

[0020] Domain adaptive mapping and covariance alignment are employed, and transfer learning is used to reduce the difference between the simulation and experimental domains. The augmentation uses a conditional generation network to generate samples according to fault type to compensate for data imbalance, resulting in the comprehensive dataset.

[0021] Preferably, the step of performing spatiotemporal feature learning on the comprehensive dataset, and extracting high-order feature vectors reflecting equipment operation and hidden fault modes through deep representation learning methods, includes:

[0022] Feature extraction is achieved by combining multi-scale temporal networks and autoencoders. This involves first extracting local temporal segment features using convolutional and residual networks, then learning long-term dependencies using long short-term memory networks or temporal convolutional networks, and finally reconstructing the error and residual features using an autoencoder together with the simulation-measured residuals to obtain a high-order feature vector.

[0023] Preferably, the step of constructing and training a data-driven security assessment model using the feature vector as input, so that the security assessment model can output the health index or failure probability distribution of devices and sites, to obtain the trained security assessment model, includes:

[0024] The security assessment model is obtained by training it using a semi-supervised or weakly supervised strategy.

[0025] The loss function for training the semi-supervised or weakly supervised strategy includes reconstruction loss, classification loss, regression loss, and a regularization term based on physical constraints, and the output is probabilistically calibrated to form a reliable failure probability distribution or health index.

[0026] Preferably, the process of processing the real-time collected data to obtain the current high-order feature vector, inputting the current high-order feature vector into the safety assessment model to calculate the current safety index or failure probability, triggering an early warning when the safety index meets preset or adaptive triggering conditions, and performing root cause analysis and maintenance strategy simulation based on the output of the safety assessment model combined with digital twin simulation to generate operation and maintenance suggestions for optimizing the maintenance plan, including:

[0027] An adaptive threshold strategy is used for early warning triggering;

[0028] The adaptive threshold is based on the exponentially weighted moving average (EWMA) or quantile sliding window and combined with model uncertainty. When the safety index or failure probability exceeds the adaptive threshold, an early warning is triggered.

[0029] The root cause analysis employs a back-mapping based on feature contribution (e.g., SHAP value) or an attention mechanism to locate possible anomalous sub-devices or operating conditions.

[0030] Preferably, the method further includes: based on the digital twin simulation, performing multi-objective simulation evaluation on multiple alternative maintenance strategies, solving the maintenance schedule with the objective of minimizing downtime and maintenance costs while maximizing system security, and generating and outputting operation and maintenance suggestions;

[0031] Incremental learning is implemented based on online observations and model predictions to adaptively update the security assessment model in response to concept drift.

[0032] Another aspect of the present invention is to provide a safety status assessment system for wind power booster substations based on digital twins, comprising:

[0033] Multi-source data acquisition and synchronous preprocessing module: used to acquire real-time observation data from the booster substation, and perform timestamp alignment, noise reduction, missing value imputation and outlier detection on the observation data to obtain a cleaned and time-aligned observation dataset;

[0034] Digital twin simulation operating condition generation module: Based on the physical parameters and operating conditions of the booster substation, it performs parameter scanning and fault simulation in the digital twin simulation model according to wind speed, load, topology and typical fault modes to obtain a labeled multi-condition simulation sample dataset.

[0035] Simulation and measured data fusion and augmentation module: used to perform domain alignment and semantic mapping between the observation dataset and the simulation sample dataset, and to augment the scarce fault samples using generative data augmentation methods to obtain a fused and augmented comprehensive dataset;

[0036] Spatiotemporal representation learning and feature extraction module: used to perform spatiotemporal feature learning on the comprehensive dataset, and extract high-order feature vectors reflecting equipment operation and hidden fault modes through deep representation learning methods;

[0037] Safety Status Assessment Model Training Module: Used to construct and train a data-driven safety assessment model with the feature vector as input, so that the safety assessment model can output the health index or failure probability distribution of the device or site, and obtain the trained safety assessment model.

[0038] The early warning triggering and operation and maintenance decision support module is used to process the real-time collected data to obtain the current high-order feature vector, input the current high-order feature vector into the safety assessment model to calculate the current safety index or failure probability, trigger an early warning when the safety index meets the preset or adaptive triggering conditions, and perform root cause analysis and maintenance strategy simulation based on the output of the safety assessment model and digital twin simulation to generate operation and maintenance suggestions and optimize the maintenance plan.

[0039] Compared with the prior art, the beneficial effects of the present invention are:

[0040] (1) This invention formally aligns high-fidelity digital twin simulation samples with field-measured samples and fuses them with physical constraints. It then employs a generative enhancement strategy based on physical constraints to construct a comprehensive sample set during the training phase that both follows physical laws and fits the field distribution. On one hand, this expands the model's learnable fault condition space, compensating for the lack of rare or extreme conditions in historical field samples. On the other hand, by introducing simulation physical constraints as training regularization, it ensures the consistency and interpretability of the model output in a physical sense, fundamentally reducing the risk of "black box" misjudgments and enhancing engineering applicability. This allows the safety status assessment method to remain robust and verifiable even when facing boundary conditions or rare faults, making it easier for engineers to trust and adopt the assessment results as a basis for operation and maintenance judgments.

[0041] (2) This invention employs multi-scale spatiotemporal representation learning to simultaneously encode instantaneous pulse features and slowly varying trend features, and simultaneously quantifies model uncertainty in the evaluation output, thereby establishing a "sensitive-reliable" dual judgment mechanism. This enables earlier identification of developing abnormal signs (such as continuous minor anomalies or trend degradation), thus achieving early warning; simultaneously, uncertainty is used as a gating criterion, allowing the system to automatically and conservatively handle situations where there are blind spots in model cognition or data deviations to suppress false alarms. This technology not only improves the timeliness of early warnings but also provides technical assurance in ensuring alarm reliability, promoting the transformation of early warning decision-making from experience-driven to risk perception-driven, thereby balancing safety and operational costs. Attached Figure Description

[0042] Figure 1This is a flowchart of a method for assessing the safety status of a wind power booster and collection station based on digital twins, according to the present invention.

[0043] Figure 2 This is a block diagram of a wind power booster and collection station safety status assessment system based on digital twins according to the present invention. Detailed Implementation

[0044] 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 some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0045] Example 1:

[0046] Please see Figure 1 As shown, a method for assessing the safety status of a wind power booster substation based on digital twins includes the following steps:

[0047] S1. Multi-source data acquisition and synchronous preprocessing: Collect real-time observation data (including voltage, current, temperature, vibration, oil temperature, partial discharge signal and environmental data such as wind speed and direction) from the booster station, and perform timestamp alignment, noise reduction, missing value imputation and outlier detection on the observation data to obtain a cleaned and time-aligned observation dataset.

[0048] In step S1, wavelet threshold denoising and outlier removal based on robust statistics are performed. Missing value imputation adopts a method combining temporal adjacent interpolation and model prediction. The differences between multi-source clocks are dynamically corrected to obtain the observation dataset with higher confidence.

[0049] S2. Generation of digital twin simulation conditions: Based on the physical parameters and operating conditions of the booster substation, parameter scanning and fault simulation are performed in the digital twin simulation model according to wind speed, load, topology and typical fault modes to obtain a labeled multi-condition simulation sample dataset.

[0050] In step S2, a multiphysics coupling model is constructed and Monte Carlo parameter sampling is carried out to cover boundary conditions; for rare fault conditions, a physics-aware GAN is used to augment the samples, resulting in the labeled multi-condition simulation sample dataset.

[0051] S3. Fusion and expansion of simulation and measured data: The observation dataset and the simulation sample dataset are aligned in domain and semantically mapped. Generative data augmentation methods are used to expand the scarce fault samples to obtain a fused and expanded comprehensive dataset.

[0052] In step S3, domain adaptive mapping and covariance alignment are adopted, and transfer learning is used to reduce the difference between the simulation and experimental domains. The expansion adopts a conditional generation network to generate samples according to the fault type to make up for the data imbalance, thus obtaining the comprehensive dataset.

[0053] S4. Spatiotemporal representation learning and feature extraction: Spatiotemporal feature learning is performed on the comprehensive dataset, and high-order feature vectors reflecting equipment operation and hidden fault modes are extracted through deep representation learning methods.

[0054] In step S4, features are extracted using a combination of multi-scale temporal networks and autoencoders. This includes: first, using convolutional and residual networks to extract local temporal segment features; then, using long short-term memory networks or temporal convolutional networks to learn long-term dependencies; and finally, using autoencoders to reconstruct errors and residual features together with simulation-measured residuals to obtain high-order feature vectors.

[0055] S5. Safety Status Assessment Model Training: Using the feature vector as input, construct and train a data-driven safety assessment model so that the safety assessment model can output the health index or failure probability distribution of equipment and sites, and obtain the trained safety assessment model.

[0056] In step S5, the security assessment model is trained using a semi-supervised or weakly supervised strategy.

[0057] The loss function for training the semi-supervised or weakly supervised strategy includes reconstruction loss, classification loss, regression loss, and a regularization term based on physical constraints, and the output is probabilistically calibrated to form a reliable failure probability distribution or health index.

[0058] S6. Online estimation, early warning triggering and operation and maintenance decision support: The real-time collected data is processed in the same way as in steps S1 to S4 to obtain the current high-order feature vector. The current high-order feature vector is input into the safety assessment model to calculate the current safety index or failure probability. When the safety index meets the preset or adaptive triggering conditions, an early warning is triggered. Based on the output of the safety assessment model, root cause analysis and maintenance strategy simulation are performed in combination with digital twin simulation to generate operation and maintenance suggestions and use them to optimize the maintenance plan.

[0059] In step S6, an adaptive threshold strategy is used to trigger the early warning.

[0060] The adaptive threshold is based on the exponentially weighted moving average (EWMA) or quantile sliding window and combined with model uncertainty. When the safety index or failure probability exceeds the adaptive threshold, an early warning is triggered.

[0061] The root cause analysis employs a back-mapping based on feature contribution (e.g., SHAP value) or an attention mechanism to locate possible anomalous sub-devices or operating conditions.

[0062] Specifically, the method further includes: based on the digital twin simulation, performing multi-objective simulation evaluation on multiple alternative maintenance strategies, solving the maintenance schedule with the objective of minimizing downtime and maintenance costs while maximizing system security, and generating and outputting operation and maintenance suggestions;

[0063] Incremental learning is implemented based on online observations and model predictions to adaptively update the security assessment model in response to concept drift.

[0064] As shown above, by formally aligning high-fidelity digital twin simulation samples with field-measured samples and fusing them with physically constrained data, and employing a generative augmentation strategy based on physical constraints, a comprehensive sample set that both follows physical laws and fits the field distribution can be constructed during the training phase. On the one hand, this expands the fault condition space that the model can learn, compensating for the lack of rare or extreme conditions in historical field samples; on the other hand, by introducing simulation physical constraints as training regularization, it ensures the consistency and interpretability of the model output in a physical sense, fundamentally reducing the risk of "black box" misjudgments and enhancing engineering applicability. This allows the safety status assessment method to remain robust and verifiable even when facing boundary conditions or rare faults, making it easier for engineers to trust and adopt the assessment results as a basis for operation and maintenance judgments.

[0065] Example 2:

[0066] Engineering implementation for routine operation and maintenance optimization: Hybrid simulation enhancement + CNN-LSTM representation.

[0067] This embodiment focuses on the safety status assessment of common wind power booster and collection stations (capacity: 3×200MVA main transformer / collection capacity), employing: real-time sensors + 12 months of historical operating data + digital twin simulation data (covering typical operating conditions and faults) → domain alignment and physical constraint GAN expansion → multi-scale temporal convolution + LSTM representation → semi-supervised training of classification / regression hybrid model → EWMA adaptive threshold early warning + simulation-driven operation and maintenance suggestions simulation comparison.

[0068] Objective: To achieve early warning, reduce false alarms, and optimize maintenance plans.

[0069] Step S1, Data Acquisition and Preprocessing:

[0070] Data types and sampling:

[0071] Voltage, current, power factor: sampling rate 1Hz;

[0072] Transformer oil temperature and winding temperature: sampling rate 0.1Hz (every 10s);

[0073] Partial discharge (PD) envelope and insulation resistance (periodic sampling): PD high-frequency envelope sampling at 10kHz (used for burst discharge capture, long-term sampling based on events);

[0074] Vibration sensing (bearings / accessories): 100Hz sampling rate (only in critical sub-devices);

[0075] Environment: Wind speed / wind direction / temperature / humidity: 1Hz.

[0076] Time synchronization and cleaning solutions:

[0077] All channels are unified to UTC timestamps, and the main acquisition is synchronized with GPS (NTP+ correction is used if GPS is unavailable);

[0078] Denoising methods: Wavelet thresholding (db4 wavelet, soft thresholding) is used for the electrical signal channel, and bandpass filtering (10–500Hz) is applied to the vibration channel.

[0079] Missing value imputation: Short missing values ​​(<5× sampling interval) are imputed by linear interpolation, and long missing values ​​are imputed by model prediction based on the nearest time period (ARIMA backoff);

[0080] Output: Cleaned and aligned dataset (size: 12 months, total sample size of major sensors approximately 1.1 × 10^7 records / channel, event truncation to form label windows).

[0081] Quality control: Calculate SNR and missing rate for each channel; remove sensor channels with a missing rate >20%.

[0082] Step S2, digital twin simulation working condition generation:

[0083] Digital twin modeling: Constructing an electromagnetic-thermal-mechanical coupled model (transformer electromagnetic model + thermodynamic heat transfer model + mechanical vibration module) and parameterizing it;

[0084] Simulation conditions: Parameter scans are performed based on wind speed (0–25 m / s) grid, load (no load–full load), external faults (such as grounding, short circuit, abnormal tap commutation) and common degradation modes (insulation aging, increased contact resistance).

[0085] Fault categories and quantities: Construct 8 typical fault scenarios (such as enhanced partial discharge, local overheating of windings, increased vibration before tripping, tap fault, etc.); simulate 500 time series samples for each type → label the complete simulation sample set, with an initial of about 4,000 samples (including time series + labels);

[0086] Output: A fully annotated simulation sample set (including normal / abnormal condition labels and fault start and end times).

[0087] Step S3, Fusion and expansion of simulation and measured data:

[0088] Domain alignment: Semantically map the dimensions of the cleaned and aligned dataset to the fully labeled simulation sample set (e.g., map the simulated winding temperature value from the simulation output to the measured temperature sensor channel); perform covariance alignment (CORAL) on the feature distribution to reduce domain differences.

[0089] Expansion: Physics-aware GAN (physical constraint loss term) was applied to the scarce fault samples in the fully labeled simulation sample set to generate an additional 3,000 fault samples; after merging, the dataset was obtained (total of approximately 1.2 × 10^4 labeled samples + several tens of thousands of unlabeled measured sequences for semi-supervised).

[0090] Step S4, Spatiotemporal Representation Learning and Feature Extraction:

[0091] Method: First, 1D convolution (Conv1D) is used to extract local temporal segment features (convolutional layer: 64→128 filters, kernel=5), and then bidirectional LSTM (hidden=128) is used to capture long-term dependencies; at the same time, an autoencoder is used to learn the reconstruction error (latent dim=64) and as an anomaly-sensitive feature.

[0092] Output feature dimension: 256-dimensional high-order feature vectors are obtained for each time window (length 60s).

[0093] Normalization: Batch standardization and PCA (dimensionality reduction to principal component 64) are performed on the 256-dimensional high-order feature vector to decouple redundant information.

[0094] Step S5, Safety Status Assessment Model Training:

[0095] Model structure: Fusion reconstruction autoencoder + classification network (fully connected 64→32→output), the output includes: fault probability distribution pfault (multi-class softmax) and regression health score hraw (linear regression output);

[0096] Loss function: L = α * cross-entropy + β * regression MSE + γ * reconstruction error + δ * physical constraint regularization term (α..δ weights are set to 1.0, 0.5, 0.3, 0.2);

[0097] Training strategy: Semi-supervised (1.2×10^4 labeled samples, unlabeled measured sequences with consistency regularization); Training set / validation / test split 70 / 15 / 15; Optimizer Adam lr = 1e-3, batch = 256, epochs = 80, early stopping patience = 10.

[0098] Probability calibration: Perform Platt scaling (or temperature scaling) on the validation set to calibrate pfault.

[0099] Output: Trained model (for online inference).

[0100] Step S6, Online estimation, warning trigger and operation and maintenance decision-making:

[0101] Online process: Real-time data is processed according to Step S1 to obtain the current features, and the current features are input into the trained safety status evaluation model to obtain pfault, hraw;

[0102] Define the health H = 100×sigmoid(αh(1 - hrawnorm)), or H = 100×(1 - pfaultmax) for the single most likely fault;

[0103] Trigger strategy: Calculate the moving average μt and standard deviation σt of EWMA (window 6h), adaptive threshold Tt = μt + k·σt (k = 3 commonly used), if the safety index (or health) < Tt, then trigger a warning; at the same time, calculate the model uncertainty U (MC-dropout or softmax entropy), when U is high, increase the warning confidence threshold to reduce false alarms.

[0104] Root cause location: Use SHAP values to calculate feature contributions and locate the most likely abnormal sub-device (e.g., the contribution ratio of winding temperature rise is 45% → indicating local overheating of the transformer).

[0105] Operation and maintenance simulation decision-making: Call the digital twin to simulate three maintenance plans (emergency shutdown / fixed-point replacement / online repair) based on the "current state" and calculate: expected downtime, direct cost, residual risk, and output operation and maintenance suggestions.

[0106] Data and test design:

[0107] Data volume: Observation data set with 12 months of historical time series alignment (about 1.1×10^7 records per channel, the training sample size after selecting key channels is about 1.2×10^4 labeled windows + 5×10^5 unlabeled windows).

[0108] Fault injection and test set composition: For 8 types of faults, there are 80 actual factory historical events (of different severity) and 1200 simulated events; the test set contains 200 independent events (including 2000 normal windows).

[0109] Evaluation indicators and calculations:

[0110] Precision = TP / (TP+FP); Recall = TP / (TP+FN); F1 = 2·Precision·Recall / (Precision+Recall).

[0111] False alarm rate (FPR) = FP / (TN + FP).

[0112] LeadTime = Alarm Time - Event Occurrence Time (positive if the alarm occurs before the event).

[0113] Maintenance cost savings are calculated as follows: If malfunctions are avoided through advance maintenance, the savings from downtime costs are combined with the savings from optimized maintenance sequence.

[0114] Training hyper-participation computing resources: GPU (NVIDIA V100×1), training time approximately 6 hours.

[0115] As shown above, multi-scale spatiotemporal representation learning simultaneously encodes instantaneous pulse features and slowly varying trend features, and quantifies model uncertainty in the evaluation output, thereby establishing a "sensitive-reliable" dual judgment mechanism. This enables earlier identification of developing abnormal signs (such as continuous minor anomalies or trend degradation), thus achieving early warning. Simultaneously, uncertainty is used as a gating criterion, allowing the system to automatically and conservatively handle situations in model cognitive blind spots or data deviations to suppress false alarms. This technology improves the timeliness of early warnings and provides technical assurance for alarm reliability, promoting a shift in early warning decision-making from experience-driven to risk-awareness-driven, thereby balancing security and operational costs.

[0116] Example 3:

[0117] For high-precision early warning and uncertainty quantification, high sampling + ensemble model + active learning:

[0118] This embodiment further enhances upon Embodiment 2 by employing: a higher sampling rate for key channels, a more comprehensive sensor layout (more vibration / partial discharge points), a model using a hybrid TCN / Transformer approach with model ensemble integration, and the introduction of uncertainty estimation (MC-dropout + deep ensemble). Simultaneously, an active learning strategy is used for on-site annotation and iterative training of rare faults. The goal is to further improve detection accuracy and early warning time within a limited cost.

[0119] Step S1, Data Acquisition and Preprocessing:

[0120] Improvements: Vibration sampling of key parts is increased to 200Hz; high-frequency original waveforms (10kHz) of partial discharge events are retained for a long time and extracted according to events; infrared thermal imaging timed sampling (every 5 minutes) is added as an auxiliary temperature channel.

[0121] Other preprocessing is the same as in Example 2, but with the addition of: Dynamic Time Warping (DTW) as an alignment aid during timing alignment (for multi-rate channel alignment).

[0122] Output the measured observation dataset: a 12-month high-frequency mixed dataset (the size has increased by about 1.8 times).

[0123] Step S2: Digital twin simulation case generation (generating labeled simulation sample dataset)

[0124] Improvements: The simulation is expanded to 12 types of fault scenarios, and Monte Carlo sampling is performed 5000 times for extreme boundary conditions; the number of synthetic fault samples is expanded to 6000 using physics-aware GAN (richer than in Example 1).

[0125] Output labeled simulation sample dataset: approximately 7000+ labeled simulation samples.

[0126] Step S3: Fusion and expansion of simulation and measured data

[0127] Improvements: A Domain-Adversarial Neural Network is used to reduce the simulation-measurement discrepancy; and a conditional generation network is used to synthesize samples in a fine-grained manner according to fault subtypes. The synthesized dataset contains 2.4 × 10^4 labeled samples plus a large number of unlabeled windows.

[0128] Step S4, Spatiotemporal Representation Learning and Feature Extraction

[0129] Improvements: Temporal Convolutional Network (TCN) is used to extract long sequence dependencies, and Transformer layers are used for cross-channel attention fusion; the latent dim of the autoencoder is increased to 128.

[0130] Output high-order feature vector dimension: 512-dimensional, compressed principal components are 128-dimensional.

[0131] Step S5, Safety Status Assessment Model Training:

[0132] Model: The output set of three sub-models (TCN-Transformer, CNN-LSTM, and Graph-based TemporalNetwork) is used to predict and the final pfault and uncertainty U are obtained by Bayesian averaging or weighted averaging.

[0133] Training strategy: Semi-supervised + active learning: Submit 1000 events with high model uncertainty for manual annotation and iterative training.

[0134] Probabilistic calibration: Bayesian calibration using temperature scaling and post-processing.

[0135] Output safety status assessment model: a highly robust model with integration and uncertainty estimation.

[0136] Step S6: Online estimation, early warning triggering, and operation and maintenance decision-making

[0137] Triggering improvement: The threshold adopts EWMA + uncertainty weighting (increase the k value to reduce false alarms when U is high).

[0138] Root cause localization: In addition to SHAP, an attention weight visualization localization subsystem is introduced.

[0139] Operation and maintenance optimization: Perform multi-objective Pareto optimization in the digital twin (objectives: minimum downtime, minimum maintenance cost, minimum residual risk), and output a set of candidate operation and maintenance plans for scheduling selection.

[0140] Data and Test Design (Example 3):

[0141] Data volume: High-frequency data increases the number of training samples and unlabeled windows several times over; labeled samples are approximately 2.4 × 10^4.

[0142] Fault injection and test set: 12 types of faults, a total of 400 historical / field events + 2400 simulated / synthetic events; 400 independent test set events.

[0143] Training and computational resources: Multiple GPUs (2×V100), training time approximately 18 hours; computational resources for model ensemble and uncertainty estimation are slightly higher.

[0144] IV. Calculation method of evaluation indicators:

[0145] 1) Health score H calculation:

[0146] Let the model output a probability vector p = [p1, ..., pn] for n types of faults, and define the principal fault probability pmax = max i pi; then the health level can be defined:

[0147] H = 100×(1 - pmax) (If using the regression output h{raw}, a linear mapping H = 100×sigmoid(-α·(h{raw}-β)) can be done. Here, pmax is used for illustration)

[0148] 2) EWMA Adaptive Threshold:

[0149] EWMA Mean μ t Recursion:

[0150] μ t = λHt+(1 - λ)μ t-1 , λ = 0.2

[0151] The moving standard deviation σt is calculated similarly (or estimated using a moving window).

[0152] Threshold:

[0153] Tt = μt + kσt, k = 3

[0154] Judgment: If Ht < Tt and the uncertainty Ut < U{th}, then confirm the early warning (U{th} is, for example, 0.3).

[0155] 3) Early Warning Lead Time Calculation

[0156] For each test event i, define the actual fault occurrence time t{event,i} (defined by the protection device trigger or the device status log), the alarm time t{alarm,i} (the model first continuously triggers for Nmin = 3 consecutive windows), then LeadTimei = t{event,i} - t{alarm,i} (if it is positive, it is in advance).

[0157] Overall LeadTime = meani(LeadTimei).

[0158] 4) Estimation of Maintenance Cost Savings:

[0159] Assume that the average direct downtime loss for a single fault is Cidle (for example: 25000 yuan per incident), the cost of an emergency repair Cemg (10000 yuan), and the cost of a planned preventive maintenance Cplan (6000 yuan). If an emergency shutdown is avoided through early warning, the savings is approximately Cidle + Cemg - Cplan.

[0160] Statistically calculate the mean for multiple events and calculate the percentage savings.

[0161] V. Test Results (Simulation + On-site Verification) and Comparative Analysis:

[0162] To provide a comparative analysis, three sets of experiments were conducted: a control group (monitoring the existing rule threshold), Example 2, and Example 3. All methods were evaluated on the same test set; statistical significance was assessed using an independent samples t-test (p < 0.05 was considered significant).

[0163] Table 1 below provides a comprehensive comparison (the control group is a commonly used industrial threshold rule system; Examples 1 and 2 are two implementations of the scheme in this application). The values ​​in the table are the mean values ​​from multiple independent trials (confidence intervals are omitted for brevity), and the improvement is expressed as a percentage change relative to the control group. "Reduction in downtime" and "reduction in maintenance costs" are both expressed as a percentage relative to the control group.

[0164] VI. Comprehensive Comparison Table (Control Group vs. Example 2 vs. Example 3)

[0165]

[0166]

[0167] The "reduction in maintenance costs" and "reduction in unplanned downtime" are both percentage estimates relative to the annual statistics of the control group (the control group consists of the current threshold monitoring and passive maintenance strategies). For example, if the total annual maintenance cost of the control group is 1,000,000 yuan, then Example 2 is estimated to save 90,000 yuan / year, and Example 3 is estimated to save 130,000 yuan / year (the values ​​are estimated based on the local downtime costs and maintenance expenses).

[0168] As can be seen from the above, this invention tightly couples online evaluation results with digital twin simulation, mapping the real-time state back to the twin model. Based on the current state, it simulates multiple maintenance or emergency response plans in parallel, and uses multi-objective evaluation to rank the plans and output quantitative operation and maintenance recommendations. On the one hand, it transforms the prediction results into verifiable and comparable simulation decision-making basis, making the operation and maintenance recommendations physically verifiable and traceable. On the other hand, by comparing the risks and costs of different strategies through simulation (within a multi-objective framework), it achieves optimized selection of maintenance timing and response plans, thereby reducing unnecessary downtime and maintenance costs while ensuring equipment safety. In addition, the response results serve as a feedback loop input into the data and model update process, realizing a closed loop of prediction-response-learning, and continuously improving evaluation and decision-making capabilities.

[0169] Example 4:

[0170] A safety status assessment system for wind power booster substations based on digital twins, comprising:

[0171] Multi-source data acquisition and synchronous preprocessing module: used to acquire real-time observation data from the booster substation, and perform timestamp alignment, noise reduction, missing value imputation and outlier detection on the observation data to obtain a cleaned and time-aligned observation dataset;

[0172] Digital twin simulation operating condition generation module: Based on the physical parameters and operating conditions of the booster substation, it performs parameter scanning and fault simulation in the digital twin simulation model according to wind speed, load, topology and typical fault modes to obtain a labeled multi-condition simulation sample dataset.

[0173] Simulation and measured data fusion and augmentation module: used to perform domain alignment and semantic mapping between the observation dataset and the simulation sample dataset, and to augment the scarce fault samples using generative data augmentation methods to obtain a fused and augmented comprehensive dataset;

[0174] Spatiotemporal representation learning and feature extraction module: used to perform spatiotemporal feature learning on the comprehensive dataset, and extract high-order feature vectors reflecting equipment operation and hidden fault modes through deep representation learning methods;

[0175] Safety Status Assessment Model Training Module: Used to construct and train a data-driven safety assessment model with the feature vector as input, so that the safety assessment model can output the health index or failure probability distribution of the device or site, and obtain the trained safety assessment model.

[0176] The early warning triggering and operation and maintenance decision support module is used to process the real-time collected data to obtain the current high-order feature vector, input the current high-order feature vector into the safety assessment model to calculate the current safety index or failure probability, trigger an early warning when the safety index meets the preset or adaptive triggering conditions, and perform root cause analysis and maintenance strategy simulation based on the output of the safety assessment model and digital twin simulation to generate operation and maintenance suggestions and optimize the maintenance plan.

[0177] The beneficial effects of this method are the same as those of the embodiment of the safety status assessment method for wind power booster substations based on digital twins, and will not be repeated here.

[0178] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for assessing the safety status of a wind power booster and collection station based on digital twins, characterized in that, Includes the following steps: Real-time observation data from the booster and collection station is collected, and the observation data is subjected to timestamp alignment, noise reduction, missing value imputation, and outlier detection to obtain a time-aligned observation dataset. Based on the physical parameters and operating conditions of the booster substation, parameter scanning and fault simulation are performed in the digital twin simulation model according to wind speed, load, topology and typical fault modes to obtain a labeled multi-condition simulation sample dataset. The time-aligned observation dataset and the simulation sample dataset are domain aligned and semantically mapped. Generative data augmentation methods are used to expand the scarce fault samples to obtain a fused and expanded comprehensive dataset. Spatiotemporal feature learning is performed on the comprehensive dataset, and high-order feature vectors reflecting equipment operation and hidden fault modes are extracted through deep representation learning methods; Using the feature vector as input, a data-driven security assessment model is constructed and trained, enabling the security assessment model to output the health index or failure probability distribution of devices and sites, thus obtaining the trained security assessment model. The real-time collected data is processed to obtain the current high-order feature vector. The current high-order feature vector is input into the safety assessment model to calculate the current safety index or failure probability. When the safety index meets the preset or adaptive triggering conditions, an early warning is triggered. Based on the output of the safety assessment model, root cause analysis and maintenance strategy simulation are performed in combination with digital twin simulation to generate operation and maintenance suggestions and optimize the maintenance plan.

2. The method for assessing the safety status of a wind power booster and collection station based on digital twins as described in claim 1, characterized in that, The real-time observation data from the collected booster and collection station is then processed by timestamp alignment, noise reduction, missing value imputation, and outlier detection to obtain a cleaned and time-aligned observation dataset, including: Wavelet thresholding denoising and robust statistical outlier removal are performed. Missing value imputation adopts a method combining temporal adjacent interpolation and model prediction. Dynamic correction is performed on multi-source clock differences to obtain the observation dataset with higher confidence. The observation data includes voltage, current, temperature, vibration, oil temperature, partial discharge signals, and environmental data such as wind speed and direction.

3. The method for assessing the safety status of a wind power booster and collection station based on digital twins as described in claim 1, characterized in that, Based on the physical parameters and operating conditions of the booster substation, parameter scanning and fault simulation are performed in a digital twin simulation model according to wind speed, load, topology, and typical fault modes, resulting in a labeled multi-condition simulation sample dataset, including: A multiphysics coupling model is constructed and Monte Carlo parameter sampling is carried out to cover boundary conditions; physical constraint generative adversarial network is used to enhance the samples for rare fault conditions, and the labeled multi-condition simulation sample dataset is obtained.

4. The method for assessing the safety status of a wind power booster substation based on digital twins as described in claim 1, characterized in that, The process involves performing domain alignment and semantic mapping between the observed dataset and the simulated sample dataset, and then using generative data augmentation methods to augment scarce fault samples to obtain a fused and augmented comprehensive dataset, including: Domain adaptive mapping and covariance alignment are adopted, and transfer learning is used to reduce the difference between the simulation and experimental domains. The augmentation uses a conditional generation network to generate samples according to the fault type to make up for the data imbalance, thus obtaining the comprehensive dataset.

5. The method for assessing the safety status of a wind power booster and collection station based on digital twins according to claim 1, characterized in that, The process of performing spatiotemporal feature learning on the comprehensive dataset, and extracting high-order feature vectors reflecting equipment operation and hidden fault modes through deep representation learning methods, includes: Feature extraction is achieved by combining multi-scale temporal networks and autoencoders. This involves first extracting local temporal segment features using convolutional and residual networks, then learning long-term dependencies using long short-term memory networks or temporal convolutional networks, and finally reconstructing the error and residual features using an autoencoder together with the simulation-measured residuals to obtain a high-order feature vector.

6. The method for assessing the safety status of a wind power booster and collection station based on digital twins according to claim 1, characterized in that, The process of constructing and training a data-driven security assessment model using the feature vector as input, enabling the security assessment model to output health indicators or failure probability distributions of devices and sites, and obtaining the trained security assessment model includes: The security assessment model is obtained by training it using a semi-supervised or weakly supervised strategy. The loss function for training the semi-supervised or weakly supervised strategy includes reconstruction loss, classification loss, regression loss, and a regularization term based on physical constraints, and the output is probabilistically calibrated to form a reliable failure probability distribution or health index.

7. The method for assessing the safety status of a wind power booster and collection station based on digital twins according to claim 1, characterized in that, The process involves processing the real-time collected data to obtain a current high-order feature vector, inputting this high-order feature vector into the safety assessment model to calculate the current safety index or failure probability, triggering an early warning when the safety index meets preset or adaptive triggering conditions, and performing root cause analysis and maintenance strategy simulation based on the output of the safety assessment model combined with digital twin simulation to generate operation and maintenance suggestions for optimizing maintenance plans, including: An adaptive threshold strategy is used for early warning triggering; The adaptive threshold is based on an exponentially weighted moving average or a quantile sliding window and combined with model uncertainty. An early warning is triggered when the safety index or failure probability exceeds the adaptive threshold. The root cause analysis employs a back-mapping method based on feature contribution or attention mechanisms to locate possible anomalous sub-devices or operating conditions.

8. The method for assessing the safety status of a wind power booster substation based on digital twins according to claim 1, characterized in that, The method further includes: based on the digital twin simulation, performing multi-objective simulation evaluation on multiple alternative maintenance strategies, solving the maintenance schedule with the objective of minimizing downtime and maintenance costs while maximizing system security, and generating and outputting operation and maintenance suggestions; Incremental learning is implemented based on online observations and model predictions to adaptively update the security assessment model in response to concept drift.

9. A safety status assessment system for wind power booster substations based on digital twins, characterized in that, include: The multi-source data acquisition and synchronous preprocessing module is used to acquire real-time observation data from the booster substation and perform timestamp alignment, noise reduction, missing value imputation, and outlier detection on the observation data to obtain a cleaned and time-aligned observation dataset. The digital twin simulation condition generation module is used to perform parameter scanning and fault simulation in the digital twin simulation model based on the physical parameters and operating conditions of the booster substation, according to wind speed, load, topology and typical fault modes, to obtain a labeled multi-condition simulation sample dataset. The simulation and measured data fusion and augmentation module is used to perform domain alignment and semantic mapping between the observation dataset and the simulation sample dataset, and to augment the scarce fault samples using a generative data augmentation method to obtain a fused and augmented comprehensive dataset. The spatiotemporal representation learning and feature extraction module is used to perform spatiotemporal feature learning on the comprehensive dataset and extract high-order feature vectors that reflect equipment operation and hidden fault modes through deep representation learning methods. The safety status assessment model training module is used to construct and train a data-driven safety assessment model with the feature vector as input, so that the safety assessment model can output the health index or failure probability distribution of the device or site, and obtain the trained safety assessment model. The early warning triggering and operation and maintenance decision support module is used to process the real-time collected data to obtain the current high-order feature vector, input the current high-order feature vector into the safety assessment model to calculate the current safety index or failure probability, trigger an early warning when the safety index meets the preset or adaptive triggering conditions, and perform root cause analysis and maintenance strategy simulation based on the output of the safety assessment model and digital twin simulation to generate operation and maintenance suggestions and optimize the maintenance plan.

Images