Recognition method for microgrid switching misoperations based on multi-source heterogeneous data fusion

Abstract

To enhance the safety of microgrid switching and the identification of misoperations, we propose Time-Synchronized Misoperation Recognition (TS-MR), a method tailored to switching operations. The approach performs rule-based pre-screening grounded in operating procedures and anti-misoperation interlocking, achieves millisecond-level time synchronization of multi-source heterogeneous data via a two-stage scheme that combines variational Bayesian inference with a UKF, and employs a fusion of a Transformer, a TCN, and a GNN for cross-modal representation with interpretable discrimination. Laboratory records constitute the training, validation, and test sets; HIL data are used solely for independent cross-validation; and public datasets are used only for cross-domain robustness calibration, and none contributes to training, validation, or threshold tuning. Under a unified evaluation protocol, TS-MR attains 94.69% accuracy and an AUC of 0.977 in typical switching scenarios; end-to-end latency is about 80 ms; the core forward-pass latency is about 42 ms; the per-inference latency is 55 ms; and computational complexity is about 3.4 GFLOPs. Compared with CNN-BiLSTM, ConvLSTM, and GAT under identical preprocessing, time synchronization, and fixed random seeds, TS-MR improves accuracy by 0.9 to 3.7% points and AUC by 0.024 to 0.057. These results indicate that TS-MR provides high-confidence misoperation recognition and interpretable assessment for microgrid switching while satisfying engineering-grade real-time constraints.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-27120-8.

Introduction

With the large-scale integration of renewable energy, microgrids exhibit stronger state coupling and higher uncertainty during critical operations such as grid-connected and islanded switching, topology reconfiguration, energy-storage coordination, and load transfer¹. Deviations in the operating sequence, interlocking logic, or state recognition can trigger misoperations, leading to voltage excursions, frequency oscillations, and relay malfunctions; in severe cases, cascading outages may ensue^2,3. In practice, the combined effects of communication jitter, network latency, meteorological disturbances, and equipment failures make cross-source time mismatch a major source of false judgments. Experimental and replay studies indicate that when multi-source time-synchronization errors reach the millisecond scale, event-level recognition performance degrades sharply^4,5. Consequently, traditional mechanical interlocks, electrical blocking, and manual procedural verification are insufficient for data-intensive, disturbance-prone microgrids that are increasingly instrumented with edge intelligence^6,7.

Existing research has advanced along two main directions. (i) Rule–data collaboration. Operating procedures and interlocking logic are formalized as non-negotiable safety boundaries to rapidly remove explicit violation sequences. In parallel, probabilistic risk priors are introduced to characterize environmental uncertainty and operational risk intensity; their role is confined to weighting during training and threshold calibration and does not replace rule judgments or interlocking logic, preserving both auditability and adaptability^8,9. (ii) Two-stage time synchronization. Engineering practice commonly employs IEEE 1588 Precision Time Protocol (PTP) and timestamping based on phasor measurement units (PMUs); however, when jitter, packet loss, and sampling-rate mismatch coexist, single-stage synchronization leaves residual bias^10,11. A two-stage strategy is therefore adopted: fixed or slowly varying offsets are first removed, and then time-varying residuals are compensated using variational Bayesian inference and an Unscented Kalman Filter (UKF). Synchronization accuracy and robustness are summarized with quantile statistics to capture central tendency and tail risk, alongside magnitude comparisons to standard synchronization references for engineering interpretability^12,13.

On representation and discrimination, early single-modality time-series models (for example, hybrids of CNN and BiLSTM, or a convolutional network coupled with an LSTM) showed some robustness for power-event detection but remained limited in capturing semantics across modalities and long-range dependencies¹⁴. Attention-centric multimodal frameworks have since emerged: the Transformer aligns semantics across modalities and weights features; the Temporal Convolutional Network (TCN) efficiently models sequential dynamics; and the Graph Neural Network (GNN) explicitly encodes equipment and circuit topology, strengthening causal association and localization^15,16. Recent work tends to fuse these approaches to form a unified representation of semantics, temporal patterns, and topology, together with interpretability tools such as attention heatmaps, permutation importance, and checks for evidence consistency, thereby improving auditability and traceability for operations and maintenance^17,18.

Regarding data and evaluation transparency, best practices favor real and laboratory measurements as the primary data, supplemented by hardware-in-the-loop (HIL) and simulated data to cover boundary conditions, with open datasets for cross-domain robustness calibration. It is advisable to disclose, by partition, each subset’s source, proportion, time window, sampling rate, and labeling protocol; to define training, validation, and test splits that preclude leakage; and to report metrics separately for real versus augmented or public subsets to ensure extrapolatability and reproducibility^19–21. At the engineering-deployment level, collaboration between the cloud and the edge is increasingly prevalent: the edge executes low-latency, real-time inference, while the cloud handles retraining, evaluation, and model distribution. Evaluation typically reports computational complexity (GFLOPs), end-to-end latency, and quantile-based stability, guiding model compression, distillation, and operator-level optimization^22–24. In parallel, data-driven cyber-protection has progressed (for example, false-data-injection detection and localization in DC microgrids using subspace identification and residual generators), but its scope—cyberattacks and data tampering—differs from the switching-misoperation task addressed here, which hinges on multi-source time synchronization and procedural-logic embedding²⁵.

To address these challenges, we propose Time-Synchronized Misoperation Recognition (TS-MR), following the principle of rule-driven filtering, two-stage time synchronization, and multimodal cooperative discrimination. Procedural and interlocking logic provide deterministic screening; time synchronization uses a dual-stage scheme that combines variational Bayesian inference with a UKF to suppress cross-domain timestamp bias, with robustness summarized by quantile statistics and compared in magnitude with industrial synchronization references. The discrimination layer integrates the complementary strengths of the Transformer, TCN, and GNN to jointly model contextual semantics and equipment topology. To avoid replacing deterministic safety rules with statistical risk, a risk module based on Monte Carlo simulation is used only as a training-phase prior for sample weighting and threshold scheduling and is excluded from rule or interlock compliance decisions.

Contributions

1. Decoupled rule filtering and probabilistic priors. We build a rule-driven anti-misoperation and interlocking module and restrict Monte Carlo usage to training-phase risk priors for sample weighting and threshold scheduling, ensuring independence from deterministic rule judgments.

2. Two-stage time synchronization with engineering comparators. A variational Bayesian and UKF scheme achieves millisecond-level multi-source synchronization; we report alignment-error quantiles (central tendency and tail risk) and provide magnitude comparisons against IEEE 1588 (PTP) and PMU-based references.

3. Multimodal fusion with interpretability. A collaborative framework that integrates the Transformer, TCN, and GNN unifies semantic information and topology; attention heatmaps, permutation importance, and node-contribution scores support operator-level interpretation and traceability.

4. Partitioned data and a unified evaluation protocol. We curate a dataset with real and laboratory measurements as the main body and public data for augmentation; we report results separately for real versus augmented subsets and adopt a reproducible protocol that discloses dataset versions, time windows, preprocessing and synchronization workflows, and random seeds, while reporting efficiency indicators, including GFLOPs and end-to-end latency, for comparability.

Mathematical modelling

TS-MR comprises four stages: (i) rule-driven pre-screening, (ii) two-stage time synchronization, (iii) multimodal representation and fusion, and (iv) decision and constrained optimization. The overall workflow is shown in Fig. 1; the solution workflow is illustrated in Fig. 2.

1. Rule-driven pre-screening: Using interlocking logic, authorization consistency, temporal safety margins, and load and communication boundaries, this stage establishes verifiable operational constraints and risk descriptors. It directly outputs three labels—Pass, Violation, Review—and preserves traceable evidence. Hydrogen-related anomalies are incorporated into the risk channel as indicators with weights to quantify their impact on equipment stability and multi-energy coupling.

2. Two-stage time synchronization: A global mean drift is first removed, followed by dynamic residual compensation to suppress desynchronization caused by switching transients; a synchronization confidence measure is also provided. Magnitude comparisons with IEEE 1588 and GPS/UTC-disciplined PMU synchronization are reported in the Methods and Results as engineering references rather than replacements for industrial time synchronization.

3. Multimodal representation and fusion: Cross-modal attention with a Transformer aligns and weights information; TCNs extract the dynamics of operational sequences; GNNs encode equipment topology and interlocking constraints. The fused representation supports discrimination and yields interpretable Top-K evidence (for example, attention heatmaps and node-contribution scores) to facilitate operations and maintenance verification and traceability.

4. Decision and constrained optimization: Within the feasible domain defined by rule-consistency and synchronization-consistency constraints, the method enforces budgets on end-to-end latency and computational complexity (GFLOPs) and outputs the misoperation probability and risk level, balancing accuracy and robustness. The Monte Carlo risk module is used only as a training-phase prior for risk weighting and sample reweighting; it does not participate in rule screening or real-time compliance decisions.

Fig. 1.

TS-MR architecture. Stage (1) rule-driven pre-screening; Stage (2) two-stage time synchronization (mean-bias calibration and dynamic residual compensation); Stage (3) multimodal representation and fusion (Transformer, TCN, GNN); Stage (4) decision and optimization under feasibility, rule-consistency, and synchronization-consistency constraints with latency and GFLOPs budgets. Monte Carlo-based risk assessment is used only as a training prior.

Fig. 2.

TS-MR solution workflow. The left column shows constraint–feature modeling (multi-source); the right column shows the PGD-based optimization. The outputs are the misoperation probability and the corresponding risk level . Model coefficients are calibrated offline using a hold-out validation and k-fold cross-validation protocol; see Supplementary Table S2.

Rule-driven anti-misoperation model

To quantify the rule-compliance risk of switching commands, five components are defined: equipment state, operational sequence, voltage–communication coordination, assembly and disassembly compliance, and multi-energy coupling. The overall objective function is

1

where denotes the overall rule-compliance risk; the -th risk component; and a non-negative weight calibrated from historical statistics and operation-maintenance experience.

1. Equipment-state risk

   2

   where is the equipment-state risk; is the -th sub-risk item and its weight; is the expected sampling or sequencing jitter; is the weight of the cloud–edge interaction cost ; is the hydrogen-energy anomaly-intensity index; and is its influence weight within the equipment-state channel.
2. Operational-sequence risk

   3

   where is the operation-sequencing risk; is the weight of sub-component ; is the -th operation-sequencing sub-risk, reflecting the deviation from the prescribed operation order; is the disturbance term of the control chain; and is the influence weight of hydrogen anomalies in the sequence-risk component.
3. Voltage–communication coordination risk

   4

   where is the integrated risk of voltage–communication coordination; denotes the risk of voltage or frequency limit violation; the risk induced by delay and jitter in the communication chain; and the transformer–breaker coordination risk. The coefficients ,, weight the respective sub-risks.
4. Assembly and disassembly procedure–compliance risk

   5

   where is the compliance risk of assembly and disassembly procedures; is the -th checklist sub-risk; is the corresponding item weight; denotes the compliance-related disturbance; is the influence weight of the hydrogen anomaly within the compliance risk; and denotes a generic disturbance term.
5. Multi-energy coupling risk

   6

   where is the multi-energy coupling risk; is the set of energy types; is the availability probability of energy type at node ; is the node weight; and represents the uncertainty induced by power surges or output fluctuations.
6. Hydrogen-anomaly propagation mechanism

   7

   where is the Sigmoid function; and are hyperparameters fitted on the validation set to ensure that decreases monotonically as increases. Here is the hydrogen anomaly intensity index, obtained by normalizing and weighting multi-signal evidence after time synchronization. Relevant evidence includes leakage or concentration alarms, abrupt drops in hydrogen-storage pressure, reduced or abnormally falling PEM fuel-cell stack voltage, electrolyzer trips, and large power fluctuations. As rises, decreases, thereby reducing in (6) and increasing , reflecting the adverse impact of hydrogen anomalies on system stability.

7. Rule and engineering constraints

   1. Temporal constraint
      8

      where and are the off and on switching instants of device at operation step ; is the minimum safe interval between steps and ; denotes the operation uncertainty of device ; is the set of all devices; and is the set of operation steps.
8. Interlocking constraint

   9

   where is the binary state of circuit breaker (0 for open, 1 for closed); is the interlocking constraint associated with step ; is the interlock-risk weight; is the interlock threshold bias; is the hydrogen-anomaly weight within the interlocking constraint; and is the permissible interlock tolerance.
9. Load constraint

   10

   where is the real-time load of device ; is the rated load limit of device ; is the load-disturbance term; and is the hydrogen-anomaly weight for voltage–load coordination scenarios.
10. Time-synchronization deviation constraint

    11

    where is the synchronization deviation at time ; is the allowable upper bound at time ; is the error-compensation term at time ; and is the hydrogen-related time-synchronization weight.
11. Feature-consistency constraint

    12

    where is the feature-space distance between and ; is the difference in the -th feature dimension; is the feature-consistency threshold; denotes the threshold disturbance term, reflecting the uncertainty in threshold setting. is the expansion disturbance for feature-consistency deviation; and is the hydrogen-anomaly weight within the feature-consistency constraint.
12. Historical-consistency constraint

    13

    where is the execution time of the current operation; and are the historical mean and standard deviation; is the tolerance coefficient for historical deviation; is the disturbance term associated with ; and is the hydrogen-related weight under the historical-consistency scenario.

13. Assembly and disassembly authorization constraint

    14

    where is the authorization requirement for the -th operation; is the authorization record for the -th operation; is the global safety-margin disturbance; is the hydrogen-energy weight under the authorization scenario; and denotes the disturbance due to misjudgment or repeated authorization.
14. Complexity and real-time constraint

    15

    where is the per-inference FLOPs of the model at parameter set ; is the upper bound of allowable complexity; is the end-to-end inference latency; and is the latency limit; denotes the model parameter set. These bounds serve as engineering-level constraints to guide deployment–performance matching; they are set with reference to engineering practice in microgrid control and grid interconnection (e.g., IEEE 2030.7 and IEEE 1547) and are used solely for benchmarking, not as compliance declarations.^26,27

Two-stage time synchronization

During switching operations, discrepancies in sampling frequency and communication paths across SCADA, PMU, video, voice, and electronic operation tickets introduce temporal deviations of approximately 100–500 ms²⁸. To obtain a unified timestamp across modalities, a two-stage time synchronization mechanism is adopted. In the first stage, mean bias is eliminated via statistical averaging; in the second stage, dynamic residuals are compensated by combining Kalman filtering (UKF where applicable) with variational Bayesian inference. The entire synchronization process follows the precision-time specifications of IEEE 1588, IEC 61850–9-3, and GPS-PMU, meeting engineering-grade consistency requirements^29–31.

Before executing the two stages, an adaptive gating coefficient is defined to modulate the correction strength and to provide robust regularization information for the second stage:

16

where is the adaptive correction strength at time ; is the residual-correction weight; is the instantaneous deviation of the multi-source signals with respect to the PMU at time ; is the jitter-error correction term; is the system-state indicator at time ; is the Monte Carlo regularization weight; and is the Monte-Carlo-based robust regularizer used to enhance stability rather than for rule screening.

Stage 1: Mean-bias correction

17

where is the average time offset of channel ; and are the raw and PMU timestamps of the -th sample, respectively; is the total number of samples; and denotes the unified timestamp (for channel , sample ) after mean-bias correction.

Stage 2: Dynamic residual compensation

18

where is the synchronized timestamp of channel at sample ; is the AR(1) residual-model coefficient; is the previous-step residual estimate; is the prior covariance; and are the process-noise and observation-noise covariance matrices; is the current observation; and is the compensation term for time jitter and sampling quantization error.

Multimodal representation and fusion

19

where denotes the fused feature representation at layer ; is the cross-node attention weight, measuring the contribution of node to node ; and are the feature vectors of nodes and at layer ; is the risk-label weight for class ; is the indicator matrix for class ; is the projection weight matrix of layer ; is the bias vector; represents scenario-level uncertainty disturbance in multimodal settings; is the hydrogen-scenario fusion factor; and denotes the multi-scenario robustness expectation of the Monte Carlo term.

Model decision and optimization

This section completes the closed loop from probabilistic decision to constrained optimization. First, the misoperation probability is obtained by Eq. (20) to form a binary decision; then the global objective within the feasible engineering domain (Eqs. (22)–(23)) is solved using Projected Gradient Descent (PGD)³².

1. Probabilistic decision

   20

   where is the Sigmoid activation; is the attention weight; , are the feature vectors of nodes and ; is the bias; is the time-weighted disturbance; is the misoperation-risk factor under hydrogen scenarios; is the Monte Carlo–based scene-verification term; and is the generalized error term.

   21

   where is the risk level of node ; is the grade index; is the total number of grades; and are the probability thresholds.
2. PGD-based constrained optimization

   22

   where is the total loss; the classification loss; the Frobenius-norm regularization (); ,,, are the weights of the corresponding terms; is the compliance vector computed from Eqs. (8)–(14); is the rule-consistency loss; the time-synchronization penalty; the MCRA-prior regularizer; and the cloud-resource penalty.

   23

   where is the feasible engineering domain jointly defined by Eqs. (8)–(15); is the validation accuracy and its lower bound. To avoid interference with relay protection and in-station communication, the upper bound of inference latency is set to the engineering-grade deployment limit (batch = 1). Computational complexity is bounded by . The threshold is determined from the ROC operating point (, , F1-max). Regularization and latency limits reference IEC 61,850–5 and IEEE 1547–2018 time scales.

PGD-based optimization procedure

1. Initialization: Set , learning rate , and iteration index .

2. Gradient update

   24

   where is the unprojected update; the parameters at iteration ; the learning rate; the gradient; and the update-noise disturbance.
3. Projection mapping

   25

   where is the projection operator onto ; is the projection-uncertainty disturbance.
4. Learning-rate adjustment

   26

   where is the learning rate at iteration ; , are its bounds; is the annealing period; the adjustment-noise term; the hydrogen-scenario learning-rate correction; and the multi-scenario consistency term.

5. Early stop: If validation accuracy shows no improvement for consecutive windows and , terminate optimization and output the optimal parameter set .

6. Output. Using for inference, obtain the misoperation probability and the corresponding risk level for node (see Eqs. (20)–(21)). See Table1 for an end-to-end example.

Table 1.

Example of end-to-end calculation for device-state risk.

Step	Symbol	Description	Example value	Corresponding Eq
Step-1	–	Post-synchronization analysis window (the period after unified time alignment by Eqs. (16)–(18)), actual event window	1 s	(16)–(18)
Step-2		Sub-risk 1: load deviation (computed by Eq. (10), inserted into Eq. (2))	0.067	(2)
Step-3		Sub-risk 2: position inconsistency (computed by Eq. (12), inserted into Eq. (2))	0.020	(2)
Step-4		Sub-risk 3: interlock deviation (computed by Eq. (9), inserted into Eq. (2))	0.015	(2)
Step-5		Weights of the three sub-risks (normalized so that )		(2)
Step-6		Expected value of sampling and timing jitter	0.0015	(2)
Step-7		Cloud–edge interaction term	0.0020	(2)
Step-8		Impact of hydrogen anomalies on device-state risk (modeled and inserted via Eqs. (2) and (7))	0.0040	(2),(7)
Step-9		Aggregated device-state risk (combined by Eq. (2))	0.0505	(2)
Step-10		Weight of each risk type; here	0.40	(1)
Step-11	,,,	Remaining risk categories (this window not triggered)	0	(3)–(6)
Step-12		Aggregated rule-level risk (combined by Eq. (1))	0.0202	(1)
Step-13	–	Engineering-grade thresholds (Pass/Review/Block)	Pass < 0.010; Review 0.010–0.030; Block > 0.030	–

Experimental setup and datasets

Experimental environment

Computational experiments were run on an Intel Xeon Gold 6230 processor (20 cores, 2.1 GHz) with 128 GB RAM and an NVIDIA RTX A6000 GPU (48 GB VRAM). Simulation modeling was performed in MATLAB R2024b, and data processing and statistical analysis were conducted in Python 3.9 with PyTorch 2.0.1. At the software layer, an IEEE 1588 Precision Time Protocol (PTP) time synchronization link was established using ptpd v2.3.1 to evaluate time alignment and quantify synchronization errors; the resulting comparisons are provided as engineering references only and do not constitute a standards-compliance statement^33,34. To ensure reproducibility, the random seed was fixed at 12,345.

The physical platform is the distribution-network prototype system of the National Key Laboratory for Power Grid Disaster Prevention and Mitigation at Changsha University of Science and Technology. It adopts a single-bus architecture with approximately 20 grid-connected units (photovoltaics, wind generation, hydrogen storage, electrochemical storage, and typical loads). The system voltage level is 10 kV with a 0.63 MVA transformer; the total grid-connected rated capacity is about 300 kW, including PV 60 kW, WT 30 kW, BESS 150 kW, and H₂ 60 kW. PMUs and SCADA are deployed at major interconnection points and feeder outlets for phasor, frequency, and status acquisition as well as event alignment. Communication and time-synchronization interfaces include Modbus-TCP, IEC 61,850-GOOSE, and PTP .

Data sources and uses

This study employs three categories of data: laboratory real data for model training, validation, and testing; HIL data for independent cross-validation and robustness enhancement; and public datasets for cross-domain robustness calibration and distribution comparison. The composition and roles of each category are summarized in Table 2.

Table 2.

Data sources and roles.

Mode	Records(n)	Sampling rate and time coverage	Key features and description	Role in this study
SCADA and PMU	32,000	SCADA: 1 Hz (± 120 ms jitter); PMU: 60 sps; coverage: 2025–09-03 08:00–16:53 (approximately 8.9 h); both streams resampled to 1 Hz after the two-stage time-synchronization pipeline	Voltage and current phasors (magnitude and angle), system frequency, active and reactive power, switching status, and other operational quantities	Primary data source for model training, validation, and testing
Operation tickets	3200	Event-level; coverage: 2025–09-03 08:00–2025–09-04 07:59	Operation tasks, step sequence, device labels, scheduled actions, execution time, and authorization levels	Operation tasks, step sequence, device labels, scheduled actions, execution time, and authorization levels
Operator command speech	3200	Event-level; original 16 kHz audio; only event timestamps and recognition confidence retained	Voice commands and operator intent; trigger words and confidence scores	Used for semantic intent recognition and multimodal verification
Video events	3200	Event-level; original 30 fps video; only event timestamps and scene summaries recorded	Operator actions and scene context; visual events and contextual summaries	Used for visual-assisted verification and scene-consistency analysis
Environment and working conditions	288	5-min sampling; coverage: 2025–09-03 08:00–2025–09-04 07:55	Irradiance, wind speed, temperature, airflow, and load-monitoring variables of the external environment	Used for operational-boundary modeling and external-disturbance analysis
Risk labels	3200	Label-level; coverage: 2025–09-03 08:00–2025–09-04 07:59	Misoperation categories and risk levels (high, medium, and low)	Used for model supervision and evaluation benchmarking

1. Laboratory real data

   Data were collected from the laboratory microgrid platform and include SCADA and PMU records together with paired operation tickets, video, and voice events. The core SCADA and PMU window comprises 32 000 samples at 1 Hz with approximately ± 120 ms jitter, covering 2025–09-03 08:00 to 16:53; other modalities extend to 2025–09-04 07:59. All data were aligned under unified PTP or GPS/UTC time synchronization using a two-stage alignment mechanism. The original PMU signals (60 fps) were aligned to 1 Hz, and the SCADA 1 Hz sequence served as the sampling index. Video and voice streams retain only event timestamps and confidence scores. This subset constitutes the primary source for model training, validation, and testing.

2. Hardware-in-the-loop data

   Generated on the StarSim-HIL platform, this subset provides approximately 9 100 multi-modal records (about 20%) and is used solely for independent cross-validation and robustness enhancement.

3. Public datasets

   Two open datasets were used: the NREL Hydrogen Microgrid Dataset v1.2 (containing hydrogen operation data) and “Power-Data-from-Microgrid” (Kaggle, v2024, including power, voltage, frequency, temperature, and other microgrid/IoT sensor data), totaling about 4 600 records (approximately 10%)^35,36. These data are used only for cross-domain robustness calibration and distribution comparison and are excluded from training, validation, and testing. No SMOTE is applied to the public datasets.

Statistical characteristics and distribution comparison

After temporal windowing, the main electrical quantities are summarized by mean, standard deviation, minimum, and maximum in Table 3. Distribution differences across laboratory real data, HIL data, and public datasets are compared in Table 4.

Table 3.

Summary statistics of key electrical variables (windowed samples, n = 32 000).

Feature	Mean	Standard deviation	Minimum	Maximum
Voltage magnitude (p.u.)	0.9998	0.0198	0.9709	1.0367
Current magnitude (p.u.)	0.499	0.288	0.1534	0.989
Frequency (Hz)	60.029	0.0502	59.875	60.105
Active power (kW)	50.175	28.890	0.002	99.996
Reactive power (kvar)	− 0.071	28.936	− 49.999	50.000

Table 4.

Comparative statistics across data sources.

Data type	Mean voltage (p.u.)	Voltage standard deviation (p.u.)	Frequency range (Hz)	Average hydrogen flow rate
Laboratory real data	0.9998	0.32	59.875–60.105	30.07
Hardware-in-the-loop (HIL) data	1.001	0.32	59.88–60.10	28.85
Public dataset	1.005	0.35	59.70–60.30	31.02

Preprocessing workflow and class imbalance handling

1. Data cleaning. Outliers are removed using the three-sigma criterion together with scenario rules. Windows with a missing-value ratio ≥ 1% are discarded; remaining gaps are repaired by forward filling and linear interpolation.

2. Time synchronization. PMU timestamps serve as the unified time baseline. A two-stage mechanism (VB followed by UKF) is employed. On the validation set, the mean synchronization error is 1.8 ms, the standard deviation is 0.5 ms, and is 3.0 ms. Comparative evaluation is conducted against PTP and GPS-PMU references.

3. Feature engineering. Electrical quantities are standardized; video streams provide event counts and confidence; speech yields keywords and ASR confidence; operation tickets are parsed into operation sequences and authorization levels; environmental variables retain their original engineering units.

4. Windowing and alignment. Event-centered synchronized slicing is applied with a fixed window and step .

5. Partitioning and balancing. Laboratory real data are split 8:1:1 into training, validation, and test sets. SMOTE is applied only within the training set to balance minority classes. Public/open data are used solely for cross-domain robustness calibration and distribution comparison, and do not participate in main training, validation, or threshold setting.

Time-synchronization parameters and stopping criteria

A two-stage VB + UKF synchronization scheme is adopted with the following settings.

1. VB stage (mean-offset correction): Convergence threshold ; maximum 100 iterations; initial variance . Process and measurement noise variances are selected by grid search: ,.

2. UKF stage (dynamic residual compensation): Unscented-transform parameters: , . Sliding-window length ; the residual purification threshold is adaptively tuned.

3. Global stopping rule: Training stops when the change in the mean residual over five consecutive windows satisfies and the 95th-percentile error shows no further decrease.

HIL connectivity and interface verification

To evaluate the deployability of TS-MR under engineering-analogous conditions, a StarSim-HIL testbed was constructed to establish bidirectional coupling with the compute side, forming a hardware-in-the-loop verification environment (Fig. 3). The experimental platform comprises FPGA-based acquisition, timing-control, and host-controller units. The entire link adopts IEEE 1588 for a unified time base, and interfaces support Modbus-TCP and IEC 61,850-GOOSE.

Fig. 3.

StarSim-HIL testbed and interfaces.

I/O capabilities: 16 analog channels (1 MS/s, 16-bit, ± 10 V) and 32 digital channels (5 V TTL and ± 25 V). Test scripts cover scenarios such as forced interlock override, grounding-conductor removal, and unauthorized entry into an energized interval. During testing, I/O waveforms, event timestamps, and synchronization step offsets are recorded to verify connectivity and temporal-sequence consistency.

Experimental validation and analysis

Network architecture and training configuration

TS-MR adopts a multi-branch encoding–fusion–classification architecture. The TCN extracts temporal features from SCADA and PMU; the Transformer performs cross-modal semantic fusion; and the GNN models topology-aware interactions among equipment nodes. After feature extraction and concatenation, the fused representation is flattened by an MLP and passed to a softmax layer for class-weighted probability prediction. The baseline models (CNN-BiLSTM, ConvLSTM, and GAT) are trained under a unified search strategy with minor hyperparameter adjustments to ensure fair comparability. Key network and training parameters are summarized in Table 5. The average end-to-end inference latency per batch (batch = 1) is 80 ms (including input normalization and feature concatenation) . The core forward time is 42 ms ( is approximately 55 ms), and the computational complexity is 3.4 GFLOPs, consistent with Table 9 and used for engineering-grade benchmarking.

Table 5.

TS-MR network architecture and training configuration summary.

Parameter category	Parameter name	Value
Network architecture	Transformer layers	4
	Attention heads	8
	TCN convolutional layers	4
	GNN layers	3
	Hidden dimension	128
Training configuration	Learning rate	0.001
	Batch size	64
	Epochs	50
	Optimizer	Adam
	Dropout ratio	0.2
Efficiency metrics	Core forward latency (mean and )	42 ms and 55 ms
	End-to-end inference latency (batch = 1)	80 ms
	Peak memory usage	2.8 GB
	Computational complexity (GFLOPs)	3.4

Table 9.

Comprehensive performance and efficiency comparison under the unified evaluation protocol.

Model	Accuracy (%)	Precision (%)	Recall (%)	F1 (%)	AUC	End-to-end latency (ms)	GFLOPs
TS-MR	94.69	92.5	95.1	93.8	0.977	80	3.4
CNN-BiLSTM	93.8	92.4	93.2	92.8	0.953	108	2.1
ConvLSTM	92.3	91.2	91.8	91.5	0.941	102	2.7
GAT	91.0	89.8	90.5	90.1	0.920	92	1.4

Rule consistency and risk suppression

1. Risk distribution suppression after rule embedding

   According to the risk terms defined in Eqs. (1)–(7), the test-set tail distribution shifts left after applying the rule-consistency mechanism. The 95th-percentile risk score decreases from 0.65 to 0.55, indicating effective suppression of high-tail risks and mitigation of low-probability hazards (Fig. 4).

2. Classification improvement under a unified threshold.

   Using a fixed threshold derived from the “pre-rule” ROC curve at TPR = 0.90, overall accuracy increases from 87.1% to 92.5% after applying the rule constraints, with both the false positive rate (FPR) and the false negative rate (FNR) reduced. A McNemar test shows the improvement is statistically significant (). Detailed results are provided in Table 6.

   Table 7.

   Per-class metrics for typical misoperation scenarios.

   Operation type	Precision (%)	Recall (%)	F1 (%)	AUC
   Unauthorized grounding disconnection	93.2	95.1	94.1	0.985
   Forced interlock override	92.5	94.8	93.6	0.987
   Unauthorized command execution	94.1	93.2	93.6	0.963
   Load disconnection under power	90.4	95.3	92.8	0.912
   Busbar configuration error	90.6	94.5	92.5	0.974
   Unauthorized entry into an energized interval	93.8	94.1	93.9	0.905
   Equipment grounding omission	91.7	93.9	92.8	0.968

Fig. 4.

Comparison of risk distributions before and after rule embedding. Dashed lines denote the 95th percentile (): Before = 0.65; After = 0.55.

Table 6.

Recognition performance before and after rule constraints.

Model stage	Recognized misoperations (samples)	False positives	False negatives	Recognized normal operations	Accuracy (%)
Before rule embedding	8500	300	1500	3700	87.1
After rule embedding	9200	250	800	3750	92.5

Validation of the rule-driven anti-misoperation model

A fixed-threshold evaluation was adopted to assess time-synchronization adequacy and classification effectiveness. We first analyzed the pre-rule ROC and selected a unified cutoff on that curve (target TPR≈0.90; realized TPRs: pre 0.85, post 0.92). Using this cutoff, we computed confusion matrices for the pre- and post-rule conditions and quantified the changes in FPR, FNR, precision, and recall (in percentage points). To mitigate class-imbalance effects, the confusion matrices were row-normalized (Fig. 5).

Fig. 5.

Effectiveness of rule-consistency filtering. (a) ROC curves; the solid marker denotes the unified cutoff chosen on the pre-rule ROC, (b) Row-normalized confusion matrices (rows: true labels; columns: predictions); row totals are conserved, (c) Changes in FPR, FNR, precision, and recall (post minus pre) with 95% confidence intervals.

Under this cutoff, rule embedding markedly reduces FPR and improves both precision and recall, with changes remaining controllable and consistent with the engineering objective of prioritizing false-alarm suppression while maintaining recall. The figures validate the marginal effectiveness of the rule-consistency module rather than a system-wide optimum under a globally tuned threshold. Comprehensive system-level results and efficiency comparisons are reported in Table 9.

To ensure consistent timing for comparison, two-stage synchronization (VB followed by UKF) was applied to the validation set. Residual statistics: mean 1.8 ms; standard deviation 0.5 ms; 3.0 ms; maximum < 6 ms. The mean residual is close to zero, indicating no observable system drift. These values serve as engineering-grade timing references; IEEE 1588 (PTP) and GPS-PMU references are used for comparison only and do not constitute a compliance statement or a classification criterion.

Time-synchronization performance

1. Training Convergence and Stability Analysis

   To assess convergence and stability, three independent training runs were conducted under a fixed configuration, differing only in random seeds; early stopping was triggered by the validation loss. Figure 6 shows the trajectories of training/validation accuracy and loss. The validation set achieved a peak accuracy of 94.69%, and the validation loss was approximately 0.20, indicating no evident overfitting.

2. Multi-Class Recognition Performance

   Figure 7 presents the row-normalized confusion matrix for nine operational categories on the test set. Class-wise recall ranges from 93.2% to 95.2%, corresponding to a macro-averaged precision of 96.50%, macro-averaged recall of 94.40%, and a macro-averaged F1 of approximately 95%. Minor errors mainly arise from weak-normal samples occasionally predicted as “normal operation” (approximately 1–3% of cases) and mild confusions between insufficient interlock safety and signal delay. Overall, the model distinguishes the nine scenarios with high accuracy, though weak-normal conditions remain improvable.

3. Comparison across Typical Misoperation Scenarios

   To evaluate discriminative ability under representative misoperation scenarios, per-class Precision, Recall, F1, and AUC are listed in Table 7. TS-MR performs best on forced interlock override and unauthorized grounding disconnection, with AUC of 0.987 and 0.985, and corresponding F1 scores of 93.6% and 94.1%. For unauthorized entry into an energized interval, partial feature overlap with normal operation slightly degrades performance (AUC 0.905, F1 93.9%). Overall, the recognition performance satisfies engineering application requirements.

4. Correlation between Risk Levels and Misoperation Labels

   To verify the consistency between the risk score computed by Eqs. (1)–(7) and recognition outcomes, we adopt of the normal-state risk-score distribution on the validation set as the reference boundary. Risk thresholds are defined as: high for score ≥ 0.85, medium for 0.65 ≤ score < 0.85, and low for score < 0.65. As shown in Table 8, forced interlock override (0.88) and unauthorized grounding disconnection (0.85) fall into the high-risk category, whereas load disconnection under power scores 0.82 and is classified as medium risk. This mapping aligns with engineering expectations and supports alarm prioritization. Note: these thresholds apply to the task-level risk score and are distinct from the rule-layer engineering thresholds.

Fig. 6.

Training and validation curves. Blue lines denote step-wise training trajectories; black dots indicate epoch-wise validation aggregation. Vertical axes: accuracy (%) and loss; horizontal axis: training iterations.

Fig. 7.

Row-normalized confusion matrix on the test set.

Table 8.

Mapping between misoperation labels and risk levels.

Label ID	Misoperation type	Risk level	Reference risk score
1	Unauthorized grounding disconnection	High	0.85
2	Forced interlock override	High	0.88
3	Unauthorized command execution	Medium	0.73
4	Load disconnection under power	Medium	0.82
5	Busbar configuration error	Medium	0.75
6	Unauthorized entry into an energized interval	Medium	0.70

Recognition performance and explainability

To improve diagnostic efficiency, an XAI visualization module is integrated into the inference stage to output three types of information: (1) attention heatmaps highlighting Top-K evidence; (2) modality contribution quantified by permutation importance; and (3) a comparison of the risk-probability curve with the preset decision threshold, as shown in Fig. 8.

Fig. 8.

XAI-based explainability analysis. (a) Attention with Top-K evidence, (b) Modality contribution (permutation importance), (c) Evidence timeline: risk probability vs threshold; Top-K hit = 0.40; TLE = N/A.

In the illustrated case, the modality-level contribution weights are: SCADA 0.56, PMU 0.08, Video 0.10, Voice 0.19, and Ticket 0.07. The Top-K hit rate is 0.40 (i.e., 2 of 5 evidence points match the ground-truth labels). The risk-probability curve does not exceed the threshold anywhere within the window; therefore, the event is judged as not triggering a misoperation, and TLE is recorded as N/A.

The explainability outputs are consistent with the operations log: the decision relies primarily on SCADA, with voice and operation tickets providing auxiliary cues, while PMU contributes weakly in this scenario. Overall, the XAI results align with the model’s predictions and help operations and maintenance (O&M) personnel rapidly trace evidence and interpret decisions.

Risk-response latency analysis

To evaluate the temporal responsiveness of TS-MR, the risk curve was computed according to Eqs. (1)–(7), with thresholds set to and . The first threshold-crossing moment was denoted , and the actual lockout trigger moment ; the response delay was defined as . Figure 9 shows that the curve remains below the threshold in the low-risk scenario; sparse events occur in the medium-risk scenario; and in the high-risk scenario, lockouts are frequent while the system maintains stable tracking. Cross-scenario statistics indicate that mainly falls within 1.5–3.0 s, with , , and These results demonstrate stable and predictable temporal response across different risk intensities, meeting engineering application requirements.

Fig. 9.

Risk-response latency analysis. (a) Risk curve and threshold under the low-risk scenario; (b) misoperation events under the medium-risk scenario (red triangles mark event timestamps); (c) lockout-state sequence under the high-risk scenario (blue vertical lines indicate lockout triggers).

Independent HIL cross-validation

To verify deployability, we conducted an independent check in a hardware-in-the-loop (HIL) environment focusing on interface connectivity and temporal consistency; these results are not included in the main evaluation. Figure 10(a) shows the HIL risk timeline aligned to the offline time base without applying a decision threshold. Figure 10(b) compares ROC shapes of offline and HIL runs; the AUC values (0.540 offline, 0.516 HIL) are reported only as a connectivity sanity check and are not used as classification performance. Figure 10(c) reports end-to-end latency with batch size 1: the offline distribution mainly lies within 60–100 ms (mean about 80 ms); HIL shifts slightly to the right (median about 90 ms) with a few samples exceeding 120 ms. Overall, interfaces and timing remain stable in HIL, supporting engineering-grade deployability.

Fig. 10.

Independent HIL cross-validation. (a) HIL risk timeline aligned to the offline time base, (b) ROC curves, offline versus HIL; AUCs are shown only to verify connectivity and label alignment, not as performance metrics, (c) End-to-end latency (batch size 1).

Comprehensive performance comparison under the unified evaluation protocol

To ensure comparability, all models adopt identical preprocessing, time alignment and window configurations, and a unified protocol for training, validation, and testing. Random seeds are fixed; class imbalance is handled by weighted cross-entropy; and early stopping is applied to avoid overfitting. Three baselines (CNN-BiLSTM, ConvLSTM, GAT) are tuned within the same search space: learning rate , network depth , dropout , and hidden dimension . The best configurations are used for testing to avoid bias from “insufficient baseline tuning.” Under this unified protocol, TS-MR attains accuracy 94.69%, F1-score 93.8%, and AUC 0.977, outperforming all baselines. End-to-end latency (batch = 1) is about 80 ms, the core forward pass is about 42 ms, and computational complexity is 3.4 GFLOPs, achieving a strong accuracy–efficiency trade-off. Compared with the three baselines, accuracy improves by 0.9–3.7 percentage points and AUC increases by 0.024–0.057, consistent with previous statistical checks.

Conclusion

To address temporal misalignment in multi-source heterogeneous data, complex operational logic, and the risk of misoperations during microgrid switching, this study proposes a time-synchronized misoperation recognition framework (TS-MR). The framework integrates procedural and interlocking constraints with dual-stage temporal alignment and multimodal fusion, forming an end-to-end pipeline comprising rule-based screening, temporal alignment, feature discrimination, and constraint optimization. This enables reliable misoperation detection and quantitative risk assessment suitable for data-intensive, disturbance-prone edge-intelligent environments.

1. Temporal alignment and feature modeling: A dual-stage synchronization mechanism combining variational Bayesian inference and UKF reduces cross-modal temporal bias and enhances sensitivity to key triggers. On this basis, the Transformer aligns multimodal semantics, the TCN models sequential dynamics, and the GNN captures equipment topology and interlocking dependencies, yielding stable performance in high-risk scenarios such as forced interlock override and unauthorized entry into energized intervals. Temporal alignment robustness is benchmarked against IEEE 1588 PTP and PMU timing referenced to GPS-disciplined UTC for magnitude comparison and robustness characterization.

2. Rule screening and risk priors: The rule module performs deterministic compliance checks to ensure that operation sequences satisfy procedural and interlocking constraints. Monte Carlo-based risk assessment is used only as an independent probabilistic prior during training for sample weighting and priority scheduling; it does not participate in rule judgments or real-time compliance processes. Together with consistency checks and perturbation analysis—supported by attention heatmaps, permutation importance, and node-contribution evidence—the model provides auditable support for suppressing false alarms and improving robustness.

3. Comprehensive performance and efficiency: Under a unified evaluation protocol, TS-MR outperforms comparative models in both accuracy and efficiency. End-to-end latency meets engineering real-time requirements, and computational complexity is constrained and reported using GFLOPs. The engineering standards and link-budget references are used solely for quantitative comparison and benchmarking, and do not constitute a compliance statement.

4. Limitations and outlook: This study primarily uses real and laboratory measurements; HIL data are reserved for independent cross-validation, and public datasets are used only for cross-domain robustness calibration, without involvement in training, validation, or testing. Synthetic samples are employed in the training set solely for imbalance handling. Although StarSim-HIL testbed validation has been completed, larger-scale field verification with extended PMU and SCADA datasets is still required to assess long-term generalization and engineering feasibility. Future work will target multi-station coordination, operator behavior prediction, and strategy co-optimization, incorporating additional HIL and field data within a cloud-edge collaborative architecture to enhance transferability and real-time deployability.

In summary, TS-MR achieves a balanced trade-off among accuracy, latency, and interpretability, providing a reusable technical pathway for trustworthy intelligent operation in microgrid switching scenarios.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Hongzhao Yang was responsible for the overall research framework design and manuscript writing. Zhan Zhang handled data processing and model simulation. Shijie Zhang contributed to the development and analysis of the time synchronization algorithm. Rui Liang was in charge of the anti-misoperation rule modeling and experimental validation. All authors have read and approved the final version of the manuscript.

Funding

This work was supported by Hunan Provincial Natural Science Foundation of China (No. 2024JJ8025) and the National Natural Science Foundation of China (Nos. 7217010719, 72171026).

Data availability

The datasets generated and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.