High-precision electric energy metering and fault diagnosis method adaptive to multiple working conditions

Abstract

The invention provides a high-precision electric energy metering and fault diagnosis method adaptive to multiple working conditions, and the method comprises the following steps: carrying out the multi-rate synchronous sampling of an electric parameter through a double-AD parallel sampling architecture, collecting a fundamental wave voltage / current signal through a low-speed AD module at a sampling rate of 4kHz, capturing a high-frequency harmonic wave and a transient signal through a high-speed AD module at a sampling rate of 256kHz, and carrying out the fault diagnosis of the high-frequency harmonic wave and the transient signal; a sampling mode is dynamically switched according to the load harmonic content; a sliding window mechanism is adopted to carry out multiple times of repeated detection on the same electric parameter, five times of independent measurement are completed in a 200ms time window, an outlier is eliminated based on a Pauta criterion, then a weighted mean value of a signal-to-noise ratio is calculated, AD reference voltage is calibrated in real time through a PT100 temperature sensor, and the accuracy of the AD reference voltage is improved. Carrying out nonlinear dynamic compensation by combining a cubic spline interpolation method of a pre-stored load characteristic curve; and constructing a composite electric energy calculation model, summing the fundamental wave electric energy and the harmonic electric energy weighted according to the standard to generate a total electric energy value, and comparing a metering result with a theoretical value in real time by using a digital phase-locked loop.

Description

Technical Field

[0001] The present invention belongs to the field of fault diagnosis, and in particular relates to a high-precision electric energy metering and fault diagnosis method adaptable to multiple working conditions. Background Art

[0002] Currently, current electricity metering and fault diagnosis technologies face significant technical bottlenecks in complex and changing operating conditions. Traditional electricity metering devices are primarily designed for steady-state operation and struggle to adapt to dynamic power environments such as nonlinear loads, harmonic pollution, and voltage sags. This results in metering errors generally exceeding ±1.0%, and even as high as ±5% under extreme operating conditions (such as inverter loads and arc furnaces). Existing fault diagnosis methods primarily rely on threshold alarms or simple rule-based judgments, lacking the ability to capture transient signals (such as microsecond voltage swells and high-frequency harmonics), resulting in a fault miss-report rate exceeding 15%, and are unable to distinguish between similar fault types (such as line short circuits and equipment insulation aging).

[0003] In terms of data acquisition, traditional technologies use A / D converters with a fixed sampling rate (typically below 4kHz), which cannot simultaneously and accurately measure the fundamental wave and high-frequency harmonics. Harmonic amplitude and phase errors can reach ±3%, seriously affecting the accuracy of power quality analysis. Existing metering system communication protocols (such as DL / T645 and IEC 61850) lack support for real-time data, with data transmission delays exceeding 200ms, making it difficult to meet the millisecond-level fault response requirements of smart grids. Furthermore, traditional fault diagnosis models (such as support vector machines and backpropagation neural networks) rely on manual feature extraction and have poor generalization capabilities. When faced with unknown load types or new power electronic equipment (such as photovoltaic inverters and energy storage converters), diagnostic accuracy drops by over 20%.

[0004] Regarding energy efficiency management, existing systems lack the ability to integrate and analyze heterogeneous data from multiple sources (voltage, current, temperature, and vibration). Energy efficiency assessment errors exceed ±8%, making precise energy optimization impossible. These technical deficiencies render existing energy metering and fault diagnosis systems unreliable in complex scenarios such as Industry 4.0 and the integration of new energy sources into the grid. There is an urgent need to overcome the technical bottlenecks of adaptability to multiple operating conditions, high-precision measurement, and intelligent diagnosis. Summary of the Invention

[0005] The present invention proposes a high-precision electric energy metering and fault diagnosis method that is adaptable to multiple working conditions. This method solves the problems of low accuracy and slow dynamic response of traditional electric energy metering systems under complex harmonic and transient working conditions through dual AD multi-rate sampling, dynamic compensation, composite electric energy calculation and intelligent fault diagnosis model. At the same time, it overcomes the defects of inaccurate fault location and delayed knowledge base update, and realizes high-precision metering and real-time fault analysis in multiple scenarios.

[0006] The technical solution of the present invention is achieved as follows: a high-precision electric energy metering and fault diagnosis method suitable for multiple working conditions, the method comprising the following steps: performing multi-rate synchronous sampling of electric parameters through a dual AD parallel sampling architecture, wherein a low-speed AD module collects fundamental voltage / current signals at a 4kHz sampling rate, and a high-speed AD module captures high-frequency harmonics and transient signals at a 256kHz sampling rate, and dynamically switches the sampling mode according to the load harmonic content;

[0007] The energy metering process involves repeated detection and dynamic analysis, using a dual-ADC parallel sampling architecture for multi-rate synchronous sampling and data preprocessing. The low-speed ADC (16-bit, 4kHz) is used for precise measurement of fundamental (50 / 60Hz) voltage and current, ensuring an error of ≤±0.2% under steady-state conditions. The high-speed ADC (14-bit, 256kHz) captures high-frequency harmonics (up to 128th order) and transient signals (such as μs-level voltage sags), with harmonic amplitude and phase errors of ≤±1%. Dynamic sampling rate switching automatically switches to high-speed sampling mode based on load characteristics (e.g., when harmonic content >5% is detected).

[0008] A sliding window mechanism is used to perform multiple repeated detections of the same electrical parameter, completing five independent measurements within a 200ms time window. The signal-to-noise ratio-weighted mean is calculated after eliminating outliers based on the Laida criterion. The AD reference voltage is calibrated in real time using a PT100 temperature sensor, and nonlinear dynamic compensation is performed using cubic spline interpolation of pre-stored load characteristic curves. Multiple repeated detection and data fusion are also performed.

[0009] Sliding Window Repeat Detection: Within each 200ms time window, the same electrical parameter (e.g., voltage RMS) is measured five times independently, eliminating outliers (using the 3σ criterion) and taking a weighted average (weighted based on the signal-to-noise ratio). Multi-dimensional Dynamic Compensation: Temperature Compensation: A PT100 sensor is used to calibrate the A / D conversion reference voltage in real time, with a temperature drift error of ≤±0.05%. Nonlinearity Correction: Based on pre-stored load characteristic curves (e.g., the V / V nonlinearity of the inverter), cubic spline interpolation is used to dynamically correct measurement results.

[0010] A composite energy calculation model is constructed, summing fundamental energy with standard-weighted harmonic energy to generate a total energy value. A digital phase-locked loop (DPLL) is used to compare the measurement results with theoretical values ​​in real time. When the difference between the measurement results and theoretical values ​​exceeds a set threshold, a retest is triggered. The composite energy calculation model: fundamental energy (low-speed AD data) + harmonic energy (high-speed AD data, weighted according to the IEC 61000-4-7 standard) = total energy, with a combined error of ≤±0.5%. A closed-loop verification mechanism is implemented: a digital phase-locked loop (DPLL) is used to compare the measurement results with theoretical values ​​in real time, with a retest triggered by abnormal data.

[0011] Extract the spatiotemporal features of N key nodes in the power grid topology. These features include the time domain mean, frequency domain FFT energy spectrum, and time-frequency wavelet packet entropy. A dynamic graph neural network model is constructed based on these spatiotemporal features. Nodes are represented as multidimensional feature vectors, and edges represent electrical coupling relationships. A hierarchical attention mechanism is used to implement fault location and propagation path analysis.

[0012] Multi-source data feature extraction and point mapping encode spatiotemporal features into spatial points. The grid topology is decomposed into N key nodes (such as transformer outlets and load access points), with voltage, current, temperature, vibration, and other data associated with each node. Time series analysis extracts 128-dimensional features from each point data point in the time domain (mean, kurtosis), frequency domain (FFT energy spectrum), and time-frequency domain (wavelet packet entropy). Graph Neural Network (GNN)-driven fault diagnosis is also implemented.

[0013] Nodes in dynamic graph construction represent key points in the power grid (including eigenvectors). Edges represent the electrical coupling relationships between nodes (impedance, power flow). Node-level attention in the hierarchical attention mechanism is used to filter high-weight features (e.g., a sudden increase in the harmonic distortion rate of the current at a certain point). Graph-level attention is used to locate fault propagation paths (e.g., abnormal temperature gradients between adjacent nodes caused by insulation aging).

[0014] Generate embedding vectors for new fault cases based on contrastive learning, automatically update the fault knowledge base and fine-tune the diagnosis model at regular intervals, and automatically reset and adjust when the diagnostic accuracy decays by ≥±1%.

[0015] Existing energy metering technologies typically use a single AD sampling architecture (e.g., a fixed 1kHz sampling rate), which is unable to accurately capture both the fundamental wave and high-frequency harmonics. This results in harmonic energy measurement errors as high as 5%-10%. This solution, however, uses a dual AD parallel architecture (4kHz + 256kHz) to dynamically switch sampling modes. This automatically enables high-speed sampling when the load harmonic content exceeds 15%, compressing the harmonic energy error to less than 1%.

[0016] Traditional methods rely on static calibration (e.g., fixed temperature compensation coefficients). Under nonlinear load conditions, temperature drift can cause reference voltage deviations of up to 0.2V, leading to significant cumulative measurement errors. This solution introduces a PT100 temperature sensor for real-time calibration of the AD reference voltage, combined with dynamic compensation using cubic spline interpolation, to suppress temperature drift errors to below 0.02V. Existing energy calculation models often ignore harmonic components or use linear superposition (e.g., IEC standard harmonic weighting), which fails to reflect the interaction between harmonics and the fundamental wave in actual power grids. The composite energy model constructed in this solution uses standard weighting coefficients to correct for harmonic energy contributions, ensuring that the total energy value deviates from the true value by ≤0.5%.

[0017] Existing fault diagnosis systems use rule-based feature matching (such as threshold alarms), resulting in an error rate exceeding 30% for fault location in complex power grids. This solution uses a dynamic graph neural network to extract spatiotemporal features (time domain mean, FFT energy spectrum, and wavelet packet entropy) and employs a hierarchical attention mechanism to analyze electrical coupling relationships. In tests on industrial power grids with over 50 nodes, fault location accuracy was improved to 95%. Traditional fault knowledge bases rely on manual annotation and updating, resulting in a high rate of missed detection of new fault types, up to 40%. This solution generates new fault embedding vectors based on contrastive learning and automatically triggers model fine-tuning when diagnostic accuracy fluctuates by 1% or more. This solution has improved the recognition rate of new arc faults from 70% to 92% in renewable energy access scenarios.

[0018] As a preferred embodiment, the low-speed AD module and high-speed AD module of the dual AD are dynamically switched through a parallel sampling architecture. The triggering condition for dynamic switching is: when the total harmonic distortion rate is detected to be ≥5%, the high-speed AD mode is automatically enabled, and the fundamental wave and harmonic components within 128 times are synchronously collected, and the harmonic amplitude phase error is ≤±1%.

[0019] As a preferred embodiment, the chip temperature is monitored in real time by a PT100 sensor, and the sampling value is dynamically corrected based on the temperature-reference voltage mapping table, with a temperature drift error of ≤±0.05%; the input of the cubic spline interpolation method is a pre-stored inverter and arc furnace typical load VI characteristic curve, and the output is a nonlinear correction coefficient.

[0020] As a preferred embodiment, the total electric energy error of the composite electric energy calculation model is ≤±0.5%, and the phase tracking accuracy of the digital phase-locked loop is ≤±0.1°, and the re-measurement trigger threshold is triggered when the deviation between the measurement result and the theoretical value exceeds ±1%.

[0021] As a preferred embodiment, the hierarchical attention mechanism includes node-level attention to screen the current harmonic distortion rate and temperature gradient high-weight features; locate the fault propagation path through graph-level attention, and generate a fault probability heat map by combining electrical impedance and power flow.

[0022] As a preferred embodiment, the contrastive learning generates a difference embedding vector by calculating the similarity between the spatiotemporal feature vector and the knowledge base sample, and the scale of the expanded training set during contrastive learning is not less than 10% of the fault cases of the original data.

[0023] After adopting the above technical solution, the beneficial effects of the present invention are: this method shows significant advantages in smart grid and industrial power scenarios:

[0024] 1) In terms of measurement accuracy, dual AD dynamic sampling makes the fundamental electric energy measurement error ≤0.2 level (traditional is 0.5 level), and the harmonic electric energy measurement error ≤1.5%;

[0025] 2) In terms of real-time performance, the sliding window mechanism (five measurements within 200ms) and the signal-to-noise ratio weighted algorithm reduce data update latency to 50ms under dynamic load fluctuations, improving response speed by four times compared to traditional methods (200-500ms).

[0026] 3) In terms of fault diagnosis capabilities, the hierarchical attention mechanism of the dynamic graph neural network can accurately locate hidden short-circuit faults in microgrids (such as high-resistance faults with impedance > 10Ω), with a positioning error of ≤ 2 meters and a false alarm rate reduced from 15% to 3%;

[0027] 4) In terms of adaptability, knowledge base updates driven by comparative learning have reduced the system's model iteration cycle from 30 days to 72 hours in the context of sudden harmonic changes in photovoltaic power plants, and kept the decay rate of fault diagnosis accuracy to less than 0.8% per month. During a shock load test on a steel mill, this method successfully identified 12 transient voltage sag events that were missed by traditional systems and accurately located the feeder-level fault point, ensuring the stability of continuous production. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0029] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0030] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0031] Example:

[0032] like Figure 1 As shown in FIG, the working principle and process of a high-precision electric energy measurement and fault diagnosis method adapted to multiple working conditions are as follows:

[0033] Taking the high-power motor load driven by an inverter in an industrial power grid as an example, the system first performs multi-rate synchronous sampling of electrical parameters through a dual AD parallel sampling architecture. The low-speed AD module stably acquires fundamental voltage / current signals at a 4kHz sampling rate, ensuring that the power frequency power metering error is ≤0.2 levels. The high-speed AD module captures high-frequency harmonics (such as 15th order and above) and voltage transient distortion signals at the moment of motor start and stop in real time at a sampling rate of 256kHz. When the harmonic content exceeds 15%, it automatically switches to high-speed mode. Compared with traditional single AD fixed-rate sampling (such as 1kHz), the harmonic power metering error is reduced from 5% to 1.5%.

[0034] A sliding window mechanism was then used to perform five independent measurements of the same electrical parameter within 200ms. Outliers due to electromagnetic interference were eliminated based on the Laida criterion (3σ principle), and the mean was calculated using a weighted signal-to-noise ratio. A PT100 temperature sensor was used to calibrate the AD reference voltage in real time (with a temperature drift compensation accuracy of ±0.02°C). Cubic spline interpolation, superimposed on pre-stored load curves, was used for nonlinear dynamic compensation, addressing the 0.5V voltage deviation that can occur with traditional static calibration when the motor load changes suddenly. A composite energy calculation model was then constructed, summing the fundamental energy with harmonic energy weighted according to IEC standards (e.g., a weight of 0.3 for the fifth harmonic and 0.2 for the seventh harmonic) to generate the total energy value. A digital phase-locked loop (DSP achieving μs-level synchronization) was used to compare the measurement results with the theoretical values ​​in real time. A re-measurement mechanism was triggered when the deviation exceeded 0.5%, resulting in a fourfold improvement in accuracy compared to the traditional linear superposition model (error ≥ 2%).

[0035] During the fault diagnosis phase, the system extracts the spatiotemporal characteristics of 50 key nodes in the power grid topology: the time domain mean reflects the steady-state deviation of voltage / current, the frequency domain FFT energy spectrum (resolution 0.1Hz) identifies harmonic distribution anomalies, and the time-frequency wavelet packet entropy (using the db4 wavelet basis) captures transient disturbance characteristics. A dynamic graph neural network model is constructed, in which nodes are encoded as 12-dimensional feature vectors (including parameters such as voltage, current, phase, and harmonics), and edge weights represent electrical connection impedance and power flow. The fault point is located through a hierarchical attention mechanism (layer 1 focuses on local topological associations, and layer 2 analyzes global propagation paths). In motor winding short-circuit tests, a positioning accuracy of 2 meters is achieved, which is 5 times higher than the traditional threshold alarm method (error ≥ 10 meters).

[0036] Finally, based on contrastive learning, new fault case embedding vectors (such as arc fault signatures) are generated. The fault knowledge base is automatically updated daily, and model parameters are fine-tuned. When diagnostic accuracy fluctuates by 1% or more (e.g., due to feature drift caused by new energy integration), the system automatically resets the optimizer and retrains, increasing the recognition rate of PV inverter harmonic faults from 70% to 92%. Traditional manual labeling requires an update cycle of up to 30 days and a missed detection rate exceeding 40%. Through the synergistic effect of this technology chain, this method successfully identified 12 transient voltage sag events that were missed by traditional systems during a steel mill impact load scenario, and accurately located the fault point at the feeder level, ensuring continuous production stability.

[0037] The dual AD system dynamically switches between its low-speed and high-speed AD modules via a parallel sampling architecture. The dynamic switching is triggered when a total harmonic distortion (THD) of 5% or higher is detected, automatically enabling high-speed AD mode. This synchronizes the fundamental and harmonic components up to the 128th order, with harmonic amplitude and phase errors ≤±1%. In a steel mill load scenario, when a total harmonic distortion (THD) of 7.2% is generated at motor startup, the system immediately switches to high-speed AD mode (256kHz), fully capturing the 5th, 7th, and 11th harmonics (with an amplitude error of 0.8% and a phase error of 0.5°). The traditional single AD mode (4kHz) misses over 30% of these high-frequency harmonics. In arc furnace operation, the high-speed AD module reduces harmonic power measurement error from the conventional 3% to 0.7% for the 17th and 19th interharmonics (frequencies of 950-1150Hz) using an anti-aliasing filter and a 24-bit Σ-Δ ADC. Meanwhile, fundamental sampling is maintained by the low-speed AD module, ensuring continuous and stable power frequency energy measurement.

[0038] A PT100 sensor monitors chip temperature in real time, and dynamically corrects sampled values ​​based on a temperature-reference voltage mapping table, achieving a temperature drift error of ≤±0.05%. The cubic spline interpolation method uses pre-stored VI characteristic curves for typical loads of inverters and arc furnaces as input, and outputs nonlinear correction coefficients. In a high-power inverter scenario at a chemical plant, when the ambient temperature rises from 25°C to 65°C, the PT100 sensor detects the ADC chip temperature change in real time. A 256-point calibration mapping table (with a resolution of 0.1°C) dynamically compensates for reference voltage drift (measured temperature drift of 0.03%) using a 256-point calibration mapping table (with a resolution of 0.1°C). For the nonlinear current waveform (THD 12%) output by the inverter, the system uses 300 pre-stored VI characteristic curves and cubic spline interpolation to generate dynamic correction coefficients, reducing the nonlinear error in current sampling from 0.1% to 0.02%. In particular, the distortion correction effect in the current zero-crossing region is eight times better than traditional linear compensation.

[0039] The composite power calculation model has a total power error of ≤±0.5%, and the phase tracking accuracy of the digital phase-locked loop is ≤±0.1°. The retest trigger threshold is triggered when the metering result deviates from the theoretical value by more than ±1%. In data center UPS system testing, the model's calculation error for distorted power containing 15th harmonics (8%) was only 0.3%, a 2.2 percentage point reduction compared to the IEC standard algorithm. The digital phase-locked loop uses a second-order loop filter, maintaining a phase synchronization accuracy of 0.05° even when the voltage drops by 20%. When a 1.2% deviation between the metering value and the theoretical value is detected (such as a capacitor switching disturbance), the system initiates a retest process within 10ms. Three sliding window reviews increase the abnormal data rejection rate to 99%. Traditional methods require manual intervention and have a response delay exceeding 500ms.

[0040] The hierarchical attention mechanism uses node-level attention to filter high-weight features such as current harmonic distortion and temperature gradients. Graph-level attention locates fault propagation paths and generates a fault probability heat map based on electrical impedance and power flow. In microgrid islanding fault diagnosis, node-level attention assigns a weight of 0.85 to the 11th harmonic distortion rate (sudden increase to 5.8%) at the photovoltaic inverter node, while also identifying abnormal temperature gradients (ΔT ≥ 15K) at cable joints. Graph-level attention analyzes impedance changes (ΔZ ≥ 30%) and reverse power flow in each branch, generating a heat map that accurately pinpoints a poor contact fault in a circuit breaker 300 meters away. This location is achieved in just 2.3 seconds (compared to over 15 minutes using traditional impedance methods). Furthermore, the misjudgment rate for fault direction caused by backflow from distributed generation (DGs) is reduced from 25% to 3%.

[0041] The comparative learning method generates a differential embedding vector by calculating the similarity between the obtained spatiotemporal feature vectors and samples from the knowledge base. The expanded training set for comparative learning must contain at least 10% of the original data. When a new high-frequency oscillation fault (1.2kHz) is detected on the wind farm's collector line, the system extracts its time-frequency wavelet packet entropy (6.8 bits) and FFT spectrum peak features, compares them with 800 sets of cases in the knowledge base using cosine similarity, and generates a 128-dimensional differential vector containing oscillation attenuation characteristics. By automatically adding 82 new cases (12% of the original database) for comparative learning fine-tuning, the recognition rate for this type of fault has increased from an initial 68% to 94%. The model update process does not require manual labeling, reducing maintenance costs by 90% compared to traditional expert systems.

[0042] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A high-precision electric energy measurement and fault diagnosis method suitable for multiple working conditions, characterized in that: The method includes the following steps: performing multi-rate synchronous sampling of electrical parameters through a dual AD parallel sampling architecture, wherein a low-speed AD module collects fundamental voltage / current signals at a 4kHz sampling rate, and a high-speed AD module captures high-frequency harmonics and transient signals at a 256kHz sampling rate, and dynamically switches the sampling mode according to the load harmonic content; A sliding window mechanism is used to repeatedly test the same electrical parameter, completing five independent measurements within a 200ms time window. The signal-to-noise ratio-weighted mean is calculated after eliminating outliers based on the Laida criterion. The AD reference voltage is calibrated in real time using a PT100 temperature sensor, and nonlinear dynamic compensation is performed using cubic spline interpolation of pre-stored load characteristic curves. A composite electric energy calculation model is constructed to sum the fundamental electric energy with the standard-weighted harmonic electric energy to generate a total electric energy value. A digital phase-locked loop is then used to compare the measurement results with the theoretical values ​​in real time. When the difference between the measurement results and the theoretical values ​​exceeds a set threshold, a retest of the abnormal data is triggered. Extract the spatiotemporal features of N key nodes in the power grid topology. These features include the time domain mean, frequency domain FFT energy spectrum, and time-frequency wavelet packet entropy. A dynamic graph neural network model is constructed based on these spatiotemporal features. Nodes are represented as multidimensional feature vectors, and edges represent electrical coupling relationships. A hierarchical attention mechanism is used to implement fault location and propagation path analysis. Generate embedding vectors for new fault cases based on contrastive learning, automatically update the fault knowledge base and fine-tune the diagnosis model at regular intervals, and automatically reset and adjust when the diagnostic accuracy decays by ≥±1%.

2. The high-precision electric energy metering and fault diagnosis method adapted to multiple working conditions according to claim 1, characterized in that: The low-speed AD module and high-speed AD module of the dual AD are dynamically switched through a parallel sampling architecture. The triggering condition for dynamic switching is: when the total harmonic distortion rate is detected to be ≥5%, the high-speed AD mode is automatically enabled, and the fundamental wave and harmonic components within 128 times are synchronously collected, and the harmonic amplitude phase error is ≤±1%.

3. The high-precision electric energy metering and fault diagnosis method adapted to multiple working conditions according to claim 1, characterized in that: The chip temperature is monitored in real time by a PT100 sensor, and the sampling value is dynamically corrected based on a temperature-reference voltage mapping table, with a temperature drift error of ≤±0.05%. The input of the cubic spline interpolation method is a pre-stored inverter and arc furnace typical load VI characteristic curve, and the output is a nonlinear correction coefficient.

4. The high-precision electric energy metering and fault diagnosis method adapted to multiple working conditions according to claim 1, characterized in that: The total electric energy error of the composite electric energy calculation model is ≤±0.5%, and the phase tracking accuracy of the digital phase-locked loop is ≤±0.1°. The re-measurement trigger threshold is triggered when the deviation between the measurement result and the theoretical value exceeds ±1%.

5. The high-precision electric energy metering and fault diagnosis method adapted to multiple working conditions according to claim 1, characterized in that: The hierarchical attention mechanism includes node-level attention to screen the high-weight features of current harmonic distortion rate and temperature gradient; locate the fault propagation path through graph-level attention, and generate a fault probability heat map by combining electrical impedance and power flow.

6. The high-precision electric energy metering and fault diagnosis method adapted to multiple working conditions according to claim 1, characterized in that: The contrastive learning generates a difference embedding vector by calculating the similarity between the spatiotemporal feature vector and the knowledge base sample. The scale of the expanded training set during the contrastive learning is not less than 10% of the fault cases of the original data.

Images