A multi-harmonic emission level analysis method fusing impedance constraint and knowledge transfer

Abstract

The application discloses a multi-harmonic emission level analysis method fusing impedance constraint and knowledge migration, and comprises the following steps: firstly, target power distribution network data is acquired to construct a basic sample set; source domain harmonic samples are generated based on a typical nonlinear load model, and feature enhancement and physical feasibility screening are performed on the source domain samples and target domain measured samples; a space-time migration network is constructed; the model is jointly trained through data fitting loss, impedance consistency loss, topology consistency loss, time continuity loss, domain alignment loss and knowledge rule loss; finally, the harmonic emission level, responsibility contribution rate and confidence under each node, each time and each harmonic order are output, and the output result is physically corrected according to the impedance residual error. According to the application, the physical consistency, cross-scene adaptability and responsibility division reliability of the multi-harmonic emission level analysis result can be improved under the condition of a small amount of measured labeled samples.

Description

Technical Field

[0001] This invention relates to the technical field of power quality analysis and harmonic responsibility allocation in power systems, and particularly to a multi-harmonic emission level analysis method that integrates impedance constraints and knowledge transfer. Background Technology

[0002] To provide theoretical support for harmonic characteristic analysis, harmonic source tracing, responsibility allocation, and harmonic mitigation in power systems, scholars both domestically and internationally have conducted extensive research on harmonic source modeling and harmonic emission level determination. Existing methods mainly fall into two categories: physical model methods and data-driven methods. Physical model methods primarily include constant current source models, harmonic Norton models, and harmonic coupling models. These methods can analyze the propagation characteristics and contribution relationships of different frequency harmonics from the perspective of circuit mechanisms and network structures, providing a relatively clear explanation of the mechanism of harmonic emission levels. However, these methods are typically constructed based on current operating parameters and spatial network attributes, resulting in problems such as numerous modeling parameters, strong parameter coupling, and high computational complexity. Furthermore, they lack the ability to characterize the dynamic fluctuation characteristics of harmonics in the time dimension, making them unsuitable for the refined analysis needs of multi-load, multi-operating-condition, and time-varying operating scenarios.

[0003] With the digitalization and networking of new power systems, a large amount of operational monitoring data has accumulated in distribution networks, making data-driven harmonic modeling methods a research hotspot. These methods, through data fitting, statistical learning, or deep learning, can effectively characterize the temporal variation of harmonics and demonstrate strong fitting capabilities for complex nonlinear relationships. However, purely data-driven methods often lack explicit physical constraints, making it difficult to accurately reflect the propagation mechanism and coupling relationships of multiple harmonic sources in the spatial dimension. Furthermore, in practical engineering, the number of harmonic annotation samples is often limited, and the coverage of operating conditions is incomplete, easily leading to insufficient model generalization ability and decreased analytical accuracy under small sample conditions.

[0004] Therefore, existing physical modeling methods and data-driven methods struggle to simultaneously address the temporal volatility of multiharmonic emission levels, the global propagation characteristics in the spatial dimension, and high-precision solutions in small-sample scenarios. This is especially true in new distribution networks, where multiple types of nonlinear loads, distributed power sources, and complex topologies coexist, leading to more significant harmonic coupling between different nodes. Traditional single-modeling approaches are insufficient to meet the demands for accurate spatiotemporal generalization and analysis of multiharmonic emission levels. Therefore, it is necessary to propose a novel method that integrates physical models, operational data, and knowledge-enhanced information to achieve unified modeling, accurate solutions, and responsibility contribution analysis of multiharmonic emission levels in both time and space dimensions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a multi-harmonic emission level analysis method that integrates impedance constraints and knowledge transfer. This addresses the problems in existing technologies, such as the difficulty of physical models simultaneously characterizing the dynamic fluctuations of multi-harmonic emissions in the time dimension, the difficulty of data-driven models accurately depicting the propagation mechanism of harmonics in the spatial dimension, and the limitation of model generalization ability due to insufficient monitoring samples in practical engineering. The present invention aims to solve the following technical problems: (1) Solving the problem that pure data-driven models lack impedance physical constraints and their output results easily deviate from the power grid operation mechanism; (2) Solving the problem that traditional physical models have strong parameter coupling, high computational complexity, and difficulty in depicting the dynamic changes of harmonic time series; (3) Solving the problem that insufficient measured and labeled samples in the target scenario lead to weak cross-scenario adaptability of the model; (4) Solving the problem that emission level, responsibility contribution rate, and confidence are difficult to jointly analyze under multiple nodes, multiple times, and multiple harmonic orders. To achieve the above-mentioned objectives and other advantages of the present invention, a multi-harmonic emission level analysis method integrating impedance constraints and knowledge transfer is provided, comprising: Based on impedance physical constraints, a spatiotemporal transfer network is constructed for the analysis of emission levels of multiple harmonic sources. Among them, physical simulation sample augmentation is used to provide source domain training samples for the impedance constraint model, transfer learning is used to realize parameter adaptation of the impedance constraint model in the target small sample scenario, knowledge rules are used to supplement and correct the load-harmonic mapping relationship that the impedance constraints cannot fully cover, and physical consistency correction layer is used to correct the model output based on impedance residual, topology residual and time continuity residual, so as to realize the joint analysis of emission level, responsibility contribution rate and confidence under multiple nodes, multiple time, and multiple harmonic orders.

[0006] Preferably, the process involves first collecting multi-source operational data from the target power grid or target scenario, and then constructing a basic sample set by combining this data with a small number of labeled harmonic samples. Secondly, a typical nonlinear load physical model is established, generating source domain harmonic samples under different load types and operating conditions to supplement the multi-condition samples required for impedance constraint model training. Then, a data-driven model is constructed with an impedance embedding layer and a physical consistency correction layer at its core, embedding system equivalent impedance, node coupling relationships, network topology relationships, and harmonic order information as physical priors into the network. Subsequently, the impedance constraint model is pre-trained on source domain samples and fine-tuned on a small number of measured samples in the target domain, enabling the model to adapt to the target scenario while maintaining impedance physical consistency. Simultaneously, load type-harmonic feature mapping rules, harmonic propagation mechanism rules, and empirical boundary knowledge are transformed into knowledge rule loss or output correction rules to assist the impedance constraint model in correcting unstable outputs under small sample conditions. Finally, the emission level, responsibility contribution rate, and confidence results for each node, time point, and harmonic order are output.

[0007] Compared with the prior art, the advantages and positive effects of the present invention are: 1. Impedance-constrained transfer learning: The impedance embedding layer and impedance consistency loss constitute the main physical constraints for model training, enabling the model to learn the common harmonic propagation laws under multiple loads and operating conditions during the source domain pre-training stage. During the target domain fine-tuning stage, transfer learning does not simply adjust model parameters, but adapts to scenarios while maintaining impedance consistency, topology consistency, and time continuity constraints, thereby reducing the risk of model output deviating from the physical laws of the power grid under small sample conditions.

[0008] 2. Knowledge-based rule-assisted impedance constraints: Load type-harmonic order mapping rules, propagation path rules, and empirical boundary knowledge are transformed into knowledge rule losses, rule weights, or output correction rules to supplement load characteristic differences that are difficult to express directly by impedance constraints. When the target domain sample is insufficient or the operating condition distribution is incomplete, knowledge rules provide auxiliary constraints on the model output, ensuring that the prediction results satisfy both the impedance propagation relationship and the typical load harmonic characteristics.

[0009] 3. Spatiotemporal Output Back-to-Back Verification of Physical Consistency: The temporal evolution curve, spatial distribution results, responsibility contribution rate, and confidence level of the model output are not only used as the final analytical results, but also to verify whether the prediction results meet the requirements of impedance consistency, topology consistency, and temporal continuity. When the impedance residual, topology residual, or temporal continuity residual exceeds a preset threshold, the physical consistency correction layer corrects the initial output results, forming a closed-loop mechanism of "prediction - physical verification - consistency correction". Attached Figure Description

[0010] Figure 1 A flowchart of the multi-harmonic emission level analytical method based on the fusion of impedance constraints and knowledge transfer according to the present invention; Figure 2 This is a data-driven model structure diagram of the impedance physical constraint embedding of the multi-harmonic emission level analytical method according to the present invention, which integrates impedance constraints and knowledge transfer. Figure 3 This is a flowchart of the small-sample adaptation of the multi-harmonic emission level analytical method based on the fusion impedance constraint and knowledge transfer of the present invention. Detailed Implementation

[0011] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] Table 1 Correspondence between typical load types and harmonic characteristics Rectifier load Steady-state rectification, light load, heavy load Low-order characteristics are obvious, and the order distribution is relatively stable. Provides typical harmonic source templates to facilitate the construction of basic priors. Variable frequency load Speed ​​switching, frequency change Significant frequency drift and strong time correlation Highlighting the time dimension features, we validated the model's ability to represent time series data. Impact load Intermittent start-up, sudden disturbance Large harmonic fluctuations and obvious short-term abrupt changes Examine the robustness of the model to anomalous disturbances and non-stationary operating conditions. Distributed power inverter Grid connection control switching, power fluctuation The multi-band coupling characteristics are obvious. Used for studying the complex propagation mechanism in new energy grid-connected scenarios Hybrid nonlinear composite load Multiple loads in parallel, topology changes Harmonic coupling relationships are complex and exhibit significant spatial differences. Used to evaluate the model's cross-node generalization ability. Reference Figure 1 A multi-harmonic emission level analytical method integrating impedance constraints and knowledge transfer includes: S1: Collect three-phase voltage, current, active power, reactive power, frequency, switch status, load type, network topology and key line parameters in the target power grid, and combine them with a small number of labeled harmonic samples to construct a basic sample set; S2: Establish physical models for rectifier loads, frequency converter loads, impulsive loads, distributed power inverters, and hybrid nonlinear loads, and generate source domain harmonic samples under different load states and operating conditions; the source domain harmonic samples include harmonic voltage, harmonic current, equivalent impedance, harmonic order, load type, and emission level labels, which are used to support subsequent impedance embedding layer training and impedance consistency constraints.

[0013] S3: Enhance the source domain simulation samples and target domain measured samples by performing feature scaling, phase perturbation, time stretching, spectrum masking, harmonic order combination perturbation, noise injection and sample mixing, and conduct physical feasibility screening through impedance boundary, power boundary, contribution rate normalization and time change rate constraint to construct an enhanced training dataset.

[0014] S4: Construct a data-driven model with embedded impedance physical constraints. The model includes an input encoding layer, an impedance embedding layer, a spatiotemporal feature extraction layer, a multi-task prediction layer, and a physical consistency correction layer, which are used to jointly output the multi-harmonic emission level, the source-side contribution rate, and the result confidence.

[0015] S5: The impedance constraint model is pre-trained on the source domain enhanced samples to enable the model to learn the common impedance propagation laws under multiple load types, multiple topologies and multiple harmonic orders; fine-tuning is performed on a small number of measured samples in the target domain to enable the model to maintain impedance physical consistency while adapting to the data distribution of the target scenario, and the fine-tuning process is constrained by the load type-harmonic characteristic mapping rule, harmonic propagation mechanism rule and empirical boundary knowledge.

[0016] S6: Using the trained model, solve for the emission levels of multiple nodes, multiple time points, and multiple harmonic orders in the target scene, and output the node-level harmonic emission values, time evolution curves, spatial distribution results, responsibility contribution rate ranking, and confidence interval results.

[0017] Example 1: 1. Multi-source data acquisition and basic sample construction First, multi-source operational data from the target power grid or target scenario is collected. This multi-source operational data includes at least three-phase voltage, three-phase current, active power, reactive power, frequency, switch status, load type labels, network topology, and key line parameters. Simultaneously, a small number of labeled harmonic samples are extracted. These harmonic samples include at least the amplitude, phase, harmonic current contribution rate, and equivalent impedance information for multiple harmonic orders.

[0018] In this embodiment, the equivalent impedance can be expressed as: ; in, and These represent the nodes respectively. At any moment Next Second harmonic voltage and harmonic current To prevent extremely small positive numbers with a denominator of zero, the above equivalent impedance calculation converts node voltage and current measurements into prior impedance information related to harmonic propagation characteristics, which then serves as input for subsequent impedance embedding layers and physical consistency constraints.

[0019] Based on the above multi-source operational data and a small number of labeled harmonic samples, a basic training sample set can be formed for subsequent physical model simulation, sample augmentation, model training, and result analysis.

[0020] Table 2 Definition of Variables for Multi-Source Acquisition

[0021] 2. Physical model simulation and scene generation After completing the construction of the basic samples, a physical model of a typical nonlinear load is established. The nonlinear load includes at least rectifier loads, frequency converter loads, impulsive loads, distributed power inverters, and hybrid nonlinear loads.

[0022] Let the q-th type of load be in the operating state. and parameter vector The harmonic current model is as follows: ; in, The amplitude of the h-th harmonic is... The phase of the h-th harmonic. ω is the fundamental angular frequency.

[0023] In practical implementation, various operating conditions, such as light load, rated load, heavy load, start-stop disturbance, parallel coupling, and topology switching, can be addressed by scanning parameters and combining scenarios for different load types to generate harmonic voltage, current, impedance, and emission level label samples covering a wider range of operating conditions. This step can compensate for the lack of measured samples and incomplete coverage of operating conditions.

[0024] 3. Data Augmentation through Feature Transformation To further improve the model's adaptability to complex operating conditions and small sample scenarios, feature enhancement is performed on the simulation samples and original measured samples generated in step two. The enhancement methods include, but are not limited to, amplitude scaling, phase perturbation, time scaling, spectral masking, harmonic order combination perturbation, noise injection, sample mixing, window pruning, and segment reconstruction.

[0025] Among them, amplitude scaling and phase perturbation are used to simulate load fluctuations and phase shifts; time scaling is used to simulate the differences in the rate of harmonic variation; spectrum masking and harmonic order combination perturbation are used to enhance the model's adaptability to atypical harmonic combinations; noise injection is used to simulate measurement errors and background interference; sample mixing and fragment reconstruction are used to expand the sample distribution range and improve the model's robustness to edge conditions.

[0026] After enhancement, the enhanced samples are physically screened to remove abnormal samples that do not meet impedance boundaries, power boundaries, or statistical distribution constraints, in order to ensure the validity and reliability of the training data.

[0027] Table 3 Data Augmentation Methods

[0028] 4. Data-driven model construction with impedance physical constraints embedded This invention constructs a data-driven model with embedded impedance physical constraints. The model includes at least an input encoding layer, an impedance embedding layer, a spatiotemporal feature extraction layer, a multi-task prediction layer, and a physical consistency correction layer.

[0029] The input encoding layer receives node voltage, current, power, frequency, state variables, load type, and topology information; the impedance embedding layer encodes the system's equivalent impedance matrix, frequency band parameters, and node coupling relationships into physical prior vectors; the spatiotemporal feature extraction layer extracts the dynamic variation of harmonics in the time dimension and their propagation in the spatial dimension; the multi-task prediction layer simultaneously outputs the multi-harmonic emission level, source-side contribution rate, and result confidence; and the physical consistency correction layer corrects the model output results for impedance consistency, topology consistency, and temporal continuity.

[0030] In model training, a joint optimization approach combining data loss terms and physical constraint loss terms is adopted, and its total loss function can be expressed as: ; in, For data fitting loss, For impedance uniformity loss, For topology consistency loss, For the loss of time continuity, To constrain losses based on contribution rate of responsibility, Loss of knowledge rules; to These are the weighting coefficients for each loss term.

[0031] Furthermore, the impedance uniformity loss can be expressed as: ; in, Indicates the first Subharmonic voltage vector Indicates the first Subharmonic impedance matrix This represents the equivalent harmonic emission current vector predicted by the model.

[0032] The aforementioned loss terms constrain the model training process from aspects such as data fitting, impedance consistency, topology consistency, temporal continuity, reasonableness of responsibility contribution rate, and consistency of knowledge rules, thereby preventing the model output from deviating from the actual power system operation mechanism. The spatiotemporal feature extraction layer can be implemented using structures such as Long Short-Term Memory Networks, Temporal Convolutional Networks, or Transformers. This implementation does not limit the internal calculation formulas of specific network units.

[0033] Table 4. Explanation of Model Training Constraints

[0034] Figure 2 This is a diagram of the data-driven model structure for the impedance physical constraint embedding of the present invention. The diagram illustrates the parallel input structure of the timing input branch, impedance prior branch, and topology knowledge branch, as well as the relationship between the fusion module, multi-task prediction head, and physical correction head. It is used to explain how the present invention simultaneously introduces data features, impedance physical priors, and network topology knowledge into the network structure.

[0035] 5. Transfer learning under impedance constraints and knowledge rule-assisted adaptation To enable the impedance-constrained model to be applicable to scenarios with a small number of measured samples in the target domain, this invention employs a transfer learning approach that combines source domain pre-training with target domain fine-tuning. Source domain pre-training is used to learn the common impedance propagation patterns under different load types, topologies, and harmonic orders; target domain fine-tuning is used to adjust the model parameters while maintaining impedance consistency constraints, so that the model adapts to the data distribution of the target scenario.

[0036] Based on this, knowledge rules are not used as an independent post-processing module, but rather to assist in the training and output correction of impedance-constrained models. Specifically, knowledge such as load type, harmonic order, propagation path, and impedance boundary is organized into rule matrices or rule weights, and participates in joint optimization through knowledge rule loss. When insufficient samples in the target domain lead to unstable model output, knowledge rules are used to correct results that do not conform to typical load harmonic characteristics or impedance boundaries.

[0037] Meanwhile, to reduce the difference in feature distribution between the source and target domains, domain alignment strategies such as maximum mean difference, adversarial domain adaptation, or correlation alignment can be introduced to enable the impedance propagation laws learned from the source domain simulation samples to be transferred to the actual measured scenario in the target domain. The knowledge rules can be composed of information such as load type, harmonic order, propagation path, and impedance boundary, and participate in model training and output correction in the form of rule matrices, rule weights, or knowledge rule losses, thereby improving the model's output stability and physical consistency under small sample conditions.

[0038] ; in, Indicates the first Class of load and the first Knowledge association weights between subharmonics This represents the load-harmonic order correlation function obtained from expert rules, historical sample statistics, or simulation sample statistics.

[0039] Furthermore, the loss of knowledge rules can be expressed as: ; in, The model predicts the first Class load in the first Emission level under subharmonics This represents the reference launch level obtained from physical models or historical statistics.

[0040] Figure 3 This is a flowchart illustrating the few-shot adaptation process of transfer learning and knowledge enhancement according to the present invention. The flowchart shows the implementation process of source domain pre-training, domain alignment, target domain fine-tuning, knowledge graph construction, rule guidance, and joint optimization, illustrating how the present invention improves the model's cross-domain adaptation and reasoning capabilities in few-shot scenarios.

[0041] 6. Spatiotemporal generalization analysis output of multi-harmonic emission After completing the model training, the trained model is used to solve the emission levels of each node, each time point, and each harmonic order in the target scene, and outputs the node-level harmonic emission values, time evolution curves, spatial heat maps, responsibility contribution rate ranking, and confidence intervals.

[0042] In this implementation, the output results mainly include the following categories: (1) Node-level emission results, used to describe the local and overall emission levels of different nodes under different harmonic orders; (2) Time evolution results, used to describe the trend of harmonic emission intensity over time; (3) Spatial distribution results, used to describe the differences in harmonic intensity and propagation characteristics between different nodes; (4) The responsibility contribution results are used to characterize the contribution of different harmonic sources to the harmonic emission level of the target node; (5) Result credibility evaluation, used to help judge the reliability of the model output.

[0043] Based on the above output results, multi-harmonic emission peak superposition curves, peak transfer curves, spatial thermal distribution maps, and spatiotemporal point cloud maps can be further generated to support harmonic responsibility allocation, key node identification, compensation device capacity configuration, and governance strategy optimization.

[0044] Table 5. Parsing Output Results

[0045] Module synergy verification: In a distribution network simulation example with 12 nodes, the following three configurations were tested respectively: Configuration A (no collaboration): Using only a data-driven model and transfer learning, with N=100 samples in the target domain, the mean absolute percentage error (MAPE) of the emitted values ​​is 18.7%. 2. Configuration B (partial coordination): Adding physical constraint embedding (i.e., configuration A + impedance physical constraint), MAPE decreases to 9.2% at the same N=100; when N decreases to 30, MAPE is 14.5%. 3. Configuration C (Fully Collaborative): With configuration B and knowledge enhancement, MAPE is further reduced to 7.8% under the same N=30, and the output results all meet the impedance boundary consistency (in configuration B, 12% of the samples violate the impedance constraint when N=30, while in configuration C, this proportion is reduced to 2%).

[0046] The above comparison shows that the introduction of physical constraints and knowledge enhancement not only contribute to the improvement of accuracy, but more importantly, it enables the model to reach an accuracy level close to that of a large-sample (N=100) model without physical constraints (7.8% vs 18.7%) under small sample conditions (N=30). This indicates that the impedance physical constraint and knowledge rule-assisted mechanism can improve the accuracy and physical consistency of the model output under small sample conditions.

[0047] Configuration A Data-driven + transfer learning 100 18.7% — Configuration B Configure A+ physical constraints 100 9.2% 0% (implied) Configuration B Same as above 30 14.5% 12% Configuration C Configure B+ knowledge enhancement 30 7.8% 2% The number of devices and processing scale described herein are for simplification purposes. Applications, modifications, and variations of this invention will be readily apparent to those skilled in the art. Although embodiments of the invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for this invention, and further modifications can be readily implemented by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, this invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A multi-harmonic emission level analytical method integrating impedance constraints and knowledge transfer, characterized in that, Includes the following steps: S1. Construct basic samples by collecting multi-source data; S2. Establish physical models for rectifier loads, frequency converter loads, impulsive loads, distributed power inverters, and hybrid nonlinear loads, and generate source domain harmonic samples under different load states and operating conditions. S3. Enhance the features of source domain simulation samples and target domain measured samples by mixing, and screen the physical feasibility through impedance boundary, power boundary, contribution rate normalization and time change rate constraints to construct an enhanced training dataset. S4. Construct a data-driven model with embedded impedance physical constraints; S5. Pre-train the impedance constraint model on the source domain enhanced samples to enable the model to learn the common impedance propagation law under multiple load types, multiple topologies and multiple harmonic orders; fine-tune the model on a small number of measured samples in the target domain to make the model adapt to the data distribution of the target scene while maintaining the physical consistency of impedance, and use the load type-harmonic characteristic mapping rule, harmonic propagation mechanism rule and empirical boundary knowledge to constrain the fine-tuning process. S6. Using the trained model, solve for the emission levels of multiple nodes, multiple time points, and multiple harmonic orders in the target scene, and output the node-level harmonic emission values, time evolution curves, spatial distribution results, responsibility contribution rate ranking, and confidence interval results.

2. The analytical method for multi-harmonic emission levels that integrates impedance constraints and knowledge transfer as described in claim 1, characterized in that, The multi-source data in step S1 includes three-phase voltage, current, active power, reactive power, frequency, switch status, load type, network topology, and key line parameters of the target power grid; the basic sample also includes labeled harmonic samples.

3. The analytical method for multi-harmonic emission levels that integrates impedance constraints and knowledge transfer as described in claim 1, characterized in that, The source domain harmonic samples in step S2 include harmonic voltage, harmonic current, equivalent impedance, harmonic order, load type, and emission level label, which are used to support subsequent impedance embedding layer training and impedance consistency constraints.

4. The analytical method for multi-harmonic emission levels that integrates impedance constraints and knowledge transfer as described in claim 1, characterized in that, The feature enhancement process in step S3, including mixing the source domain simulation samples and the target domain measured samples, is as follows: The simulation samples in the source domain and the measured samples in the target domain are enhanced with features such as amplitude scaling, phase perturbation, time stretching, spectrum masking, harmonic order combination perturbation, noise injection, and sample mixing.

5. The analytical method for multi-harmonic emission levels that integrates impedance constraints and knowledge transfer as described in claim 2, characterized in that, The harmonic sample includes at least the amplitude, phase, harmonic current contribution rate, and equivalent impedance information of multiple harmonic orders.

6. The analytical method for multi-harmonic emission levels that integrates impedance constraints and knowledge transfer as described in claim 1, characterized in that, The data-driven model includes at least an input encoding layer, an impedance embedding layer, a spatiotemporal feature extraction layer, a multi-task prediction layer, and a physical consistency correction layer. The input encoding layer receives node voltage, current, power, frequency, state variables, load type, and topology information; the impedance embedding layer encodes the system's equivalent impedance matrix, frequency band parameters, and node coupling relationships into physical prior vectors; the spatiotemporal feature extraction layer extracts the dynamic variation of harmonics in the time dimension and their propagation in the spatial dimension; the multi-task prediction layer simultaneously outputs the multi-harmonic emission level, source-side contribution rate, and result confidence; and the physical consistency correction layer corrects the model output results for impedance consistency, topology consistency, and temporal continuity.

Images