A low-voltage diagnosis deduction and achievement backtracking method and system

Abstract

The present application relates to the technical field of low-voltage treatment of power distribution network, and discloses a low-voltage diagnosis deduction and effect backtracking method and system, comprising: obtaining multi-modal monitoring data of each node of the power distribution network, extracting features, and then performing cross-modal fusion to obtain fusion features of each node; taking the fusion features as input, embedding the power flow equation as a physical constraint into a neural network loss function, setting the equivalent impedance of the line as a trainable parameter for optimization training, and outputting the backtracking impedance of each node; calculating the deviation of the backtracking impedance and the nominal impedance, and jointly determining the hidden defect node according to the infrared thermal spot feature; limiting the reconstruction coding search space according to the hidden defect node, performing scheme simulation deduction under the constraint of the fund pool, and determining the optimal reconstruction scheme and the deduction expected income; comparing the actual income after reconstruction with the deduction expected income to determine the income error, and feeding back the neural network parameters for correction. The present application breaks through the full-service closed loop of low-voltage treatment, and effectively improves the diagnosis accuracy and investment effectiveness.

Description

Technical Field

[0001] This invention relates to the field of low voltage management technology in power distribution networks, and in particular to a method and system for low voltage diagnosis, deduction, and effectiveness retrospective analysis. Background Technology

[0002] Low voltage management in distribution networks is a core business for ensuring power supply quality and user experience, and has long been a key focus for power supply companies. With the deepening of the construction of new power systems, a large number of distributed photovoltaic systems are being connected to the distribution network. In addition, problems such as long power supply radii, small wire diameters, and heavy overload in rural and remote areas are intertwined and superimposed, making the causes of low voltage increasingly complex and significantly increasing the difficulty of management.

[0003] However, existing low-voltage management technologies rely solely on single electrical timing data in the diagnostic phase, lacking physical mechanism constraints and multimodal feature fusion, making it difficult to accurately locate hidden physical defects such as abnormal contact resistance. In the scheme development phase, single-scheme trial-and-error simulation verification is used, which cannot systematically compare multiple technology combinations, lacks quantitative assessment of the side effects of front-end overvoltage caused by new technologies such as reactive power compensation, and is not linked to the limited funding pool, causing the optimal scheme to fail due to economic infeasibility. In the effectiveness evaluation phase, it simply compares the absolute voltage values ​​before and after the transformation, failing to isolate external confounding factors such as weather fluctuations and load growth to quantify the true causal benefits. Furthermore, when the transformation effect does not meet expectations, there is a lack of full-chain reverse tracing methods to locate management breakpoints, and the massive case data accumulated on-site cannot be fed back to the front-end diagnostic model, causing similar errors to occur repeatedly and diagnostic accuracy to stagnate for a long time. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method and system for low-voltage diagnostic deduction and effectiveness retrospective analysis, comprising: Acquire multimodal monitoring data of each node in the distribution network, extract features from the multimodal monitoring data and perform cross-modal fusion to obtain the fused features of each node; Using the fusion features as input and embedding the distribution network power flow equation as a physical constraint into the neural network loss function, the equivalent impedance of the line is optimized and trained, and the backtracking impedance of each node is output. The deviation between the traceback impedance and the nominal impedance is determined to obtain the impedance deviation of each node, and the hidden defect node is determined based on the infrared hot spot feature in the joint fusion feature of the impedance deviation. Based on the implicit defect node-limited renovation scheme, the scheme simulation is carried out, and under the constraint of the capital pool, the optimal renovation scheme and the expected return of each node are determined based on the simulation results. The actual benefits after the optimal modification scheme are compared with the projected expected benefits to determine the benefit error, and the parameters in the neural network are corrected based on the benefit error.

[0005] Furthermore, the acquisition of multimodal monitoring data from each node of the distribution network, and the extraction of features from the multimodal monitoring data for cross-modal fusion to obtain the fused features of each node, includes: Acquire multimodal monitoring data of each node in the distribution network. The multimodal monitoring data includes electrical time-series data, visual environment data, topology and text data, and extract the single-modal features of the multimodal monitoring data respectively. Based on the cross-attention mechanism, each single modal feature is used as a key-value pair for cross-modal semantic space alignment. Through contrastive learning constraints, the single modal features of the same physical entity are aggregated in the latent space, and the fused features of each node are output.

[0006] Furthermore, the cross-attention mechanism aligns each single-modal feature as a key-value pair across the semantic space, and uses contrastive learning constraints to aggregate the single-modal features of the same physical entity in the latent space, outputting the fused features of each node, including: The electrical time-series features are linearly mapped to obtain the query matrix, and the visual features and structured features are linearly mapped to obtain the key matrix and value matrix. Attention weights are obtained by normalizing the product of the transpose of the query matrix and the key matrix, divided by the square root of the dimension, and then multiplied by the value matrix to calculate the fusion features of each node. Positive and negative sample pairs are constructed based on each unimodal feature, and the cosine similarity between each unimodal feature in the fused feature is calculated using the InfoNCE loss function to force the positive sample pairs to cluster and the negative sample pairs to disperse.

[0007] Furthermore, the process of using fused features as input and embedding the distribution network power flow equation as a physical constraint into the neural network loss function to optimize and train the equivalent impedance of the line, and outputting the backtracking impedance of each node, includes: The fusion characteristics of each node are used as input, and the power flow equation of the distribution network is embedded into the loss function of the neural network as a physical constraint. The equivalent impedance of the line is set as a trainable parameter of the neural network. In backpropagation, the equivalent impedance of the line is automatically optimized and trained based on the joint gradient of physical loss and data loss to obtain the backtracking impedance of each node. The total loss function is composed of a weighted sum of physical loss and data loss. The data loss is the mean square error between the predicted voltage and the measured voltage, and the physical loss is the sum of the squares of the active power residual, the squares of the reactive power residual, and the squares of the voltage residual for all branches.

[0008] Furthermore, determining the deviation between the traceback impedance and the nominal impedance to obtain the impedance deviation of each node, and identifying latent defect nodes based on the infrared hotspot features in the joint fusion features of the impedance deviation, includes: Determine the nominal impedance and calculate the difference between the traceback impedance and the nominal impedance to obtain the impedance deviation of each node; Determine the preset deviation threshold and the infrared hotspot features in the fusion features of each node; When the impedance deviation of a node exceeds a preset deviation threshold and infrared hotspot characteristics indicate the presence of heating, the node is identified as having a latent high impedance defect and is thus designated as a latent defect node.

[0009] Furthermore, the simulation and deduction of the modification scheme based on the implicit defect nodes, and under the constraint of the capital pool, the optimal modification scheme and the expected return of each node are determined based on the simulation results, including: Determine the topological coordinates or topological node codes of the latent defect nodes, and limit the coding search space of the modification scheme based on the topological coordinates or topological node codes of the latent defect nodes; Chromosome encoding is performed on the modification schemes within the encoding search space, and after decoding, the power system simulation engine is called to perform power flow calculations on each modification scheme. The calculation constructs an objective function with the optimization objectives of minimizing the end voltage deviation, minimizing the side effects of the front voltage exceeding the limit, and minimizing the total life cycle cost. Under the constraint of the available funding pool, the optimal modification scheme for each node is generated through multi-objective genetic iteration, and the expected return of the optimal modification scheme is output.

[0010] Furthermore, the formulas for calculating the objective functions of minimizing the terminal voltage deviation, minimizing the front-end voltage over-limit side effects, and minimizing the total lifecycle cost are as follows: , in, This is the penalty value for the terminal voltage deviation. The number of end nodes, This is the lower limit of the acceptable voltage. Calculate the voltage magnitude for the power flow at the i-th node; , in, This is the penalty value for front-end overvoltage side effects. The number of front-end nodes. Calculate the voltage magnitude of the power flow at the j-th node. This is the upper limit of the acceptable voltage. , in, For total lifecycle cost, For initial investment costs, The project's economic evaluation period, The operating and maintenance cost in year t is... Let the line loss cost be in year t. is the discount rate.

[0011] Furthermore, the comparison between the actual benefits after the optimal modification scheme and the projected expected benefits to determine the benefit error, and the correction of the parameters in the neural network based on the benefit error, includes: A dataset is constructed based on the data after the optimal transformation scheme, and external confounding variables in the dataset are removed based on the dual machine learning residual method to quantify the real benefits of implementing the optimal transformation scheme for each node. Calculate the difference between the actual return and the projected return to obtain the return error, and determine the preset error threshold. When the revenue error exceeds the preset error threshold, the business chain reverse tracing based on the knowledge graph is triggered to locate the business breakpoints in the design, materials, and construction stages, and the revenue error is fed back into the neural network as a negative sample weight to correct the impedance backtracking parameters.

[0012] Furthermore, the dataset constructed based on the data after the optimal transformation scheme is used to remove external confounding variables from the dataset using a dual machine learning residual method, in order to quantify the real benefits of implementing the optimal transformation scheme for each node, including: Based on the data after the optimal modification scheme, intervention variables, outcome variables, and confounding variables were determined, and a dataset was constructed based on these variables. The dataset is randomly divided into K folds. Based on the XGBoost model, the conditional expectations of the outcome variable and the intervention variable with respect to the confounding variable are fitted on the training set for each fold. The residuals of the outcome variable and the intervention variable are calculated on the validation set. The residuals of all folds are then summed. Ordinary least squares regression was performed on the residuals of the aggregated outcome variables and the residuals of the intervention variables, and the regression coefficients in the least squares regression were determined as the real benefits after the implementation of the optimal transformation plan for each node.

[0013] This invention also provides a low-voltage diagnostic deduction and effectiveness retrospective system, comprising: The acquisition module is used to acquire multimodal monitoring data of each node in the distribution network, extract the features of the multimodal monitoring data and perform cross-modal fusion to obtain the fusion features of each node; The training module is used to optimize the equivalent impedance of the line by taking the fused features as input and embedding the power flow equation of the distribution network as a physical constraint into the loss function of the neural network, and outputting the backtracking impedance of each node. The determination module is used to determine the deviation between the backtracking impedance and the nominal impedance, obtain the impedance deviation of each node, and determine the hidden defect node based on the infrared hot spot feature in the joint fusion feature of the impedance deviation. The simulation module is used to simulate and extrapolate the transformation scheme based on the implicit defect nodes, and under the constraint of the capital pool, determine the optimal transformation scheme and the expected return of each node based on the simulation results. The correction module is used to compare the actual benefits after the optimal modification scheme with the projected expected benefits, determine the benefit error, and correct the parameters in the neural network based on the benefit error.

[0014] Compared with existing technologies, the low-voltage diagnosis deduction and effectiveness retrospective method and system of this invention have the following advantages: This invention utilizes a cross-attention mechanism to precisely align electrical timing anomalies with infrared hotspots and topological structures in a unified feature space, overcoming the limitations of single electrical data and accurately locating hidden defect nodes. This invention embeds the power flow equation of the distribution network as a physical hard constraint into a neural network and sets the impedance as a trainable parameter, so that the impedance backtracking strictly follows Kirchhoff's laws and gives the diagnostic results physical interpretability. This invention limits the coding search space by using hidden defect nodes, while introducing multi-dimensional fund pool constraints to automatically screen the optimal solution that is technically feasible, has no side effects, and is economically compliant, thus preventing blind line switching and abuse of new technologies. This invention utilizes a dual machine learning residual method to remove external confounding variables such as weather and load growth, accurately quantifies the true causal benefits of the transformation, and feeds the benefit error as a negative sample weight back to the diagnostic neural network. It adaptively increases the distrust level of nominal impedance, enabling the model to more sensitively capture hidden defects in the next iteration, thus forming an adaptive closed loop. This fundamentally solves the industry problems of unclear cause diagnosis, blind scheme deduction, inaccurate effect evaluation, and lack of management closed loop in low voltage governance, achieving the dual goals of precise governance investment and continuous self-evolution of the diagnostic model. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the process structure of the low voltage diagnosis deduction and effectiveness retrospective method in this embodiment of the invention; Figure 2 This is a schematic diagram of the composition of the low voltage diagnosis deduction and effectiveness retrospective system in an embodiment of the present invention. Detailed Implementation

[0016] The specific embodiments of this application will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0017] like Figure 1 As shown in the embodiments of this application, a method for low-voltage diagnosis deduction and effectiveness retrospective analysis is provided, including: S100: Acquire multimodal monitoring data of each node in the distribution network, extract the features of the multimodal monitoring data and perform cross-modal fusion to obtain the fused features of each node.

[0018] In a specific implementation process, the acquisition of multimodal monitoring data from each node of the distribution network, and the extraction of features from the multimodal monitoring data for cross-modal fusion to obtain the fused features of each node, includes: Acquire multimodal monitoring data of each node in the distribution network. The multimodal monitoring data includes electrical time-series data, visual environment data, topology and text data, and extract the single-modal features of the multimodal monitoring data respectively. Based on the cross-attention mechanism, each single modal feature is used as a key-value pair for cross-modal semantic space alignment. Through contrastive learning constraints, the single modal features of the same physical entity are aggregated in the latent space, and the fused features of each node are output.

[0019] Specifically, the system acquires electrical time-series data from each node of the distribution network, visual environment data collected by drone inspections, and GIS topology and text data. For each modality, a suitable feature extractor is used. The time-series modality is processed by a one-dimensional convolutional and temporal convolutional network to extract dynamic change patterns. The visual modality is processed by a visual Transformer to extract hot spots and equipment appearance features. The topology and text modality are processed by a graph sampling aggregation network and a BERT model to extract structural and semantic features, respectively. After obtaining heterogeneous single-modal features, a cross-modal semantic space alignment is performed using electrical time-series features as query vectors and visual and structured features as key-value pairs based on a cross-attention mechanism. This automatically establishes spatial semantic associations between electrical anomalies, infrared hot spots, and topological weak points in a unified feature space. Simultaneously, the InfoNCE contrastive learning loss function is introduced, forcing multimodal features from the same transformer area at the same time to approach each other in the latent space, while features from different transformer areas or at different times move away from each other, thus outputting fused features of each node rich in physical semantics. This step completely breaks through the limitations of traditional solutions that rely solely on single electrical data for perception. It enables subsequent diagnostic models not only to detect low voltage values, but also to understand comprehensive contextual information such as whether the node has oxidized and heated joints or is located at the end of the power supply radius. This achieves a qualitative leap in multi-source heterogeneous data from physical acquisition to semantic alignment and feature convergence, providing a complete and physically interpretable input foundation for the accurate backtracking of subsequent hidden defects.

[0020] In a specific implementation, the cross-attention mechanism aligns the single-modal features as key-value pairs across the semantic space, and uses contrastive learning constraints to aggregate the single-modal features of the same physical entity in the latent space, outputting the fused features of each node, including: The electrical time-series features are linearly mapped to obtain the query matrix, and the visual features and structured features are linearly mapped to obtain the key matrix and value matrix. Attention weights are obtained by normalizing the product of the transpose of the query matrix and the key matrix, divided by the square root of the dimension, and then multiplied by the value matrix to calculate the fusion features of each node. Positive and negative sample pairs are constructed based on each unimodal feature, and the cosine similarity between each unimodal feature in the fused feature is calculated using the InfoNCE loss function to force the positive sample pairs to cluster and the negative sample pairs to disperse.

[0021] Specifically, the electrical time-series features are linearly mapped to obtain a query matrix, and the visual and structured features are linearly mapped to obtain a key matrix and a value matrix, respectively. The product of the query matrix and the transpose of the key matrix is ​​scaled by dividing by the square root of the dimension and then normalized to obtain an attention weight distribution reflecting the semantic correlation between electrical anomalies and visual / topological features. The weights are then multiplied by the value matrix to output the fused features of each node. This allows the diagnostic model to automatically focus on the infrared hotspot location corresponding to the moment of most severe voltage fluctuation and the weakest power supply section in the topology, achieving on-demand focusing of cross-modal information rather than simple splicing. On this basis, single-modal features from the same transformer area at the same time are used as positive sample pairs, and features from different transformer areas or different times are used as negative sample pairs. The cosine similarity is calculated in the fused feature space using the InfoNCE loss function, forcing positive sample pairs to cluster closely in the latent space and negative sample pairs to move away from each other. This enables the electrical anomaly signal and the infrared heating signal from the same physical entity to achieve semantic locking at the feature level. This mechanism uses cross-attention to ensure that heterogeneous data from different sources no longer exist in isolation but corroborate each other. Even if a certain modality has noise or missing information, other modalities can still compensate for the lack of information through attention weights. Furthermore, the convergence effect of contrastive learning greatly enhances the robustness of fused features to individual differences in transformer substations. This allows the subsequent impedance backtracking model to stably use the triple semantic anchor points of electrical anomaly + visual heating + topological weakness as diagnostic criteria when facing different operating conditions, thereby significantly improving the recall and accuracy of latent defect localization.

[0022] S200: Taking the fused features as input and embedding the power flow equation of the distribution network as a physical constraint into the neural network loss function, it optimizes and trains the equivalent impedance of the line and outputs the backtracking impedance of each node.

[0023] In a specific implementation, the process involves taking fused features as input and embedding the distribution network power flow equation as a physical constraint into the neural network loss function to optimize and train the equivalent impedance of the line, outputting the backtracking impedance of each node, including: The fusion characteristics of each node are used as input, and the power flow equation of the distribution network is embedded into the loss function of the neural network as a physical constraint. The equivalent impedance of the line is set as a trainable parameter of the neural network. In backpropagation, the equivalent impedance of the line is automatically optimized and trained based on the joint gradient of physical loss and data loss to obtain the backtracking impedance of each node. The total loss function is composed of a weighted sum of physical loss and data loss. The data loss is the mean square error between the predicted voltage and the measured voltage, and the physical loss is the sum of the squares of the active power residual, the squares of the reactive power residual, and the squares of the voltage residual for all branches.

[0024] Specifically, the fused features of each node and the measured voltage are used as network inputs, and the equivalent impedance of the line is explicitly set as a trainable parameter of the neural network to participate in backpropagation iteration. At the same time, the radial DistFlow power flow equation of the distribution network is embedded in the loss function as an insurmountable physical hard constraint. The total loss is composed of the weighted sum of data loss and physical loss. Among them, the data loss constrains the predicted voltage to approximate the measured value to ensure the fitting accuracy, while the physical loss requires that the active, reactive, and voltage relationships of any adjacent nodes in the network strictly follow the DistFlow residual derived from Kirchhoff's laws, forcing the network training process to always proceed within the physically feasible solution space. In backpropagation, the Adam optimizer automatically updates the impedance parameters according to the guidance of the joint gradient of data loss and physical loss, so that it gradually converges from the nominal initial value to the optimized value that reflects the actual physical state, and finally outputs the backtracking impedance of each node. This step, while using a purely data-driven model, may yield mathematically fit but physically absurd diagnostic results under conditions not covered by the training samples. The existence of physical loss ensures that impedance optimization is always constrained by the fundamental law that the power flow of the distribution network must be balanced, guaranteeing that the diagnostic results do not violate basic electrical principles under any operating scenario, and significantly improving the model's extrapolation and generalization capabilities. By setting impedance as a trainable parameter and optimizing it in reverse under physical constraints, the algorithm can start from the appearance of low voltage at the end and accurately reverse the actual increase in resistance of each section along the line, effectively overcoming the black-box limitations of traditional methods and providing a quantitative basis for locating hidden defects that combines physical interpretability and numerical accuracy.

[0025] S300: Determine the deviation between the traceback impedance and the nominal impedance, obtain the impedance deviation of each node, and determine the hidden defect node based on the infrared hot spot feature in the impedance deviation joint fusion feature.

[0026] In a specific implementation process, determining the deviation between the traceback impedance and the nominal impedance, obtaining the impedance deviation of each node, and identifying latent defect nodes based on the infrared hotspot features in the joint fusion features of the impedance deviations, includes: Determine the nominal impedance and calculate the difference between the traceback impedance and the nominal impedance to obtain the impedance deviation of each node; Determine the preset deviation threshold and the infrared hotspot features in the fusion features of each node; When the impedance deviation of a node exceeds a preset deviation threshold and infrared hotspot characteristics indicate the presence of heating, the node is identified as having a latent high impedance defect and is thus designated as a latent defect node.

[0027] Specifically, using the nominal impedance of the lines provided by the distribution network GIS ledger as a benchmark, the absolute difference between the traceback impedance output after PINN physical constraint optimization training and the nominal impedance is calculated to obtain the impedance deviation of the line segments associated with each node. This deviation quantitatively characterizes the abnormal increase in the actual impedance of the segment compared to its healthy state. Based on the fusion features of each node, the semantic features of infrared hot spots are extracted. These features reflect the actual heating state of the equipment captured by the UAV inspection. When a node simultaneously meets the two conditions of impedance deviation being greater than the preset deviation threshold and infrared hot spot features indicating that the node has obvious heating phenomena, it is determined that the node has a hidden high impedance defect caused by joint oxidation, bolt loosening, or contact burning, and it is marked as a hidden defect node. Its spatial coordinates are output as the spatial positioning basis for subsequent scheme deduction. This step addresses the issue that relying solely on impedance deviation for judgment can easily misjudge design factors such as excessively long lines or thin wire diameters as construction defects like poor contact. However, the introduction of infrared hotspots effectively eliminates non-defect scenarios such as inherently large voltage drops in long lines. Although the line has high impedance, it does not generate heat, thus avoiding false alarms. When there are only infrared hotspots without impedance deviation evidence, the heat generation may be due to non-contact reasons such as excessive load or poor heat dissipation, which also will not trigger defect judgment. The "AND" logic relationship between the two achieves a balance between extremely high diagnostic accuracy and extremely low false alarm rate, giving the hidden defect nodes output by the system a credible label that is both abnormal and verifiable. This provides a reliable spatial positioning basis for subsequent scheme deduction to accurately allocate limited governance resources to the physical locations that truly need repair.

[0028] S400: Based on the implicit defect node limitation transformation scheme, the scheme simulation is performed, and under the constraint of the capital pool, the optimal transformation scheme and the expected return of each node are determined based on the simulation results.

[0029] In a specific implementation process, the modification scheme based on the implicit defect node constraint is simulated and deduced, and under the constraint of the capital pool, the optimal modification scheme and the expected return of each node are determined based on the deduction results, including: Determine the topological coordinates or topological node codes of the latent defect nodes, and limit the coding search space of the modification scheme based on the topological coordinates or topological node codes of the latent defect nodes; Chromosome encoding is performed on the modification schemes within the encoding search space, and after decoding, the power system simulation engine is called to perform power flow calculations on each modification scheme. The calculation constructs an objective function with the optimization objectives of minimizing the end voltage deviation, minimizing the side effects of the front voltage exceeding the limit, and minimizing the total life cycle cost. Under the constraint of the available funding pool, the optimal modification scheme for each node is generated through multi-objective genetic iteration, and the expected return of the optimal modification scheme is output.

[0030] Specifically, the topological coordinates or topological node codes of latent defect nodes are extracted and used as spatial anchors to limit the search space of chromosome coding for modification schemes. This allows the genetic algorithm's gene position mutation and crossover operations to focus on the feasible measures set for the segment where the latent defect is located and its downstream affected nodes, fundamentally preventing the investment of ineffective modification resources in defect-free areas. Within the limited search space, chromosome coding is performed on each modification scheme. After decoding, power flow calculations are performed using power system simulation engines such as OpenDSS or BPA. Simultaneously, a fitness function is constructed with three objectives: minimizing end-voltage deviation, minimizing the side effects of front-end voltage exceeding limits, and minimizing the total life cycle cost. During optimization iterations, it is verified in real time whether the initial investment in the total life cycle cost of each scheme exceeds the available amount of the current urban and rural power grid or technical modification fund pool. Schemes exceeding the limits are automatically eliminated by imposing an infinite penalty. Pareto optimal frontier is generated through fast non-dominated sorting and congestion distance calculation using a non-dominated sorting genetic algorithm. Based on the actual fund pool attributes, the comprehensive optimal scheme is automatically recommended and its expected return is output. This step limits the coding search space by using the spatial coordinates of hidden defect nodes, transforming traditional brute-force simulation into targeted and precise inference, which improves the algorithm's convergence speed and avoids wasting resources on healthy lines. The explicit quantification of front-end overvoltage side effects effectively curbs the risk of voltage exceeding limits caused by excessive reactive power compensation or improper energy storage charging. The linkage between the full life cycle cost and the capital pool limit ensures that the solution is not only technically feasible but also economically feasible for engineering implementation. This completely changes the blindness and impracticality of traditional single-solution trial-and-error simulation, providing a globally optimal decision-making basis that is technically feasible, economically compliant, and without side effects for low-voltage governance investment.

[0031] In a specific implementation process, the formulas for calculating the objective functions of minimizing the terminal voltage deviation, minimizing the front-end voltage over-limit side effects, and minimizing the total lifecycle cost are as follows: , in, This is the penalty value for the terminal voltage deviation. The number of end nodes, This is the lower limit of the acceptable voltage. Calculate the voltage magnitude for the power flow at the i-th node; , in, This is the penalty value for front-end overvoltage side effects. The number of front-end nodes. Calculate the voltage magnitude of the power flow at the j-th node. This is the upper limit of the acceptable voltage. , in, For total lifecycle cost, For initial investment costs, The project's economic evaluation period, The operating and maintenance cost in year t is... Let the line loss cost be in year t. is the discount rate.

[0032] S500: Compare the actual benefits after the optimal modification scheme with the projected expected benefits to determine the benefit error, and correct the parameters in the neural network based on the benefit error.

[0033] In a specific implementation process, the actual benefit after the optimal modification scheme is compared with the projected expected benefit to determine the benefit error, and the parameters in the neural network are corrected based on the benefit error, including: A dataset is constructed based on the data after the optimal transformation scheme, and external confounding variables in the dataset are removed based on the dual machine learning residual method to quantify the real benefits of implementing the optimal transformation scheme for each node. Calculate the difference between the actual return and the projected return to obtain the return error, and determine the preset error threshold. When the revenue error exceeds the preset error threshold, the business chain reverse tracing based on the knowledge graph is triggered to locate the business breakpoints in the design, materials, and construction stages, and the revenue error is fed back into the neural network as a negative sample weight to correct the impedance backtracking parameters.

[0034] Specifically, an effectiveness evaluation dataset is constructed based on actual operational data after the implementation of the optimal transformation plan. A dual machine learning residualization solution method is introduced, and the dataset is randomly divided into K folds for cross-fitting. In the first stage, XGBoost is used to fit the conditional expectations of the outcome variable and the intervention variable with respect to the confounding variable on the training set and calculate the residuals. In the second stage, ordinary least squares regression is performed on the aggregated residuals, and the obtained regression coefficients are the average treatment effect after removing external confounding factors such as meteorological fluctuations and natural load growth, thereby quantifying the real causal benefits after the implementation of the optimal transformation plan at each node. The difference between the real benefit and the projected expected benefit is calculated to obtain the benefit error. When the benefit error exceeds a preset error threshold... When the value is reached, the knowledge graph reverse tracing mechanism based on the Neo4j graph database is triggered. Cypher queries are performed along the entire chain from project to design drawings to material batches to construction logs to acceptance photos. The actual wire diameter of the material batch is automatically compared with the nominal diameter on the design drawings. The image recognition model is used to compare the equipment nameplate in the acceptance photos with the material batch. The root causes of business breakpoints such as construction downgrades or material mismatches are accurately located and rectification work orders are generated. At the same time, the error is fed back as a negative sample weight to the physical information neural network and added to the physical constraint weight coefficient of the total loss function. This forces the model to adaptively increase the distrust of nominal impedance in the next iteration of training and more sensitively identify similar hidden defects. The dual machine learning in this step elevates causal inference from correlation to causality through residual processing, completely solving the industry pain point that high temperatures or load increases in traditional evaluations can mask or amplify the true effectiveness of governance, making the evaluation of transformation results objective and credible. The knowledge graph reverse tracing accurately breaks down the technical level of unsatisfactory benefits into management level design-materials-construction breakpoints, preventing similar errors from recurring due to lack of evidence. The benefit error drives the self-evolution of the PINN model in the form of negative sample weights, so that the success or failure of each governance experience is transformed into an improvement in the immunity of the diagnostic model, constructing a complete business closed loop of diagnosis-deduction-implementation-evaluation-feedback-evolution, fundamentally solving the long-standing industry problems of neglect after governance, management without evaluation, and evaluation without feedback in traditional low voltage governance.

[0035] In a specific implementation process, the data is constructed based on the data after the optimal transformation scheme. External confounding variables are removed from the dataset using a dual machine learning residual method to quantify the actual benefits of implementing the optimal transformation scheme at each node. This includes: Based on the data after the optimal modification scheme, intervention variables, outcome variables, and confounding variables were determined, and a dataset was constructed based on these variables. The dataset is randomly divided into K folds. Based on the XGBoost model, the conditional expectations of the outcome variable and the intervention variable with respect to the confounding variable are fitted on the training set for each fold. The residuals of the outcome variable and the intervention variable are calculated on the validation set. The residuals of all folds are then summed. Ordinary least squares regression was performed on the residuals of the aggregated outcome variables and the residuals of the intervention variables, and the regression coefficients in the least squares regression were determined as the real benefits after the implementation of the optimal transformation plan for each node.

[0036] Specifically, based on the actual operational data after the optimal transformation scheme, intervention variables, outcome variables, and confounding variables are clearly defined. A performance evaluation dataset containing multidimensional variables is constructed accordingly. After randomly dividing the dataset into K folds, a two-stage residual calculation strategy using dual machine learning is employed. In the first stage, the XGBoost model is used to fit the conditional expectation functions of the outcome variable and intervention variable with respect to the confounding variables on the training set for each fold. The influence of the confounding variables is then removed on the validation set to calculate the residuals of the outcome variable and intervention variable. The residuals of all folds are then aggregated to form a biased residual dataset. In the second stage, ordinary least squares regression is performed on the aggregated residuals of the outcome variable and intervention variable. The regression coefficient θ in this regression equation is determined as the true causal benefit after the implementation of the optimal transformation scheme at each node. This coefficient statistically represents the independent change in the outcome variable caused by each unit change in the intervention variable, eliminating path interference from all known confounding variables. Compared to traditional effectiveness evaluation methods that simply compare the absolute voltage values ​​before and after the upgrade, this step is prone to misjudging the surge in load caused by high summer temperatures or the voltage rebound caused by light winter loads as upgrade results. However, the dual machine learning of this invention, through residual processing, mathematically and rigorously separates the effect of the upgrade itself from the effect of external environmental fluctuations. This is equivalent to constructing a parallel counterfactual comparison that is independent of weather and load, making the final output regression coefficient θ a pure upgrade benefit stripped of all external interference. At the same time, XGBoost, as a base learner, can automatically capture the complex nonlinear relationship between confounding variables and outcome variables, avoiding the estimation bias caused by incorrect model settings in traditional linear regression models. This provides a statistically rigorous and engineering-reliable quantitative basis for determining whether the actual benefit is lower than expected, effectively supporting the scientific setting of subsequent business breakpoint tracing thresholds.

[0037] like Figure 2 As shown in the embodiments of this application, a low-voltage diagnosis simulation and effectiveness retrospective system is provided, including: The acquisition module 100 is used to acquire multimodal monitoring data of each node in the distribution network, extract the features of the multimodal monitoring data and perform cross-modal fusion to obtain the fusion features of each node; Training module 200 is used to optimize the line equivalent impedance by taking the fused features as input and embedding the distribution network power flow equation as a physical constraint into the neural network loss function, and outputting the backtracking impedance of each node. The determination module 300 is used to determine the deviation between the traceback impedance and the nominal impedance, obtain the impedance deviation of each node, and determine the hidden defect node based on the infrared hot spot feature in the impedance deviation joint fusion feature. The simulation module 400 is used to perform simulation simulations of the transformation scheme based on the implicit defect nodes, and under the constraint of the capital pool, to determine the optimal transformation scheme and the expected return of each node based on the simulation results. The correction module 500 is used to compare the actual benefits after the optimal modification scheme with the projected expected benefits, determine the benefit error, and correct the parameters in the neural network based on the benefit error.

[0038] In summary, this invention provides a method and system for low-voltage diagnosis, simulation, and effectiveness retrospective analysis. The method includes: acquiring multi-modal monitoring data from each node of the distribution network; extracting features and performing cross-modal fusion to obtain fused features for each node; using the fused features as input, embedding the power flow equation as a physical constraint into a neural network loss function, setting the line equivalent impedance as a trainable parameter for optimization training, and outputting the retrospective impedance of each node; calculating the deviation between the retrospective impedance and the nominal impedance, and using infrared hotspot features to determine latent defect nodes; limiting the transformation coding search space with latent defect nodes, and performing multi-objective scheme simulation under a funding pool constraint to determine the optimal transformation scheme and the expected return; comparing the actual return after transformation with the expected return to determine the return error, and feeding back to correct the neural network parameters. This invention establishes a complete closed-loop system for low-voltage governance, effectively improving diagnostic accuracy and investment effectiveness.

[0039] Finally, it should be noted that those skilled in the art can obviously make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for low-voltage diagnosis deduction and effectiveness retrospective analysis, characterized in that, include: Acquire multimodal monitoring data of each node in the distribution network, extract features from the multimodal monitoring data and perform cross-modal fusion to obtain the fused features of each node; Using the fusion features as input and embedding the distribution network power flow equation as a physical constraint into the neural network loss function, the equivalent impedance of the line is optimized and trained, and the backtracking impedance of each node is output. The deviation between the traceback impedance and the nominal impedance is determined to obtain the impedance deviation of each node, and the hidden defect node is determined based on the infrared hot spot feature in the joint fusion feature of the impedance deviation. Based on the implicit defect node-limited renovation scheme, the scheme simulation is carried out, and under the constraint of the capital pool, the optimal renovation scheme and the expected return of each node are determined based on the simulation results. The actual benefits after the optimal modification scheme are compared with the projected expected benefits to determine the benefit error, and the parameters in the neural network are corrected based on the benefit error.

2. The low-voltage diagnosis deduction and effectiveness retrospective method according to claim 1, characterized in that, The process involves acquiring multimodal monitoring data from each node of the distribution network, extracting features from the multimodal monitoring data, and performing cross-modal fusion to obtain the fused features of each node, including: Acquire multimodal monitoring data of each node in the distribution network. The multimodal monitoring data includes electrical time-series data, visual environment data, topology and text data, and extract the single-modal features of the multimodal monitoring data respectively. Based on the cross-attention mechanism, each single modal feature is used as a key-value pair for cross-modal semantic space alignment. Through contrastive learning constraints, the single modal features of the same physical entity are aggregated in the latent space, and the fused features of each node are output.

3. The low-voltage diagnosis deduction and effectiveness retrospective method according to claim 2, characterized in that, The cross-attention mechanism aligns each unimodal feature as a key-value pair across the semantic space, and uses contrastive learning constraints to aggregate the unimodal features of the same physical entity in the latent space, outputting the fused features of each node, including: The electrical time-series features are linearly mapped to obtain the query matrix, and the visual features and structured features are linearly mapped to obtain the key matrix and value matrix. Attention weights are obtained by normalizing the product of the transpose of the query matrix and the key matrix, divided by the square root of the dimension, and then multiplied by the value matrix to calculate the fusion features of each node. Positive and negative sample pairs are constructed based on each unimodal feature, and the cosine similarity between each unimodal feature in the fused feature is calculated using the InfoNCE loss function to force the positive sample pairs to cluster and the negative sample pairs to disperse.

4. The low-voltage diagnosis deduction and effectiveness retrospective method according to claim 1, characterized in that, The process involves using fused features as input and embedding the distribution network power flow equation as a physical constraint into the neural network loss function to optimize and train the equivalent impedance of the line, outputting the backtracking impedance of each node, including: The fusion characteristics of each node are used as input, and the power flow equation of the distribution network is embedded into the loss function of the neural network as a physical constraint. The equivalent impedance of the line is set as a trainable parameter of the neural network. In backpropagation, the equivalent impedance of the line is automatically optimized and trained based on the joint gradient of physical loss and data loss to obtain the backtracking impedance of each node. The total loss function is composed of a weighted sum of physical loss and data loss. The data loss is the mean square error between the predicted voltage and the measured voltage, and the physical loss is the sum of the squares of the active power residual, the squares of the reactive power residual, and the squares of the voltage residual for all branches.

5. The method for low-voltage diagnosis deduction and effectiveness retrospection according to claim 1, characterized in that, The process of determining the deviation between the traceback impedance and the nominal impedance, obtaining the impedance deviation of each node, and identifying latent defect nodes based on the infrared hotspot features in the joint fusion features of the impedance deviation includes: Determine the nominal impedance and calculate the difference between the traceback impedance and the nominal impedance to obtain the impedance deviation of each node; Determine the preset deviation threshold and the infrared hotspot features in the fusion features of each node; When the impedance deviation of a node exceeds a preset deviation threshold and infrared hotspot characteristics indicate the presence of heating, the node is identified as having a latent high impedance defect and is thus designated as a latent defect node.

6. The method for low-voltage diagnosis deduction and effectiveness retrospection according to claim 1, characterized in that, The simulation and deduction of the modification scheme based on the implicit defect node constraint, and under the constraint of the capital pool, the optimal modification scheme and the expected return of each node are determined based on the simulation results, including: Determine the topological coordinates or topological node codes of the latent defect nodes, and limit the coding search space of the modification scheme based on the topological coordinates or topological node codes of the latent defect nodes; Chromosome encoding is performed on the modification schemes within the encoding search space, and after decoding, the power system simulation engine is called to perform power flow calculations on each modification scheme. The calculation constructs an objective function with the optimization objectives of minimizing the end voltage deviation, minimizing the side effects of the front voltage exceeding the limit, and minimizing the total life cycle cost. Under the constraint of the available funding pool, the optimal modification scheme for each node is generated through multi-objective genetic iteration, and the expected return of the optimal modification scheme is output.

7. The low-voltage diagnosis deduction and effectiveness retrospective method according to claim 6, characterized in that, The formulas for calculating the objective functions of minimizing the terminal voltage deviation, minimizing the side effects of front-end voltage exceeding limits, and minimizing the total life-cycle cost are as follows: ， in, This is the penalty value for the terminal voltage deviation. The number of end nodes, This is the lower limit of the acceptable voltage. Calculate the voltage magnitude for the power flow at the i-th node; ， in, This is the penalty value for front-end overvoltage side effects. The number of front-end nodes. Calculate the voltage magnitude of the power flow at the j-th node. This is the upper limit of the acceptable voltage. ， in, For total lifecycle cost, For initial investment costs, The project's economic evaluation period, The operating and maintenance cost in year t is... Let the line loss cost be in year t. is the discount rate.

8. The method for low-voltage diagnosis deduction and effectiveness retrospection according to claim 1, characterized in that, The process of comparing the actual benefits after the optimal modification scheme with the projected expected benefits to determine the benefit error, and then correcting the parameters in the neural network based on the benefit error, includes: A dataset is constructed based on the data after the optimal transformation scheme, and external confounding variables in the dataset are removed based on the dual machine learning residual method to quantify the real benefits of implementing the optimal transformation scheme for each node. Calculate the difference between the actual return and the projected return to obtain the return error, and determine the preset error threshold. When the revenue error exceeds the preset error threshold, the business chain reverse tracing based on the knowledge graph is triggered to locate the business breakpoints in the design, materials, and construction stages, and the revenue error is fed back into the neural network as a negative sample weight to correct the impedance backtracking parameters.

9. The method for low-voltage diagnosis deduction and effectiveness retrospection according to claim 8, characterized in that, The dataset is constructed based on the data after the optimal transformation scheme. External confounding variables are removed from the dataset using a dual machine learning residual method to quantify the actual benefits of implementing the optimal transformation scheme at each node, including: Based on the data after the optimal modification scheme, intervention variables, outcome variables, and confounding variables were determined, and a dataset was constructed based on these variables. The dataset is randomly divided into K folds. Based on the XGBoost model, the conditional expectations of the outcome variable and the intervention variable with respect to the confounding variable are fitted on the training set for each fold. The residuals of the outcome variable and the intervention variable are calculated on the validation set. The residuals of all folds are then summed. Ordinary least squares regression was performed on the residuals of the aggregated outcome variables and the residuals of the intervention variables, and the regression coefficients in the least squares regression were determined as the actual benefits after the implementation of the optimal transformation plan for each node.

10. A low-voltage diagnosis and deduction system with retrospective results, characterized in that, include: The acquisition module is used to acquire multimodal monitoring data of each node in the distribution network, extract the features of the multimodal monitoring data and perform cross-modal fusion to obtain the fusion features of each node; The training module is used to optimize the equivalent impedance of the line by taking the fused features as input and embedding the power flow equation of the distribution network as a physical constraint into the loss function of the neural network, and outputting the backtracking impedance of each node. The determination module is used to determine the deviation between the traceback impedance and the nominal impedance, obtain the impedance deviation of each node, and determine the hidden defect node based on the infrared hot spot feature in the joint fusion feature of the impedance deviation. The simulation module is used to simulate and extrapolate the transformation scheme based on the implicit defect nodes, and under the constraint of the capital pool, determine the optimal transformation scheme and the expected return of each node based on the simulation results. The correction module is used to compare the actual benefits after the optimal modification scheme with the projected expected benefits, determine the benefit error, and correct the parameters in the neural network based on the benefit error.

Images