Online low-voltage deduction method and system for distribution area based on synergistic driving modeling

Abstract

The application discloses an online low-voltage deduction method and system for a power distribution area based on cooperative driving modeling, specifically: the mechanism base layer adopts linearized DistFlow for deduction, and is embedded with topology checking and power balance verification to output basic voltage distribution; the data-driven compensation layer learns the spatial coupling and time dynamic characteristics of voltage residuals by using GCN-LSTM, adopts PINN to embed power flow / physical constraints into network training, and synthesizes to obtain dynamic compensation; a hybrid deduction engine calculates adaptive weights and fuses to obtain voltage prediction; a rolling deduction and incremental learning module continuously outputs voltage situation spectrum, and triggers online adjustment of the parameters of the data-driven compensation layer when the error exceeds the threshold; a two-way feedback loop corrects the input deviation of the mechanism base layer and adjusts the data-driven compensation layer. The application greatly improves the accuracy and real-time response capability of voltage deduction while ensuring physical consistency.

Description

Technical Field

[0001] This invention relates to the field of voltage control and simulation technology, and in particular to an online low-voltage simulation method for distribution substations based on collaborative driving modeling. Background Technology

[0002] With the advancement of the "dual carbon" target, the penetration rate of distributed energy in distribution networks is constantly increasing, posing greater challenges to voltage state management in distribution substations. Voltage projection has significant strategic value in distribution network dispatching: it provides decision-making basis for active voltage control, source-grid-load-storage coordinated optimization, and distributed energy dispatching, ensuring stable grid operation. However, due to the rapid changes in distributed photovoltaic, energy storage, and flexible loads, voltage fluctuations in substations are frequent, and existing traditional voltage projection methods are no longer sufficient to meet the requirements of high timeliness and high accuracy. Current voltage projection faces several challenges: ① Inconsistent spatiotemporal data: Different data sources (such as SCADA, PMU, photovoltaic forecasts, etc.) have different timeliness and accuracy, making it difficult to effectively combine them for real-time projection; ② Strong voltage volatility: The volatility of new energy generation and the randomness of loads exacerbate rapid voltage changes, making it difficult for traditional projection methods to reflect this dynamic change in real time; ③ High requirements for computational and physical consistency: The projection results must conform to the physical laws of the distribution network and be able to adjust quickly in the event of equipment failure or topology changes to ensure the accuracy of control commands. Therefore, voltage extrapolation for distribution substations requires innovative methods that can be efficiently calculated, adapt to complex dynamics, and maintain physical consistency.

[0003] Existing voltage control and extrapolation techniques are generally divided into two categories: computational models based on physical mechanisms and data-driven methods.

[0004] (1) Model based on physical mechanism

[0005] Calculations are performed using the power grid's topology and line parameters, such as power flow calculations. While these methods offer strong physical interpretation, their high computational complexity makes them unsuitable for real-time response, especially in large-scale distribution networks. Furthermore, the calculation results of the physical model heavily rely on accurate grid parameters; therefore, the model's adaptability is poor in the event of topology changes or equipment failures, failing to respond promptly to changes in the power grid.

[0006] (2) Data-driven model

[0007] For example, deep learning and neural networks rely on historical data to extrapolate voltage, enabling relatively fast calculations. However, most of these methods lack physical consistency, leading to a significant reduction in the reliability and accuracy of predictions when the power grid undergoes topological changes or extreme weather events.

[0008] Currently, both pure physical models and pure data-driven models have their limitations, and the two have not been effectively integrated, resulting in a lack of a voltage extrapolation method that can both ensure physical consistency and quickly respond to changes in the power grid.

[0009] In summary, current technologies have significant gaps in addressing voltage prediction issues under conditions of high-proportion renewable energy integration, mainly in the following aspects:

[0010] (1) Insufficient real-time performance: In high-penetration new energy scenarios, the voltage change rate changes from minutes to seconds or even faster. The traditional calculation method based on power flow calculation is computationally intensive and slow to update, often outputting "outdated voltage". It cannot keep up with actual operation and is difficult to serve as a pre-conditioning for active voltage regulation. It cannot meet the needs of rapid changes in the power grid, especially when distributed photovoltaic power generation fluctuates greatly.

[0011] (2) Poor adaptability: The operation of transformer area switching, the switching of distributed power sources, and the charging and discharging of energy storage will change the topology and power flow. Existing methods are mostly fixed models and fixed parameters. When encountering such changes in operation mode, it is necessary to remodel or perform batch calculations. It is difficult to provide reliable results in a short period of time and cannot adapt to complex dynamic conditions such as changes in grid topology, equipment failures, and fluctuations in new energy output.

[0012] (3) Lack of physical consistency: Although the pure data-driven model is fast, it does not take into account the grid structure, line impedance, or power balance. Extreme scenarios outside the training data (sudden photovoltaic drop, communication loss, topology change) may give voltage values ​​that do not conform to the grid rules, making it difficult to guarantee the physical rationality of the output and prone to failure under complex operating conditions.

[0013] Therefore, there is an urgent need for a voltage extrapolation method that combines the advantages of physical mechanisms and data-driven approaches, which can provide high-precision and high-real-time voltage prediction while ensuring physical consistency. Summary of the Invention

[0014] The purpose of this invention is to provide an online low-voltage simulation method and system for distribution substations based on collaborative driving modeling. By introducing a collaborative driving modeling method, the mechanism is combined with a data-driven model, integrating the advantages of physical mechanisms and deep learning algorithms. While ensuring physical consistency, the accuracy and real-time response capability of voltage simulation are greatly improved.

[0015] The technical solution to achieve the purpose of this invention is as follows: an online low-voltage simulation method for distribution substations based on collaborative driving modeling. This method adopts a collaborative driving closed-loop architecture of mechanism-data-fusion-rolling-feedback, and includes a mechanism base layer, a data-driven compensation layer, a hybrid simulation engine, a rolling simulation and incremental learning module, and a bidirectional feedback loop. The specific simulation process is as follows:

[0016] Step 1: Integrate two types of inputs: one is multi-source operational and prediction data; the other is network structure and parameter information.

[0017] Step 2: In the mechanistic foundation layer, based on the topology and line parameters, linearized DistFlow is used for deduction, and topology verification and power balance verification are embedded to output the base voltage distribution. ;

[0018] Step 3: The data-driven compensation layer takes the output of the mechanistic foundation layer and the multi-source spatiotemporal characteristics as inputs, and obtains voltage residual compensation through two parallel paths: Mode A uses GCN-LSTM to learn the spatial coupling and temporal dynamic characteristics of the voltage residual, while Mode B uses PINN to embed power flow / physical constraints into the network training to ensure the physical consistency of the compensation results; the two are then adaptively synthesized through confidence evaluation to obtain the dynamic compensation amount. ;

[0019] Step 4: The hybrid inference engine adaptively calculates weights based on real-time data quality and volatility levels. ,Will and The final voltage prediction is obtained by fusion. ;

[0020] Step 5: The rolling simulation and incremental learning module updates the sliding window with a fixed step size, continuously outputs the voltage situation spectrum within a set time period in the future, and triggers online adjustment of the data-driven compensation layer parameters when the error exceeds the threshold.

[0021] Step 6: Two-way feedback loop to transmit the deduction results back: on the one hand, it is used to correct the input deviation of the mechanism base layer, and on the other hand, it is used to adjust the fusion weight and physical constraint strength of the data-driven compensation layer.

[0022] An online low-voltage simulation system for distribution substations based on collaborative-driven modeling is provided. This system implements the aforementioned online low-voltage simulation method for distribution substations based on collaborative-driven modeling. The system includes a mechanism foundation layer, a data-driven compensation layer, a hybrid simulation engine, a rolling simulation and incremental learning module, and a bidirectional feedback loop.

[0023] The mechanistic base layer, based on topology and line parameters, uses linearized DistFlow for derivation, and incorporates topology verification and power balance verification to output the base voltage distribution. ;

[0024] The data-driven compensation layer takes the output of the mechanistic foundation layer and multi-source spatiotemporal characteristics as inputs, and obtains voltage residual compensation through two parallel paths: Mode A uses GCN-LSTM to learn the spatial coupling and temporal dynamic characteristics of the voltage residual, while Mode B uses PINN to embed power flow / physical constraints into the network training to ensure the physical consistency of the compensation results; the two are then adaptively synthesized through confidence evaluation to obtain the dynamic compensation amount. ;

[0025] The hybrid inference engine adaptively calculates weights based on real-time data quality and volatility levels. ,Will and The final voltage prediction is obtained by fusion. ;

[0026] The rolling simulation and incremental learning module updates the sliding window with a fixed step size, continuously outputs the voltage situation spectrum within a set time period in the future, and triggers online adjustment of the data-driven compensation layer parameters when the error exceeds the threshold.

[0027] The bidirectional feedback loop is used to transmit the inference results back. On the one hand, it is used to correct the input deviation of the mechanism base layer, and on the other hand, it is used to adjust the fusion weight and physical constraint strength of the data-driven compensation layer.

[0028] Compared with the prior art, the significant advantages of this invention are:

[0029] (1) Improved voltage prediction accuracy: Based on the mechanism calculation, the present invention superimposes data-driven compensation, which can significantly reduce voltage prediction error. In scenarios with photovoltaic fluctuations, the average prediction error can be reduced from about 0.58% of the traditional method to about 0.28%, which can meet the accuracy requirements of voltage early warning and active voltage regulation at the distribution area level.

[0030] (2) Maintaining data utility: By adopting linearized DistFlow and lightweight spatiotemporal model, the single simulation time can be controlled to about 80 ms, which is suitable for 5-second rolling simulation scenarios. It can continuously output the voltage status for a period of time in the future, providing continuous input for real-time scheduling.

[0031] (3) Optimized robustness of operating conditions: Through the adaptive fusion of mechanism results and data compensation, as well as the incremental learning mechanism triggered by error, the stability of the inference results can still be maintained under operating conditions such as topology adjustment, data loss, and sudden changes in photovoltaic output, avoiding inference distortion caused by mismatch of a single model;

[0032] (4) Active voltage control support effect: The simulation results can be used as a prerequisite for reactive power compensation devices, distributed power inverter voltage regulation and distribution network reconfiguration strategies. It can identify possible voltage over-limit sections in advance, reduce over-limit time, reduce the frequency of adjustment actions, thereby improving the safety and economy of distribution network operation. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the online low-voltage simulation method for distribution substations based on collaborative driving modeling, as described in this invention.

[0034] Figure 2 This is a diagram of the four-layer closed-loop structure adopted in this invention: "mechanism base layer - data-driven compensation layer - hybrid inference engine - rolling update and feedback".

[0035] Figure 3 This is a comparison chart of the expected errors in voltage prediction using the pure mechanistic model, the pure data model, and the model of this invention.

[0036] Figure 4 This is a graph showing the changes in various loss values ​​during the training phase. Detailed Implementation

[0037] This invention proposes an online low-voltage simulation method for distribution substations based on collaborative driving modeling. Combining physical mechanisms and data-driven models, this method adopts a collaborative driving closed-loop architecture of "mechanism-data-fusion-rolling-feedback," comprising a mechanism base layer, a data-driven compensation layer, a hybrid simulation engine, a rolling simulation and incremental learning module, and a bidirectional feedback loop. Figure 1 As shown, the specific deduction process is as follows:

[0038] Step 1: Integrate two types of inputs: one is multi-source operational and prediction data; the other is network structure and parameter information.

[0039] Step 2: In the mechanistic foundation layer, based on the topology and line parameters, linearized DistFlow is used for deduction, and topology verification and power balance verification are embedded to output the base voltage distribution. ;

[0040] Step 3: The data-driven compensation layer takes the output of the mechanistic foundation layer and the multi-source spatiotemporal characteristics as inputs, and obtains voltage residual compensation through two parallel paths: Mode A uses GCN-LSTM to learn the spatial coupling and temporal dynamic characteristics of the voltage residual, while Mode B uses PINN (Physical Information Neural Network) to embed power flow / physical constraints into the network training to ensure the physical consistency of the compensation results; the two are then adaptively synthesized through confidence evaluation to obtain the dynamic compensation amount. ;

[0041] Step 4: The hybrid inference engine adaptively calculates weights based on real-time data quality and volatility levels. ,Will and The final voltage prediction is obtained by fusing (and potentially overlaying ARIMA trend term corrections). ;

[0042] Step 5: The rolling simulation and incremental learning module updates the sliding window with a fixed step size (e.g., 5s), continuously outputs the voltage situation spectrum within a future set time (e.g., 15min), and triggers online adjustment of the data-driven compensation layer parameters when the error exceeds the threshold, thereby enhancing the adaptability to changes in operating conditions.

[0043] Step 6: Bidirectional feedback loop, transmitting back the deduction results: on the one hand, it is used to correct the input deviation of the mechanism base layer, and on the other hand, it is used to adjust the fusion weight and physical constraint strength of the data-driven compensation layer, thereby achieving simultaneous improvement in accuracy and stability during continuous operation.

[0044] As a specific example, the multi-source operation and forecasting data mentioned in step 1 includes SCADA measurements, PMU synchronization phasors, and meteorological / photovoltaic / load forecasts; the network structure and parameter information includes topology and switch status, and line parameters. wait.

[0045] As a specific example, step 2 is as follows:

[0046] In the mechanistic foundation layer, based on the topology of the distribution network, node parameters, and distributed generation information, the voltage distribution is calculated using the linearized DistFlow equation as the basis for derivation. The following is the voltage distribution calculation process:

[0047] Using the low-voltage busbar / feeder outlet of the transformer in the distribution area as the reference node, denoted as node 0, and the voltage amplitude as the nominal value V0, a ​​radial directed tree is constructed by combining the line set E and the node set N of the low-voltage distribution area in the switch-closed state. For any branch road The circuit parameters are resistance. Reactance (This can be calculated from the GIS line length and unit impedance).

[0048] Define nodes Net active power Net reactive power Let the active power and reactive power absorbed from the network be respectively (load is taken as positive, and distributed generation output is taken as negative). Then, when the online loss term is ignored or treated as a second-order term, the branch power and node power satisfy the recursive relationship of linearized DistFlow:

[0049] Branch power recursion (bottom-up aggregation):

[0050] (1)

[0051] (2)

[0052] Voltage recursion (top-down calculation):

[0053] (3)

[0054] in, Indicates a branch Active power flowing upstream; Indicates a branch active power, yes The child node, that is, from Flow direction The active power; Indicates a branch The reactive power flowing upstream; Indicates a branch. The reactive power, that is, from Flow direction reactive power; For This is the set of downstream child nodes of the parent node; Represents a node The sum of the active power of all downstream branches. Represents a node The sum of reactive power of all downstream branches;

[0055] Following the above recursive method, the voltage of each node in the entire network can be obtained. And form the basic voltage distribution Meanwhile, to ensure that the calculation results conform to physical constraints and provide an accurate reference for subsequent compensation layers, the mechanism base layer embeds a "topology verifier" and a "power balance verifier" before and after performing DistFlow recursion, respectively.

[0056] As a specific example, the topology validator takes switch / feeder status as input (from SCADA or distribution network automation) and uses GIS topology as a reference to generate a valid map. The verification content includes:

[0057] Connectivity: Starting from node 0, perform BFS / DFS. If there are unvisited nodes, it is determined that there is an island / disconnected branch. The voltage of such nodes does not participate in the mechanism deduction in this round, or the reliable voltage of the previous moment is used as the boundary condition.

[0058] Radial (acyclic): For undirected graphs Perform a disjoint-set data structure or Depth-First Search (DFS) to find cycles; the equivalence criterion is "connected and...". " Indicates the number of sides. Indicates the number of nodes; if a loop is found, select a branch as the "edge to be disconnected" based on the switch operation record or branch confidence (e.g., communication quality, state consistency) to restore the tree structure.

[0059] Parent node uniqueness and direction consistency: After passing the acyclicity check, a directed tree is generated with node 0 as the root. Ensure that every non-root node Only one parent node This avoids power aggregation errors caused by dual power supply or incorrect pointing.

[0060] Topology check output is a directed tree The sequence of subsequent / preorder traversals is directly used as input for voltage distribution calculation (branch power summation and voltage forward propagation).

[0061] As a specific example, the power balance verifier calculates the branch power flow. With node voltage This is executed later to verify whether node power conservation and overall network power balance are satisfied. A typical approach is to calculate the power imbalance residual.

[0062] ① Nodal power conservation residual: for any non-root node ,definition

[0063] (4)

[0064] (5)

[0065] in This represents the branch loss term. Losses can be ignored in the main linearization process; during the verification phase, an approximate estimate can be used.

[0066] (6)

[0067] (7)

[0068] ② Reference node power balance: For root node 0, the injected power obtained by computer simulation:

[0069] (8)

[0070] (9)

[0071] And measure the power on the transformer side. By comparison, we can obtain

[0072] (10)

[0073] (11)

[0074] ③ Threshold discrimination and correction: If or Exceeding the preset threshold (Can be set according to area capacity / measurement accuracy), then it is determined that there is a deviation in the input power of this round (load prediction error, distributed power output loss, measurement loss, etc.). At this time, the mechanism base layer can perform rapid balance correction: for example, the unbalance of the root node. By node weight (Determined by historical load percentage or measurement reliability) Distributed to unmeasured nodes and updated using the following formula:

[0075] (12)

[0076] (13)

[0077] The DistFlow recursion is then re-executed; simultaneously, the "balanced residual" is output as a data quality indicator to the hybrid inference engine for subsequent adaptive fusion weight adjustment.

[0078] Through the above dual verification of "prior topological consistency + posterior power conservation residual", it is ensured that the basic voltage distribution generated by the mechanism base layer is physically interpretable and can be used for subsequent data compensation and fusion deduction.

[0079] As a concrete example, the data-driven compensation layer in step 3 uses GCN (Graphical Neural Network) and LSTM (Long Short-Term Memory Network) to learn the spatiotemporal characteristics of voltage residuals by combining real-time data from SCADA (Supervisory Control and Data Acquisition), PMU (Phasor Measurement Unit), and photovoltaic data. Specifically, GCN is mainly used to capture the spatial influence of grid topology on voltage coupling, while LSTM excels at capturing the temporal dynamic characteristics brought about by load changes and photovoltaic fluctuations. Through the residual learning mode, the network directly learns the deviation between the measured voltage value and the mechanistic calculation result to compensate for the nonlinear dynamic part that the mechanistic base layer cannot capture.

[0080] In the data-driven compensation layer, the Physical Information Embedding (PINN) constraint embeds the residuals of the DistFlow equation as a regularization term into the loss function of the neural network. This forces the network to not only conform to historical data during training and inference but also to follow the physical laws of the power grid. This approach avoids the "physical failure" problem that may be caused by purely data-driven models and improves the robustness of the model under extreme conditions.

[0081] The dynamic compensation quantity is generated by using the outputs of a graph neural network and an LSTM model, combined with physical constraints, to dynamically correct errors in the mechanistic voltage distribution, thereby improving the accuracy of voltage prediction. The formula can be expressed as:

[0082] (14)

[0083] in, It is the basic voltage distribution output from the mechanistic base layer. It is a dynamic compensation amount output by the data-driven compensation layer. This is the corrected voltage distribution.

[0084] As a specific example, the hybrid inference engine described in step 4 is used to adaptively fuse the base voltage distribution output by the mechanistic foundation layer with the dynamic compensation amount output by the data-driven compensation layer to obtain the final real-time voltage prediction result. Its core is:

[0085] Define and update the "fusion weight" in real time to reflect the assessment of data compensation reliability at the current moment: increase the data compensation ratio when PMU availability is high, photovoltaic fluctuations are large, and data quality is good; decrease the data compensation ratio and revert to the mechanism result when communication is interrupted, measurements are missing, or anomalies occur. The "fusion weight" refers to the weighting coefficient used in the hybrid simulation to control the contribution ratio of mechanism results and data compensation, denoted as... .

[0086] Assume the mechanism base layer at time The output node voltage (or voltage squared) vector is The dynamic compensation amount (voltage residual compensation) output by the data-driven compensation layer is: The basic fusion of the hybrid deduction is as follows:

[0087] (15)

[0088] in From the mechanism calculation in step 2, The "dynamic compensation amount" is derived from the output of the data-driven compensation layer in step 3.

[0089] Fusion weights Calculated by the "Real-time Confidence Assessment Module" based on data quality and operational fluctuation characteristics (e.g., PMU data availability, communication status, photovoltaic power fluctuation rate, etc.). A feasible engineering implementation (Sigmoid mapping + truncation):

[0090] PMU availability: (Number of valid frames in the window / Number of frames expected);

[0091] Data missing / outlier rate: (Percentage of missing or abnormal points);

[0092] Communication status: (1 = normal, 0 = interrupted);

[0093] Photovoltaic power fluctuation: (For example );

[0094] Construct confidence score:

[0095] (16)

[0096] The mapping yields the fusion weights:

[0097] (17)

[0098] (18)

[0099] in, Indicate Limit to the specified range Inside;

[0100] Intuitive meaning: high, Large (and highly volatile) data compensation is needed. Increase; Missing / Abnormal High or communication interruption but This reduces the voltage dependence mechanism, thereby ensuring stability. Large (and highly volatile) data compensation is needed. Increase; Missing / Abnormal High or communication interruption but This reduces the voltage dependence mechanism, thereby ensuring stability.

[0101] Hybrid inference engine based on weights Will and The final voltage prediction is obtained by fusion. .

[0102] As a specific example, the rolling deduction and incremental learning module described in step 5 is used to expand the "current voltage situation estimate" formed in steps 1 to 4 into a "future multi-timescale voltage situation map," and triggers online fine-tuning when the error exceeds the limit, enabling the system to continuously adapt to changes in operating characteristics caused by topology changes, equipment switching, photovoltaic sudden changes, etc. Its logical relationship is as follows: the system uses each rolling time... Using multi-source measurement / prediction data as input, the foundation voltage distribution is first calculated by the mechanistic base layer based on topology and line parameters. Then, the data-driven compensation layer outputs the voltage residual compensation amount on the same topology. The hybrid inference engine then fuses the data according to adaptive weights to obtain the predicted voltage at the current moment. The rolling deduction and incremental learning module uses this... Starting with the load / photovoltaic forecast input in the short term, the system continuously outputs the voltage status spectrum for the next 15 minutes. At the same time, when the actual measurement arrives at the subsequent time, the system calculates the inference error and decides whether to start incremental learning based on this, thus forming a closed-loop optimization.

[0103] The rolling simulation uses a sliding window mechanism, each Update the input once and output the future The prediction results correspond to the number of prediction steps as follows:

[0104] (19)

[0105] The "input data" specifically includes: SCADA measurements (voltage / power / switch status, etc.), PMU high-frequency synchronous phasor measurements (voltage / current phasors, frequency, etc.), meteorological time-series data (irradiance, etc.), and photovoltaic output and load forecasts obtained from meteorological / historical data. These data undergo time synchronization alignment and anomaly detection and repair before being entered into the window, serving as the "node injected power and topology state input" for the mechanism-based foundation layer and the "spatiotemporal feature sequence input" for the data-driven compensation layer, respectively. The latest window data is updated with each rolling step. The output of the rolling simulation is a sequence of future multi-step node voltages. This forms the voltage distribution pattern for the next 15 minutes.

[0106] "Dependency error" refers to the deviation between the predicted output voltage and the subsequently arrived actual measured voltage, within the set of nodes with available measurements. For the above calculations, the mean absolute error can be used as the online criterion:

[0107] (20)

[0108] When the error exceeds the threshold and continues Incremental learning is triggered when a scrolling window is reached. It only performs small-step online fine-tuning of the "data-driven compensation layer model," without retraining the mechanistic base layer, to ensure the feasibility of running in 5-second online scenarios. Here, "data-driven model" and "data-driven compensation layer" have the same meaning; both refer to the internal components of the compensation layer used to output dynamic compensation values. The model (including the GCN-LSTM path for residual learning and the PINN path for physical embedding) primarily uses neural network weights and biases (e.g., weight matrices for each layer of GCN, LSTM gating parameters, PINN network parameters, etc., with particular emphasis on fine-tuning the LSTM weights, which are sensitive to temporal dynamics). During online updates, the system extracts the latest samples from a sliding window to form a mini-batch. For the parameters of the compensation model Perform several steps of gradient descent:

[0109] (twenty one)

[0110] The loss function follows the training objective of the compensation layer and includes the form of "residual fitting term + physical constraint term":

[0111] (twenty two)

[0112] Through the above-mentioned rolling simulation and error-triggered incremental learning, the system can continuously correct the compensation model under operating conditions such as topology switching, photovoltaic fluctuations, and partial measurement loss, so that the voltage situation spectrum for the next 15 minutes remains stable, interpretable, and with controllable accuracy.

[0113] As a specific example, step 6 forms a closed loop with the aforementioned steps 2 to 5: step 2 outputs the mechanism base voltage, step 3 outputs the compensation amount, step 4 merges to obtain the deduction result, and step 5 outputs the result in a rolling manner and calculates the error; when the error or residual shows a continuous deviation, the bidirectional feedback loop transmits the deduction result back to the mechanism base layer and the data-driven compensation layer respectively, realizing online correction and continuous optimization.

[0114] The implementation method for feeding back the inference results to the mechanism base layer is as follows: With a set of measurement nodes... Calculate the residuals If the residuals exhibit systematic deviations across multiple consecutive windows (such as same feeder / phase sign offset), then the mechanism-side inputs (load forecasting, photovoltaic equivalent injection, etc.) are considered to have deviations, and these deviations are weighted accordingly. The residuals are amortized back to the injected power, and the inputs for the next window mechanism calculation are updated, for example...

[0115] (twenty three)

[0116] This improves the accuracy of the starting point of the next round of mechanism voltage.

[0117] The implementation method for feeding the inference results back to the data-driven compensation layer is as follows: The compensation strategy is adjusted online based on data quality and residual characteristics (increasing the proportion of residual learning when data is complete, and increasing the proportion of physical constraints when data is missing / abnormal), and small-step incremental learning is triggered when the error continuously exceeds the threshold, only fine-tuning the compensation model parameters; simultaneously, the strength of physical constraints can be adaptively adjusted so that the compensation output closely approximates the measurement while satisfying basic physical laws.

[0118] (twenty four)

[0119] Through the aforementioned mechanism-side "input correction" and data-side "weight / constraint adaptation + fine-tuning", the system continuously reduces errors and improves robustness during rolling simulations.

[0120] This invention also provides an online low-voltage simulation system for distribution substations based on collaborative-driven modeling. This system implements the aforementioned online low-voltage simulation method for distribution substations based on collaborative-driven modeling. The system includes a mechanism foundation layer, a data-driven compensation layer, a hybrid simulation engine, a rolling simulation and incremental learning module, and a bidirectional feedback loop.

[0121] The mechanistic base layer, based on topology and line parameters, uses linearized DistFlow for derivation, and incorporates topology verification and power balance verification to output the base voltage distribution. ;

[0122] The data-driven compensation layer takes the output of the mechanistic foundation layer and multi-source spatiotemporal characteristics as inputs, and obtains voltage residual compensation through two parallel paths: Mode A uses GCN-LSTM to learn the spatial coupling and temporal dynamic characteristics of the voltage residual, while Mode B uses PINN to embed power flow / physical constraints into the network training to ensure the physical consistency of the compensation results; the two are then adaptively synthesized through confidence evaluation to obtain the dynamic compensation amount. ;

[0123] The hybrid inference engine adaptively calculates weights based on real-time data quality and volatility levels. ,Will and The final voltage prediction is obtained by fusion. ;

[0124] The rolling simulation and incremental learning module updates the sliding window with a fixed step size, continuously outputs the voltage situation spectrum within a set time period in the future, and triggers online adjustment of the data-driven compensation layer parameters when the error exceeds the threshold.

[0125] The bidirectional feedback loop is used to transmit the inference results back. On the one hand, it is used to correct the input deviation of the mechanism base layer, and on the other hand, it is used to adjust the fusion weight and physical constraint strength of the data-driven compensation layer.

[0126] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0127] Example

[0128] This embodiment selects the IEEE 33-node distribution network as the benchmark system, configuring distributed photovoltaic (PV) access points and some energy storage devices within it to simulate the operation scenario of a transformer substation with a high proportion of distributed power access and significant power fluctuations. Several switchable branch switches are also set up to simulate dynamic changes in operating mode and topology. PV output uses a typical summer solar radiation curve superimposed with random disturbances to recreate rapid power fluctuations caused by cloud cover.

[0129] like Figure 2 As shown, the method in this embodiment adopts a four-layer closed-loop structure: "mechanism base layer—data-driven compensation layer—hybrid inference engine—rolling update and feedback." First, the base voltage distribution at the current moment is calculated based on the linearized DistFlow equation. Then, multi-source measurement data from SCADA / PMU / meteorological sources are input, and the spatial coupling and temporal evolution characteristics between nodes are extracted through a GCN-LSTM network to output the voltage residual compensation. Next, the hybrid inference engine adaptively weights the mechanism results and compensation results according to the real-time data quality to obtain the final voltage situation prediction. Finally, continuous rolling inference is achieved through a 5-second sliding window mechanism, and incremental learning is triggered when the error exceeds the limit to update the data-driven part online. This process can improve adaptability to rapidly fluctuating scenarios while ensuring physical consistency.

[0130] like Figure 3 As shown, the expected error of voltage prediction was compared between the pure mechanistic model, the pure data model, and the collaborative driving model of this invention. The results show that during periods of drastic changes in photovoltaic output, the error of the pure mechanistic model increases significantly, the pure data model exhibits bias at individual nodes, while the overall error level of the collaborative driving model remains in a low range. This verifies that the "mechanism + compensation" structure can effectively mitigate the mismatch problem of a single model.

[0131] like Figure 4 As shown, during the training phase, all loss values ​​exhibited a stable decrease and eventually converged, indicating that after introducing PINN constraints, the data-driven compensation layer can maintain conformity to the power flow equations while approximating actual measurements, providing a stable initial model for subsequent online simulations. In actual deployment, when the rolling simulation detects that the prediction error exceeds a preset threshold for multiple consecutive windows, incremental learning is triggered. Only the parameters of the data compensation part are updated in small steps, without retraining the mechanism layer, thus ensuring operability and real-time performance in online scenarios.

[0132] In the above scenarios, the method of the present invention can reduce the mean absolute error from about 0.58% in the pure mechanistic model to about 0.28%, and control the extrapolation cycle within 80 ms in 95% of the time, which can meet the time requirements of 5s rolling prediction and active voltage control at the distribution area level. At the same time, under simulated topology switching, photovoltaic sudden drop, and partial measurement missing conditions, the extrapolation results remain smooth and interpretable, proving that the method has good engineering applicability under complex distribution network operating conditions.

Claims

1. A method for online low-voltage simulation of distribution substations based on collaborative driving modeling, characterized in that, This method adopts a collaborative closed-loop architecture driven by mechanism, data, fusion, rolling, and feedback. It includes a mechanism base layer, a data-driven compensation layer, a hybrid inference engine, a rolling inference and incremental learning module, and a bidirectional feedback loop. The specific inference process is as follows: Step 1: Integrate two types of inputs: one is multi-source operational and prediction data; the other is network structure and parameter information. Step 2: In the mechanistic foundation layer, based on the topology and line parameters, linearized DistFlow is used for deduction, and topology verification and power balance verification are embedded to output the base voltage distribution. ; Step 3: The data-driven compensation layer takes the output of the mechanistic foundation layer and the multi-source spatiotemporal characteristics as inputs, and obtains voltage residual compensation through two parallel paths: Mode A uses GCN-LSTM to learn the spatial coupling and temporal dynamic characteristics of the voltage residual, while Mode B uses PINN to embed power flow / physical constraints into the network training to ensure the physical consistency of the compensation results; the two are then adaptively synthesized through confidence evaluation to obtain the dynamic compensation amount. ; Step 4: The hybrid inference engine adaptively calculates weights based on real-time data quality and volatility levels. ,Will and The final voltage prediction is obtained by fusion. ; The hybrid inference engine described in step 4 is used to adaptively fuse the base voltage distribution output by the mechanism base layer with the dynamic compensation amount output by the data-driven compensation layer to obtain the final real-time voltage prediction result, as follows: Define and update the fusion weights in real time to reflect the assessment of data compensation reliability at the current moment: when PMU availability is high, photovoltaic fluctuations are large, and data quality is good, increase the data compensation ratio; when communication is interrupted, measurements are missing, or anomalies occur, decrease the data compensation ratio and revert to the mechanism results; the fusion weight refers to the weighting coefficient used in the hybrid simulation to control the contribution ratio of mechanism results and data compensation, denoted as ; Assume the mechanism base layer at time The output node voltage vector is The dynamic compensation amount output by the data-driven compensation layer is The basic fusion of hybrid deduction is as follows: （15） Fusion weights Calculated based on data quality and operational fluctuation characteristics: PMU availability This indicates the number of valid frames / the number of frames that should be received within the window; Data missing / outlier rate This indicates the percentage of missing or abnormal points; Communication status A value of 1 indicates normal operation, while a value of 0 indicates interruption. Photovoltaic power fluctuation ; Construct confidence score: （16） The mapping yields the fusion weights: （17） （18） in, Indicate Limit to the specified range Inside; Hybrid inference engine based on weights Will and The final voltage prediction is obtained by fusion. ; Step 5: The rolling simulation and incremental learning module updates the sliding window with a fixed step size, continuously outputs the voltage situation spectrum within a set time period in the future, and triggers online adjustment of the data-driven compensation layer parameters when the error exceeds the threshold. Step 6: Two-way feedback loop to transmit the deduction results back: on the one hand, it is used to correct the input deviation of the mechanism base layer, and on the other hand, it is used to adjust the fusion weight and physical constraint strength of the data-driven compensation layer.

2. The online low-voltage simulation method for distribution substations based on collaborative driving modeling as described in claim 1, characterized in that, The multi-source operation and prediction data mentioned in step 1 includes SCADA measurements, PMU synchronization phasors, and meteorological / photovoltaic / load predictions; the network structure and parameter information includes topology and switch status, and line parameters.

3. The online low-voltage simulation method for distribution substations based on collaborative driving modeling as described in claim 1, characterized in that, Step 2 is described in detail below: In the mechanistic foundation layer, based on the distribution network topology, node parameters, and distributed generation information, the voltage distribution is calculated using the linearized DistFlow equation: Using the low-voltage busbar / feeder outlet of the transformer in the distribution area as the reference node, denoted as node 0, and the voltage amplitude as the nominal value V0, a ​​radial directed tree is constructed by combining the line set E and the node set N of the low-voltage distribution area in the switch-closed state. For any node To the node branch road The circuit parameters are resistance. Reactance ; Define nodes Net active power Net reactive power Let active power and reactive power be absorbed from the network, respectively. When the online loss term is ignored or treated as a second-order term, the branch power and node power satisfy the recursive relationship of linearized DistFlow: Branch power recursion from bottom to top: （1） （2） Voltage recursion is calculated from top to bottom: （3） in, Indicates a branch Active power flowing upstream; Indicates a branch active power, yes The child node, that is, from Flow direction The active power; Indicates a branch The reactive power flowing upstream; Indicates a branch. The reactive power, that is, from Flow direction reactive power; For This is the set of downstream child nodes of the parent node; Represents a node The sum of the active power of all downstream branches. Represents a node The sum of reactive power of all downstream branches; Following the above recursive method, the voltage of each node in the entire network can be obtained. And form the basic voltage distribution Meanwhile, the mechanistic base layer embeds a topology verifier and a power balance verifier before and after performing DistFlow recursion, respectively.

4. The online low-voltage simulation method for distribution substations based on collaborative driving modeling as described in claim 3, characterized in that, The topology validator takes switch / feeder status as input and GIS topology as a reference to generate a valid map. The verification content includes: Connectivity: Starting from node 0, perform BFS / DFS. If there are unvisited nodes, it is determined that there are isolated / disconnected branches. The voltage of such nodes does not participate in the mechanism deduction in this round, or the reliable voltage of the previous moment is used as the boundary condition. Radiation, or acyclicity, for undirected graphs Perform a disjoint-set data structure or a depth-first search (DFS) to find cycles; the equivalence criterion is "connected and...". " Indicates the number of sides. Indicates the number of nodes; if a loop is found, select a branch as the "edge to be disconnected" based on the switch operation record or branch confidence to restore the tree structure; Parent node uniqueness and direction consistency: After passing the acyclicity check, a directed tree is generated with node 0 as the root. Ensure that every non-root node Only one parent node To avoid power aggregation errors caused by dual power supply / incorrect orientation; Topology check output is a directed tree The sequence of traversals, along with their subsequent / preceding traversals, are directly used as input for voltage distribution calculation.

5. The online low-voltage simulation method for distribution substations based on collaborative driving modeling according to claim 4, characterized in that, The power balance verifier calculates the branch power flow. With node voltage The subsequent execution is used to verify whether node power conservation and overall network power balance are satisfied: ① Nodal power conservation residual: for any non-root node ,definition （4） （5） in This refers to branch loss. In the main linearization process, losses are ignored, and approximate estimates are used during the verification phase: （6） （7） ② Reference node power balance: For root node 0, the injected power obtained by computer simulation: （8） （9） And measure the power on the transformer side. By comparison, we can obtain （10） （11） ③ Threshold discrimination and correction: If or Exceeding the preset threshold If the input power in this round is found to be deviated, the mechanism base layer will perform a fast balance correction: the imbalance at the root node will be adjusted. By node weight When allocated to unmeasured nodes, update using the following formula: （12） （13） The DistFlow recursion is then re-executed; the balanced residuals are output as a data quality metric to the hybrid inference engine for subsequent adaptive fusion weight adjustment.

6. The online low-voltage simulation method for distribution substations based on collaborative driving modeling as described in claim 5, characterized in that, In step 3, the data-driven compensation layer uses GCN and LSTM, combined with real-time data from SCADA, PMU, and photovoltaics, to learn the spatiotemporal characteristics of voltage residuals. GCN is used to capture the spatial influence of grid topology on voltage coupling, while LSTM captures the temporal dynamic characteristics brought about by load changes and photovoltaic fluctuations. Through the residual learning mode, the network directly learns the deviation between the measured voltage value and the mechanism calculation result to compensate for the nonlinear dynamic part. In the data-driven compensation layer, the PINN constraint embeds the residuals of the DistFlow equation as regularization terms into the loss function of the neural network, so that the network outputs results that conform to historical data and follow the physical laws of the power grid during training and inference. The dynamic compensation quantity is generated by the output of the graph neural network and LSTM model, combined with physical constraints, to dynamically correct the error in the mechanism voltage distribution. The formula is expressed as: （14） in, It is the basic voltage distribution output from the mechanistic base layer. It is a dynamic compensation amount output by the data-driven compensation layer. This is the corrected voltage distribution.

7. The online low-voltage simulation method for distribution substations based on collaborative driving modeling as described in claim 6, characterized in that, Step 5 describes the rolling deduction and incremental learning module for voltage prediction. Starting from this point, and combining future load / photovoltaic forecasts, the system continuously outputs voltage status graphs for the next 15 minutes. Simultaneously, when the actual measurements arrive at subsequent times, the system calculates the extrapolation error and decides whether to initiate incremental learning, thus forming a closed-loop optimization. The rolling simulation uses a sliding window mechanism, each Update the input once and output the future The prediction results correspond to the number of prediction steps as follows: （19） The input data specifically includes: SCADA measurements, PMU high-frequency synchronous phasor measurements, meteorological time-series data, and photovoltaic power output and load forecasts obtained from meteorological / historical data; These data undergo time synchronization and alignment, as well as anomaly detection and repair, before being entered into the window. They serve as the "node injection power and topology state input" for the mechanism base layer and the "spatiotemporal feature sequence input" for the data-driven compensation layer, respectively, and are updated with the latest window data at each scroll step. The output of the rolling simulation is a sequence of future multi-step node voltages. This forms the voltage distribution pattern for the next 15 minutes; The extrapolation error refers to the deviation between the extrapolated output voltage and the subsequently arrived actual measured voltage, within the set of nodes with available measurements. The above calculation uses the mean absolute error as the online criterion: （20） When the error exceeds the threshold and continues Incremental learning is triggered when a scrolling window is reached, and the data-driven compensation layer is fine-tuned online with small steps, adjusting parameters including neural network weights and biases; during online updates, the system extracts the latest samples from the sliding window to form a mini-batch. For the parameters of the compensation model Perform several steps of gradient descent: （21） The loss function follows the training objective of the compensation layer and includes a residual fitting term plus a physical constraint term. （22）。 8. The online low-voltage simulation method for distribution substations based on collaborative driving modeling according to claim 7, characterized in that, The bidirectional feedback loop described in step 6 transmits the deduction results back to the mechanism base layer and the data-driven compensation layer respectively, realizing online correction and continuous optimization; The implementation method for feeding back the inference results to the mechanism base layer is as follows: With a set of measurement nodes... Calculate the residuals ; If the residuals exhibit systematic bias across multiple consecutive windows, then the input to the mechanism base layer is considered biased, and weights are applied accordingly. The residual is amortized back to the injected power, and the input to the next window mechanism base layer is updated: （23） The implementation method for feeding back the inference results to the data-driven compensation layer is as follows: The compensation strategy is adjusted online based on data quality and residual characteristics, and small-step incremental learning is triggered when the error continuously exceeds the threshold, fine-tuning only the parameters of the data-driven compensation layer; simultaneously, the physical constraint strength is adaptively adjusted. （24）。 9. An online low-voltage simulation system for distribution substations based on collaborative driving modeling, characterized in that, This system is used to implement the online low-voltage simulation method for distribution substations based on collaborative driving modeling as described in any one of claims 1 to 8. The system includes a mechanism foundation layer, a data-driven compensation layer, a hybrid simulation engine, a rolling simulation and incremental learning module, and a bidirectional feedback loop. The mechanistic base layer, based on topology and line parameters, uses linearized DistFlow for derivation, and incorporates topology verification and power balance verification to output the base voltage distribution. ; The data-driven compensation layer takes the output of the mechanistic foundation layer and multi-source spatiotemporal characteristics as inputs, and obtains voltage residual compensation through two parallel paths: Mode A uses GCN-LSTM to learn the spatial coupling and temporal dynamic characteristics of the voltage residual, while Mode B uses PINN to embed power flow / physical constraints into the network training to ensure the physical consistency of the compensation results; the two are then adaptively synthesized through confidence evaluation to obtain the dynamic compensation amount. ; The hybrid inference engine adaptively calculates weights based on real-time data quality and volatility levels. ,Will and The final voltage prediction is obtained by fusion. ; The rolling simulation and incremental learning module updates the sliding window with a fixed step size, continuously outputs the voltage situation spectrum within a set time period in the future, and triggers online adjustment of the data-driven compensation layer parameters when the error exceeds the threshold. The bidirectional feedback loop is used to transmit the inference results back. On the one hand, it is used to correct the input deviation of the mechanism base layer, and on the other hand, it is used to adjust the fusion weight and physical constraint strength of the data-driven compensation layer.

Images