A droop control coefficient setting method and system based on a physical information neural network

Abstract

This application relates to the field of microgrid technology and discloses a method and system for setting droop control coefficients based on a physical information neural network. The method first obtains the target steady-state output current to be set for each parallel converter in a multi-source parallel DC microgrid; then, it inputs the target steady-state output current into a pre-trained physical information neural network surrogate model. This surrogate model is an inverse mapping model that uses the converter's steady-state output current as input features and the droop control coefficient as the output target. Its training loss function includes a physical constraint loss term constructed based on Kirchhoff's voltage law for the DC microgrid topology; finally, it outputs the corresponding droop control coefficients through model forward propagation, thus completing the parameter setting. This application improves the accuracy and efficiency of droop control coefficient setting, ensuring the accuracy of current distribution and operational stability of the DC microgrid.

Description

Technical Field

[0001] This application relates to the field of DC microgrid technology, specifically to a method and system for setting droop control coefficients based on physical information neural networks. Background Technology

[0002] In the field of microgrid technology, the stable operation of multi-source parallel DC microgrids depends on the reasonable setting of the droop control coefficients of each parallel converter. As a core parameter for achieving accurate current distribution within the microgrid, the setting effect of the droop control coefficients significantly impacts the overall operating status of the DC microgrid. Currently, in the actual operation and management of multi-source parallel DC microgrids, it is difficult to quickly and accurately determine suitable droop control parameters, failing to effectively meet the system's requirement for accurate current distribution to each converter as needed. This adversely affects the stable and efficient operation of multi-source parallel DC microgrids. Summary of the Invention

[0003] To solve, or at least partially solve, the above-mentioned technical problems, this application provides a method and system for setting droop control coefficients based on physical information neural networks.

[0004] In a first aspect, this application provides a method for setting the droop control coefficient based on a physical information neural network, comprising the following steps: S1. Obtain the target steady-state output current to be set for each parallel converter in a multi-power parallel DC microgrid; S2. Input the target steady-state output current into a preset trained physical information neural network proxy model; wherein, the physical information neural network proxy model is an inverse mapping model with the converter steady-state output current as the input feature and the droop control coefficient as the output target, and the training loss function of the physical information neural network proxy model includes a physical constraint loss term constructed based on Kirchhoff's voltage law of DC microgrid topology. S3. Through the forward propagation of the physical information neural network proxy model, the corresponding droop control coefficients are output to complete the setting of the droop control parameters for each converter.

[0005] Optionally, the physical constraint loss term is a physical equation residual term constructed based on Kirchhoff's voltage law, combined with the matching relationship between the droop control coefficient of each branch of the DC microgrid, the line resistance, and the steady-state output current of the converter.

[0006] Optionally, the inverse mapping model utilizes the nonlinear fitting characteristics of neural networks to establish a mathematical mapping relationship between the steady-state output current of the converter and the droop control coefficient, and completes the calculation of the droop control coefficient through a single network forward propagation.

[0007] Optionally, the training loss function of the physical information neural network proxy model further includes a bus voltage deviation constraint loss term; The bus voltage deviation constraint loss term is based on the physical correspondence between the rated voltage of the DC microgrid bus, the actual operating voltage of the bus, the droop control coefficient of each branch converter, the steady-state output current of the converter, and the line resistance. It constructs a physical equation residual term, which together with the physical constraint loss term constitutes a composite physical loss, so that the output droop control coefficient simultaneously meets the target steady-state output current distribution requirements and the bus voltage deviation control requirements.

[0008] Optionally, the training loss function of the physical information neural network agent model further includes a converter capacity constraint loss term; The converter capacity constraint loss term is constructed based on the rated capacity parameters of each parallel converter to form a physical constraint residual term for the feasible domain boundary of the droop control coefficient, so that the output droop control coefficient falls within the safe operating range of the corresponding converter.

[0009] Optionally, the training loss function of the physical information neural network proxy model further includes a system stability damping constraint loss term; The system stability damping constraint loss term is based on the small-signal equivalent model of the DC microgrid. It constructs the physical correspondence between the droop control coefficient and the system damping characteristics, forming a damping constraint residual term. This ensures that the output droop control coefficient meets the target steady-state output current distribution requirements while also satisfying the preset stability damping margin requirements.

[0010] Optionally, the physical constraint loss term is a robust physical equation residual term constructed by combining the preset uncertainty range of the line resistance of each branch of the DC microgrid; The residual term of the robust physical equation covers the fluctuation range of the line resistance within the preset uncertainty range, so that the output droop control coefficient meets the distribution requirements of the target steady-state output current when the line resistance parameter fluctuates.

[0011] Optionally, S1 specifically includes: acquiring the set of target steady-state output currents to be set for each parallel converter corresponding to multiple different operating conditions in a multi-power parallel DC microgrid; S2 specifically includes: inputting the target steady-state output current set into the physical information neural network proxy model, and constructing corresponding physical equation residual terms for each group of target steady-state output currents in the target steady-state output current set; S3 specifically includes: through the forward propagation of the physical information neural network proxy model, batch outputting the droop control coefficients corresponding to each group of operating conditions, and completing the setting of droop control parameters for each converter under multiple operating conditions.

[0012] Optionally, the method further includes: when the updated target steady-state output current is obtained, calculating the deviation between the updated target steady-state output current and the historical target steady-state output current corresponding to the last time the droop control parameter setting was completed, and determining whether the deviation is within a preset small fluctuation range; If the deviation is within the preset small fluctuation range, based on the historical droop control coefficient of the previous output, combined with the droop control coefficient of each branch constructed by Kirchhoff's voltage law of the DC microgrid topology, the steady-state physical equation relationship between the line resistance and the target steady-state output current, the droop control coefficient corresponding to the updated target steady-state output current is obtained, and the droop control parameter setting of each converter is completed; if the deviation exceeds the preset small fluctuation range, S2 to S3 are executed.

[0013] Secondly, this application also provides a droop control coefficient setting system based on a physical information neural network, comprising: The acquisition module is used to acquire the target steady-state output current to be set for each parallel converter in a multi-power parallel DC microgrid; The input module is used to input the target steady-state output current into a preset trained physical information neural network proxy model; wherein, the physical information neural network proxy model is an inverse mapping model with the converter steady-state output current as the input feature and the droop control coefficient as the output target, and the training loss function of the physical information neural network proxy model includes a physical constraint loss term constructed based on Kirchhoff's voltage law of DC microgrid topology; The output module is used to output the corresponding droop control coefficients through the forward propagation of the physical information neural network proxy model, thereby completing the setting of the droop control parameters for each converter.

[0014] Compared with the prior art, this application has the following advantages: This application constructs an inverse mapping model with the steady-state output current of the converter as the input feature and the droop control coefficient as the output target, which replaces the complex analytical calculation and iterative solution process in the traditional droop control coefficient setting process. The droop control coefficient that adapts to the target current can be directly obtained through the forward propagation of the network, which effectively simplifies the calculation process of parameter setting and improves the response efficiency of droop control coefficient setting.

[0015] Meanwhile, this application introduces a physical constraint loss term based on Kirchhoff's voltage law for DC microgrid topology into the model's training loss function. This enables the droop control coefficients output by the model to conform to the physical operation of the DC microgrid, avoiding the problem that the output results of a purely data-driven model are not in line with physical reality. This effectively improves the accuracy of the droop control coefficient setting and ensures that the output current of each converter can match the preset target allocation requirements.

[0016] Building upon this foundation, by supplementing the loss function with corresponding physical constraint terms, the accuracy of target current allocation can be ensured while also considering practical operational requirements such as bus voltage deviation control, converter safe operating range limitations, and system stability damping margin requirements, thereby enhancing the engineering practicality of the set droop control coefficient. Furthermore, optimized designs adaptable to different application scenarios, such as line parameter fluctuations, multi-condition batch processing, and small-scale target current adjustments, further improve the environmental adaptability and scenario suitability of the parameter setting results, meeting the diverse operation and management needs of multi-source parallel DC microgrids. Attached Figure Description

[0017] Figure 1 A schematic flowchart of a droop control coefficient setting method based on a physical information neural network provided in this application embodiment; Figure 2 This application provides a schematic diagram of the overall training process of a physical information neural network proxy model. Figure 3 A single-bus DC microgrid architecture diagram provided for an embodiment of this application; Figure 4 A training loss diagram for a physical information neural network proxy model provided in this application embodiment Figure 5 A comparison chart of the predicted droop control coefficient and the actual droop control coefficient provided in this application embodiment; Figure 6 This application provides a waveform diagram of the DC output current of a converter under the droop control coefficient predicted by a physical information neural network surrogate model. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, specific embodiments of this application will be described in further detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely for explaining this application and not for limiting it. It should also be noted that, for ease of description, only the parts relevant to this application are shown in the drawings, not all of them. Before discussing exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe operations (or steps) as sequential processes, many of these operations can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but may also have additional steps not included in the drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] See Figure 1 This application provides a method for setting droop control coefficients based on a physical information neural network, including the following steps: S1. Obtain the target steady-state output current to be set for each parallel converter in a multi-power parallel DC microgrid; S2. Input the target steady-state output current into a pre-trained physical information neural network proxy model; wherein, the physical information neural network proxy model is an inverse mapping model with the converter steady-state output current as the input feature and the droop control coefficient as the output target, and the training loss function of the physical information neural network proxy model includes a physical constraint loss term constructed based on Kirchhoff's voltage law of DC microgrid topology. S3. Through the forward propagation of the physical information neural network surrogate model, the corresponding droop control coefficients are output to complete the droop control parameter setting of each converter.

[0021] Specifically, in a multi-power parallel DC microgrid, each distributed power source is connected to the DC bus through a parallel converter. Droop control is a common control method for the parallel converter to achieve autonomous distribution of output current. The accuracy of the droop control coefficient setting affects the power distribution effect and operational stability of the DC microgrid. This method focuses on the setting process of the droop control coefficient.

[0022] During implementation, the target steady-state output current of each parallel converter in the multi-power parallel DC microgrid is first obtained. The target steady-state output current is determined comprehensively based on the total load demand of the DC microgrid, the rated operation plan of each parallel converter, and the preset power allocation requirements, and is the target steady-state operating output current value that each parallel converter needs to achieve.

[0023] The acquired target steady-state output current is then input into a pre-trained physical information neural network surrogate model. This physical information neural network surrogate model is an inverse mapping model that uses the converter's steady-state output current as input and the droop control coefficient as the output target, establishing a mapping relationship from electrical operation target quantities to control parameters. The training loss function of the physical information neural network surrogate model includes a physical constraint loss term constructed based on Kirchhoff's voltage law for the DC microgrid topology. The corresponding core physical relationship is determined based on the steady-state operating characteristics of droop control and fundamental circuit laws. The steady-state equation for the droop control of a single parallel branch is: ; in, This is the DC bus voltage. The output reference voltage of the parallel converter This is the droop control coefficient. For the steady-state output current of the converter, This represents the line resistance on the output side of the parallel converter. Based on Kirchhoff's voltage law, the steady-state operation of each parallel branch must satisfy the voltage balance constraint of the same DC bus. This is used as the core to construct the physical constraint loss term, ensuring that the model output results conform to the physical operation law of the DC microgrid.

[0024] By using the forward propagation of the physical information neural network surrogate model, the corresponding droop control coefficients are output, thus completing the droop control parameter setting for each parallel converter. After the target steady-state output current is input into the surrogate model, the droop control coefficients matching the target steady-state output current are directly obtained through the forward calculation process of the surrogate model. These droop control coefficients are then sent down to the underlying controller of the corresponding parallel converter, enabling each parallel converter to complete the distribution of steady-state output current according to the preset target.

[0025] This implementation process simplifies the solution process for the droop control coefficient, reduces the computational complexity of parameter setting, improves the efficiency of parameter setting and the control accuracy of steady-state current distribution, and ensures the matching of parameter setting results with the actual operating characteristics of the DC microgrid.

[0026] In some implementations, the physical constraint loss term is a physical equation residual term constructed based on Kirchhoff's voltage law, combined with the droop control coefficient of each branch of the DC microgrid, the matching relationship between line resistance and the steady-state output current of the converter.

[0027] The overall training process of the physical information neural network agent model is as follows: Figure 2As shown, the model takes the steady-state output current of the converter as input and the droop control coefficient as output, and fits the nonlinear mapping relationship between the input and output through a fully connected neural network. During training, data loss and physical constraint loss are calculated simultaneously, and the two are combined into a total loss function. The weight parameters of the neural network are adjusted through backpropagation to iteratively optimize the model performance until the model converges. The physical constraint loss term is a core component in the training process of the physical information neural network surrogate model. It is used to embed the circuit operation law of the DC microgrid during the model iterative optimization process, constrain the fitting direction of the surrogate model, and ensure that the final output result of the surrogate model conforms to the physical operation logic in the actual engineering scenario. In the model training phase, the droop control parameters of each parallel converter in the multi-power parallel DC microgrid under multiple sets of different operating conditions, as well as the steady-state output current of the converter corresponding to each set of droop control parameters, are collected to form a paired sample dataset for model training, providing a data foundation for the surrogate model to fit the mapping relationship between electrical quantities and control parameters.

[0028] Based on Kirchhoff's voltage law, the steady-state output voltages of all parallel converters connected to the same DC bus must remain consistent and equal to the DC bus voltage. Considering the steady-state operating characteristics of droop control, the steady-state voltage balance relationship of each parallel branch can be expressed through branch parameters and operating electrical quantities. The steady-state voltage equation for a single parallel branch under droop control is shown in the example above. Based on the constraint that the voltages of all branches on the same bus are equal, the following balance relationship is satisfied between any two parallel branches: ; in, , These are the droop control coefficients for the converters corresponding to the two parallel branches. , These are the line resistances of the two branches, , These are the steady-state output currents of the converters corresponding to the two branches, respectively. This formula clarifies the matching relationship between the droop control coefficient, line resistance, and steady-state output current of each branch in a DC microgrid, and is the core physical law for achieving current distribution through droop control in a DC microgrid.

[0029] The physical constraint loss term is the residual term of the physical equations constructed based on the above matching relationship. In specific implementation, for a DC microgrid topology with multiple branches in parallel, the residual physical equations of the voltage balance relationship between all branches are constructed, and the cumulative calculation result of the physical equation residuals is used as the value of the physical constraint loss term. During the model iterative training process, the physical constraint loss term and the data fitting loss term jointly participate in the optimization and updating of model parameters, enabling the model to learn the mapping relationship of sample data while following the core physical laws of the DC microgrid, and avoiding invalid results that do not conform to the circuit operation logic.

[0030] The physical constraint loss term constructed in this way can accurately anchor the core mechanism of droop control current distribution, so that the droop control coefficients output by the trained surrogate model can accurately match the distribution requirements of the target steady-state output current, improve the accuracy and engineering practicality of parameter setting results, and ensure that the steady-state operation effect of the DC microgrid meets expectations after the parameters are issued.

[0031] In some implementations, the inverse mapping model utilizes the nonlinear fitting characteristics of neural networks to establish a mathematical mapping relationship between the steady-state output current of the converter and the droop control coefficient, and completes the calculation of the droop control coefficient through a single network forward propagation.

[0032] Specifically, the inverse mapping model is the core architecture of the physical information neural network surrogate model, built to address the limitations of the traditional droop control coefficient solution process. The traditional solution for droop control coefficients requires first pre-setting initial values, then solving the differential algebraic equations of the DC microgrid to obtain the steady-state output current of the converter under the corresponding operating condition, and finally comparing the deviation between the calculated current value and the target value, iteratively correcting the droop control coefficient. This process is cumbersome, computationally complex, and makes it difficult to quickly obtain accurate parameter results matching the target. The inverse mapping model directly transforms the logical direction of parameter solution, using the converter's steady-state output current as the input feature of the surrogate model and the droop control coefficient as the output target, directly establishing a mapping path from the target electrical quantity to the control parameter to be determined.

[0033] The inverse mapping model leverages the nonlinear fitting characteristics of neural networks to fit the complex nonlinear correspondence between the steady-state output current of the converter and the droop control coefficient in a multi-source parallel DC microgrid. During the surrogate model training process, a data fitting loss term is constructed based on paired samples of droop control coefficients under various operating conditions and their corresponding steady-state output currents of the converter. This constrains the surrogate model's learning accuracy of the mapping relationship in the sample data. Simultaneously, a physical constraint loss term based on Kirchhoff's voltage law constrains the fitting direction of the surrogate model, ensuring that the learned mapping relationship conforms to the circuit operation rules of the DC microgrid. Through multiple rounds of iterative training, the weights and bias parameters within the surrogate model are continuously optimized, ultimately forming a stable and accurate mathematical mapping relationship. This enables the surrogate model to output droop control coefficients that conform to the physical operating rules based on the input steady-state output current of the converter.

[0034] The trained inverse mapping model can calculate the droop control coefficients through a single network forward propagation. During implementation, after the target steady-state output current is input into the model, the data will sequentially perform linear and nonlinear transformations of each network layer according to the network structure trained by the model. No additional iterative solutions or parameter correction steps are required. The droop control coefficients matching the target steady-state output current can be directly output through only one forward calculation process.

[0035] This solution method significantly simplifies the process of calculating the droop control coefficient, reduces the complexity of parameter calculation, shortens the time required for parameter setting, and ensures the accuracy of parameter results. It can adapt to the rapid parameter tuning requirements when adjusting the operating conditions of DC microgrids, and improves the convenience and practicality of engineering applications.

[0036] In some implementations, the training loss function of the physical information neural network surrogate model also includes a bus voltage deviation constraint loss term; The bus voltage deviation constraint loss term is based on the physical correspondence between the rated voltage of the DC microgrid bus, the actual operating voltage of the bus, the droop control coefficient of each branch converter, the steady-state output current of the converter, and the line resistance. The residual term of the physical equation is constructed together with the physical constraint loss term to form a composite physical loss, so that the output droop control coefficient can simultaneously meet the target steady-state output current distribution requirements and the bus voltage deviation control requirements.

[0037] Specifically, the value of the droop control coefficient affects both the distribution accuracy of the converter's steady-state output current and the sag of the DC bus voltage. In the traditional parameter tuning process, it is often necessary to repeatedly weigh and adjust between current distribution accuracy and bus voltage deviation control. The parameter tuning process is cumbersome and it is difficult to meet the control requirements of both operating indicators at the same time. By adding a bus voltage deviation constraint loss term to the training loss function of the physical information neural network surrogate model, the two control requirements can be met simultaneously.

[0038] Based on the steady-state operation characteristics of droop control in DC microgrids, there is a clear physical correspondence between the actual operating voltage of the DC bus and the operating and control parameters of each parallel branch. The steady-state voltage balance equation for a single parallel branch is as described in the above embodiment. The DC bus voltage deviation is the difference between the rated voltage and the actual operating voltage of the DC bus, which is determined by the droop control coefficient of each branch, the steady-state output current of the converter, and the line resistance. It is one of the core control indicators for the steady-state operation of the DC microgrid.

[0039] The bus voltage deviation constraint loss term is constructed based on the aforementioned physical correspondence. Using the rated DC bus voltage as a benchmark, a physical equation residual term for the bus voltage deviation is constructed. In specific implementation, based on the converter steady-state output current input to the model and combined with the iterative values ​​of the droop control coefficient during the model solving process, the corresponding predicted value of the actual operating voltage of the DC bus is calculated. Then, based on the difference between the rated DC bus voltage and the predicted actual operating voltage, a voltage deviation residual is constructed. The calculation result of the voltage deviation residual is used as the core value of the bus voltage deviation constraint loss term.

[0040] The bus voltage deviation constraint loss term and the physical constraint loss term based on Kirchhoff's voltage law together constitute the composite physical loss, which jointly participates in the training and iteration process of the physical information neural network surrogate model. During the surrogate model training phase, the composite physical loss and the data fitting loss term jointly participate in the optimization and updating of the model's internal parameters, enabling the surrogate model to simultaneously constrain the deviation amplitude of the DC bus voltage while learning the mapping relationship and satisfying the physical constraints of current distribution. In the forward propagation solution process after the surrogate model has completed training, the trained surrogate model has embedded the constraint logic of the bus voltage deviation, and the output droop control coefficient can simultaneously match the distribution requirements of the target steady-state output current and the control requirements of the DC bus voltage deviation.

[0041] This implementation method takes into account both the core operational requirements of current distribution accuracy and bus voltage stability. It avoids the cumbersome process of verifying bus voltage deviation and repeatedly correcting parameters after parameter setting, which is a traditional method. It also avoids the problem of reduced current distribution accuracy caused by post-processing parameter correction. It simplifies the setting process of droop control coefficient, improves the comprehensive applicability of parameter setting results, and ensures that the DC microgrid maintains stable operation of DC bus voltage while meeting power distribution requirements.

[0042] In some implementations, the training loss function of the physical information neural network agent model also includes a converter capacity constraint loss term; The converter capacity constraint loss term is constructed based on the rated capacity parameters of each parallel converter to form a physical constraint residual term for the feasible region boundary of the droop control coefficient, so that the output droop control coefficient falls within the safe operating range of the corresponding converter.

[0043] In a multi-power parallel DC microgrid, the rated capacities of each parallel converter differ. The droop control coefficient is limited by the rated capacity of the converter. A droop control coefficient exceeding the safe range can lead to overload and output exceeding limits during converter operation, affecting equipment safety and system stability. Traditionally, droop control coefficient tuning requires first solving the parameters and then manually verifying whether the parameters match the safe operating range of the converter's rated capacity. If they exceed the range, iterative solving is necessary. This cumbersome tuning process is prone to issues such as decreased current distribution accuracy after parameter correction.

[0044] The rated operating parameters of the converter determine the feasible domain boundary of the droop control coefficient. Considering the steady-state operating characteristics of droop control, when the converter outputs its rated current, the voltage drop caused by droop control must be within a preset allowable range. Based on the converter's rated capacity and rated output voltage, the safe operating range of the droop control coefficient can be clearly defined. The corresponding core parameter relationships are as follows: ; ; in, The rated capacity of the converter, This is the rated output voltage of the converter. This is the rated output current of the converter. This is the droop control coefficient. and These are the lower and upper limits of the feasible region for the droop control coefficient, determined based on the converter's rated capacity and allowable voltage drop range. This feasible region represents the hard constraint boundary for the safe and stable operation of the converter.

[0045] The converter capacity constraint loss term is constructed based on the rated capacity parameters of each parallel converter, with the core being the physical constraint residual term of the feasible region boundary of the droop control coefficient. In specific implementation, for each parallel converter, the upper and lower limits of the feasible region for the corresponding droop control coefficient are first determined based on its rated capacity parameters, rated output voltage, and preset allowable voltage drop range. During the iterative model training process, the iterative values ​​of the droop control coefficient output by the model are checked to see if they fall within the feasible region of the corresponding converter. If the iterative value of the droop control coefficient is within the feasible region, the corresponding constraint residual is zero. If the iterative value of the droop control coefficient exceeds the upper or lower limit of the feasible region, the corresponding residual term is constructed based on the difference exceeding the boundary. Finally, the cumulative calculation result of the residuals of all parallel converters is used as the value of the converter capacity constraint loss term.

[0046] The converter capacity constraint loss term is incorporated into the training loss function of the physical information neural network surrogate model. Together with the physical constraint loss term based on Kirchhoff's voltage law and the data fitting loss term, it participates in the internal parameter optimization and update during model training. During the iterative training phase, this loss term guides the fitting direction of the surrogate model, enabling it to learn the mapping relationship between electrical quantities and control parameters, satisfy the physical laws of current distribution, and simultaneously conform to the safe operating boundary constraints of the converter equipment. In the forward propagation solution after model training, the trained surrogate model has embedded the safety constraint logic of the converter capacity, and the output droop control coefficients will naturally fall within the safe operating range of the corresponding converter, requiring no additional post-processing limiting operations.

[0047] This implementation avoids the tedious process of repeatedly verifying and correcting parameters during traditional parameter tuning, and also avoids the problems of decreased current distribution accuracy and destruction of physical consistency caused by post-processing limiting operations. It ensures the matching accuracy of the droop control coefficient to the target steady-state output current, and also ensures that the parameter setting results meet the safe operation requirements of the converter equipment, thereby improving the efficiency of parameter tuning and the safety and reliability of DC microgrid operation.

[0048] In some implementations, the training loss function of the physical information neural network surrogate model also includes a system stability damping constraint loss term; The system stability damping constraint loss term is based on the small-signal equivalent model of the DC microgrid. It constructs a physical correspondence between the droop control coefficient and the system damping characteristics, forming a damping constraint residual term. This ensures that the output droop control coefficient meets the target steady-state output current distribution requirements while also satisfying the preset stability damping margin requirements.

[0049] In the operation of a multi-source parallel DC microgrid, the value of the droop control coefficient not only determines the accuracy of the steady-state output current distribution of the converter but also affects the small-signal stability performance of the system. Traditional droop control coefficient tuning processes mostly prioritize meeting the needs of steady-state current distribution and bus voltage control. After parameter solving, the system's stability performance is verified through small-signal modeling. If the verification reveals insufficient system damping or a risk of oscillation and instability, the droop control coefficient needs to be readjusted, and steady-state calculations and stability verifications need to be performed again. This parameter tuning process is cumbersome and prone to problems where steady-state control accuracy and system dynamic stability performance cannot be simultaneously achieved.

[0050] There is a clear physical correspondence between the droop control coefficient and the system damping characteristics, which can be quantitatively expressed through a small-signal equivalent model of a DC microgrid. For parallel converters employing droop control, the droop control equations and the converter main circuit equations are first subjected to small-signal perturbation processing at the steady-state operating point to obtain a small-signal linearized model of a single converter. The small-signal equation for the droop control element of a single converter is as follows: ; in, This is a small-signal disturbance to the converter output reference voltage. This is the droop control coefficient. This represents the small-signal disturbance of the steady-state output current of the converter. By integrating the small-signal models, line impedance models, and DC bus load models of each parallel converter with the topology and line parameters of a multi-source parallel DC microgrid, a small-signal state-space model of the entire parallel system can be established, expressed as: ; in, The first derivative of the system state variable vector. Let the system be the state variable vector. Let be the system's state matrix. The droop control coefficients are part of the state matrix. The core input parameters determine the values ​​of the elements in the state matrix, and the system stability and dynamic damping characteristics of this small-signal state-space model are determined by the state matrix. The distribution of eigenvalues ​​determines the damping ratio corresponding to the eigenvalues. This ratio is the core indicator for measuring the system's stability damping margin. The magnitude of the damping ratio reflects the system's ability to suppress oscillations and determines whether the system can smoothly recover to its steady-state operating point after being disturbed.

[0051] The system stability damping constraint loss term is constructed based on the small-signal equivalent model of the aforementioned DC microgrid. Its core is the damping constraint residual term formed by the physical correspondence between the droop control coefficient and the system damping characteristics. In specific implementation, the preset stability damping margin requirement for the DC microgrid system is first determined. This requirement is the minimum damping ratio threshold required to ensure stable system operation. During the iterative training of the physical information neural network surrogate model, a small-signal state-space model of the system under the corresponding operating condition is constructed for the droop control coefficient output by the current iteration of the model. The eigenvalues ​​of the state matrix are solved, and the damping ratio corresponding to the dominant oscillation mode of the system is extracted. The calculated damping ratio is compared with the preset stability damping margin threshold to construct the corresponding constraint residual. If the calculated damping ratio meets or exceeds the preset stability damping margin requirement, the corresponding constraint residual is zero. If the calculated damping ratio is lower than the preset stability damping margin requirement, the corresponding residual term is constructed based on the difference in damping ratios. Finally, the calculated result of the residual is used as the value of the system stability damping constraint loss term.

[0052] The system stability damping constraint loss term is incorporated into the training loss function of the physical information neural network surrogate model. Together with the physical constraint loss term based on Kirchhoff's voltage law and the data fitting loss term, it participates in the internal parameter optimization and update during model training. During the iterative training phase, this loss term guides the surrogate model's fitting optimization direction, enabling it to simultaneously learn the mapping relationship between the converter's steady-state output current and the droop control coefficient, satisfy the steady-state operating physical constraints, and simultaneously meet the system's stability damping margin requirements. In the forward propagation solution after model training, the trained surrogate model has embedded the system stability damping constraint logic. The output droop control coefficients, while meeting the target steady-state output current distribution requirements, ensure that the system meets the preset stability damping margin requirements, eliminating the need for additional small-signal stability checks and parameter corrections.

[0053] This implementation method balances the system's steady-state control accuracy and dynamic stability performance, avoiding the cumbersome process of iteratively balancing steady-state and dynamic performance during traditional parameter tuning. It significantly simplifies the tuning of the droop control coefficient and effectively avoids the risk of insufficient system damping and oscillation instability caused by improper selection of the droop control coefficient, thereby improving the stability and reliability of the DC microgrid under all operating conditions.

[0054] In some implementations, the physical constraint loss term is a robust physical equation residual term constructed by combining a preset uncertainty range of the line resistance of each branch of the DC microgrid; The residual terms of the robust physical equation cover the fluctuation range of the line resistance within the preset uncertainty range, so that the output droop control coefficient meets the distribution requirements of the target steady-state output current when the line resistance parameter fluctuates.

[0055] In a multi-power parallel DC microgrid, the line resistance on the output side of the parallel converter is a key parameter affecting droop control current distribution. In actual engineering scenarios, the actual value of the line resistance is affected by factors such as measurement errors, changes in ambient temperature, long-term aging of the line, and fluctuations in contact resistance, resulting in deviations from the nominal design value. These deviations also vary with operating conditions.

[0056] For the line resistance of each parallel branch, based on the nominal design parameters of the line, the allowable error range of engineering measurements, the resistance fluctuation coefficient corresponding to the operating environment temperature, and the maximum deviation range of long-term aging of the line, the uncertainty range of the line resistance of each branch is determined in advance, and the corresponding parameters are expressed as follows: ; in, This is the actual operating value of the line resistance. This is the nominal design value of the line resistance. The maximum permissible fluctuation range of the line resistance is determined by comprehensively considering the aforementioned engineering influencing factors. and These represent the lower and upper limits of the uncertainty range for line resistance, which covers the entire fluctuation range of line resistance during normal operation throughout the entire life cycle of the project.

[0057] Based on Kirchhoff's voltage law and the steady-state operating characteristics of droop control, the steady-state output voltage of each parallel branch connected to the same DC bus must remain consistent. The voltage balance relationship between any two parallel branches is as follows: ; in, , These are the droop control coefficients for the converters corresponding to the two parallel branches. , These are the line resistances of the two branches, , These represent the steady-state output currents of the converters corresponding to the two branches. Traditional physical constraint loss terms calculate residuals based solely on the nominal value of the line resistance, failing to cover the impact of parameter fluctuations. In contrast, the robust physical equation residual terms are constructed by combining the preset uncertainty range of the line resistance, ensuring that the residuals of the aforementioned voltage balance relationship are effectively constrained for any line resistance value within the range.

[0058] In practical implementation, for a DC microgrid topology with multiple branches connected in parallel, a preset uncertainty range is first determined for the line resistance of each branch. When constructing the physical constraint loss term, for the voltage balance physical equation between each branch, multiple sampling points covering the boundary of the range and typical fluctuation scenarios are selected within the uncertainty range of the corresponding line resistance. The physical equation residuals for each sampling point with the corresponding line resistance value are calculated. Based on the residual calculation results of all sampling points, the maximum residual value is selected, or all residuals are weighted and accumulated to obtain the final robust physical equation residual term, which serves as the core value of the physical constraint loss term. This construction method allows the physical constraint to cover the entire fluctuation range of the line resistance within the preset uncertainty range, rather than only constraining the nominal parameter points.

[0059] The physical constraint loss term, composed of the residual terms of the robust physical equations, is added to the training loss function of the physical information neural network surrogate model. Together with the data fitting loss term, it participates in the internal weight and bias parameter optimization and update during model training. During the iterative training phase, this robust physical constraint guides the fitting direction of the surrogate model, ensuring that the mapping relationship between the converter's steady-state output current and the droop control coefficient learned by the surrogate model not only satisfies the physical operating rules under the nominal line resistance value but also adapts to arbitrary fluctuations in line resistance within the uncertainty range, preventing the mapping relationship from failing due to changes in line parameters. During the forward propagation solution after model training, the trained surrogate model has embedded robust constraint logic for line resistance fluctuations. Even when the line resistance parameters fluctuate within a preset range, the output droop control coefficients still ensure that the actual output current of each parallel converter matches the target steady-state output current allocation requirements, eliminating the need for additional online identification of line parameters and readjustment of the droop control coefficients.

[0060] This implementation solves the problem of decreased current distribution accuracy caused by the uncertainty of line resistance parameters. It does not require additional online detection and control links, and does not increase the complexity of the system. At the same time, it significantly improves the robustness and environmental adaptability of the droop control coefficient setting results, ensuring that the power distribution accuracy of the DC microgrid always meets the preset requirements throughout its entire life cycle, thereby improving the practicality of engineering applications and the stability of system operation.

[0061] In some implementations... S1 specifically includes: acquiring the set of target steady-state output currents to be set for each parallel converter corresponding to multiple different operating conditions in a multi-power parallel DC microgrid; S2 specifically includes: inputting the target steady-state output current set into the physical information neural network surrogate model, and constructing corresponding physical equation residual terms for each set of target steady-state output currents in the target steady-state output current set; S3 specifically includes: using the forward propagation of the physical information neural network surrogate model, batch outputting the droop control coefficients corresponding to each set of operating conditions, and completing the setting of droop control parameters for each converter under multiple operating conditions.

[0062] In actual operation, multi-power parallel DC microgrids will face various operating conditions such as light load, rated load, heavy load, distributed power source switching, and load type changes. Different operating conditions correspond to different power distribution requirements, and it is necessary to set matching droop control coefficients for each operating condition. If the solution can only be completed for a single operating condition and needs to be adapted to multiple operating conditions, the complete process of parameter acquisition, model input, and coefficient solution needs to be repeatedly executed. The operation process is cumbersome and time-consuming, which is difficult to meet the actual needs of multi-operating condition parameter pre-tuning in engineering.

[0063] In this embodiment, during implementation, the first step is to obtain the target steady-state output current sets for each parallel converter under multiple different operating conditions in a multi-power parallel DC microgrid. These multiple operating conditions are categorized based on the actual operating scenarios of the DC microgrid, including typical operating scenarios with different load power levels, different numbers of distributed power sources, and different power allocation ratios. For each defined operating condition, the target steady-state output current for each parallel converter under that condition is determined by considering the total load demand, converter operation plan, and power allocation strategy. The target steady-state output currents for all operating conditions are then integrated to form a complete set of target steady-state output currents. Each set of data in the set corresponds to an independent operating condition, and the data sets are independent of each other and do not interfere with each other.

[0064] The target steady-state output current set is input into the trained physical information neural network surrogate model. During the model's processing of the input data, the physical constraint loss term constructs corresponding physical equation residual terms for each set of target steady-state output currents within the target steady-state output current set. Each set of physical equation residual terms is independently constructed based on the target steady-state output current under the corresponding operating condition, combined with Kirchhoff's voltage law for the DC microgrid topology. The residual calculation processes for different sets are independent of each other.

[0065] Then, through forward propagation of the physical information neural network surrogate model, the droop control coefficients corresponding to each set of operating conditions are output in batches. After receiving the target steady-state output current set, the surrogate model uses its internally trained network structure to simultaneously perform forward calculations on multiple sets of input data in the set, without processing them group by group. It completes the mapping and solution of all sets of data at once, outputting droop control coefficients that correspond one-to-one with each set of operating conditions. The batch-output droop control coefficients are then stored in the parameter configuration module of each operating condition, thus completing the setting of droop control parameters for each converter under multiple operating conditions.

[0066] This implementation method can solve the droop control coefficient for multiple operating conditions in one go, eliminating the need to repeatedly execute the parameter setting process for a single operating condition, shortening the overall time for multi-condition parameter tuning, and improving the overall efficiency of parameter setting. Simultaneously, the parameters corresponding to each set of operating conditions undergo independent physical constraint verification, ensuring that the droop control coefficients under all operating conditions conform to the physical operating laws of the DC microgrid and match the target steady-state output current distribution requirements of the corresponding operating conditions. This provides stable and accurate parameter support for the DC microgrid to switch between different operating conditions, improving the convenience and practicality of engineering applications.

[0067] In some implementations, the method further includes: When the updated target steady-state output current is obtained, the deviation between the updated target steady-state output current and the historical target steady-state output current corresponding to the last time the droop control parameters were set is calculated, and it is determined whether the deviation is within the preset small fluctuation range. If the deviation is within the preset small fluctuation range, based on the historical droop control coefficient of the previous output, combined with the droop control coefficient of each branch constructed by Kirchhoff's voltage law of the DC microgrid topology, and the steady-state physical equation relationship between line resistance and target steady-state output current, the updated droop control coefficient corresponding to the target steady-state output current is obtained, and the droop control parameter setting of each converter is completed; if the deviation exceeds the preset small fluctuation range, S2 to S3 are executed.

[0068] Specifically, in the operation of multi-source parallel DC microgrids, scenarios such as small load fluctuations and fine adjustments to power distribution ratios often occur. In these scenarios, the change in the target steady-state output current of the converter is relatively small. If the complete model input and forward propagation solution process is executed every time, unnecessary computational overhead will be incurred, and the response time of parameter tuning will be prolonged. For these small fluctuation operation scenarios, an incremental correction step can be added to the original parameter setting process to improve the response efficiency of parameter tuning while ensuring parameter accuracy.

[0069] During implementation, when the operating conditions of the DC microgrid are adjusted and an updated target steady-state output current is obtained, the historical target steady-state output current corresponding to the last time the droop control parameters were set is retrieved. The updated target steady-state output current is then compared with the historical target steady-state output current, and the deviation between the two sets of data is calculated. The deviation value is used to measure the magnitude of change in the target steady-state output current and can directly reflect the degree of fluctuation in the current operating conditions.

[0070] After calculating the deviation value, it is compared with a preset small fluctuation range. The preset small fluctuation range is determined based on the actual engineering conditions such as the normal load fluctuation range and power distribution fine-tuning amplitude of the DC microgrid, and is used to distinguish between small fluctuations and large changes in the target steady-state output current. If the deviation value is within the preset small fluctuation range, it indicates that only a small adjustment has occurred in the current operating condition, and there is no need to execute the complete model solution process. An incremental correction method can be used to quickly obtain the appropriate droop control coefficient.

[0071] The incremental correction process is based on the steady-state physical equations constructed using Kirchhoff's voltage law for the DC microgrid topology. The droop control coefficients of each parallel branch, line resistance, and the steady-state output current of the converter satisfy a fixed steady-state voltage balance relationship, which will not be elaborated further. Since the target steady-state output current only fluctuates slightly, the physical correspondence between the parameters remains stable. Based on the historical droop control coefficients from the previous output, and combined with the updated target steady-state output current and line resistance parameters, numerical correction can be directly performed using the aforementioned steady-state physical equations to quickly obtain the droop control coefficients that match the updated target steady-state output current, thus completing the droop control parameter settings for each converter.

[0072] If the deviation value exceeds the preset small fluctuation range, it indicates that the current operating conditions have changed significantly and the original historical parameters and physical relationships cannot be adapted to the new operating conditions. In this case, the initial parameter setting process is continued, the updated target steady-state output current is input into the trained physical information neural network proxy model, and the corresponding droop control coefficient is output through the forward propagation calculation of the model to complete the parameter setting.

[0073] This implementation method employs differentiated processing approaches for different fluctuation amplitudes of the target steady-state output current. In scenarios with small fluctuations, it eliminates the need for the complete forward propagation calculation process of the model, significantly shortening the response time for parameter tuning and reducing the computational load on the control system. Simultaneously, it relies on inherent physical laws to complete corrections, ensuring the accuracy and physical consistency of parameter settings. In scenarios with large changes, it still uses the complete model solution process, ensuring the accuracy and adaptability of parameter results, balancing the efficiency and accuracy of parameter tuning, adapting to different amplitudes of operating condition fluctuations in DC microgrids, and improving the flexibility and practicality of the overall control process.

[0074] To more clearly demonstrate the actual implementation process and application effect of the above-mentioned droop control coefficient setting method based on physical information neural network, a single bus DC microgrid system with three machines in parallel is used as a specific application example. The actual parameter acquisition, model construction, training and engineering application are illustrated in this example. This example is only used to intuitively demonstrate the implementation process of the method and does not constitute a limitation on the application scenarios and applicable scale of this method.

[0075] like Figure 3 The diagram shown is a single-bus DC microgrid architecture provided in an embodiment of this application. The single-bus DC microgrid system in this example includes three independent AC power sources. , , Each AC power source is connected to the shared DC bus via its corresponding AC / DC converter. The converters supply power to the system load are, in sequence, converter #1, converter #2, and converter #3. Each converter's output terminal is equipped with a line resistor. , , Each converter employs a nonlinear droop control strategy, with corresponding droop control coefficients as follows: , , The output DC current of each converter is as follows: , , , , , For capacitors, the core objective of this example is to set the droop control coefficient of each converter using the above method, thereby achieving accurate distribution of the output DC current of each converter.

[0076] In the data acquisition and preprocessing stage, a parameter traversal scanning program was used to collect data from the three-machine parallel nonlinear droop simulation model, and droop control coefficients were set. , , The traversal interval is [0.2, 4]. Within this interval, 12 sampling points are uniformly selected for each droop control coefficient. A total of 12 sampling points are obtained through forward frequency sweep. 3 =1728 sets of original datasets containing droop control coefficients and corresponding steady-state output currents. Outlier cleaning was performed on the original datasets. First, invalid data rows with negative output currents were removed. Then, deep cleaning based on physical residuals was performed to remove outliers with an absolute value greater than 2 of the physical constraint formula residuals. Finally, 1700 valid data were obtained. Then, all variables after cleaning were scaled to the [0,1] interval to complete the data preprocessing.

[0077] In the proxy model construction phase, a physical information neural network proxy model adapted to the three-machine parallel system is built, and the input layer of the fully connected neural network is set as the steady-state output current vector of the microgrid system. , , The output layer is the corresponding droop control coefficient vector. , , The network contains two hidden layers. The first hidden layer has 64 neurons and uses the tanh activation function, while the second hidden layer has 32 neurons and uses the sigmoid activation function. It establishes an inverse mapping structure from the steady-state output current of the converter to the droop control coefficient.

[0078] In the physical information constraint and model training phase, a composite loss function is constructed, including data loss terms and physical loss terms. The expression for the total loss function is as follows: ,in These are the weighting coefficients for the data loss terms. For data loss items, The weighting coefficients for the physical loss term. This represents the physical loss term. Based on Kirchhoff's voltage law, the physical residual equation corresponding to the physical loss term is constructed as follows: , The physical operating principles of the microgrid are embedded as prior knowledge into the model training process. The model training process consists of 5000 iterations, employing a gradual weight adjustment strategy. The first 2000 iterations focus on data fitting, with the weight coefficients of the physical loss term adjusted accordingly. It grows linearly, after 2000 iterations Reaching the maximum value forces the network to adhere to physical constraints during training. For example... Figure 4 The image shows the training loss diagram of a PINN (Physical Information Neural Network) surrogate model provided in an embodiment of this application. After 5000 rounds of iterative training, the total loss of the surrogate model decreased to 0.0217, of which the data loss was 0.0154 and the physical loss was 0.0713. The overall goodness of fit R of the surrogate model was... 2 Up to 0.9928. For example... Figure 5 The figure shown is a comparison between the droop control coefficient predicted by the PINN surrogate model and the actual value provided in an embodiment of this application. It can be seen that the droop control coefficient predicted by the surrogate model is highly consistent with the actual value, thus completing the model training operation.

[0079] In the target control parameter solution and engineering verification stage, the target steady-state output current of the microgrid system in the actual project is input into the trained physical information neural network surrogate model. Through a single network forward propagation, the corresponding droop control coefficient can be directly output. This coefficient is then sent to the underlying controller of each converter to complete the setting of the droop control coefficient. For example... Figure 6 The figure shown is a waveform of the converter output DC current under the prediction of droop control coefficient using a PINN surrogate model according to an embodiment of this application. The different colored curves in the figure represent... , , And briefly described as , , This example sets up two actual test cases for verification. Test case 1 is a non-current sharing condition, where the target output current of the microgrid system is set to [15A, 5A, 5A]. The droop control coefficient of the model's direct predicted output is [0.7697, 2.2839, 2.2749]. After this coefficient is sent to the underlying controller, as shown... Figure 6 As shown in (a), the measured steady-state output current of the system is [14.91A, 5.066A, 5.086A], achieving accurate current distribution. and The corresponding curves almost overlap, therefore The curve is obscured; Test case 2 is also a non-uniform current condition, with the target output current of the microgrid system set as [15A, 6A, 3A], and the droop control coefficient of the model directly predicting the output as [1.2174, 3.0859, 6.1680]. After this coefficient is sent to the underlying controller, as shown... Figure 6 As shown in (b), the measured steady-state output current of the system is [15.05A, 5.981A, 3.000A], which also achieves accurate current distribution.

[0080] The above application example of a three-unit parallel DC microgrid system intuitively demonstrates the feasibility and effectiveness of the droop control coefficient setting method based on physical information neural networks in practical implementation. This method enables rapid and accurate setting of the droop control coefficient, achieving precise distribution of the converter output current. In practical engineering applications, this method can flexibly adjust the model's input / output dimensions and training parameters according to the actual number of parallel converters, system topology, power distribution requirements, and operating conditions of the multi-source parallel DC microgrid. It is adaptable to setting the droop control coefficient for multi-source parallel DC microgrid systems of different scales and operating requirements, exhibiting good engineering adaptability.

[0081] This application also provides a droop control coefficient setting system based on a physical information neural network, including: The acquisition module is used to acquire the target steady-state output current to be set for each parallel converter in a multi-power parallel DC microgrid; The input module is used to input the target steady-state output current into a pre-trained physical information neural network proxy model. The physical information neural network proxy model is an inverse mapping model with the converter steady-state output current as the input feature and the droop control coefficient as the output target. The training loss function of the physical information neural network proxy model includes a physical constraint loss term constructed based on Kirchhoff's voltage law of DC microgrid topology. The output module is used to output the corresponding droop control coefficients through the forward propagation of the physical information neural network proxy model, thereby completing the setting of the droop control parameters for each converter.

[0082] The system embodiments provided in this application have the same technical features as the method embodiments described above, and therefore can achieve the same technical effects, which will not be repeated here.

[0083] The above description is merely a preferred embodiment and the technical principles employed in this application. This application is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions that can be made by those skilled in the art will not depart from the scope of protection of this application. Therefore, although this application has been described in detail through the above embodiments, this application is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of this application.

Claims

1. A method for setting droop control coefficients based on a physical information neural network, characterized in that, Includes the following steps: S1. Obtain the target steady-state output current to be set for each parallel converter in a multi-power parallel DC microgrid; S2. Input the target steady-state output current into a preset trained physical information neural network proxy model; wherein, the physical information neural network proxy model is an inverse mapping model with the converter steady-state output current as the input feature and the droop control coefficient as the output target, and the training loss function of the physical information neural network proxy model includes a physical constraint loss term constructed based on Kirchhoff's voltage law of DC microgrid topology. S3. Through the forward propagation of the physical information neural network proxy model, the corresponding droop control coefficients are output to complete the setting of the droop control parameters for each converter.

2. The method according to claim 1, characterized in that, The physical constraint loss term is a physical equation residual term constructed based on Kirchhoff's voltage law, combined with the matching relationship between the droop control coefficient of each branch of the DC microgrid, the line resistance, and the steady-state output current of the converter.

3. The method of claim 1, wherein, The inverse mapping model utilizes the nonlinear fitting characteristics of neural networks to establish a mathematical mapping relationship between the steady-state output current of the converter and the droop control coefficient, and completes the calculation of the droop control coefficient through a single network forward propagation.

4. The method of claim 1, wherein, The training loss function of the physical information neural network proxy model also includes a bus voltage deviation constraint loss term; The bus voltage deviation constraint loss term is based on the physical correspondence between the rated voltage of the DC microgrid bus, the actual operating voltage of the bus, the droop control coefficient of each branch converter, the steady-state output current of the converter, and the line resistance. It constructs a physical equation residual term, which together with the physical constraint loss term constitutes a composite physical loss, so that the output droop control coefficient simultaneously meets the target steady-state output current distribution requirements and the bus voltage deviation control requirements.

5. The method of claim 1, wherein, The training loss function of the physical information neural network proxy model also includes a converter capacity constraint loss term; The converter capacity constraint loss term is constructed based on the rated capacity parameters of each parallel converter to form a physical constraint residual term for the feasible domain boundary of the droop control coefficient, so that the output droop control coefficient falls within the safe operating range of the corresponding converter.

6. The method of claim 1, wherein, The training loss function of the physical information neural network proxy model also includes a system stability damping constraint loss term; The system stability damping constraint loss term is based on the small-signal equivalent model of the DC microgrid. It constructs the physical correspondence between the droop control coefficient and the system damping characteristics, forming a damping constraint residual term. This ensures that the output droop control coefficient meets the target steady-state output current distribution requirements while also satisfying the preset stability damping margin requirements.

7. The method of claim 1, wherein, The physical constraint loss term is a robust physical equation residual term constructed by combining the preset uncertainty range of the line resistance of each branch of the DC microgrid. The residual term of the robust physical equation covers the fluctuation range of the line resistance within the preset uncertainty range, so that the output droop control coefficient meets the distribution requirements of the target steady-state output current when the line resistance parameter fluctuates.

8. The method of claim 1, wherein, S1 specifically includes: acquiring the set of target steady-state output currents to be set for each parallel converter under different operating conditions in a multi-power parallel DC microgrid; S2 specifically includes: inputting the target steady-state output current set into the physical information neural network proxy model, and constructing corresponding physical equation residual terms for each group of target steady-state output currents in the target steady-state output current set; S3 specifically includes: through the forward propagation of the physical information neural network proxy model, batch outputting the droop control coefficients corresponding to each group of operating conditions, and completing the setting of droop control parameters for each converter under multiple operating conditions.

9. The method according to claim 1, characterized in that, The method further includes: When the updated target steady-state output current is obtained, the deviation between the updated target steady-state output current and the historical target steady-state output current corresponding to the last time the droop control parameters were set is calculated, and it is determined whether the deviation is within the preset small fluctuation range. If the deviation is within the preset small fluctuation range, based on the historical droop control coefficient of the previous output, combined with the droop control coefficient of each branch constructed by Kirchhoff's voltage law of the DC microgrid topology, the steady-state physical equation relationship between the line resistance and the target steady-state output current, the droop control coefficient corresponding to the updated target steady-state output current is obtained, and the droop control parameter setting of each converter is completed; if the deviation exceeds the preset small fluctuation range, S2 to S3 are executed.

10. A droop control coefficient setting system based on a physical information neural network, characterized in that, include: The acquisition module is used to acquire the target steady-state output current to be set for each parallel converter in a multi-power parallel DC microgrid; The input module is used to input the target steady-state output current into a preset trained physical information neural network proxy model; wherein, the physical information neural network proxy model is an inverse mapping model with the converter steady-state output current as the input feature and the droop control coefficient as the output target, and the training loss function of the physical information neural network proxy model includes a physical constraint loss term constructed based on Kirchhoff's voltage law of DC microgrid topology; The output module is used to output the corresponding droop control coefficients through the forward propagation of the physical information neural network proxy model, thereby completing the setting of the droop control parameters for each converter.

Images