A method and system for determining nonlinear droop control coefficients of a direct current microgrid

Abstract

The application provides a direct-current micro-grid nonlinear droop control coefficient determination method and system, the method comprising: constructing an average simulation model of a multi-source parallel direct-current micro-grid adopting a nonlinear droop control strategy, and performing multi-dimensional grid scanning sampling to determine a training data set; eliminating negative current samples, samples greater than the rated current upper limit of the parallel converter, and non-convergent output samples in the training data set, and performing data cleaning processing based on Kirchhoff's voltage law; constructing and training an inverse mapping artificial neural network with the target steady-state output current as the input and the nonlinear droop control coefficient as the output; inputting the actual target steady-state output current into the trained network to determine the corresponding nonlinear droop control coefficient. The application improves the physical consistency of the training data through two-stage data cleaning, and completes the determination of the droop coefficient through single forward propagation and without relying on line impedance, effectively guaranteeing the precise current sharing and stable operation of the direct-current micro-grid.

Description

Technical Field

[0001] This application relates to the field of power system control technology, specifically to a method and system for determining the nonlinear droop control coefficient of a DC microgrid. Background Technology

[0002] In multi-source parallel DC microgrids, the architecture of heterogeneous distributed power sources such as photovoltaics, energy storage, fuel cells, and shore power being connected in parallel to the DC bus via converters is widely used in power supply scenarios such as new energy grid connection, ports, ships, and multi-electric aircraft. Droop control, with its characteristics of no communication, modularity, and decentralization, has become the mainstream power distribution scheme for this system.

[0003] However, droop control faces two major problems in engineering applications: First, the line impedance of each parallel branch varies due to factors such as wiring, conductor characteristics, environment, and aging, and is difficult to measure accurately. Impedance mismatch will cause the current distribution to deviate from the expected value. Second, conventional linear droop control has inherent contradictions. Under heavy load, a small droop coefficient is needed to constrain voltage drop, but this will amplify the distribution deviation. Under light load, a large droop coefficient is needed to ensure distribution accuracy, but this will exacerbate voltage drop under heavy load. It is difficult to balance distribution accuracy and bus voltage stability.

[0004] Nonlinear droop control (including id-Vdc coupled type, piecewise type, exponential type, adaptive type, etc.) extends the droop relationship from linear to nonlinear, which can better balance the light load distribution accuracy and the bus voltage stability under heavy load. It can also eliminate the outer loop of bus voltage in conventional dual-loop control and shorten the dynamic response time. It has become an important research and application direction for droop control of multi-source parallel DC microgrids.

[0005] However, in nonlinear droop control, the control coefficient has a much greater impact on current distribution and bus voltage than in linear droop control. The coefficient of linear droop is approximately linearly related to the output current, while the coefficient of nonlinear droop affects both the characteristic curve shape and the steady-state operating point; even small changes can cause nonlinear fluctuations in current distribution. This makes the accuracy of coefficient tuning directly determine the control effect and places higher demands on parameter settings.

[0006] Existing methods for setting droop control coefficients mainly include the following categories: (1) Analytical compensation methods, which analyze and correct the droop coefficient based on the direct measurement or online identification of the impedance of each branch line. For example, in the patent "CN114156932A; An adaptive design method for droop coefficient of DC distribution network based on oscillation frequency", the system oscillation frequency is identified online, the relationship between the droop coefficient and the oscillation frequency is analyzed and derived, the coefficient is corrected in real time, and the oscillation caused by the time-varying impedance is suppressed. However, relying on accurate impedance information makes it difficult to adapt to engineering scenarios where impedance changes with temperature, aging, and contact state; (2) Heuristic optimization methods, such as particle swarm optimization (PSO) and NSGA-II meta-heuristic algorithms, perform multi-objective optimization of droop coefficients, but their computational overhead is large, and the fitness function and weight coefficients need to be carefully set. When the operating conditions change or the system topology changes, they need to be readjusted; (3) Forward proxy artificial neural network method, first trains a forward proxy model with droop coefficient as input and steady-state output current as output offline, then constructs a fitness function for the design index and performs exhaustive search or gradient search on the candidate coefficient grid. It still retains the complete post-training optimization search link, which increases the computational load of multi-converter systems rapidly and is accompanied by the burden of manual adjustment of weight coefficients.

[0007] Existing methods generally suffer from the following shortcomings when facing the high-precision setting requirements of nonlinear droop control coefficients: ① They rely to varying degrees on prior information about line impedance and are sensitive to impedance identification errors and time-varying characteristics; ② Existing data generation or sampling methods lack verification of circuit physical constraints, leading to invalid / erroneous samples contaminating the training set and thus affecting the accuracy of neural network mapping; ③ After training, post-processing steps such as optimization search, fitness function construction, or parameter limiting are still required to obtain the final droop coefficients, resulting in lengthy deployment links; ④ Under nonlinear droop control, the impact of the droop coefficient on system performance is significantly amplified, and the accuracy of existing methods is often insufficient to meet the accuracy requirements of nonlinear droop control; ⑤ It is difficult to simultaneously consider the flexibility of operating condition switching and the rapid response requirements of online tuning.

[0008] In summary, the key to the engineering application of nonlinear droop control lies in the setting of high-precision coefficients. However, existing methods are limited by impedance dependence, insufficient network mapping accuracy, deployment complexity, and dynamic adaptability. There is an urgent need for a new, efficient, and reliable online tuning method to facilitate its engineering implementation. Summary of the Invention

[0009] In view of one of the defects in the prior art, the purpose of this application is to provide a method and system for determining the nonlinear droop control coefficient of a DC microgrid.

[0010] A first aspect of this application provides a method for determining the nonlinear droop control coefficient of a DC microgrid, comprising: An averaged simulation model of a multi-source parallel DC microgrid with a preset nonlinear droop control strategy is constructed, and multi-dimensional grid scanning sampling is performed to determine the training dataset, which includes the target steady-state output current and the nonlinear droop control coefficient. The training dataset is cleaned by removing samples with negative current, samples with current greater than the rated current limit of the parallel converter, and samples with non-convergent output, and the data is cleaned based on Kirchhoff's voltage law to determine the effective training dataset. Construct an inverse mapping artificial neural network, using the target steady-state output current in the effective training dataset as input and the nonlinear droop control coefficient as the desired output, and train the inverse mapping artificial neural network. The actual target steady-state output current of the multi-source parallel DC microgrid is input into a trained inverse mapping artificial neural network to determine the nonlinear droop control coefficient corresponding to the actual target steady-state output current.

[0011] Optionally, the construction of an averaged simulation model of a multi-source parallel DC microgrid employing a preset nonlinear droop control strategy, and the performing of multi-dimensional grid scanning sampling to determine the training dataset, includes: An averaged simulation model of the multi-source parallel DC microgrid employing a preset nonlinear droop control strategy was built in the simulation platform. Based on the preset nonlinear droop control strategy, determine the range of values ​​for the preset nonlinear droop control coefficient. Within the preset range of nonlinear droop control coefficient values, each nonlinear droop control coefficient corresponding to each parallel converter is uniformly sampled to determine the combination of nonlinear droop control coefficients. Using each set of nonlinear droop control coefficients in the aforementioned combination, the averaged simulation model is run to a preset steady-state time to determine the target steady-state output current corresponding to each set of nonlinear droop control coefficients, thereby constructing the training dataset.

[0012] The data cleaning process based on Kirchhoff's voltage law, which determines the effective training dataset, includes: For each sample in the training dataset that excludes samples with negative current, samples with current greater than the rated current limit of the parallel converter, and samples with non-convergent output, calculate the branch voltage balance residual between any two branches. Obtain the measurement error term, parameter error term, and steady-state error term; The measurement error term, the parameter error term, and the steady-state error term are summed to determine a preset threshold. If the absolute value of the voltage balance residual of any branch is not less than the preset threshold, the corresponding sample is deleted, and the effective training dataset is determined.

[0013] Optionally, the inverse mapping artificial neural network includes an input layer, a hidden layer, and an output layer connected in sequence; The input layer includes N neurons, the number of which is equal to the number of parallel converters in the multi-source parallel DC microgrid, and the number of N neurons is an integer not less than 2. The N neurons of the input layer are used to receive the target steady-state output current of the N parallel converters respectively. The hidden layer includes a first hidden layer and a second hidden layer. The first hidden layer includes multiple neurons and uses a hyperbolic tangent activation function. The second hidden layer includes multiple neurons and uses a logarithmic sigmoid activation function. The output layer comprises N neurons, each of which is used to output the nonlinear droop control coefficients corresponding to the N parallel converters, and the output layer uses a purely linear activation function.

[0014] Optionally, training the inverse mapping artificial neural network by using the target steady-state output current from the effective training dataset as input and the nonlinear droop control coefficient as the desired output includes: The effective training dataset is divided into a training set, a test set, and a validation set; The target steady-state output current in the training set is input into the back-mapping artificial neural network, which outputs the nonlinear droop control coefficient. The Levenberg-Marquardt algorithm is used for backpropagation training, and the training is performed for a preset number of iterations to determine the trained back-mapping artificial neural network. During the model training process, the validation set is used to validate the back-mapping artificial neural network in each iteration. If the mean square error of the back-mapping artificial neural network is less than a preset mean square error threshold or the mean square error of the back-mapping artificial neural network increases for a consecutive preset number of iterations, an early stopping mechanism is activated to stop training.

[0015] Optionally, the method for determining the actual target steady-state output current of the multi-source parallel DC microgrid includes: Based on the DC power, convert the total load power into the total load current; Based on the rated power of the converters in the multi-source parallel DC microgrid, determine the load ratio undertaken by each converter; The actual target steady-state output current of each converter is determined based on the load ratio undertaken by each converter and the total load current.

[0016] Optionally, the step of inputting the actual target steady-state output current of the multi-source parallel DC microgrid into a trained inverse mapping artificial neural network to determine the nonlinear droop control coefficient corresponding to the actual target steady-state output current includes: Invoke the trained inverse mapping artificial neural network; The actual target steady-state output current of the multi-source parallel DC microgrid is input into the trained inverse mapping artificial neural network for a single forward propagation to determine the nonlinear droop control coefficient corresponding to the actual target steady-state output current of the multi-source parallel DC microgrid.

[0017] Optionally, the preset nonlinear droop control strategy includes any one of the following: id-Vdc coupled nonlinear droop control strategy, piecewise nonlinear droop control strategy, exponential nonlinear droop control strategy, and quadratic nonlinear droop control strategy.

[0018] Optionally, the maximum total system load current and the maximum DC bus voltage sag of the multi-source parallel DC microgrid are used as constraint boundaries, and the nonlinear droop control coefficient is checked in conjunction with the equivalent total resistance of the corresponding branch of each converter to determine the nonlinear droop control coefficient that meets the voltage sag constraint. The nonlinear droop control coefficients that meet the voltage drop constraint are sent to the underlying nonlinear droop control loop of the parallel converter for parameter setting.

[0019] A second aspect of this application provides a system for determining the nonlinear droop control coefficient of a DC microgrid, comprising: The training dataset construction module is used to construct an averaged simulation model of a multi-source parallel DC microgrid using a preset nonlinear droop control strategy, and to perform multi-dimensional grid scanning sampling to determine the training dataset, which includes the target steady-state output current and the nonlinear droop control coefficient. The data cleaning module is used to remove samples with negative current, samples with current greater than the rated current limit of the parallel converter, and samples with non-convergent output from the training dataset, and to perform data cleaning based on Kirchhoff's voltage law to determine the effective training dataset. The model training module is used to construct an inverse mapping artificial neural network, taking the target steady-state output current as input and the nonlinear droop control coefficient as the desired output, and training the inverse mapping artificial neural network. The nonlinear droop control coefficient determination module is used to input the actual target steady-state output current of the multi-source parallel DC microgrid into a trained inverse mapping artificial neural network to determine the nonlinear droop control coefficient corresponding to the actual target steady-state output current.

[0020] This application performs a first data cleaning by removing samples with negative current, samples exceeding the rated current limit of the parallel converter, and samples with non-convergent output from the training dataset. A second data cleaning is then performed based on Kirchhoff's voltage law. This ensures that the training samples for the inverse mapping artificial neural network simultaneously satisfy both device operating boundaries and DC branch voltage balance constraints, improving the mapping accuracy of the inverse mapping artificial neural network. By using an effective training dataset to train the inverse mapping artificial neural network, a direct inverse mapping is established, making the tuning process of the nonlinear droop control coefficient completely independent of line impedance measurement or identification, avoiding accuracy degradation in scenarios with unknown impedance, time-varying impedance, or aging. After training, the trained inverse mapping artificial neural network outputs the corresponding nonlinear droop control coefficient through a single forward propagation based on the actual target steady-state output current of the input multi-source parallel DC microgrid. This eliminates the need for exhaustive post-training or gradient optimization, significantly shortening the tuning path and achieving a rapid dynamic response where the steady-state current smoothly reaches the target value without circulating current oscillation.

[0021] Other technical effects resulting from the additional features will be further illustrated in the corresponding embodiments. Attached Figure Description

[0022] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating a method for determining the nonlinear droop control coefficient of a DC microgrid according to an exemplary embodiment.

[0023] Figure 2 This is a schematic diagram of a three-source parallel DC microgrid topology according to an exemplary embodiment.

[0024] Figure 3 This is a schematic diagram of a nonlinear droop control structure according to an exemplary embodiment.

[0025] Figure 4 This is a comparative schematic diagram of a forward mapping artificial neural network and a backward mapping artificial neural network according to an exemplary embodiment.

[0026] Figure 5 This is a training regression scatter plot of a back-mapping artificial neural network according to an exemplary embodiment.

[0027] Figure 6 This is a schematic diagram illustrating a sample-by-sample comparison of the predicted and true values ​​of a trained back-mapping artificial neural network on a test set, according to an exemplary embodiment.

[0028] Figure 7This is a schematic diagram of the DC current waveform of a converter output under typical operating conditions, according to an exemplary embodiment.

[0029] Figure 8 This is a schematic diagram comparing the training data with and without Kirchhoff's voltage law secondary data cleaning and the branch voltage balance constraints of the network output, according to an exemplary embodiment.

[0030] Figure 9 This is a schematic diagram illustrating the structure of a nonlinear droop control coefficient determination system for a DC microgrid according to an exemplary embodiment. Detailed Implementation

[0031] The present application will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application. These all fall within the protection scope of the present application.

[0032] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature.

[0033] In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0034] The terms "comprising" and "having," and any variations thereof, in the embodiments of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or devices.

[0035] Existing techniques for determining the nonlinear droop control coefficients of DC microgrids mainly include analytical compensation methods that rely on line impedance measurement or identification, heuristic optimization algorithms with high computational overhead and the need for parameter retuning, data generation or sampling methods that lack verification of circuit physical constraints, leading to invalid / erroneous samples contaminating the training set and affecting the accuracy of neural network mapping, and forward surrogate neural network methods that still require exhaustive search or gradient search after training. Their common drawbacks are sensitivity to prior information about line impedance, difficulty in adapting to time-varying impedance scenarios, and lengthy and complex tuning links. Their accuracy is insufficient to meet the high requirements of nonlinear droop control, failing to balance the flexibility of operating condition switching and the need for rapid online response, thus hindering the engineering implementation of nonlinear droop control. Based on the above problems, this application provides a method for determining the nonlinear droop control coefficients of DC microgrids to solve the aforementioned problems.

[0036] Figure 1 This is a flowchart illustrating a method for determining the nonlinear droop control coefficient of a DC microgrid according to an exemplary embodiment.

[0037] Reference Figure 1 As shown in one embodiment of this application, a method for determining the nonlinear droop control coefficient of a DC microgrid is provided, which can be implemented using S11 to S14.

[0038] S11. Construct an averaged simulation model of a multi-source parallel DC microgrid using a preset nonlinear droop control strategy, and perform multi-dimensional grid scanning sampling to determine the training dataset.

[0039] Specifically, the training dataset includes the target steady-state output current and the nonlinear droop control coefficients.

[0040] S12, Remove samples with negative current, samples with current greater than the rated current limit of the parallel converter, and samples with non-convergent output from the training dataset, and perform data cleaning based on Kirchhoff's voltage law to determine the effective training dataset.

[0041] S13, construct an inverse mapping artificial neural network, using the target steady-state output current in the effective training dataset as input and the nonlinear droop control coefficient as the desired output, and train the inverse mapping artificial neural network.

[0042] S14 inputs the actual target steady-state output current of the multi-source parallel DC microgrid into the trained inverse mapping artificial neural network to determine the nonlinear droop control coefficient corresponding to the actual target steady-state output current.

[0043] This application performs a first data cleaning by removing samples with negative current, samples exceeding the rated current limit of the parallel converter, and samples with non-convergent output from the training dataset. A second data cleaning is then performed based on Kirchhoff's voltage law. This ensures that the training samples for the inverse mapping artificial neural network simultaneously satisfy both device operating boundaries and DC branch voltage balance constraints, improving the mapping accuracy of the inverse mapping artificial neural network. By using an effective training dataset to train the inverse mapping artificial neural network, a direct inverse mapping is established, making the tuning process of the nonlinear droop control coefficient completely independent of line impedance measurement or identification, avoiding accuracy degradation in scenarios with unknown impedance, time-varying impedance, or aging. After training, the trained inverse mapping artificial neural network outputs the corresponding nonlinear droop control coefficient through a single forward propagation based on the actual target steady-state output current of the input multi-source parallel DC microgrid. This eliminates the need for exhaustive post-training or gradient optimization, significantly shortening the tuning path and achieving a rapid dynamic response where the steady-state current smoothly reaches the target value without circulating current oscillation.

[0044] In some specific embodiments of this application, the preset nonlinear droop control strategy may include, but is not limited to, any one of the following: id-Vdc coupled nonlinear droop control strategy, piecewise nonlinear droop control strategy, exponential nonlinear droop control strategy, and quadratic nonlinear droop control strategy. It can also be adapted to other droop control strategies with nonlinear droop characteristic curves.

[0045] The nonlinear droop control coefficient for DC microgrids provided in this application can be adapted to multi-source parallel DC microgrids of different parallel scales, and has strong engineering adaptability.

[0046] To obtain the training dataset, in some specific embodiments of this application, for S11, an averaged simulation model of a multi-source parallel DC microgrid using a preset nonlinear droop control strategy is constructed, and multi-dimensional grid scanning sampling is performed to determine the training dataset. This can be implemented using S111 to S114.

[0047] S111, an averaged simulation model of a multi-source parallel DC microgrid using a preset nonlinear droop control strategy is built in the simulation platform.

[0048] Specifically, the simulation platform can be MATLAB / Simulink.

[0049] S112, Based on the preset nonlinear droop control strategy, determine the range of values ​​for the preset nonlinear droop control coefficient.

[0050] For example, the preset range of nonlinear droop control coefficient can be [0.2, 4], and can be adaptively adjusted according to the specific nonlinear droop control strategy.

[0051] S113, uniformly sample each nonlinear droop control coefficient corresponding to each parallel converter within the preset range of nonlinear droop control coefficient values ​​to determine the combination of nonlinear droop control coefficients.

[0052] For example, multiple sampling points are uniformly sampled for the N droop control coefficients corresponding to the N parallel converters. For instance, 12 sampling points can be selected for each dimension to generate an N-dimensional combination space of nonlinear droop control coefficients.

[0053] Each parallel converter corresponds to a nonlinear droop control coefficient, forming a set of nonlinear droop control coefficients, which in turn form multiple sets of nonlinear droop control coefficients.

[0054] S114. Using each set of nonlinear droop control coefficients in the combination of nonlinear droop control coefficients, run the averaged simulation model to the preset steady-state time, determine the target steady-state output current corresponding to each set of nonlinear droop control coefficients, and construct a training dataset.

[0055] Specifically, for each set of nonlinear droop control coefficients, the averaged simulation model constructed in step S111 is run to the preset steady-state time, and the steady-state output current corresponding to each parallel converter is recorded as the target steady-state output current. Each set of nonlinear droop control coefficients is paired with the target steady-state output current and stored in the training dataset.

[0056] For example, the preset steady-state time can be t=0.5s after running the averaged simulation model, or it can be adaptively adjusted according to the specific needs of the scenario.

[0057] The embodiments described above in this application efficiently construct a standardized training dataset covering the entire parameter space by uniformly scanning the droop coefficients of each converter based on a preset value range and extracting the steady-state current in a simulation platform, thus providing reliable data support for the subsequent training of a high-precision inverse mapping model.

[0058] To filter unstable training samples, in some specific embodiments of this application, for S12, samples with negative current, samples with current greater than the rated current limit of the parallel converter, and samples with non-convergent output are removed from the training dataset, and data cleaning is performed based on Kirchhoff's voltage law to determine the effective training dataset. This can be done using S121 to S122.

[0059] S121, Remove samples with negative current, samples with current exceeding the rated current limit of the parallel converter, and samples with non-convergent output from the training dataset to determine the training dataset after the first data cleaning process.

[0060] Specifically, the sample includes the target steady-state output current and the nonlinear droop control coefficient corresponding to the target steady-state output current.

[0061] Non-convergent output samples represent invalid or abnormal data samples that failed to converge to the steady-state operating point.

[0062] S122, based on Kirchhoff's voltage law, a second data cleaning process is performed on the training dataset that has undergone the first data cleaning process to determine the effective training dataset.

[0063] Specifically, step S122 involves the core Kirchhoff voltage physical constraint for droop control of a DC microgrid: during steady-state operation, the output voltages of any two parallel branches connected to the same DC bus must be equal. Theoretically, the branch voltage balance residual derived from this should approach zero. If the residual of a sample data exceeds a preset threshold, it indicates an anomaly in the data (such as excessive sensor measurement noise, unstable transient processes, line parameter drift, equipment failure, etc.), and it should be filtered out. Ensuring the physical consistency of the training dataset from the source significantly improves the training efficiency and generalization ability of subsequent neural network models.

[0064] Step S122 can be taken as S1221 to S1224.

[0065] S1221, for each sample in the training dataset after removing samples with negative current, samples with current exceeding the rated current limit of the parallel converter, and samples with non-convergent output, calculate the branch voltage balance residual between any two branches: Calculation based on the formula for calculating inter-branch voltage balance residuals: For any sample in the training dataset after the first data cleaning process Calculate the voltage balance residual between any two branches m and n:

[0066] in, This represents the voltage balance residual between the m-th branch and the n-th branch in the i-th sample. This represents the nonlinear droop control coefficient corresponding to the m-th branch. This represents the line resistance of the m-th branch. Let represent the steady-state output current of the m-th branch in the i-th sample. This represents the nonlinear droop control coefficient corresponding to the nth branch. This represents the line resistance of the nth branch. This represents the steady-state output current of the nth branch in the i-th sample.

[0067] The above and The design parameters are set and recorded along with the training dataset when constructing an offline averaged simulation model of a multi-source parallel DC microgrid, and are used for offline physical consistency verification of the training dataset.

[0068] During the online inference phase, only the actual target steady-state output current is input into the trained back-mapping artificial neural network, without requiring online measurement or identification of the current line impedance of the operating system.

[0069] S1222, obtain the measurement error term, parameter error term, and steady-state error term.

[0070] Specifically, ; ;

[0071] in, This represents the measurement error term. Indicates the accuracy of the current sensor. This represents the rated current of the nth branch. Indicates the parameter error term. This indicates the error in the droop coefficient; Represents the steady-state error term. Represents the steady-state error coefficient. V bus,rated This indicates the rated voltage of the DC bus.

[0072] Measurement error term The accuracy is determined by the measurement accuracy of the current sensor and the voltage sensor. The accuracy of the current sensor is... ; Parameter error term The droop factor error is determined by the error in setting the droop factor and the error in measuring the line resistance. The error in setting the droop factor is... The line resistance error is ; steady-state error term The value is determined by the steady-state fluctuations of the system and the deviation of the sampling time, and is usually taken as 0.1% to 0.5% of the rated bus voltage.

[0073] S1223, sum the measurement error term, the parameter error term, and the steady-state error term to determine a preset threshold.

[0074] Specifically, preset threshold The value of directly determines the effectiveness of data cleaning and needs to be determined comprehensively based on the actual system parameters and measurement accuracy. The following quantitative calculation methods can be used:

[0075] in, Indicates the preset threshold. Indicates the measurement error term. Indicates the parameter error term. This represents the steady-state error term.

[0076] S1224, If the absolute value of any set of branch voltage balance residuals is not less than the preset threshold, delete the corresponding sample and determine the effective training dataset.

[0077] Specifically, for each sample Calculate the absolute value of the residuals for all pairwise branch combinations. If there exists any pairwise residual whose absolute value is not less than a preset threshold... If the sample is found to be abnormal, it will be removed from the training dataset D after the first processing. , making Then delete the sample. ;

[0078] For batch processing of multi-branch systems: For a DC microgrid system with N branches, a total of [number] calculations are required. The residuals between branches are considered invalid if any set of residuals exceeds the threshold.

[0079] For example, taking a three-source parallel DC microgrid system (N=3) as an example, the following three sets of residuals need to be calculated: 1. Branch Road 1 and Branch Road 2:

[0080] 2. Branch Road 1 and Branch Road 3:

[0081] 3. Branch Road 2 and Branch Road 3:

[0082] like or or If the i-th sample is deleted, then the i-th sample will be deleted.

[0083] The embodiments described above in this application effectively filter out invalid data that violates the basic laws of circuits and device constraints by eliminating physically infeasible or abnormal samples such as negative current, over-rated current, and non-convergent output, and by performing secondary verification in conjunction with Kirchhoff's voltage law. This ensures that the effective training dataset simultaneously meets the device operating boundary and DC branch voltage balance constraints, thereby significantly improving the physical consistency and reliability of the training dataset. Using the above-processed effective training dataset to train the inverse mapping artificial neural network ensures that the constructed nonlinear droop control model has correct steady-state characteristics and engineering practicality, avoids the model learning incorrect mapping relationships that could lead to control failure or system instability, and improves the mapping accuracy of the inverse mapping artificial neural network.

[0084] In some specific embodiments of this application, the inverse mapping artificial neural network includes an input layer, a hidden layer, and an output layer connected in sequence.

[0085] The input layer consists of N neurons, the number of which is equal to the number of parallel converters in the multi-source parallel DC microgrid, and the number of N neurons is an integer not less than 2. The N neurons in the input layer are used to receive the target steady-state output current of the N parallel converters respectively.

[0086] Specifically, the input layer consists of N neurons, where N is the number of parallel converters in the multi-source parallel DC microgrid.

[0087] For example, in a three-source parallel DC microgrid, the input layer of the corresponding inverse mapping artificial neural network includes three neurons.

[0088] The hidden layers include a first hidden layer and a second hidden layer. The first hidden layer contains multiple neurons and uses a hyperbolic tangent activation function (tansig). The second hidden layer contains multiple neurons and uses a log-sig activation function (logsig).

[0089] For example, the number of neurons in the first hidden layer can be 64, and the number of neurons in the second hidden layer can be 32.

[0090] The output layer consists of N neurons, which are used to output the nonlinear droop control coefficients corresponding to the N parallel converters. The output layer uses a pure linear activation function.

[0091] Specifically, the output layer consists of N neurons, where N is the number of parallel converters in the multi-source parallel DC microgrid.

[0092] For example, in a three-source parallel DC microgrid, the output layer of the corresponding inverse mapping artificial neural network includes three neurons.

[0093] The embodiments described above in this application reverse the input-output mapping direction of the neural network compared to existing forward proxy schemes. By strictly matching the number of neurons in the input and output layers to the number of parallel converters, and using the desired target steady-state output current vector of each parallel converter as the input of the neural network, and the nonlinear droop control coefficient vector as the output, a precise end-to-end mapping from the target steady-state current to the nonlinear droop coefficient is achieved. This allows the reverse mapping neural network to directly output nonlinear droop control coefficients adapted to the target steady-state output current in a single forward propagation, without any post-training optimization search, without constructing a fitness function, and without... The actual values ​​of the impedance of each branch line need to be obtained or identified, thereby significantly shortening the setting link and improving the setting accuracy and efficiency of the droop control coefficient under nonlinear droop control. The hidden layer adopts a cascaded combination of hyperbolic tangent and logarithmic sigmoid activation functions, which enhances the fitting ability and feature extraction depth of the inverse mapping artificial neural network to the strong nonlinear characteristics of the microgrid. The output layer adopts a pure linear activation function, which ensures the distortion-free output of the droop control coefficient in the continuous domain and avoids the loss of control accuracy caused by activation function saturation or truncation. Thus, the accuracy and engineering applicability of the inverse mapping are significantly improved while ensuring physical compliance.

[0094] To effectively train the inverse mapping capability of the inverse mapping artificial neural network, in some specific embodiments of this application, for S13, the target steady-state output current in the effective training dataset is input, and the nonlinear droop control coefficient is used as the desired output to train the inverse mapping artificial neural network, which can be implemented by S131 to S134.

[0095] S131 divides the effective training dataset into a training set, a test set, and a validation set.

[0096] For example, the effective training set can be divided according to a preset ratio, such as 50%:25%:25%, and the division ratio can be adjusted according to the specific model training requirements.

[0097] S132: Input the target steady-state output current in the training set into the inverse mapping artificial neural network, output the nonlinear droop control coefficient, use the Levenberg–Marquardt algorithm for backpropagation training, iterate for a preset number of rounds, and determine the trained inverse mapping artificial neural network.

[0098] Specifically, the maximum number of preset rounds can be set to 500 rounds.

[0099] Specifically, in the model training phase of step S132, a validation set can be used to validate the trained back-mapping artificial neural network in each iteration. If the mean square error of the back-mapping artificial neural network is less than a preset mean square error threshold or the mean square error of the back-mapping artificial neural network increases for a consecutive preset number of iterations, an early stop mechanism is activated to stop training.

[0100] Specifically, the preset mean square error threshold can be set to 1×10. -4 The number of consecutive preset iteration rounds can be set to 6 consecutive rounds. The mean square error threshold and the number of consecutive preset iteration rounds can be adaptively adjusted.

[0101] The early stop mechanism means terminating training prematurely, before the preset number of rounds has been reached.

[0102] After training is complete, the trained back-mapping artificial neural network can be tested using a test set to determine the model evaluation results of the trained back-mapping artificial neural network.

[0103] Specifically, the target steady-state output current of each group in the test set is input into a trained inverse mapping artificial neural network to obtain the corresponding predicted value of the nonlinear droop control coefficient. The predicted value is then compared with the corresponding real nonlinear droop control coefficient in the test set on a sample-by-sample basis, and the maximum relative error (MRE) and coefficient of determination (R²) are calculated as model evaluation indicators. If the coefficient of determination is higher than 0.9990, and considering the steady-state characteristics of nonlinear droop control, the maximum relative error of current distribution under current sharing conditions can be controlled below 0.31%, and the maximum relative error of current distribution under non-current sharing conditions does not exceed 4.65%, which meets the methodological requirements of high-precision tuning of droop control coefficients for nonlinear droop control, then the model is determined to meet the accuracy requirements for engineering applications. Otherwise, the process is returned to readjust the network structure or supplement training data for training.

[0104] The embodiments described above in this application utilize the Levenberg-Marquardt algorithm for efficient training combined with an early stopping mechanism on the validation set. This ensures rapid model convergence while effectively preventing overfitting, thereby obtaining a back-mapping artificial neural network with strong generalization ability and reliable accuracy.

[0105] In order to set the nonlinear droop control coefficient of the DC microgrid, in some specific embodiments of this application, for S14, the actual target steady-state output current of the multi-source parallel DC microgrid is input into the trained inverse mapping artificial neural network to determine the nonlinear droop control coefficient corresponding to the actual target steady-state output current, which can be achieved by S141 to S142.

[0106] S141 invokes the trained inverse mapping artificial neural network.

[0107] S142, the actual target steady-state output current of the multi-source parallel DC microgrid is input into the trained inverse mapping artificial neural network, and a single forward propagation is performed to determine the nonlinear droop control coefficient corresponding to the actual target steady-state output current of the multi-source parallel DC microgrid.

[0108] Specifically, the actual target steady-state output current of a multi-source parallel DC microgrid can be determined by: Step 1: Convert the total load power to the total load current using the DC power formula: , in, Indicates the total load current of the system. This represents the total load power of the system. This indicates the rated voltage of the DC bus.

[0109] Specifically, if the total current is known, this step can be skipped.

[0110] Step 2: Determine the load ratio of each converter based on the rated power of the converters in the multi-source parallel DC microgrid.

[0111] Specifically, firstly, the preset power allocation ratio of each converter is determined based on the rated power of the converters in the multi-source parallel DC microgrid. Then, the preset power allocation ratio is normalized to determine the load ratio undertaken by each converter. , , n=1,2,…,N in, This indicates the rated power of the first converter. This indicates the rated power of the second converter. This indicates the rated power of the third converter. This represents the rated power of the nth converter. This indicates the preset power allocation ratio of the first converter. This indicates the preset power allocation ratio of the second converter. This indicates the preset power allocation ratio of the third converter. This represents the preset power allocation ratio of the nth converter. This indicates the load proportion undertaken by the first converter. This indicates the load proportion undertaken by the second converter. This indicates the load proportion undertaken by the third converter. This indicates the load proportion undertaken by the nth converter, and N represents the number of parallel converters.

[0112] When the rated power of each converter is different, the preset power allocation ratio is proportional to the rated power.

[0113] For example, if the rated power of the three converters are 10kW, 20kW, and 30kW respectively, then the preset ratio is 1:2:3, corresponding to... =1 / 6、 =1 / 3、 =1 / 2.

[0114] Step 3: Determine the actual target steady-state output current of each converter based on the load ratio of each converter and the total load current:

[0115] in, This represents the target steady-state output current of the nth converter. This indicates the load proportion undertaken by the nth converter. This represents the total load current of the system.

[0116] To address the need for batch tuning under multiple operating conditions, the target steady-state output current vectors corresponding to multiple operating conditions can be packaged and input into a trained inverse mapping artificial neural network at one time. The trained inverse mapping artificial neural network then outputs the set of nonlinear droop control coefficients corresponding to each operating condition in parallel during a single batch forward propagation.

[0117] The embodiments described above in this application can quickly and accurately obtain and issue nonlinear droop control coefficients based on the actual target steady-state output current through a single forward propagation of a trained back-mapping artificial neural network. This eliminates the need for any optimization search, table lookup interpolation, or parameter limiting post-processing, significantly shortening the tuning link and enabling real-time adaptive tuning of control parameters for multi-source parallel DC microgrids. Furthermore, the entire tuning process does not rely on any form of line impedance measurement or identification, enabling stable application in real engineering scenarios where line impedance is unknown, time-varying, or aging-related. This avoids the accuracy degradation problem of analytical compensation methods under impedance identification errors.

[0118] In some specific embodiments of this application, a method for determining the nonlinear droop control coefficient of a DC microgrid further includes steps S15 to S16.

[0119] S15, using the maximum total system load current and the maximum DC bus voltage sag of the multi-source parallel DC microgrid as constraint boundaries, and combining them with the equivalent total resistance of the corresponding branch of each converter, the nonlinear droop control coefficients are checked to determine the nonlinear droop control coefficients that meet the voltage sag constraints: ; in, This represents the nonlinear droop control coefficient of the nth parallel converter. This represents the equivalent total resistance of the branch corresponding to the nth converter. This represents the maximum total system load current. This indicates the maximum DC bus voltage drop.

[0120] Specifically, the maximum DC bus voltage drop value is the maximum allowable DC bus voltage drop value for a multi-source parallel DC microgrid system. It is used to limit the bus voltage drop of the multi-source parallel DC microgrid under the maximum total load condition, and to prevent the DC bus voltage from falling below the minimum allowable operating voltage of the load and the parallel converter.

[0121] Maximum DC bus voltage This is a preset system-level voltage quality constraint.

[0122] Nonlinear droop control coefficient This represents the nonlinear droop control coefficient of the inverse mapping neural network for the output of the nth parallel converter.

[0123] Equivalent total resistance of the corresponding branch of the converter The preset line parameters, line design parameters, or conservative upper limits of line resistance in the averaged simulation model can be used.

[0124] Maximum total system load current These are preset system-level design parameters.

[0125] The principle behind the formula for checking voltage sag constraints on nonlinear droop control coefficients is as follows: Determine the maximum equivalent total resistance of the system based on the allowable voltage drop: The DC bus voltage sag is determined by the equivalent total resistance (droop factor + line resistance) of each converter and the output current. The maximum system voltage sag occurs under the condition of maximum total load and current distribution according to the preset ratio.

[0126] Based on the droop control characteristics, the voltage drop of a single converter is:

[0127] in, Indicates the first n Voltage sag value of the corresponding branch of the converter. Indicates the first n The output current of the converter Indicates the first n The equivalent total resistance of the corresponding branch of the converter. Indicates the first n Nonlinear droop control coefficient of the converter Indicates the firstn The line resistance of the corresponding branch of the converter. Since all converters share the same bus voltage, ,in, This indicates the DC bus voltage drop value.

[0128] To ensure that the DC bus voltage drop does not exceed the maximum DC bus voltage The system's equivalent total conductance (the sum of the conductances of all branches) must satisfy the following:

[0129] in, This represents the equivalent total resistance of the branch corresponding to the first converter. This represents the equivalent total resistance of the branch corresponding to the second converter. This represents the equivalent total resistance of the branch corresponding to the third converter. This indicates the maximum total load current designed for the system. This indicates the maximum allowable DC bus voltage drop value of the system.

[0130] S16, the nonlinear droop control coefficient that meets the voltage drop constraint is sent to the underlying nonlinear droop control loop of the parallel converter for parameter setting.

[0131] The nonlinear droop control coefficients that meet the voltage drop constraint are sent to the underlying nonlinear droop control loop of each parallel converter to complete the setting of the nonlinear droop control parameters of each parallel converter.

[0132] Nonlinear droop control coefficients that do not meet the voltage sag constraint are identified as candidate control coefficients that do not meet the engineering constraint, and nonlinear droop control coefficients that meet the voltage sag constraint are re-determined.

[0133] The embodiments described above incorporate the maximum total system load current, the maximum DC bus voltage sag, and the branch equivalent resistance into a unified verification framework, thereby achieving quantitative constraint tuning of the nonlinear droop control coefficient and effectively avoiding the risk of bus voltage exceeding limits due to improper parameter selection. At the same time, the verified nonlinear droop control coefficient is directly sent down to the underlying control loop, forming a closed-loop parameter configuration mechanism from system-level constraints to device-level execution, which significantly improves the voltage stability and operational reliability of multi-source parallel DC microgrids under extreme conditions.

[0134] The method for determining the nonlinear droop control coefficient of a DC microgrid provided in this application solves the technical problems of the significantly amplified impact of the droop control coefficient on system performance under nonlinear droop control, the high requirement for setting accuracy, and the reliance of existing setting methods on line impedance information or post-training optimization search.

[0135] The embodiments described above in this application, by removing samples with negative current, samples exceeding the rated current limit of the parallel converter, and samples with non-convergent output from the training dataset and performing a second data cleaning using Kirchhoff's voltage law, ensure that the training samples for training the inverse mapping artificial neural network simultaneously satisfy the device operating boundary and DC branch voltage balance constraints, thereby improving the mapping accuracy of the inverse mapping artificial neural network. The input-output mapping direction of the neural network surrogate model is reversed from the existing forward surrogate scheme's droop coefficient → steady-state output current to the target steady-state output current → droop coefficient. This allows the tuning parameters to be directly obtained through a single forward propagation after training, completely eliminating the post-training search stages such as fitness function construction, candidate coefficient grid exhaustion, or gradient optimization required in existing forward surrogate artificial neural network schemes. Furthermore, the tuning process does not require obtaining or identifying the current actual values ​​of the impedance of each branch during the training and inference phases, and the droop control coefficient... The implicit fitting relationship of line impedance is obtained entirely through the nonlinear fitting capability of the training samples to the inverse mapping artificial neural network, which is different from all analytical compensation methods and adaptive droop methods that require measurement or estimation of line impedance. It is not limited to a specific form of nonlinear droop control strategy, but has universality for nonlinear droop control strategies. It only requires multi-dimensional grid scanning sampling of the simulation model of the corresponding nonlinear droop control strategy during the offline training stage. The structure, training process and inference process of the inverse mapping proxy model are consistent. The nonlinear fitting capability of the inverse mapping artificial neural network directly establishes an accurate inverse mapping from the target steady-state output current to the droop control coefficient, which perfectly matches the methodological requirement of high-precision tuning of the droop control coefficient in nonlinear droop control. The inverse mapping artificial neural network also supports single-batch forward propagation and parallel output of multiple sets of droop control coefficients corresponding to multiple operating conditions, which is convenient for pre-tuning of typical operating conditions and parameter table storage and retrieval in engineering practice.

[0136] This application provides a method for determining the nonlinear droop control coefficient of a DC microgrid. In engineering deployment, it only includes two stages: offline grid scanning + data cleaning + network training and online single forward inference + distribution. It eliminates the need for table lookup interpolation, online optimization, impedance identification, and inter-converter communication, thus simplifying the engineering deployment process.

[0137] The preferred features in the above embodiments can be used individually in any embodiment, or in any combination thereof, provided they do not conflict with each other. Furthermore, parts not described in detail in the embodiments can be implemented using existing technologies.

[0138] The following examples and comparative examples will be used to further illustrate this application in order to better understand the above-mentioned technical solutions. It should be understood that the following are only some examples and are not intended to limit this application.

[0139] This paper uses a three-source parallel DC microgrid and an id-Vdc coupled nonlinear droop control strategy as an example to illustrate a method for determining the nonlinear droop control coefficient of a DC microgrid provided in this application. It should be noted that the id-Vdc coupled nonlinear droop control is only a typical implementation of the nonlinear droop control strategy of this application. The method provided in this application is also applicable to the setting of droop control coefficients for other nonlinear droop control strategies, such as piecewise nonlinear droop, exponential nonlinear droop, and quadratic nonlinear droop.

[0140] The id-Vdc coupled nonlinear droop control strategy is as follows: using the converter d-axis current reference value idi As the direct output of the nonlinear droop control loop, it eliminates the need for the bus voltage PI outer loop, which is located outside the current inner loop in conventional dual-loop droop control.

[0141] The id-Vdc coupled nonlinear droop control strategy can be expressed as: i di = i d,ref - k i ·( V dc - V dc ) in, k i This represents the nonlinear droop control coefficient to be tuned. i d,ref This indicates the reference value of the unloaded rated d-axis current. V dc Indicates the reference value of DC voltage at the bus. V dc This represents the actual value of the DC voltage at the bus.

[0142] Figure 2 This is a schematic diagram of a three-source parallel DC microgrid topology according to an exemplary embodiment.

[0143] Reference Figure 2As shown, the multi-source parallel DC microgrid includes three independent power sources G1, G2, and G3. Converters #1, #2, and #3 are voltage source converters VSC1, VSC2, and VSC3, respectively. The three independent power sources are connected in parallel to a shared DC bus (DC BUS) via their respective converters. The line resistances of cables #1, #2, and #3 on the three branches are Rc1, Rc2, and Rc3, respectively. The line resistances of the three branches are set to unequal engineering values, such as Rc1 = 0.021Ω, Rc2 = 0.011Ω, and Rc3 = 0.032Ω, to represent non-ideal wiring conditions with impedance mismatch. A resistive load Rload is connected to the bus, which in this embodiment is Rload = 10Ω. A DC capacitor Ci, i = 1, 2, 3, is connected in parallel on the bus side; in this embodiment, Cb = 2 mF.

[0144] Each converter adopts an id-Vdc coupled nonlinear droop control strategy, and its droop control strategy is as follows: i di = i d,ref - k i ·( V dc - V dc ) in, k i (i = 1, 2, 3) represents the nonlinear droop control coefficient to be tuned.

[0145] It is particularly important to emphasize that this application does not require the actual values ​​of Rc1, Rc2, and Rc3 during the tuning process; the above-mentioned line resistance values ​​exist only as real environmental parameters of the offline simulation model, and the inverse mapping artificial neural network of this application keeps the line resistance values ​​completely black box during the training and inference phases.

[0146] Figure 3 This is a schematic diagram of a nonlinear droop control structure according to an exemplary embodiment.

[0147] Reference Figure 3 As shown, an averaged simulation model of the three-source parallel DC microgrid using the id-Vdc coupled nonlinear droop control strategy was built on the MATLAB / Simulink platform.

[0148] in, (i=1,2,3) represent independent power sources. Indicates the instantaneous value of three-phase AC voltage ( (three phases) Indicates the instantaneous value of the three-phase alternating current ( (three phases) This indicates the DC input current of the converter; This represents the DC-side voltage of the converter (i.e., the capacitor). (Voltage at both ends) This indicates that a DC capacitor is connected in parallel on the bus side to stabilize the DC bus voltage. This indicates the total current output to the load or bus. This represents the phase angle of the phase-locked loop output, used for coordinate transformation; The reference angular frequency is typically 2π×50 or 60 rad / s; L represents the filter inductance value.

[0149] In converter vector control: This represents a stationary three-phase coordinate system. dq Indicates a synchronously rotating coordinate system. d Axis: Typically oriented in the direction of the mains voltage vector. q Axis: Perpendicular to d Axis; PLL stands for Phase-Locked Loop, used to detect the phase and frequency of the mains voltage, and outputs an angle. .

[0150] , These represent the actual measured currents at... d shaft and q The components of the axis (feedback value); , They represent d shaft and q Shaft current reference value (given value); , These represent the converter output voltage at... d shaft and q The axis component (the controller's output, which is inversely transformed to generate PWM). This indicates the grid electromotive force (or the voltage at point PCC) at... d The axis components are locked by the PLL. d Shaft, usually =0.

[0151] In nonlinear droop control: express d The deviation or disturbance signal of the shaft current is used as the input for droop control; It represents the reciprocal of the droop coefficient or gain coefficient; The droop characteristic is based on DC voltage. and d-axis current The nonlinear droop control module, as shown in the diagram, uses a folded-out icon to represent a nonlinear functional relationship, i.e., based on... Dynamic adjustment of current reference value PI stands for Proportional-Integral Controller, used for tracking current commands without steady-state error. The PI output above corresponds to... d Part of the shaft voltage component, corresponding to the PI output below. q A portion of the shaft voltage component; modulation indicates SVPWM or SPWM modulation module, which will , It is converted into a drive signal for the switching transistor.

[0152] Reference Figure 3 As shown, its logical flow is: Acquisition , , ; Obtained via PLL ,Will Transform into , ; Outer ring (non-linear droop) according to Calculate the droop characteristic curve and the preset curve. d Shaft current reference value ( It is usually set to 0 to maintain unity power factor, or it is given by reactive power control. Inner loop (vector control): Calculate current error ( ), Indicates the reference current. This indicates the measured current, which is input to the PI controller; the PI controller output is then fed forward decoupling term (…). and ) and grid voltage feedforward ( ); final generation , ; Will , Inverse transformation back Coordinate system, driving converter switch.

[0153] In this embodiment, the droop control coefficient is... , , The value range is set to [0.2, 4]. Within this value range, 12 sampling points are evenly selected for each droop control coefficient to form 12 × 12 × 12 = 1728 combinations of droop control coefficients.

[0154] For each combination of droop control coefficients, an averaged simulation model was run for 0.8 s, and the output currents of the three converters were recorded at the steady-state time of t = 0.5 s. , , Each droop control coefficient is paired with its corresponding steady-state output current and stored in the training dataset, with each row containing 6 tuples. , , , , , ].

[0155] The training dataset is cleaned row by row, removing any sample row with negative current value or saturated sample row that exceeds the rated current capacity of the converter. The data is then cleaned again based on Kirchhoff's voltage law, and finally the cleaned effective training dataset is obtained.

[0156] Negative current values ​​indicate that the converter has entered an abnormal operating range or the simulation has not converged.

[0157] Figure 4 This is a comparative schematic diagram of a forward mapping artificial neural network and a backward mapping artificial neural network according to an exemplary embodiment.

[0158] Reference Figure 4 As shown, this demonstrates the bidirectional application process of using a fully connected artificial neural network (ANN) for system modeling and parameter inversion.

[0159] The upper part is the forward surrogate model, also known as the forward mapping artificial neural network, which aims to establish a model based on control parameters ( From system response (DC current) arrive The first part is a fast mapping of DC current to replace complex physical simulations and assist in optimization searches; the second part is the reverse application, which constructs an inverse mapping artificial neural network to achieve the desired system state index (DC current). arrive Directly reverse-engineer the required control parameters ( This allows for the reverse design or parameter tuning of inverse artificial neural networks.

[0160] Specifically, a three-layer feedforward neural network is constructed as a back-mapping artificial neural network: the input layer contains 3 neurons, corresponding to the input vector [ , , The first hidden layer contains 64 neurons and uses the tansig activation function; the second hidden layer contains 32 neurons and uses the logsig activation function; the output layer contains 3 neurons, corresponding to the output vector []. , , The purelin activation function is used.

[0161] The effective training dataset was randomly divided into training, validation, and test sets at a ratio of 50%:25%:25%. Backpropagation training was performed using the training set and the Levenberg-Marquardt algorithm. A maximum of 500 iterations and a mean squared error (MSE) convergence threshold of <1×10⁻⁶ were set. -4 The early stopping mechanism is enabled, and training is stopped when the mean square error of the validation set increases for six consecutive rounds.

[0162] In this embodiment, the trained back-mapping artificial neural network has an overall coefficient of determination R² = 0.9990 and a mean squared error MSE = 9.79 × 10⁻⁶ on the test set. -5 .

[0163] Figure 5 This is a training regression scatter plot of a back-mapping artificial neural network according to an exemplary embodiment. Figure 6 This is a schematic diagram illustrating a sample-by-sample comparison of the predicted and true values ​​of a trained back-mapping artificial neural network on a test set, according to an exemplary embodiment.

[0164] Figure 5 This demonstrates the regression performance of the back-mapping artificial neural network during the training phase. The predicted values ​​and the true target values ​​in the figure highly coincide near the ideal diagonal (Y=T), and the coefficient of determination... A value as high as 0.9990 indicates that the trained back-mapping artificial neural network has a high degree of accuracy in training the parameters. , , The learning fit is extremely high.

[0165] Figure 6 This further validates the model's generalization ability on the test set. Sample-by-sample comparison shows that, regardless of whether the variation is stepwise... serrated wavy It is still a high-frequency oscillation. The prediction curves (orange dashed lines) of the ANN closely follow the actual value curves (blue solid lines), proving that the network can accurately deduce the corresponding control parameters based on the system state.

[0166] Reference Figure 5 and Figure 6 As shown, the trained back-mapping artificial neural network achieves an engineering-usable level of fitting accuracy for back-mapping relationships.

[0167] When setting droop control coefficients for this three-source parallel DC microgrid in practical engineering, the desired steady-state output current vector only needs to be directly fed into the trained back-mapping artificial neural network. After one forward propagation, the corresponding droop control coefficient vector can be directly output. The droop control coefficient vector is then sent down to the underlying id-Vdc coupled nonlinear droop control loops of VSC1, VSC2, and VSC3 respectively, thus completing the droop control parameter setting. The entire online tuning process does not involve any post-training optimization search, does not construct any fitness function, and does not require obtaining the current value of the line impedance.

[0168] In this embodiment, the method provided in this application is verified by way of example for seven representative operating conditions (including flow equalization condition and non-flow equalization condition).

[0169] Typical verification result example: ① The target steady-state output current is [8 A, 8 A, 8 A], the droop control coefficient predicted by the inverse mapping artificial neural network is [2.306, 2.314, 2.309], and the measured steady-state current after distribution is [8.012 A, 8.017 A, 7.975 A], with a maximum relative error of 0.31%; ② The target steady-state output current [15 A, 6 A, 3 A], the droop control coefficient predicted by the inverse mapping artificial neural network is [1.227, 3.086, 6.165], the measured steady-state current is [14.940 A, 6.036 A, 3.024 A], and the maximum relative error is 0.60%; ③ The target steady-state output current is [18 A, 2 A, 4 A]. The droop control coefficient predicted by the inverse mapping artificial neural network is [0.960, 8.500, 4.429]. The measured steady-state current is [18.04 A, 2.093 A, 3.994 A]. The maximum relative error is 4.65%.

[0170] Under all seven operating conditions, the maximum current distribution relative error does not exceed 4.65%. The larger relative error occurs in the branch with the smallest target current (such as 2 A or 3 A), and the corresponding absolute deviation still does not exceed 0.1 A. The relative error is amplified by the small denominator.

[0171] Figure 7 This is a schematic diagram of the DC current waveform of a converter output under typical operating conditions, according to an exemplary embodiment.

[0172] Reference Figure 7 As shown, Figure 7(a) and (b) in the figure are the DC current waveforms of the converter output under the current sharing condition (8 / 8 / 8 A) and the non-current sharing condition (15 / 6 / 3 A), respectively. It can be seen that the output current of each converter converges smoothly to the target steady-state value within about 50 ms, without continuous oscillation or circulating current between branches, which verifies the effectiveness and closed-loop stability of the method provided in this application.

[0173] The id-Vdc coupled nonlinear droop control law in the above implementation process can be replaced with other forms of nonlinear droop control strategies, such as: ① Piecewise nonlinear droop—using droop characteristics with different slopes in different current ranges; ② Exponential nonlinear droop—the droop characteristic curve is an exponential function of the converter output current; ③ Quadratic nonlinear droop—the droop characteristic curve is a quadratic function of the converter output current. When using other nonlinear droop control strategies, only the underlying nonlinear droop control strategy needs to be replaced in the averaging simulation model. The structure, training process, data cleaning, inference, and deployment process of the inverse mapping artificial neural network remain consistent. Therefore, it can be seen that the inverse mapping structure and the end-to-end tuning path without line impedance identification provided in this application are universally applicable to various nonlinear droop control strategies.

[0174] In the embodiments of piecewise nonlinear, exponential nonlinear, and quadratic nonlinear droop control strategies, the data acquisition and reverse mapping process in the embodiments corresponding to the above-mentioned id-Vdc coupled nonlinear droop control strategy can all be used: For three parallel converters, the three nonlinear droop control coefficients to be tuned in the corresponding strategy are limited to [0.2, 4], and 12 sampling points are uniformly taken in each dimension to form 12³=1728 sets of coefficient combinations. The averaged simulation model is run for 0.8 s for each set, and the steady-state output current of the three channels is recorded at t=0.5 s. After two-stage data cleaning, the inverse mapping artificial neural network is trained using 3 input neurons, 64 / 32 hidden layer neurons and 3 output neurons.

[0175] Different strategies only replace the nonlinear droop function and its coefficients to be tuned in the averaged simulation model. The actual target steady-state output current is used as the input, the nonlinear droop control coefficient is used as the output, and the two-stage data cleaning, network training and single forward inference steps are kept consistent.

[0176] When the number of parallel converters in a multi-source parallel DC microgrid is expanded from N = 3 to N > 3, it is only necessary to adjust the number of neurons in the input and output layers of the back-mapping artificial neural network from 3 to N, and correspondingly increase the number of sampling combinations in the droop control coefficient dimension; the network hidden layer structure, activation function selection, training algorithm, and data cleaning criteria remain unchanged; it can be seen that the method provided in this application has good scalability for multi-source parallel DC microgrids of different parallel scales.

[0177] Taking N=4 as an example, the input and output layers of the inverse mapping artificial neural network each have 4 neurons, which receive the target steady-state output current of 4 parallel converters and output 4 nonlinear droop control coefficients respectively; if each dimension still uniformly takes 12 sampling points in the range of [0.2,4], then 12 4 =20736 combinations of four-dimensional coefficients. The hidden layer can continue to use 64 / 32 neurons, and the training set, validation set and test set are divided at 50% / 25% / 25%, with a maximum of 500 training epochs.

[0178] Figure 8 This is a schematic diagram comparing the training data with and without Kirchhoff's voltage law secondary data cleaning and the branch voltage balance constraints of the network output, according to an exemplary embodiment.

[0179] Figure 8 (a) in the diagram represents a comparison of the proportion of training data and network output that violate KVL constraints before and after Kirchhoff's Voltage Law (KVL) secondary data cleaning. Figure 8 (b) in the figure represents the empirical cumulative distribution function curve of the maximum branch voltage balance residual corresponding to the network output before and after the second data cleaning of Kirchhoff's voltage law.

[0180] To verify the effectiveness of the second data cleaning based on Kirchhoff's voltage law, a comparative experiment with and without Kirchhoff's voltage law-based second data cleaning was conducted on a 375 V three-source parallel DC microgrid averaging simulation model.

[0181] The circuit resistances of the three branches are 0.021 Ω, 0.011 Ω, and 0.032 Ω, respectively. Each nonlinear droop control coefficient is selected from 11 values ​​at equal intervals of 0.4 to 4.4, and the nonlinear droop control coefficients are determined as: {0.4, 0.8, 1.2, 1.6, 2.0, 2.4, 2.8, 3.2, 3.6, 4.0, 4.4}, resulting in a total of 1331 scan samples.

[0182] Reference Figure 8 As shown in (a), the voltage balance residuals of the three pairs of branches were calculated according to the preset threshold Δ=0.761029 V. Among the 1331 samples that had undergone the first data cleaning but had not yet undergone the second cleaning of Kirchhoff's voltage law, there were 12 samples whose voltage balance residuals of at least one pair of branches were not less than the preset threshold, that is, there were 12 samples that exceeded the limit, accounting for 0.9016%. After deleting the 12 samples, 1319 valid samples were retained, and the proportion of training data that exceeded the limit after the second cleaning of Kirchhoff's voltage law was reduced to 0.

[0183] Using the same 64 / 32 hidden layer structure, tansig / logsig activation function, purelin output layer, trainlm training algorithm, and 50% / 25% / 25% partitioning method to divide the training set, validation set and test set, and with a fixed random seed, the output of the back-mapping artificial neural network was compared and tested on 329 common test samples.

[0184] Reference Figure 8 As shown, at a preset threshold Δ=0.761029 V, when the two sets of inverse mapping artificial neural networks output nonlinear droop control coefficients for the test samples, the branch voltage balance residuals corresponding to their output results are all strictly less than or equal to the preset threshold, and the over-limit ratio of the nonlinear droop control coefficients is 0. After secondary cleaning using Kirchhoff's voltage law, the maximum branch voltage balance residual decreased from 0.70741 V to 0.63299 V, a reduction of 10.52%, the P95 residual decreased from 0.36497 V to 0.29799 V, a reduction of 18.35%, and the mean square error of the droop coefficient decreased by 30.27%.

[0185] Among them, the P95 residual characterizes the residual value located at the 95th quantile after sorting the absolute values ​​of the branch voltage balance residuals of all test samples from smallest to largest.

[0186] Reference Figure 8 As shown in (b), after secondary data cleaning using Kirchhoff's voltage law, the cumulative distribution curve of the maximum branch voltage balance residual in the network output shifts towards a smaller residual, indicating that the residual distribution and physical consistency margin are improved.

[0187] The above results show that the second data cleaning of Kirchhoff's voltage law eliminates default samples in the training data and improves the physical margin of the network output relative to the preset threshold. In this embodiment, the over-limit rate of both sets of network outputs is 0 under the preset threshold, so it is not described as a decrease in the over-limit rate.

[0188] Figure 9 This is a schematic diagram illustrating the structure of a nonlinear droop control coefficient determination system for a DC microgrid according to an exemplary embodiment.

[0189] Reference Figure 9 As shown, another embodiment of this application provides a nonlinear droop control coefficient determination system 100 for DC microgrids, which may include a training dataset construction module 110, a data cleaning module 120, a model training module 130, and a nonlinear droop control coefficient determination module 140.

[0190] The training dataset construction module 110 is used to construct an averaged simulation model of a multi-source parallel DC microgrid using a preset nonlinear droop control strategy, and to perform multi-dimensional grid scanning sampling to determine the training dataset, which includes the target steady-state output current and the nonlinear droop control coefficient. The data cleaning module 120 is used to remove samples with negative current, samples with current greater than the rated current limit of the parallel converter, and samples with non-convergent output from the training dataset, and to perform data cleaning processing based on Kirchhoff's voltage law to determine the effective training dataset. The model training module 130 is used to construct an inverse mapping artificial neural network, taking the target steady-state output current as input and the nonlinear droop control coefficient as the desired output, and training the inverse mapping artificial neural network. The nonlinear droop control coefficient determination module 140 is used to input the actual target steady-state output current of the multi-source parallel DC microgrid into a trained inverse mapping artificial neural network to determine the nonlinear droop control coefficient corresponding to the actual target steady-state output current.

[0191] In the embodiments described above, the first data cleaning is performed by removing samples with negative current, samples exceeding the rated current limit of the parallel converter, and samples with non-convergent output from the training dataset. A second data cleaning is then performed based on Kirchhoff's voltage law. This ensures that the training samples for the inverse mapping artificial neural network simultaneously satisfy both device operating boundaries and DC branch voltage balance constraints, improving the mapping accuracy of the inverse mapping artificial neural network. Using an effective training dataset to train the inverse mapping artificial neural network establishes a direct inverse mapping, making the tuning process of the nonlinear droop control coefficient completely independent of line impedance measurement or identification, avoiding accuracy degradation under unknown impedance, time-varying, or aging scenarios. After training, the trained inverse mapping artificial neural network outputs the corresponding nonlinear droop control coefficient through a single forward propagation based on the actual target steady-state output current of the input multi-source parallel DC microgrid. This eliminates the need for exhaustive post-training or gradient optimization, significantly shortening the tuning path and achieving a rapid dynamic response where the steady-state current smoothly reaches the target value without circulating current oscillation.

[0192] Regarding the embodiments of the above system, the specific ways in which each module performs operations have been described in detail in the embodiments of the method, and will not be elaborated here.

[0193] Based on the same technical concept, in some specific embodiments of this application, a terminal includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it can be used to execute a method for determining the nonlinear droop control coefficient of a DC microgrid.

[0194] Based on the same technical concept, in some specific embodiments of this application, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, can be used to perform a method for determining the nonlinear droop control coefficient of a DC microgrid.

[0195] Optionally, the memory is used to store programs; the memory may include volatile memory, such as random-access memory (RAM), such as static random-access memory (SRAM), double data rate synchronous dynamic random-access memory (DDR SDRAM), etc.; the memory may also include non-volatile memory, such as flash memory. The memory is used to store computer programs (such as application programs and functional modules that implement the above methods), computer instructions, etc., and the aforementioned computer programs and computer instructions can be partitioned and stored in one or more memories. Furthermore, the aforementioned computer programs, computer instructions, data, etc., can be accessed by the processor.

[0196] The aforementioned computer programs, computer instructions, etc., can be stored in partitions within one or more memory locations. Furthermore, the aforementioned computer programs, computer instructions, data, etc., can be accessed by a processor.

[0197] A processor is used to execute a computer program stored in memory to implement the various steps of the methods involved in the above embodiments. For details, please refer to the relevant descriptions in the preceding method embodiments.

[0198] The processor and memory can be separate structures or integrated structures. When the processor and memory are separate structures, they can be coupled together via a bus.

[0199] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0200] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0201] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0202] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0203] The foregoing has described some specific embodiments of this application. It should be understood that this application is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the substantive content of this application. The above-described preferred features can be used in any combination without conflict.

Claims

1. A method for determining the nonlinear droop control coefficient of a DC microgrid, characterized in that, include: An averaged simulation model of a multi-source parallel DC microgrid with a preset nonlinear droop control strategy is constructed, and multi-dimensional grid scanning sampling is performed to determine the training dataset, which includes the target steady-state output current and the nonlinear droop control coefficient. The training dataset is cleaned by removing samples with negative current, samples with current greater than the rated current limit of the parallel converter, and samples with non-convergent output, and the data is cleaned based on Kirchhoff's voltage law to determine the effective training dataset. Construct an inverse mapping artificial neural network, using the target steady-state output current in the effective training dataset as input and the nonlinear droop control coefficient as the desired output, and train the inverse mapping artificial neural network. The actual target steady-state output current of the multi-source parallel DC microgrid is input into a trained inverse mapping artificial neural network to determine the nonlinear droop control coefficient corresponding to the actual target steady-state output current.

2. The method for determining the nonlinear droop control coefficient of a DC microgrid according to claim 1, characterized in that, The averaged simulation model of the multi-source parallel DC microgrid, which employs a preset nonlinear droop control strategy, is constructed, and multi-dimensional grid scanning sampling is performed to determine the training dataset, including: An averaged simulation model of the multi-source parallel DC microgrid employing a preset nonlinear droop control strategy was built in the simulation platform. Based on the preset nonlinear droop control strategy, determine the range of values ​​for the preset nonlinear droop control coefficient. Within the preset range of nonlinear droop control coefficient values, each nonlinear droop control coefficient corresponding to each parallel converter is uniformly sampled to determine the combination of nonlinear droop control coefficients. Using each set of nonlinear droop control coefficients in the aforementioned combination, the averaged simulation model is run to a preset steady-state time to determine the target steady-state output current corresponding to each set of nonlinear droop control coefficients, thereby constructing the training dataset.

3. The method for determining the nonlinear droop control coefficient of a DC microgrid according to claim 1, characterized in that, The data cleaning process based on Kirchhoff's voltage law, which determines the effective training dataset, includes: For each sample in the training dataset after removing samples with negative current, samples exceeding the rated current limit of the parallel converter, and samples with non-convergent output, calculate the branch voltage balance residual between any two branches: ； in, This represents the voltage balance residual between the m-th branch and the n-th branch in the i-th sample. This represents the nonlinear droop control coefficient corresponding to the m-th branch. This represents the line resistance of the m-th branch. Let represent the steady-state output current of the m-th branch in the i-th sample. This represents the nonlinear droop control coefficient corresponding to the nth branch. This represents the line resistance of the nth branch. This represents the steady-state output current of the nth branch in the i-th sample; Obtain the measurement error term, parameter error term, and steady-state error term: ； ； ； in, This represents the measurement error term. Indicates the accuracy of the current sensor. This represents the rated current of the nth branch. Indicates the parameter error term. This indicates the error in the droop coefficient; Represents the steady-state error term. Represents the steady-state error coefficient. V bus,rated Indicates the rated voltage of the DC bus; The measurement error term, the parameter error term, and the steady-state error term are summed to determine a preset threshold. If the absolute value of the voltage balance residual of any branch is not less than the preset threshold, the corresponding sample is deleted, and the effective training dataset is determined.

4. The method for determining the nonlinear droop control coefficient of a DC microgrid according to claim 1, characterized in that, The inverse mapping artificial neural network includes an input layer, a hidden layer, and an output layer connected in sequence. The input layer includes N neurons, the number of which is equal to the number of parallel converters in the multi-source parallel DC microgrid, and the number of N neurons is an integer not less than 2. The N neurons of the input layer are used to receive the target steady-state output current of the N parallel converters respectively. The hidden layer includes a first hidden layer and a second hidden layer. The first hidden layer includes multiple neurons and uses a hyperbolic tangent activation function. The second hidden layer includes multiple neurons and uses a logarithmic sigmoid activation function. The output layer comprises N neurons, each of which is used to output the nonlinear droop control coefficients corresponding to the N parallel converters, and the output layer uses a purely linear activation function.

5. The method for determining the nonlinear droop control coefficient of a DC microgrid according to claim 1, characterized in that, The step of training the inverse mapping artificial neural network by taking the target steady-state output current in the effective training dataset as input and the nonlinear droop control coefficient as the desired output includes: The effective training dataset is divided into a training set, a test set, and a validation set; The target steady-state output current in the training set is input into the back-mapping artificial neural network, which outputs the nonlinear droop control coefficient. The Levenberg-Marquardt algorithm is used for backpropagation training, and the training is performed for a preset number of iterations to determine the trained back-mapping artificial neural network. During the model training process, the validation set is used to validate the back-mapping artificial neural network in each iteration. If the mean square error of the back-mapping artificial neural network is less than a preset mean square error threshold or the mean square error of the back-mapping artificial neural network increases for a consecutive preset number of iterations, an early stopping mechanism is activated to stop training.

6. The method for determining the nonlinear droop control coefficient of a DC microgrid according to claim 1, characterized in that, The method for determining the actual target steady-state output current of the multi-source parallel DC microgrid includes: Based on the DC power formula, the total load power is converted into the total load current: ； in, Indicates the total load current of the system. This represents the total load power of the system. Indicates the rated voltage of the DC bus; Based on the rated power of the converters in the multi-source parallel DC microgrid, determine the load proportion undertaken by each converter: ； ； ,n=1,2,…,N; in, This indicates the rated power of the first converter. This indicates the rated power of the second converter. This indicates the rated power of the third converter. This represents the rated power of the nth converter. This indicates the preset power allocation ratio of the first converter. This indicates the preset power allocation ratio of the second converter. This indicates the preset power allocation ratio of the third converter. This represents the preset power allocation ratio of the nth converter. This indicates the load proportion undertaken by the first converter. This indicates the load proportion undertaken by the second converter. This indicates the load proportion undertaken by the third converter. This indicates the load proportion undertaken by the nth converter, where N represents the number of parallel converters; Based on the load proportion undertaken by each converter and the total load current, determine the actual target steady-state output current of each converter: ； in, This represents the target steady-state output current of the nth converter. This indicates the load proportion undertaken by the nth converter. This represents the total load current of the system.

7. The method for determining the nonlinear droop control coefficient of a DC microgrid according to claim 1, characterized in that, The step of inputting the actual target steady-state output current of the multi-source parallel DC microgrid into a trained inverse mapping artificial neural network to determine the nonlinear droop control coefficient corresponding to the actual target steady-state output current includes: Invoke the trained inverse mapping artificial neural network; The actual target steady-state output current of the multi-source parallel DC microgrid is input into the trained inverse mapping artificial neural network for a single forward propagation to determine the nonlinear droop control coefficient corresponding to the actual target steady-state output current of the multi-source parallel DC microgrid.

8. The method for determining the nonlinear droop control coefficient of a DC microgrid according to claim 1, characterized in that, The preset nonlinear droop control strategy includes any one of the following: id-Vdc coupled nonlinear droop control strategy, piecewise nonlinear droop control strategy, exponential nonlinear droop control strategy, and quadratic nonlinear droop control strategy.

9. The method for determining the nonlinear droop control coefficient of a DC microgrid according to claim 8, characterized in that, The method further includes: Using the maximum total system load current and the maximum DC bus voltage sag of the multi-source parallel DC microgrid as constraint boundaries, and combining them with the equivalent total resistance of the corresponding branch of each converter, the nonlinear droop control coefficients are checked to determine the nonlinear droop control coefficients that meet the voltage sag constraints. ； in, This represents the nonlinear droop control coefficient of the nth parallel converter. This represents the equivalent total resistance of the branch corresponding to the nth converter. This represents the maximum total system load current. This indicates the maximum DC bus voltage sag. The nonlinear droop control coefficients that meet the voltage drop constraint are sent to the underlying nonlinear droop control loop of the parallel converter for parameter setting.

10. A system for determining the nonlinear droop control coefficient of a DC microgrid, characterized in that, include: The training dataset construction module is used to construct an averaged simulation model of a multi-source parallel DC microgrid using a preset nonlinear droop control strategy, and to perform multi-dimensional grid scanning sampling to determine the training dataset, which includes the target steady-state output current and the nonlinear droop control coefficient. The data cleaning module is used to remove samples with negative current, samples with current greater than the rated current limit of the parallel converter, and samples with non-convergent output from the training dataset, and to perform data cleaning based on Kirchhoff's voltage law to determine the effective training dataset. The model training module is used to construct an inverse mapping artificial neural network, taking the target steady-state output current as input and the nonlinear droop control coefficient as the desired output, and training the inverse mapping artificial neural network. The nonlinear droop control coefficient determination module is used to input the actual target steady-state output current of the multi-source parallel DC microgrid into a trained inverse mapping artificial neural network to determine the nonlinear droop control coefficient corresponding to the actual target steady-state output current.

Images