Direct-current micro-grid system modeling method, system and equipment based on physical structured neural network, medium and program product

Abstract

The invention belongs to the technical field of DC micro-grid modeling, and discloses a DC micro-grid system modeling method based on a physical structured neural network. The method comprises the following steps: firstly, decomposing a DC micro-grid into a plurality of subsystems according to a DC micro-grid topological structure; secondly, extracting key features of each subsystem by using a long short-term memory network; then, integrating subsystem characteristics and system input into a unified attention mechanism model; and finally, when a new load is accessed to the system, realizing extensible modeling through feature extraction and physical structured neural network weight sharing. Compared with a traditional data driving model which can only aim at a certain determined topology modeling and cannot process a dynamic structure change, the method provided by the invention can realize rapid modeling under the operation condition of the dynamic change of the DC micro-grid topology structure, and endows the DC micro-grid data driving model with physical interpretability and expandability; the method is of great significance for mastering the running state of the DC micro-grid in real time and ensuring safe and stable running of the system.

Description

Technical Field

[0001] The present invention belongs to the technical field of DC microgrid modeling methods, and specifically relates to a DC microgrid system modeling method, system, equipment, medium, and program product based on a physical structured neural network. Background Art

[0002] DC microgrids offer advantages such as easy integration of DC components, higher energy transmission efficiency, and the elimination of reactive power compensation. They are a technical solution for the efficient grid connection of renewable energy, energy storage systems, and DC loads. Control and protection strategies are crucial for maintaining the safe and stable operation of DC microgrids, but their performance is ultimately constrained by the accuracy of the available system models. Therefore, establishing accurate DC microgrid models is crucial. However, current knowledge-based modeling approaches for DC microgrids involve extensive high-dimensional matrix operations, struggle to handle nonlinear systems, and are highly sensitive to parameter uncertainty. Traditional data-driven approaches are unable to establish universal models that can adapt to changes in system topology. Therefore, an accurate dynamic modeling approach for DC microgrids is urgently needed.

[0003] Patent application publication number CN119761085A discloses a method and system for rapid modeling of a DC microgrid converter. The method involves fitting the harmonic components of a discontinuous current converter switching function using a trigonometric Fourier series to generate a fitted switching model. Based on the fitted switching model, a maximum power point tracking algorithm is used to obtain an improved photovoltaic time-domain model. An improved energy storage time-domain model is constructed using the fitted switching model. A bipolar switching function is defined based on the three-phase inverter circuit topology, and an improved inverter time-domain model is obtained based on the bipolar switching function. Finally, the improved photovoltaic time-domain model, the improved energy storage time-domain model, and the improved inverter time-domain model are integrated into a DC microgrid architecture to construct a joint simulation model of a photovoltaic-storage DC microgrid. However, because this method uses Fourier series to linearize the switching function, it is difficult to simulate the strongly nonlinear characteristics of the switching process of power electronic devices, and the model accuracy is highly dependent on parameters such as the modulation ratio. Summary of the Invention

[0004] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a DC microgrid system modeling method, system, device, medium, and program product based on a physical structured neural network. Specifically, the present invention establishes a physical structured neural network model based on the DC microgrid topology structure, decomposes it into multiple subsystems, and uses a long short-term memory network to extract the key features of the subsystem for each subsystem. The key features of the subsystem and the system input are respectively input into an attention mechanism network to construct a physical structured neural network model, and ultimately obtain the predicted value of the DC microgrid system state quantity. When a new load is connected to the system, it is only necessary to divide the newly connected part into a new subsystem. By extracting the new load features and sharing the attention mechanism network weights, scalable modeling is achieved. The method proposed in the present invention gives the physical structured neural network model physical interpretability and scalability, avoids a large number of high-dimensional matrix operations, has good accuracy and strong practicality, and can achieve wide applicability of the model when the system topology changes. It provides a new idea for scalable modeling of DC microgrids and ensures rapid response of the system.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A DC microgrid system modeling method based on a physical structured neural network includes the following steps:

[0007] Step 1: Decompose the DC microgrid system into multiple subsystems according to the DC microgrid topology and determine the input and output of each subsystem;

[0008] Step 2: Use the input and output of each subsystem to train the long short-term memory network and extract the key features of each subsystem;

[0009] Step 3: The key features of the subsystems are concatenated and input into an attention mechanism network. The input of the DC microgrid system is also input into another attention mechanism network to jointly construct a physical structured neural network model. The output of the physical structured neural network model is subjected to matrix operations to obtain the predicted value of the DC microgrid system state quantity.

[0010] Step 4: When a new load is connected to the DC microgrid system, the new load is divided into a new subsystem through step 1, and then the characteristics of the new subsystem are extracted separately through step 2, and then the process goes to step 3.

[0011] In step 1, the DC microgrid topology is that the power supply and load of the DC microgrid are connected to the same bus through converters and cables. Based on this topology, the DC microgrid system is divided into several maximum fixed topology units with relatively fixed and indivisible topologies as subsystems.

[0012] In step 2, the input and output of each subsystem are used to train the corresponding long short-term memory network, and the output before the linear layer of the long short-term memory network is used as the key feature of the subsystem.

[0013] In step 3, the matrix operation process of the two attention mechanism network outputs is:

[0014]

[0015] Among them, x pred is the predicted value of the state quantity, u is the input quantity of the DC microgrid system, and output u The output of the attention mechanism network that takes the input of the DC microgrid system as input, output feature is the output of the attention mechanism network that takes the key features of the subsystem as input.

[0016] In step 4, when a new load is connected to the DC microgrid system, the system is connected to the bus as a maximum fixed topology unit, and the new part is divided into a new subsystem. For the DC microgrid system, the atomic system structure has not changed, only a new subsystem is added. Since a new subsystem is added to the topology structure of the DC microgrid system, a new subnetwork is added to the neural network. It is necessary to train a small long short-term memory network for the new subsystem, extract its key features, and splice the key features with the feature matrix before the new load is connected, and share the network weights corresponding to the subsystems with the same load properties. For the DC microgrid system, only one linear layer in the attention mechanism network needs to be trained to obtain the data-driven network model corresponding to the new load after the new load is connected, and state estimation is performed through the data-driven network model.

[0017] A DC microgrid system modeling system based on physical structured neural network, comprising:

[0018] DC microgrid system division module: decomposes the DC microgrid system into multiple subsystems according to the DC microgrid topology and determines the input and output of each subsystem;

[0019] Feature extraction module: Use the input and output of each subsystem to train the long short-term memory network and extract the key features of each subsystem;

[0020] Model construction module: The key features of the subsystem are input into an attention mechanism network through splicing, and the input quantity of the DC microgrid system is input into another attention mechanism network to jointly construct a physical structured neural network model. The output of the physical structured neural network model is subjected to matrix operation to obtain the predicted value of the DC microgrid system state quantity; when a new load is connected to the DC microgrid system, the new load is divided into a new subsystem through the DC microgrid system division module, and then the new subsystem features are separately extracted through the feature extraction module and transferred to the model construction module.

[0021] A DC microgrid system modeling device based on a physical structured neural network, comprising:

[0022] Memory: used for storing a computer program for implementing a DC microgrid system modeling method based on a physical structured neural network;

[0023] Processor: used to implement a DC microgrid system modeling method based on a physical structured neural network when executing the computer program.

[0024] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a direct current microgrid system modeling method based on a physical structured neural network.

[0025] A computer program product includes a computer program, which, when executed by a processor, implements a direct current microgrid system modeling method based on a physical structured neural network.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] 1. Simple model and small training volume: The present invention uses a long short-term memory network to extract the key features of each subsystem. Whenever the system is connected to a new load, only a small long short-term memory network and a linear layer of the attention mechanism network need to be trained offline for the subsystem, effectively reducing the number of network training parameters and the demand for system operation data.

[0028] 2. Strong applicability: The present invention constructs a physical structured neural network based on the actual topology of the DC microgrid, which is flexibly applicable to the variable topology modeling of the DC microgrid system, and provides a new adaptive modeling method for the DC microgrid system with constantly changing operating parameters and topological structures in practice.

[0029] 3. Good accuracy: Compared with traditional data-driven methods, this method adopts a physical structured neural network, uses a long short-term memory network to extract subsystem features, and integrates the model through different attention mechanism networks to ensure data prediction accuracy while adapting to changes in the DC microgrid topology.

[0030] In summary, the present invention can achieve adaptive rapid modeling with small parameters under operating conditions where the DC microgrid topology changes dynamically, ensure the modeling accuracy of the DC microgrid system, and endow the DC microgrid data-driven model with physical interpretability and scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a structural diagram of the DC microgrid system of the present invention.

[0032] Figure 2 This is a structural diagram of the physical structured neural network model of the present invention.

[0033] Figure 3(a) is a fitting curve diagram of the traditional attention mechanism model.

[0034] FIG3( b ) is a fitting curve diagram of the physical structured neural network model of the present invention. DETAILED DESCRIPTION

[0035] The present invention will be described in detail below with reference to the accompanying drawings.

[0036] A DC microgrid system modeling method based on a physical structured neural network includes the following steps:

[0037] Step 1: Decompose the DC microgrid system into multiple subsystems according to the DC microgrid topology and determine the input and output of each subsystem;

[0038] Step 2: Use the input and output of each subsystem to train the long short-term memory network and extract the key features of each subsystem;

[0039] Step 3: The key features of the subsystems are concatenated and input into an attention mechanism network. The input of the DC microgrid system is also input into another attention mechanism network to jointly construct a physical structured neural network model. The output of the physical structured neural network model is subjected to matrix operations to obtain the predicted value of the DC microgrid system state quantity.

[0040] Step 4: When a new load is connected to the DC microgrid system, the new load is divided into a new subsystem through step 1, and then the characteristics of the new subsystem are extracted separately through step 2, and then the process goes to step 3.

[0041] In step 1, if Figure 1As shown in the figure, the DC microgrid topology is a DC microgrid where both the power source and load are connected to the same busbar via converters and cables. Based on this topology, the DC microgrid system is divided into several subsystems with relatively fixed and indivisible maximum fixed topology units. In a DC microgrid, the topological connection relationship between two directly connected cables and the internal components of the converter remains constant, and the internal structure does not change when a new load is connected. Therefore, to handle dynamic topological changes in the DC microgrid, the converter and cable directly connected to the power source or load are divided into a subsystem.

[0042] In step 2, the input and output of each subsystem are used to train the corresponding long short-term memory network, and the output before the linear layer of the long short-term memory network is used as the temporal feature of the subsystem.

[0043] In step 3, all key features extracted in step 2 are concatenated into a feature matrix, which is used as the input of the attention mechanism network. The input u of the DC microgrid system shown in formula (1) is used as the input of another attention mechanism network, as shown in Figure 2 shown.

[0044] u=[v in i L d bo d bu1 d bu2 v bus ] T (1)

[0045] Where u is the input of the DC microgrid system, v in is the input voltage of the DC microgrid system, i L is the load current of the battery load in the DC microgrid system, d bo d bu1 d bu2 are the duty cycles of boost, buck1, and buck 2 converters, respectively, and v bus is the bus voltage of the DC microgrid system.

[0046] The outputs of the two attention mechanism networks are finally transformed through matrix calculation to obtain the transformation matrix, which is applied to the DC microgrid system input. The matrix operation process of the outputs of the two attention mechanism models is shown in formula (2):

[0047]

[0048] Among them, x pred is the predicted value of the state quantity, u is the input quantity of the DC microgrid system, and output u The output of the attention mechanism network that takes the input of the DC microgrid system as input, outputfeature is the output of the attention mechanism network that takes the key features of the subsystem as input.

[0049] The final DC microgrid system state x is generated as shown in formula (3):

[0050] x=[i L,bo v C,bo i L,bu1 v C,bu1 v in,bu1 i L,bu2 v C,bu2 v in,bu2 i 1a i 1b i 2a i 2b i 3a i 3b ] T (3)

[0051] Among them, x is the system state quantity, i L,bo is the inductor current of the boost converter, v C,bo is the capacitor voltage of the boost converter, i L,bu1 is the inductor current of buck 1 converter, v C,bu1 is the filter capacitor voltage of buck 1 converter, v in,bu1 is the input capacitor voltage of the buck 1 converter, i L,bu2 is the inductor current of the buck2 converter, v C,bu2 is the filter capacitor voltage of the buck 2 converter, v in,bu2 is the input capacitor voltage of the buck2 converter, i 1a is the inductor current at the left end of cable 1, i 1b is the inductor current at the right end of cable 1, i 2a is the inductor current at the left end of cable 2, i 2b is the inductor current at the right end of cable 2, i 3a is the inductor current at the left end of cable 3, i 3b is the inductor current at the right end of cable 3.

[0052] In step 4, the scalability of the DC microgrid system is achieved by treating the new load as an independent subsystem. Figure 1As shown in the figure, when a new load Buck 3 is connected to the DC microgrid system, the system can be regarded as a maximum fixed topology unit connected to the busbar. Therefore, the new part is divided into a new subsystem. For the DC microgrid system, the atomic system structure does not change, only a new subsystem is added. Since the DC microgrid system topology has added the load Buck 3, the neural network corresponds to the long short-term memory network of the new load Buck 3. It is necessary to train the long short-term memory network of the load Buck 3 offline, extract its time series features, and splice the time series features with the feature matrix before the new load is connected, as shown in the figure. Figure 2 As shown in Figure 2. Since both Buck 3 and Buck 1 are DC loads, they can directly share the weight matrix of Buck 1. This model adapts to the changed topology with minimal new data by fine-tuning a linear layer in the attention module, thereby making accurate predictions under the modified system structure. The input and state of the new system are shown in Equations (4) and (5).

[0053] u=[v in i L d bo d bu1 d bu2 d bu3 v bus ] T (4)

[0054] x=[i L,bo v C,bo i L,bu1 v C,bu1 v in,bu1 i L,bu2 v C,bu2 v in,bu2 i L,bu3 v C,bu3 v in,bu3 i 1a i 1b i 2a i 2b i 3a i 3b i 4a i 4b ] T (5)

[0055] Where u is the input of the DC microgrid system, v in is the input voltage of the DC microgrid system, i L is the load current of the battery load in the DC microgrid system, d bo d bu1 d bu2 d bu3are the duty cycles of converter boost, buck 1, buck 2, and buck 3, respectively, and v bus is the bus voltage of the DC microgrid system, x is the system state quantity, i L,bo is the inductor current of the boost converter, v C,bo is the capacitor voltage of the boost converter, i L,bu1 is the inductor current of buck 1 converter, v C,bu1 is the filter capacitor voltage of buck1 converter, v in,bu1 is the input capacitor voltage of the buck 1 converter, i L,bu2 is the inductor current of the buck 2 converter, v C,bu2 is the filter capacitor voltage of the buck 2 converter, v in,bu2 is the input capacitor voltage of the buck 2 converter, i L,bu3 is the inductor current of the buck 3 converter, v C,bu3 is the filter capacitor voltage of the buck 3 converter, v in,bu3 is the input capacitor voltage of the buck3 converter, i 1a is the inductor current at the left end of cable 1, i 1b is the inductor current at the right end of cable 1, i 2a is the inductor current at the left end of cable 2, i 2b is the inductor current at the right end of cable 2, i 3a is the inductor current at the left end of cable 3, i 3b is the inductor current at the right end of cable 3, i 4a is the inductor current at the left end of cable 4, i 4b is the inductor current at the right end of cable 4.

[0056] A DC microgrid system modeling system based on physical structured neural network, comprising:

[0057] A DC microgrid system partitioning module is configured to decompose the DC microgrid system into multiple subsystems according to the DC microgrid topology, and determine the input and output of each subsystem to implement step 1 of the modeling method of the present invention.

[0058] Feature extraction module: using the input and output of each subsystem to train a long short-term memory network and extract key features of each subsystem for implementing step 2 of the modeling method of the present invention;

[0059] Model construction module: The key features of the subsystem are input into an attention mechanism network through splicing, and the input quantity of the DC microgrid system is input into another attention mechanism network to jointly construct a physical structured neural network model. The output of the physical structured neural network model is subjected to matrix operation to obtain the predicted value of the DC microgrid system state quantity; when a new load is connected to the DC microgrid system, the new load is divided into a new subsystem through the DC microgrid system division module, and then the new subsystem features are separately extracted through the feature extraction module and transferred to the model construction module to implement steps 3 and 4 of the modeling method described in the present invention.

[0060] A DC microgrid system modeling device based on a physical structured neural network, comprising:

[0061] Memory: used for storing a computer program for implementing a DC microgrid system modeling method based on a physical structured neural network;

[0062] Processor: used to implement a DC microgrid system modeling method based on a physical structured neural network when executing the computer program.

[0063] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a direct current microgrid system modeling method based on a physical structured neural network.

[0064] A computer program product includes a computer program, which, when executed by a processor, implements a direct current microgrid system modeling method based on a physical structured neural network.

[0065] Comparative experiment:

[0066] First, use Figure 1 The operational data of the Zhongyuan system is used to train the physically structured DC microgrid data-driven model, and then the new DC load operational data is used to extract Figure 1 The characteristics of the Buck 3 subsystem are analyzed to verify the applicability of the proposed modeling method under changes in the system topology. The system simulation time is 0.02s, and the new battery load is connected to Buck 2 within 0.01s.

[0067] In order to evaluate the impact of the proposed modeling method on the model accuracy, this study used simulated data to train the traditional attention model and the proposed method respectively, with the Boost inductor current i L,boFor example, the comparison between the prediction curve of the traditional attention mechanism model and the actual operation curve is shown in Figure 3(a), and the comparison between the prediction curve of the physical structured neural network model described in the present invention and the actual operation curve is shown in Figure 3(b). It can be seen that whether before or after the 0.01s load change, the model constructed by the method proposed in the present invention can accurately predict the change of state variables, and its fitting ability is not reduced compared to the traditional attention mechanism model. Therefore, the model constructed by the method proposed in the present invention has the ability to accurately predict the dynamic changes of the system.

[0068] Table 1 shows the relative fitting error between the model constructed using the proposed method and the traditional attention model using mean squared error. As can be seen, the error for both models is less than 2%. This demonstrates that the proposed modeling approach has superior accuracy compared to traditional data-driven models. These results demonstrate the robustness and adaptability of the proposed modeling approach, which can accurately capture the dynamic changes in the DC microgrid state variables while adapting to changes in the DC microgrid system topology.

[0069] Table 1 Comparison of modeling errors of different networks

[0070]

[0071]

Claims

1. A DC microgrid system modeling method based on physical structured neural network, characterized in that: The steps include: Step 1: Decompose the DC microgrid system into multiple subsystems according to the DC microgrid topology and determine the input and output of each subsystem; Step 2: Use the input and output of each subsystem to train the long short-term memory network and extract the key features of each subsystem; Step 3: The key features of the subsystems are concatenated and input into an attention mechanism network. The input of the DC microgrid system is also input into another attention mechanism network to jointly construct a physical structured neural network model. The output of the physical structured neural network model is subjected to matrix operations to obtain the predicted value of the DC microgrid system state quantity. Step 4: When a new load is connected to the DC microgrid system, the new load is divided into a new subsystem through step 1, and then the characteristics of the new subsystem are extracted separately through step 2, and then the process goes to step 3.

2. A DC microgrid system modeling method based on physical structured neural network according to claim 1, characterized in that: In step 1, the DC microgrid topology is that the power supply and load of the DC microgrid are connected to the same bus through converters and cables. Based on this topology, the DC microgrid system is divided into several maximum fixed topology units with relatively fixed and indivisible topologies as subsystems.

3. A DC microgrid system modeling method based on physical structured neural network according to claim 1, characterized in that: In step 2, the input and output of each subsystem are used to train the corresponding long short-term memory network, and the output before the linear layer of the long short-term memory network is used as the key feature of the subsystem.

4. A DC microgrid system modeling method based on physical structured neural network according to claim 1, characterized in that: In step 3, the matrix operation process of the two attention mechanism network outputs is: Among them, x pred is the predicted value of the state quantity, u is the input quantity of the DC microgrid system, and output u The output of the attention mechanism network that takes the input of the DC microgrid system as input, output feature is the output of the attention mechanism network that takes the key features of the subsystem as input.

5. A DC microgrid system modeling method based on physical structured neural network according to claim 1, characterized in that: In step 4, when a new load is connected to the DC microgrid system, the system is connected to the bus as a maximum fixed topology unit, and the newly added part is divided into a new subsystem. For the DC microgrid system, the atomic system structure remains unchanged, only a new subsystem is added, a small long short-term memory network is trained for the new subsystem, its key features are extracted, the key features are concatenated with the feature matrix before the new load is connected, and the network weights corresponding to the subsystems with the same load properties are shared. A linear layer in the attention mechanism network is trained for the DC microgrid system to obtain a data-driven network model corresponding to the new load after connection, and state estimation is performed through the data-driven network model.

6. A DC microgrid system modeling system based on a physical structured neural network and the modeling method according to any one of claims 1 to 5, characterized in that: include: DC microgrid system division module: decomposes the DC microgrid system into multiple subsystems according to the DC microgrid topology and determines the input and output of each subsystem; Feature extraction module: Use the input and output of each subsystem to train the long short-term memory network and extract the key features of each subsystem; Model construction module: The key features of the subsystem are input into an attention mechanism network through splicing, and the input quantity of the DC microgrid system is input into another attention mechanism network to jointly construct a physical structured neural network model. The output of the physical structured neural network model is subjected to matrix operation to obtain the predicted value of the DC microgrid system state quantity; when a new load is connected to the DC microgrid system, the new load is divided into a new subsystem through the DC microgrid system division module, and then the new subsystem features are separately extracted through the feature extraction module and transferred to the model construction module.

7. A DC microgrid system modeling device based on physical structured neural network, characterized in that: include: Memory: used to store a computer program for implementing the DC microgrid system modeling method based on a physical structured neural network according to any one of claims 1 to 5; Processor: configured to implement the DC microgrid system modeling method based on physical structured neural network according to any one of claims 1 to 5 when executing the computer program.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the DC microgrid system modeling method based on a physical structured neural network are implemented as described in any one of claims 1 to 5.

9. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the DC microgrid system modeling method based on a physical structured neural network according to any one of claims 1 to 5 is implemented.

Images