Power system component dynamic modeling method based on MMOE-DAE model

Abstract

The invention discloses a power system component dynamic modeling method based on an MMOE-DAE model, and belongs to the technical field of electrical digital data processing, high precision and three-phase unified modeling of a power system dynamic component: the MMOE-DAE can realize accurate modeling of a complex high-order dynamic component in a power system in the absence of a detailed physical model. By integrating the neural ODE and the automatic order lifting mechanism, the modeling ability of the dynamic behaviors of the complex store assembly in various disturbance scenes is effectively improved, and it is ensured that the method still has high precision and robustness under the conditions of strong nonlinearity and high-order dynamic. The unified modeling of the three-phase circuit can process the dynamic characteristics of each phase circuit at the same time under the same framework, and processes the coupling relationship between phases through a multi-task learning mechanism, thereby significantly reducing the error caused by phase separation processing in the traditional modeling method, and improving the overall simulation efficiency.

Description

Technical Field

[0001] The invention belongs to the field of electrical digital data processing, and in particular relates to a method for dynamic modeling of power system components based on a MMOE-DAE model. Background Art

[0002] Due to the integration of large-scale power electronic devices, power systems present complex dynamics, and there is an urgent need to develop highly reliable dynamic equivalent models using limited component measurement data. Electromagnetic transient simulation is based on the state of the system instantaneous value and can describe microsecond processes involving switching devices. It is essential for accurate stability analysis and control of modern power systems. Traditional electromagnetic transient simulation relies on detailed physical models of components (such as differential equations, algebraic equations, etc.), which usually require accurate physical parameters and operating conditions. However, in actual power systems, due to technical confidentiality, measurement limitations or component complexity, the physical models of some components may not be fully available, forming so-called "black box" or "grey box" models, making it difficult for traditional modeling methods to provide reliable dynamic predictions. Combining artificial intelligence technology, especially deep learning, with electromagnetic transient simulation is a hot research point at present, and this technology still has many bottlenecks and challenges. First, the system measurement equipment can only capture low-dimensional discrete data samples with limited frequency, which causes the traditional neural network model to perform well under known conditions, but has poor dynamic performance under unknown operating conditions; second, it is not clear whether the data-driven model can make full use of historical data and adaptively adjust parameters according to the operating conditions of the system to improve the robustness and reliability of the simulation model; in addition, although some studies have combined the physical characteristics of power systems with deep learning, the ability of existing models to handle high-order differential dynamics and adapt to diverse scenarios still needs to be improved; finally, despite significant progress in data-driven modeling, it is still rare to integrate deep learning models into unified simulations of electromagnetic transient simulations in commercial operations.

[0003] With the large-scale integration of renewable energy and power electronic loads, large power systems show increasingly complex nonlinear dynamic and coupling characteristics, and the requirements for accurate stability analysis and control of modern power systems are also increasing. However, in the absence of prior knowledge, due to technical protection and measurement limitations, system access equipment usually has black box or gray box models, in which the unmodeled components greatly weaken the reliability of operators' system operation decisions. In addition, even with a comprehensive white box model, it is still a major challenge to achieve high-order dynamic response in large-scale systems within acceptable time and computing resource constraints. Therefore, it is necessary to dynamically model the black box components and use electromagnetic transient simulation for calculation and analysis.

[0004] Existing dynamic modeling technologies for electromagnetic transient (EMT) simulation have many problems in dealing with the dynamic behavior of complex components in modern power systems. First, there is a lack of physical models. Due to technical confidentiality, measurement limitations, and the complexity of system components, many power system components lack detailed internal models (i.e., "black box" or "gray box" models), making it difficult for traditional electromagnetic transient simulation to provide accurate dynamic predictions without a complete physical description, reducing the simulation accuracy and robustness of the system. Second, although deep learning technology has broad application prospects in simulation, it is highly dependent on high-quality measurement data. However, measurement equipment in power systems can usually only collect low-dimensional discrete data samples with limited frequency, making it difficult for deep learning models developed from these data to generate reliable, continuous, and accurate dynamic equivalent models. Third, there is a lack of adaptive capabilities. Although existing deep learning models perform well under known training data sets, their dynamic characteristics cannot be adaptively adjusted when faced with unseen operating conditions or faults, resulting in poor simulation performance in complex and changing scenarios. Fourth, the traditional deep learning model has limited ability to capture high-order dynamic characteristics. It is difficult to maintain high dynamic tracking accuracy when dealing with strong coupling and nonlinear disturbances in the system. In particular, the model is unstable when dealing with strong disturbances in large-scale power systems. Fifth, it is very rare to integrate deep learning models into existing commercial electromagnetic transient simulation platforms for unified simulation. This is mainly due to the lack of effective data interfaces in existing technologies, which makes it difficult to achieve seamless connection between deep learning models and traditional simulation platforms, thus limiting the application of these technologies in actual engineering.

[0005] Therefore, at this stage, it is necessary to design a dynamic modeling method, system and storage medium for power system components based on the MMOE-DAE model to solve the above problems. Summary of the invention

[0006] The purpose of the present invention is to provide a method, system and storage medium for dynamic modeling of power system components based on the MMOE-DAE model, so as to solve the technical problems existing in the above-mentioned prior art.

[0007] To achieve the above object, the technical solution of the present invention is:

[0008] The dynamic modeling method of power system components based on the MMOE-DAE model includes the following steps:

[0009] S1: A multi-gate hybrid differential algebraic equation expert network MMOE-DAE is proposed. Multiple gated neural differential algebraic equation DAEs are used to simulate the dynamic characteristics of three-phase circuits and establish the relationship between phases.

[0010] S2: An automatic order raising module is proposed, which uses multiple independent continuous numerical integration channels and related state variable storage to automatically infer the optimal order for dynamic component modeling through a fully gated network;

[0011] S3: MMOE-DAE accurately simulates the dynamic relationship between different phases in the same model, realizes unified modeling of three-phase circuits, and makes up for the shortcomings of traditional models in handling coupling relationships between phases; MMOE-DAE is seamlessly integrated into the electromagnetic transient simulation platform, realizing real-time simulation and dynamic analysis in actual industrial application scenarios, thereby improving the practicality and reliability of electromagnetic transient simulation in power systems.

[0012] Furthermore, in step S1:

[0013] The architecture of the MMOE-DAE network for modeling three-phase circuits with dynamic components is divided into three parts: the input part, the underlying model, the tower model, and the gated network module;

[0014] The input part receives the historical current signal i of each phase circuit. A (t),i B (t),i C (t), historical voltage signal u of each phase circuit A (t),u B (t),u C (t), and external signals; the above signals are transmitted to the underlying neural ODE expert module through the gating network to predict the current and voltage changes in the next time step;

[0015] The underlying model includes multiple neural ODE expert modules, each of which is responsible for processing different dynamic characteristics and determining which modules participate in the current modeling task through the scheduling of the gating network; the neural ODE expert module is used to approximate the differential equations of the dynamic components and model the relationship between the current, voltage and time of each phase circuit; the input signal i A (t),i B (t),i C (t) After the neural ODE, the current i of each phase circuit in the next time step is obtained A (t+1),i B (t+1),i C (t+1);

[0016] The gated network module is responsible for the selection and scheduling of the neural ODE expert module. It selects the neural ODE expert suitable for the current task according to the characteristics of the input signal, and assigns weights through the softmax function to schedule the neural ODE expert to participate in modeling. Its output will be transmitted to the corresponding DAE tower according to the situation of each phase circuit.

[0017] The tower model is divided into multiple DAE towers, each tower is used to process the corresponding phase current in the three-phase circuit, that is, i A (t+1),i B (t+1),i C (t+1); Based on the DAE tower, the MMOE-DAE model can handle the multi-task modeling of three-phase circuits simultaneously and can establish the relationship between the currents of different phases.

[0018] Furthermore, in step S2, the automatic order enhancement module is specifically as follows:

[0019] Through multiple numerical integration channels, the optimal dynamic order of components is automatically inferred, allowing neural ODE experts to handle high-order dynamic behaviors; when facing components that lack detailed physical information, MMOE-DAE can still dynamically adjust the modeling order according to the data to ensure modeling accuracy;

[0020] The automatic order raising module processes dynamic behaviors of different orders through multiple independent integral channels and automatically selects the optimal output through the gating network and the backward Euler method;

[0021] The multi-channel integration module has multiple independent integration channels, each channel is used for derivative processing of different orders; each channel will perform numerical integration based on the previous state to calculate the derivative value of the corresponding order; the first-order channel calculates the rate of change of the system state, that is, the derivative, to predict the change of the system state in the next time step; the second-order channel processes the first-order derivative and the second-order derivative, and calculates the acceleration or higher-order changes of the system through the transformation of the state space; the third-order channel processes the rate of change of higher orders, which is used to capture the high-order dynamic behavior in complex systems;

[0022] A state storage unit is also introduced in the automatic order raising module to store the state information calculated for each channel, ensuring that when performing multiple integrations, the system can access the previous state and derivative values ​​to ensure the continuity and accuracy of each channel;

[0023] The gating network is used to select the best order output from the outputs of multiple integration channels; through a trainable weight matrix W go , use the softmax function to determine which channel's output is most appropriate;

[0024] Among them, the principle of the multi-channel backward Euler upgrade module with dynamic component order identification is:

[0025] The input initial state h0, first-order derivative h0′ and second-order derivative h0″ are upgraded by backward Euler method to obtain the first-order, second-order and third-order prediction values; the backward Euler method is used to perform numerical integration of each order derivative to deduce the state change of each time step, and finally the state of the next time step is obtained.

[0026] Furthermore, step S3 is specifically as follows:

[0027] Electromagnetic transient simulation replaces each component with a Norton equivalent circuit and constructs network equations based on the equivalent circuit; avoids analyzing all DAEs of the system as discrete entities;

[0028] The component simulation interface includes the following four steps:

[0029] (1) Calculate the state variable x(t+Δt)

[0030] Measurements based on component ports and To infer the state variable x(t+Δt); use the trapezoidal integration method to calculate x(t+Δt):

[0031]

[0032] Among them, the variable x(t) is an implicit function of the port measurement value, which is constructed by ODE expert neurons, that is,

[0033] (2) Derivation of the Norton equivalent current i(t+Δt) of the component

[0034] Based on the formula

[0035]

[0036] Where i is the component injection current, g(·) is the algebraic equation describing the component current; the port current of the dynamic component is calculated using the AE tower module

[0037]

[0038] Steps (1) and (2) are completely executed within the MMOE-DAE network, and their functions are consistent with the external characteristics of general components;

[0039] (3) Calculate the voltage vector V(t+Δt): According to the node connection relationship, the current I of the data-driven model component is converted into eql (t+Δt) and the current I of conventional components gen (t+Δt) is combined to obtain

[0040] I(t+Δt)=[I eql (t+Δt)+I gen (t+Δt)]

[0041] Substituting into the formula GV=I, we can get

[0042] GV(t+Δt)=I(t+Δt)

[0043] (4) Calculate the branch current I ij (t+Δt): To complete the historical current iteration process of conventional components, it is necessary to calculate I ij (t+Δt), the current passes through V(t+Δt) and the corresponding admittance G ij Calculated

[0044] I ij (t+Δt)=G ij (V i (t+Δt)-V j (t+Δt)).

[0045] A power system component dynamic modeling system based on the MMOE-DAE model adopts the power system component dynamic modeling method based on the MMOE-DAE model as mentioned above to perform power system component dynamic modeling.

[0046] A storage medium having a computer program stored thereon, wherein the computer program, when executed, executes the above-mentioned method for dynamic modeling of power system components based on the MMOE-DAE model.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] High-precision and three-phase unified modeling of power system dynamic components: MMOE-DAE can achieve accurate modeling of complex high-order dynamic components in power systems in the absence of detailed physical models. By integrating neural ODE and automatic order-raising mechanism, the modeling capability of the dynamic behavior of complex components under various disturbance scenarios is effectively improved, ensuring high accuracy and robustness under strong nonlinear and high-order dynamic conditions. The unified modeling of three-phase circuits can simultaneously process the dynamic characteristics of each phase circuit under the same framework, and handle the coupling relationship between phases through a multi-task learning mechanism, which significantly reduces the errors caused by phase separation processing in traditional modeling methods and improves the overall simulation efficiency.

[0049] The GNN model is used to aggregate information between adjacent nodes and combine system topology information to achieve better performance.

[0050] Comparing the model parameters and performance when using multidimensional data matrix and oscillation current time series data as model input, the results show that using multidimensional data matrix as model input not only has better timeliness but also has better positioning performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Architecture diagram for modeling a three-phase circuit with dynamic components for an MMOE-DAE network.

[0052] Figure 2 Schematic diagram of the automatic upgrade mechanism. DETAILED DESCRIPTION

[0053] A dynamic modeling method for power system components based on MMOE-DAE model is proposed.

[0054] First, MMOE-DAE is a data-driven multi-expert model that uses measurement data (such as voltage and current) at the system ports to infer the dynamic behavior of components. Although a complete physical model is lacking, the input-output relationship of the component can be captured through the measurement signal, reflecting the dynamic characteristics of the component under different operating conditions. Therefore, even if the physical model is incomplete, MMOE-DAE can still use the measurement data to build an equivalent dynamic model and achieve high-precision modeling of the dynamic characteristics of the component.

[0055] Secondly, the dynamic behavior of components in modern power systems is usually expressed as high-order nonlinear differential equations. This patent introduces an automatic order improvement module that can automatically identify the dynamic order of components and thus enhance the simulation capability of high-order dynamic behavior of components.

[0056] Third, MMOE-DAE can accurately simulate the dynamic relationship between different phases in the same model, realize the unified modeling of three-phase circuits, and make up for the shortcomings of traditional models in handling the coupling relationship between phases.

[0057] Fourthly, MMOE-DAE can be seamlessly integrated into the existing electromagnetic transient simulation platform CloudPSS, ensuring efficient real-time simulation and dynamic analysis in actual industrial application scenarios, thereby improving the practicality and reliability of electromagnetic transient simulation in power systems.

[0058] Architecture of MMOE-DAE network for modeling three-phase circuits with dynamic components:

[0059] The MMOE-DAE network can effectively learn the dynamic characteristics of components by combining multiple neural differential equation (Neural-ODE) expert modules. The neural ODE module regards the state of the system (such as current, voltage, etc.) as a function of time, and approximates the differential changes of these states over time by fitting the change trends of these states (i.e., approximating the traditional differential equations). This is equivalent to learning a differential equation that is similar to the real dynamics of the system, and can approximate complex nonlinear system behaviors with high accuracy. Therefore, even without detailed physical equations, the input-output relationship of the component can still be learned from the data to establish a dynamic equivalent model. When the historical current and voltage signals of each phase circuit enter, the MMOE-DAE will automatically select the most suitable neural ODE expert module to handle the current modeling task through the gating network, that is, each expert module is assigned a special task, focusing on a certain type of dynamic characteristics, to ensure that the model can perform well in different scenarios. Therefore, the MMOE-DAE can not only simulate a single component, but also process different components or different dynamic characteristics at the same time through multiple expert modules, forming a unified multi-task learning framework. In this way, even if some components do not have complete physical models, the network can learn the complex dynamic behaviors of components from limited port data by sharing the underlying neural network and collaboration between different tasks.

[0060] The architecture of MMOE-DAE network for dynamic component three-phase circuit modeling is as follows Figure 1 As shown:

[0061] The architecture is mainly divided into three parts: input part, underlying model (including neural ODE expert module), tower model (DAE tower) and gate network module. The input part receives the historical current signal i of each phase circuit. A (t),i B (t),i C (t), historical voltage signal u of each phase circuit A (t),u B (t),u C (t), and external signals. These input signals are transmitted to the underlying neural ODE expert module through the gating network to predict the current and voltage changes in the next time step. The underlying model consists of multiple neural ODE expert modules, each of which is responsible for processing different dynamic characteristics and determining which modules participate in the current modeling task through the scheduling of the gating network. The neural ODE module is used to approximate the differential equations of the dynamic components and model the relationship between the current, voltage and time of each phase circuit. The input signal i A (t),i B (t),i C (t) After the neural ODE, the current i of each phase circuit in the next time step is obtained A (t+1),iB (t+1),i C (t+1). The gating network is responsible for expert module selection and scheduling. It selects the neural ODE experts suitable for the current task according to the characteristics of the input signal, and assigns weights through the softmax function to schedule these experts to participate in modeling. Its output will be transmitted to the corresponding DAE tower according to the situation of each phase circuit. The tower model is divided into multiple DAE towers, each of which is used to process the corresponding phase current in the three-phase circuit, that is, i A (t+1),i B (t+1),i C (t+1). Through these DAE towers, the MMOE-DAE model is able to handle the multi-task modeling of three-phase circuits simultaneously and is able to establish the mutual relationship between different phase currents.

[0062] Automatic upgrade mechanism:

[0063] High-order dynamic behaviors (such as high-frequency oscillations, nonlinear perturbations, etc.) are usually difficult to capture for traditional neural network models. The automatic order improvement mechanism introduced in this patent can automatically infer the optimal dynamic order of the component through multiple numerical integration channels, thereby enabling neural ODE experts to handle high-order dynamic behaviors. This mechanism enables MMOE-DAE to dynamically adjust the modeling order according to the data when facing complex components that lack detailed physical information, ensuring high-precision modeling. Specifically, the module processes dynamic behaviors of different orders through multiple independent integration channels, and automatically selects the optimal output through a gating network and the backward Euler method, greatly improving the modeling accuracy in high-order dynamic systems, and is particularly suitable for processing complex, nonlinear and multi-path coupled systems.

[0064] The multi-channel integration module has multiple independent integration channels, each of which is used for derivative processing of different orders. Each channel will perform numerical integration based on the previous state to calculate the derivative value of the corresponding order. The first-order channel calculates the rate of change of the system state, that is, the derivative, which is used to predict the change of the system state in the next time step; the second-order channel processes the first-order derivative and the second-order derivative, and calculates the acceleration or higher-order changes of the system through the transformation of the state space; the third-order channel processes higher-order rates of change to capture high-order dynamic behaviors in complex systems. The module also introduces a state storage unit to store the state information calculated by each channel, ensuring that the system can access the previous state and derivative values ​​when performing multiple integrations, ensuring the continuity and accuracy of each channel.

[0065] The gating network is used to select the best order output from the outputs of multiple integration channels. Through a trainable weight matrix W go , use the softmax function to determine which channel's output is most appropriate.

[0066] The principle of the multi-channel backward Euler upgrade module with dynamic component order identification is as follows Figure 2 :

[0067] The input initial state h0, the first-order derivative h0′ and the second-order derivative h0″ are subjected to backward Euler order upgrading to obtain the first-order, second-order and third-order prediction values. Specifically, the backward Euler method performs numerical integration of derivatives of various orders to deduce the state change at each time step and finally obtain the state of the next time step.

[0068] Dynamic simulation interface:

[0069] Electromagnetic transient simulation replaces each component with a Norton equivalent circuit and constructs network equations based on these equivalent circuits. This approach avoids analyzing all DAEs of the system as discrete entities, thereby improving modeling efficiency. The component simulation interface includes the following four steps:

[0070] (1) Calculate the state variable x(t+Δt)

[0071] Measurements based on component ports and To infer the state variable x(t+Δt). Use the trapezoidal integration method to calculate x(t+Δt):

[0072]

[0073] Among them, the variable x(t) is an implicit function of the port measurement value, which is constructed by ODE expert neurons, that is,

[0074] (2) Derivation of the Norton equivalent current i(t+Δt) of the component

[0075] Based on the formula

[0076]

[0077] Where i is the component injection current and g(·) is the algebraic equation describing the component current. Use the AE Tower module to calculate the port current of a dynamic component

[0078]

[0079] Steps (1) and (2) are completely executed within the MMOE-DAE network, and their functions are consistent with the external characteristics of general components.

[0080] (3) Calculate the voltage vector V(t+Δt): According to the node connection relationship, the current I of the data-driven model component is converted into eql (t+Δt) and the current I of conventional components gen (t+Δt) is combined to obtain

[0081] I(t+Δt)=[I eql (t+Δt)+I gen (t+Δt)]

[0082] Substituting into the formula GV=I, we can get

[0083] GV(t+Δt)=I(t+Δt)

[0084] (4) Calculate the branch current I ij (t+Δt): In order to complete the historical current iteration process of conventional components, it is necessary to calculate I ij (t+Δt), the current passes through V(t+Δt) and the corresponding admittance G ij Calculated

[0085] I ij (t+Δt)=G ij (V i (t+Δt)-V j (t+Δt)).

[0086] Application links:

[0087] 1) Data preparation and processing

[0088] Data collection: Establish a simulation example, obtain the example topology, set the position disturbance source in the example, and collect the current timing operation data of each line during the system fault and the disturbance source location label.

[0089] Signal processing: Preprocess the collected time series signals, such as extracting discrete components, filtering, denoising, etc., to improve the accuracy and efficiency of model training. Then extract the preprocessed attenuated DC component and perform Taylor expansion, extract derivatives of various orders to construct the model input matrix.

[0090] (2) Model design and training

[0091] Model architecture design: Design the GNN model architecture and build a broadband oscillation disturbance source localization model.

[0092] Model training: Use the prepared dataset to train the model and adjust the model parameters to optimize performance.

[0093] Verification and testing: The model is verified and tested to evaluate its performance in locating the source of broadband oscillation disturbances.

[0094] (3) Online applications and real-time performance

[0095] Deploy model: Deploy the trained model to the actual power system monitoring platform.

[0096] Real-time performance optimization: Research and implement strategies to optimize the model’s disturbance source localization performance, including reducing latency, improving accuracy, and improving response speed.

[0097] Continuous learning: The implementation mechanism enables the model to be continuously updated and optimized based on newly collected data, improving the generalization of the model.

[0098] (4) Evaluation and iteration

[0099] Performance evaluation: Comprehensively evaluate the positioning performance of the model and establish a confusion matrix to evaluate the model performance.

[0100] Iterative optimization: Iteratively optimize the model based on the model performance evaluation results to improve its stability and reliability.

[0101] In summary: High-precision and three-phase unified modeling of dynamic components in power systems: MMOE-DAE can achieve accurate modeling of complex high-order dynamic components in power systems in the absence of detailed physical models. By integrating neural ODE and automatic order-raising mechanism, the modeling capability of the dynamic behavior of complex components under various disturbance scenarios is effectively improved, ensuring high accuracy and robustness under strong nonlinear and high-order dynamic conditions. The unified modeling of three-phase circuits can simultaneously process the dynamic characteristics of each phase circuit under the same framework, and process the coupling relationship between phases through a multi-task learning mechanism, which significantly reduces the errors caused by phase separation processing in traditional modeling methods and improves the overall simulation efficiency.

[0102] The GNN model is used to aggregate information between adjacent nodes and combine system topology information to achieve better performance.

[0103]

[0104] Comparing the model parameters and performance when using multidimensional data matrix and oscillation current time series data as model input, the results show that using multidimensional data matrix as model input not only has better timeliness but also has better positioning performance.

[0105] index Multidimensional Data Matrix Oscillation current timing data Parameter quantity 27378 711828 Inference speed 2.429ms 4.318ms Accuracy 97.2％ 96.8％

Claims

1. A dynamic modeling method for power system components based on the MMOE-DAE model, characterized in that: The following steps are involved: S1: A multi-gate hybrid differential algebraic equation expert network MMOE-DAE is proposed. Multiple gated neural differential algebraic equation DAEs are used to simulate the dynamic characteristics of three-phase circuits and establish the relationship between phases. S2: An automatic order raising module is proposed, which uses multiple independent continuous numerical integration channels and related state variable storage to automatically infer the optimal order for dynamic component modeling through a fully gated network; S3: MMOE-DAE accurately simulates the dynamic relationship between different phases in the same model, realizes unified modeling of three-phase circuits, and makes up for the shortcomings of traditional models in handling coupling relationships between phases; MMOE-DAE is seamlessly integrated into the electromagnetic transient simulation platform, realizing real-time simulation and dynamic analysis in actual industrial application scenarios, thereby improving the practicality and reliability of electromagnetic transient simulation in power systems.

2. The method for dynamic modeling of power system components based on the MMOE-DAE model according to claim 1, characterized in that: In step S1: The architecture of the MMOE-DAE network for modeling three-phase circuits with dynamic components is divided into three parts: the input part, the underlying model, the tower model, and the gated network module; The input part receives the historical current signal i of each phase circuit. A (t),i B (t),i C (t), historical voltage signal u of each phase circuit A (t),u B (t),u C (t), and external signals; the above signals are transmitted to the underlying neural ODE expert module through the gating network to predict the current and voltage changes in the next time step; The underlying model includes multiple neural ODE expert modules, each of which is responsible for processing different dynamic characteristics and determining which modules participate in the current modeling task through the scheduling of the gating network; the neural ODE expert module is used to approximate the differential equations of the dynamic components and model the relationship between the current, voltage and time of each phase circuit; the input signal i A (t),i B (t),i C (t) After the neural ODE, the current i of each phase circuit in the next time step is obtained A (t+1),i B (t+1),i C (t+1); The gated network module is responsible for the selection and scheduling of the neural ODE expert module. It selects the neural ODE expert suitable for the current task according to the characteristics of the input signal, and assigns weights through the softmax function to schedule the neural ODE expert to participate in modeling. Its output will be transmitted to the corresponding DAE tower according to the situation of each phase circuit; The tower model is divided into multiple DAE towers, each tower is used to process the corresponding phase current in the three-phase circuit, that is, i A (t+1),i B (t+1),i C (t+1); Based on the DAE tower, the MMOE-DAE model can handle the multi-task modeling of three-phase circuits simultaneously and can establish the relationship between the currents of different phases.

3. The method for dynamic modeling of power system components based on the MMOE-DAE model according to claim 2 is characterized in that: In step S2, the automatic order enhancement module is specifically as follows: Through multiple numerical integration channels, the optimal dynamic order of components is automatically inferred, allowing neural ODE experts to handle high-order dynamic behaviors; when facing components that lack detailed physical information, MMOE-DAE can still dynamically adjust the modeling order according to the data to ensure modeling accuracy; The automatic order raising module processes dynamic behaviors of different orders through multiple independent integral channels and automatically selects the optimal output through the gating network and the backward Euler method; The multi-channel integration module has multiple independent integration channels, each channel is used for derivative processing of different orders; each channel will perform numerical integration based on the previous state to calculate the derivative value of the corresponding order; the first-order channel calculates the rate of change of the system state, that is, the derivative, to predict the change of the system state in the next time step; the second-order channel processes the first-order derivative and the second-order derivative, and calculates the acceleration or higher-order changes of the system through the transformation of the state space; the third-order channel processes the rate of change of higher orders, which is used to capture the high-order dynamic behavior in complex systems; A state storage unit is also introduced in the automatic order raising module to store the state information calculated for each channel, ensuring that when performing multiple integrations, the system can access the previous state and derivative values ​​to ensure the continuity and accuracy of each channel; The gating network is used to select the best order output from the outputs of multiple integration channels; through a trainable weight matrix W go , use the softmax function to determine which channel's output is most appropriate; Among them, the principle of the multi-channel backward Euler upgrade module with dynamic component order identification is: The input initial state h0, first-order derivative h0′ and second-order derivative h0″ are upgraded by backward Euler method to obtain the first-order, second-order and third-order prediction values; the backward Euler method is used to perform numerical integration of each order derivative to deduce the state change of each time step, and finally the state of the next time step is obtained.

4. The method for dynamic modeling of power system components based on the MMOE-DAE model according to claim 3 is characterized in that: Step S3 is as follows: Electromagnetic transient simulation replaces each component with a Norton equivalent circuit and constructs network equations based on the equivalent circuit; avoids analyzing all DAEs of the system as discrete entities; The component simulation interface includes the following four steps: (1) Calculate the state variable x(t+Δt) Measurements based on component ports and To infer the state variable x(t+Δt); use the trapezoidal integration method to calculate x(t+Δt): Among them, the variable x(t) is an implicit function of the port measurement value, which is constructed by ODE expert neurons, that is, (2) Derivation of the Norton equivalent current i(t+Δt) of the component Based on the formula Where i is the component injection current, g(·) is the algebraic equation describing the component current; the port current of the dynamic component is calculated using the AE tower module Steps (1) and (2) are completely executed within the MMOE-DAE network, and their functions are consistent with the external characteristics of general components; (3) Calculate the voltage vector V(t+Δt): According to the node connection relationship, the current I of the data-driven model component is converted into eql (t+Δt) and the current I of conventional components gen (t+Δt) is combined to obtain I(t+Δt)=[I eql (t+Δt)+I gen (t+Δt)] Substituting into the formula GV=I, we can get GV(t+Δt)=I(t+Δt) (4) Calculate the branch current I ij (t+Δt): To complete the historical current iteration process of conventional components, it is necessary to calculate I ij (t+Δt), the current passes through V(t+Δt) and the corresponding admittance G ij Calculated I ij (t+Δt)=G ij (V i (t+Δt)-V j (t+Δt))。 5. A power system component dynamic modeling system based on MMOE-DAE model, characterized in that: The power system component dynamic modeling method based on the MMOE-DAE model as described in any one of claims 1 to 4 is used to perform dynamic modeling of the power system component.

6. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed, the method for dynamic modeling of power system components based on the MMOE-DAE model as described in any one of claims 1 to 4 is executed.

Images