Power system digital twin bidirectional mapping construction method and system, medium and product

Abstract

The application discloses a kind of electric power system digital twin bidirectional mapping construction method, system, medium and product, belong to electric power system digital twin field, the method is: real-time acquisition electric power system multi-source heterogeneous data and equipment parameter, constructs digital mirror model;According to the strength of physical law, it is shunted as strong, weak law data, respectively by embedding the first neural network of physical constraint and the second neural network of pure data driving mapping, cooperatively corrects and updates model with digital twin state;Based on the model, control instruction virtual simulation is generated, and executable control instruction is issued to execute, and response data is obtained;Response and estimated state are compared bidirectionally, and residual compensation signal is generated, mapping correction parameter is updated, and digital mirror model after parameter correction is obtained.By implementing the application, the problem of insufficient new power system control precision and unexecutable instruction caused by lack of physical-information space bidirectional closed-loop mapping mechanism in existing digital twin technology can be solved.

Description

Technical Field

[0001] This invention belongs to the field of power system digital twins, and relates to a method, system, medium and product for constructing a two-way mapping of power system digital twins. Background Technology

[0002] With the advancement of the "dual carbon" target and the large-scale integration of renewable energy, the new power system exhibits characteristics such as a high proportion of distributed resources, multi-timescale regulation needs, and strong uncertainty. It is urgent to achieve accurate perception and proactive regulation of the coordinated operation status of source-grid-load-storage through high-fidelity, real-time synchronous digital means. Digital twin technology, due to its ability to construct dynamic mapping between physical entities and information systems, is regarded as a key path to support the intelligent operation of the new power system.

[0003] Existing technologies for constructing digital twins of power systems typically employ a unified modeling strategy to process various types of operational data, failing to differentiate between the physical interpretability of different data. Furthermore, their mapping process is often limited to one-way state synchronization from the physical system to the information system, lacking a reverse path to effectively translate high-level control intentions into actionable actions of terminal equipment. Additionally, no feedback verification mechanism has been established between the execution results and the model state. Consequently, the digital twin only possesses monitoring and simulation capabilities, making it difficult to support the feasibility assurance of instructions and the maintenance of continuous accuracy required for closed-loop control. Summary of the Invention

[0004] This application provides a method, system, medium, and product for constructing a two-way mapping of digital twins for power systems, which can solve the problems of insufficient control accuracy and unexecutable commands in new power systems caused by the lack of a two-way closed-loop mapping mechanism between physical and information spaces in existing digital twin technologies.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for constructing a bidirectional digital twin mapping for a power system, comprising: Real-time acquisition of multi-source heterogeneous data and equipment parameters of the power system, and construction of a digital mirror model based on the equipment parameters; The multi-source heterogeneous data is divided into strong-law data and weak-law data according to the strength of physical laws. The strong-law data and weak-law data are mapped by a first neural network embedded with physical constraints and a second data-driven neural network, respectively. The digital twin estimation state is obtained through collaborative correction and the digital mirror model is updated to obtain the updated digital mirror model. The control target instruction is obtained, and the control target instruction is virtually simulated and deduced based on the updated digital mirror model to generate executable control instructions and send them to the power system for execution, thereby obtaining response data. The response data is compared with the estimated state of the digital twin in two directions to generate a residual compensation signal and update the parameters used to correct the mapping deviation in the updated digital mirror model, thus obtaining a digital mirror model with corrected parameters.

[0006] Compared to existing technologies, the embodiments of this application have the following beneficial effects: Real-time acquisition of multi-source heterogeneous data and equipment parameters of the power system, and construction of a digital mirror model based on the equipment parameters, provides a digital foundation with physical structural constraints for subsequent mapping and control, ensuring the interpretability of state estimation within the physical boundaries of the equipment; further, by dividing multi-source heterogeneous data into strongly regular and weakly regular data according to the strength of physical laws, and mapping them respectively using a first neural network embedded with physical constraints and a data-driven second neural network, the electrical measurement data that obeys the power grid physical equations is constrained by physical laws and maintains consistency during the mapping process. Simultaneously, non-electrical data representing the behavior of environmental and market actors adapts to uncertainty through a flexible data-driven approach, thus balancing the physical reliability and data adaptability of the modeling; after collaborative correction, the digital twin estimated state is obtained and the digital mirror model is updated. Dynamic complementary correction is achieved by fusing the deviation information from the two types of mapping outputs, suppressing the single modeling paradigm from operating condition deviations or insufficient samples. The system addresses performance degradation over time. Simultaneously, by acquiring control target commands and performing virtual simulations based on the updated digital twin model, the high-level control intentions can be pre-enacted in a high-fidelity digital space, generating executable control commands before being sent to the physical system. This avoids unverified commands directly impacting the power grid and causing over-limits or instability. Furthermore, by comparing response data with the estimated state from the digital twin in both directions, residual compensation signals are generated and the parameters in the model used to correct mapping deviations are updated. This achieves joint sensing and feedback correction of state prediction deviations and command execution deviations, enabling the digital twin model to continuously correct its mapping deviations based on operational feedback and maintain state consistency over long-term operation. These features work synergistically to construct a complete link: a high-fidelity uplink mapping from physical to information, a feasible downlink inversion from information to physical, and a two-way closed-loop feedback correction. This solves the problems of insufficient control accuracy and unexecutable commands caused by the lack of a two-way closed-loop mapping mechanism between physical and information spaces in existing digital twin technologies.

[0007] In some embodiments of the first aspect of this application, the step of dividing the multi-source heterogeneous data into strongly regular data and weakly regular data according to the strength of physical laws includes: The multi-source heterogeneous data is subjected to spatiotemporal alignment processing to generate standardized fused feature data; Electrical measurement data that conform to the power grid physical equations in the standardized fusion feature data are considered as strong regularity data, while non-electrical measurement data that characterize environmental factors and the behavior of market participants are considered as weak regularity data.

[0008] Compared with existing technologies, the above embodiments have the following beneficial effects: Spatiotemporal alignment processing of multi-source heterogeneous data generates standardized fused feature data, eliminating the dimensional mismatch in time sampling frequency and spatial measurement location of data from the source, grid, load, and storage sides, providing a consistent input for subsequent strong and weak current distribution according to physical laws; furthermore, by taking electrical measurement data that obeys the physical equations of the power grid as strong regularity data and non-electrical measurement parts that characterize environmental factors and market behavior as weak regularity data in the standardized fused feature data, the technical basis for current distribution criteria is clarified, giving subsequent differentiated modeling a clear physical boundary and improving the rationality and feasibility of the mapping strategy.

[0009] In some embodiments of the first aspect of this application, the first neural network is a multilayer fully connected neural network trained with a loss function weighted by data loss terms and physical loss terms; the second neural network is a deep residual network containing skip connections; wherein, the data loss term is an error measure between the mapped output value and the actual measured value; and the physical loss term is characterized by calculating the sum of the squares of the active power imbalance and reactive power imbalance in the power flow equation of the power grid.

[0010] Compared with existing technologies, the above embodiments have the following beneficial effects: the data loss term ensures fitting accuracy by measuring the error between the mapped output value and the actual measured value; the physical loss term quantifies the degree of violation of Kirchhoff's laws by calculating the sum of the squares of the active power imbalance and reactive power imbalance in the power flow equation; the weighted synergy of the two enables the network output to approximate the measured data while forcibly satisfying the basic physical constraints of the power grid; the second neural network is a deep residual network containing skip connections, which alleviates the gradient vanishing problem and improves the ability to represent complex nonlinear relationships in weakly regular data through the residual structure, thereby enhancing the robustness and generalization performance of the overall mapping model.

[0011] In some embodiments of the first aspect of this application, the method of mapping the strong regularity data and the weak regularity data using a preset first neural network embedded with physical constraints and a data-driven second neural network, respectively, and then obtaining the digital twin estimation state through collaborative correction, includes: The strong regularity data is input into the first neural network to calculate the first state component representing the electrical state of the power grid, and the weak regularity data is input into the second neural network to calculate the second state component representing the environmental and market behavior state. Calculate the first deviation between the first state component and the corresponding electrical measurement value in the standardized fusion feature data, and the second deviation between the second state component and the corresponding environmental and behavioral measurement values ​​in the standardized fusion feature data; Based on the first bias, the second bias, and the standardized fusion feature data, the residual signal is obtained by using an online extreme learning machine and fitting it using the recursive least squares method. The first state component and the second state component are compensated and fused according to the residual signal to output the estimated state of the digital twin.

[0012] Compared with existing technologies, the above embodiments have the following beneficial effects: Strongly regular data is input into a first neural network to calculate a first state component representing the electrical state of the power grid; weakly regular data is input into a second neural network to calculate a second state component representing the environmental and market behavior states, thus achieving decoupled modeling and independent output of the two types of states; further, by calculating the first deviation between the first state component and the corresponding electrical measurement value, and the second deviation between the second state component and the corresponding environmental and behavioral measurement values, the local errors of the two types of models are quantified respectively; then, based on the two types of deviations and standardized fused feature data, an online extreme learning machine is used and a recursive least squares method is used to fit the residual signal, enabling the correction process to have real-time performance and lightweight parameters; finally, the two state components are compensated and fused according to the residual signal, and a digital twin estimated state is output, realizing the dynamic collaborative fusion of two heterogeneous states under error-driven conditions, improving the accuracy and consistency of the overall state estimation.

[0013] In some embodiments of the first aspect of this application, the step of performing virtual simulation and deduction of the control target instruction based on the updated digital mirror model to generate executable control instructions includes: Based on the control target instruction, and combined with the static parameters and current operating status of each terminal device in the digital mirror model, the control target instruction is decomposed into preliminary control instructions for each terminal device. Based on the preset response behavior prediction model of each terminal device in the digital mirror model, the corresponding preliminary control instructions are predicted and queried to obtain the acceptance probability and expected response deviation of each market entity to the preliminary control instructions. Based on the acceptance probability and expected response deviation, the preliminary control instructions are modified to generate device-level control instructions, which are then used as executable control instructions.

[0014] Compared to existing technologies, the above embodiments have the following beneficial effects: Based on the control target instruction and the static parameters and current operating status of each terminal device in the digital mirror model, the control instruction is decomposed into preliminary control instructions, ensuring that the instruction decomposition conforms to the physical capability boundaries of the devices; further, the preliminary control instructions are predicted and queried based on the preset response behavior prediction model of the market actors to which each terminal device belongs, obtaining the acceptance probability and expected response deviation of each actor, explicitly incorporating the uncertainty of multi-agent behavior into the instruction generation process; then, the preliminary control instructions are corrected based on the acceptance probability and expected response deviation, generating device-level control instructions as executable control instructions, making the final issued instructions closer to the actual execution intentions and capabilities, thus improving the success rate and reliability of the control instruction implementation.

[0015] In some embodiments of the first aspect of this application, the generation of each device-level control command includes: Each of the device-level control commands is substituted back into the digital mirror model to perform steady-state power flow deduction and generate the system steady-state state. The operating constraint boundaries of each terminal device are read from the digital mirror model. It is determined whether the steady-state state of the system satisfies the operating constraint boundaries. If not, the control quantity of each device-level control instruction is iteratively adjusted using an optimization algorithm with the goal of minimizing the deviation of the control target and the operating constraint boundaries as constraints. The adjusted control quantities are then substituted back into the digital mirror model for secondary deduction until the steady-state state of the system satisfies the operating constraint boundaries. Finally, the adjusted device-level control instruction is output as the final executable control instruction.

[0016] Compared with existing technologies, the above embodiments have the following beneficial effects: Control commands at each device level are substituted back into the digital mirror model for steady-state power flow simulation, generating a system steady-state state to predict the grid operation conditions after command execution; further, the operating constraint boundaries of each terminal device are read from the digital mirror model, and it is determined whether the system steady-state state meets these boundaries, achieving a preliminary verification of command feasibility; if not, the control quantity of each device-level control command is iteratively adjusted using an optimization algorithm with the goal of minimizing the deviation of the control target and the operating constraint boundaries as constraints, and the adjusted control quantity is substituted back for secondary simulation until the constraints are met, outputting the final executable control command. This ensures that the issued commands are strictly executable in the physical system and do not cause grid safety exceedances, enhancing the safety and robustness of control.

[0017] In some embodiments of the first aspect of this application, the digital mirror model includes a runtime layer that dynamically updates the system state using a recursive state estimation algorithm; The step of performing a bidirectional error comparison between the response data and the estimated digital twin state, generating a residual compensation signal, and updating the parameters used to correct the mapping bias in the updated digital mirror model to obtain a parameter-corrected digital mirror model includes: The state quantities directly related to the executable control instructions are extracted from the estimated state of the digital twin as predicted values, and the actual measurement data are extracted from the response data as actual values. Based on each predicted value and actual value, the state prediction error is calculated. The actual control quantity of each terminal device in the power system when executing executable control instructions is obtained as the execution value. The control target value of each terminal device is extracted from the executable control instructions as the instruction value. Based on the execution value and the instruction value, the instruction execution error is calculated. By concatenating the state prediction errors and instruction execution errors, a comprehensive error vector is obtained. Based on the comprehensive error vector, a residual compensation signal is generated, and based on the residual compensation signal, the weight coefficients of the physical constraints in the first neural network, the output bias parameters in the second neural network, and the parameters of the recursive state estimation algorithm are updated to obtain a digital mirror model with corrected parameters.

[0018] Compared with existing technologies, the above embodiments have the following beneficial effects: State quantities directly related to executable control commands are extracted from the digital twin estimated state as predicted values, and actual measurement data is extracted from the response data as actual values. State prediction errors are calculated to reflect uplink mapping deviations. Simultaneously, the actual control quantities during terminal device execution are obtained as execution values, and control target values ​​are extracted from executable control commands as command values. Command execution errors are calculated to reflect downlink execution deviations. The two types of errors are then concatenated to obtain a comprehensive error vector, fully characterizing the overall mismatch degree of bidirectional mapping. Finally, a residual compensation signal is generated based on the comprehensive error vector, and the physical constraint weight coefficients in the first neural network, the output bias parameters in the second neural network, and the parameters of the recursive state estimation algorithm are updated accordingly. This achieves multi-level synchronous correction of model parameters, improving the adaptive accuracy and stability of the system under long-term operation.

[0019] Secondly, the present invention also provides a power system digital twin bidirectional mapping construction system, comprising: a data acquisition module, a first update module, a response module, and a second update module; The data acquisition module is used to collect multi-source heterogeneous data and equipment parameters of the power system in real time, and to construct a digital mirror model based on the equipment parameters. The first update module is used to divide the multi-source heterogeneous data into strong regularity data and weak regularity data according to the strength of physical laws, and use a preset first neural network embedded with physical constraints and a data-driven second neural network to map the strong regularity data and weak regularity data respectively, and then obtain the digital twin estimation state through collaborative correction and update the digital mirror model to obtain the updated digital mirror model. The response module is used to acquire the control target instruction, perform virtual simulation and deduction of the control target instruction based on the updated digital mirror model, generate executable control instructions and send them to the power system for execution, and obtain response data. The second update module is used to perform a bidirectional error comparison between the response data and the estimated state of the digital twin, generate a residual compensation signal, and update the parameters used to correct the mapping deviation in the updated digital mirror model to obtain a digital mirror model with corrected parameters.

[0020] Compared with existing technologies, the above embodiments of this application have the following beneficial effects: Real-time acquisition of multi-source heterogeneous data and equipment parameters of the power system, and construction of a digital mirror model based on the equipment parameters, provides a digital foundation with physical structural constraints for subsequent mapping and control, ensuring the interpretability of state estimation within the physical boundaries of the equipment; further, by dividing multi-source heterogeneous data into strongly regular data and weakly regular data according to the strength of physical laws, and mapping them respectively using a first neural network embedded with physical constraints and a data-driven second neural network, the electrical measurement data that obeys the power grid physical equations is constrained by physical laws and maintains consistency during the mapping process. Simultaneously, non-electrical data representing the behavior of environmental and market actors adapts to uncertainty through a flexible data-driven approach, thus balancing the physical reliability and data adaptability of the model; after collaborative correction, the digital twin estimated state is obtained and the digital mirror model is updated. Dynamic complementary correction is achieved by fusing the deviation information of the two types of mapping outputs, suppressing the deviation of a single modeling paradigm due to operating condition shifts or sample inconsistencies. The system effectively mitigates performance degradation over time. Simultaneously, by acquiring control target commands and performing virtual simulations based on the updated digital twin model, the high-level control intentions can be pre-enacted in a high-fidelity digital space, generating executable control commands before being sent to the physical system. This avoids unverified commands directly impacting the power grid and causing over-limit or instability. Furthermore, by comparing response data with the estimated state from the digital twin in both directions, residual compensation signals are generated and the parameters in the model used to correct mapping deviations are updated. This achieves joint perception and feedback correction of state prediction deviations and command execution deviations, enabling the digital twin model to continuously correct its mapping deviations based on operational feedback and maintain state consistency over long-term operation. These features work synergistically to construct a complete link: a high-fidelity uplink mapping from physical to information, a feasible downlink inversion from information to physical, and a two-way closed-loop feedback correction. This solves the problems of insufficient control accuracy and unexecutable commands caused by the lack of a two-way closed-loop mapping mechanism between physical and information spaces in existing digital twin technologies.

[0021] Thirdly, the present invention also provides a computer program product, including a computer program or instructions, which, when executed, implement any one of the power system digital twin bidirectional mapping construction methods of the present invention.

[0022] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any one of the power system digital twin bidirectional mapping construction methods of the present invention. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating a method for constructing a bidirectional digital twin mapping for a power system, provided in some embodiments of the present invention.

[0024] Figure 2 This is a schematic diagram of the structure of a power system digital twin bidirectional mapping construction system provided in some embodiments of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1: Please refer to Figure 1 To address the issues of insufficient control precision and unexecutable commands in new power systems caused by the lack of a two-way closed-loop mapping mechanism between physical and information spaces in existing digital twin technologies, an embodiment of the present invention provides a method for constructing a two-way mapping of a power system digital twin, comprising steps S1 to S4: Step S1: Collect multi-source heterogeneous data and equipment parameters of the power system in real time, and construct a digital mirror model based on the equipment parameters; Step S2: The multi-source heterogeneous data is divided into strong regularity data and weak regularity data according to the strength of physical laws. The strong regularity data and weak regularity data are mapped by a first neural network with embedded physical constraints and a second data-driven neural network, respectively. The digital twin estimation state is obtained through collaborative correction and the digital mirror model is updated to obtain the updated digital mirror model. Furthermore, in step S2, the diversion can be implemented through the following preferred embodiments, including steps S21-S22: S21: Perform spatiotemporal alignment processing on the multi-source heterogeneous data to generate standardized fused feature data; S22: The electrical measurement data that obeys the power grid physical equation in the standardized fusion feature data are taken as strong regularity data, and the non-electrical measurement part that represents environmental factors and market behavior entities in the standardized fusion feature data is taken as weak regularity data.

[0027] In existing technologies, multi-source heterogeneous data are often directly input into a unified model for processing without considering the essential differences in physical interpretability between electrical measurement data and environmental and behavioral data. This leads to a sharp drop in model accuracy due to physical mismatch when operating conditions deviate, or insufficient generalization ability when behavioral samples are sparse.

[0028] In this embodiment, step S21 performs spatiotemporal alignment of multi-source heterogeneous data, employing flexible time window interpolation and a spatial mapping matrix based on grid topology impedance to effectively eliminate scale mismatch and spatial parallax between millisecond-level PMU data and hourly-level meteorological data. This provides high-quality, spatiotemporally consistent standardized fusion feature data for subsequent flow separation. Step S22 then performs strong-weak flow separation based on whether the data conforms to the basic physical equations of the power grid (such as Kirchhoff's laws and power flow equations). This keeps strong-pattern data within the physical constraint framework, while weak-pattern data is flexibly modeled by the data-driven module, thus laying a logical foundation for differentiated mapping and avoiding the dual loss of accuracy and robustness caused by "one-size-fits-all" modeling.

[0029] Furthermore, the first neural network is a multilayer fully connected neural network trained with a loss function constructed by weighting data loss terms and physical loss terms; the second neural network is a deep residual network containing skip connections; wherein, the data loss term is an error measure between the mapped output value and the actual measured value; the physical loss term is characterized by calculating the sum of the squares of the active power imbalance and reactive power imbalance in the power flow equation of the power grid.

[0030] Traditional neural networks often face the problem of lack of physical consistency in power system modeling. While purely data-driven models can fit complex nonlinear relationships, their outputs may violate basic circuit laws, leading to erroneous state estimates from twins outside the safety boundary. In this embodiment, the first neural network introduces a physical loss term constructed from the residuals of the power flow equations, embedding the sum of squares of the active / reactive power imbalance of the grid as a soft constraint into the training process. This allows the network to approximate measured data while its output naturally tends to satisfy physical laws, significantly improving the reliability of the model when there is parameter mismatch or operating condition extrapolation. The second neural network adopts a deep residual structure with skip connections, effectively alleviating the gradient vanishing problem in deep network training and enhancing the ability to capture complex patterns such as time-series dependence and economic sensitivity in the response behavior of market participants. The two neural networks work together to achieve an organic unity of "physical credibility" and "data adaptability".

[0031] Furthermore, in step S2, the mapping and co-correction can be implemented through the following preferred embodiments, including steps S22-S25: S22: Input the strong regularity data into the first neural network to calculate the first state component representing the electrical state of the power grid, and input the weak regularity data into the second neural network to calculate the second state component representing the environmental and market behavior state. S23: Calculate the first deviation between the first state component and the corresponding electrical measurement value in the standardized fusion feature data, and the second deviation between the second state component and the corresponding environmental and behavioral measurement values ​​in the standardized fusion feature data; S24: Based on the first deviation, the second deviation, and the standardized fusion feature data, the residual signal is obtained by using an online extreme learning machine and fitting it using the recursive least squares method; S25: Compensate and fuse the first state component and the second state component according to the residual signal respectively, and output the digital twin estimated state.

[0032] Existing digital twin systems often employ simple weighted averaging or fixed-rule fusion during the state fusion stage, making it difficult to dynamically balance the confidence levels of the mechanistic model and the data-driven model. This embodiment constructs a dynamic collaborative correction mechanism through steps S22–S25: First, the deviations between the two types of state components and their corresponding measurements are calculated separately to reflect their respective local errors; then, an online extreme learning machine combined with recursive least squares is used to fit the residual signal in real time with low computational overhead, avoiding the delays of traditional batch training; finally, the two state components are compensated and fused based on the residual signal, achieving error-driven dynamic weight adjustment. This mechanism allows the system to rely on a high-precision mechanistic model in steady state, and automatically enhance the correction effect of the data-driven module during transients or behavioral abrupt changes, thereby maintaining a high-precision and highly robust digital twin estimation state under all operating conditions.

[0033] Step S3: Obtain the control target instruction, perform virtual simulation and deduction of the control target instruction based on the updated digital mirror model, generate executable control instructions and send them to the power system for execution, and obtain response data; Furthermore, in step S3, generating executable control instructions can be implemented through the following preferred embodiments, including steps S31-S33: S31: Based on the control target instruction, and combined with the static parameters and current operating status of each terminal device in the digital mirror model, the control target instruction is decomposed into preliminary control instructions for each terminal device; S32: Based on the preset response behavior prediction model of each terminal device in the digital mirror model, perform prediction query on each corresponding preliminary control command to obtain the acceptance probability and expected response deviation of each market entity to the preliminary control command; S33: Based on the acceptance probability and expected response deviation, each of the preliminary control instructions is modified to generate each device-level control instruction as the executable control instruction.

[0034] Traditional decomposition of control instructions typically ignores the autonomous decision-making characteristics of distributed resource entities, directly allocating control quantities proportionally, resulting in low actual execution rates and diminished control effectiveness. This embodiment, building upon the initial decomposition in step S31, introduces steps S32–S33 to explicitly invoke the preset response behavior prediction models of the market actors to which each terminal device belongs, quantifying their probability of accepting instructions and the deviation from expected responses. Subsequently, based on this behavioral feedback, the allocation scheme is dynamically adjusted, prioritizing the assignment of control tasks to entities with high responsiveness and strong performance reliability, while ensuring that the allocation quantity is within its physical and economic feasible range. This strategy incorporates the "human" factor into the digital twin closed loop, improving the executability and success rate of instructions from the source, avoiding the dilemma of "paper scheduling, no results on-site."

[0035] Furthermore, after step S34, steps S35-S36 are also included: S35: Substitute each of the device-level control commands back into the digital mirror model to perform steady-state power flow deduction and generate the system steady-state state; S36: Read the operating constraint boundaries of each terminal device from the digital mirror model, and determine whether the steady-state state of the system satisfies the operating constraint boundaries; if not, take the minimum deviation of the control target as the objective and the operating constraint boundaries as the constraint conditions, use the optimization algorithm to iteratively adjust the control quantities of each device-level control instruction, and substitute the adjusted control quantities back into the digital mirror model for secondary deduction until the steady-state state of the system satisfies the operating constraint boundaries, and output the adjusted device-level control instruction as the final executable control instruction.

[0036] Existing simulations often stop at power flow calculations, lacking proactive verification and correction mechanisms for command feasibility. This can easily lead to safety risks such as voltage overruns and equipment overloads caused by issued commands. This embodiment constructs a "simulation-verification-correction" closed loop through steps S35–S36: First, the equipment-level control commands are substituted back into the digital mirror model for steady-state power flow simulation to generate the system's steady-state state. Then, key indicators such as equipment output, node voltage, and SOC are rigorously compared to see if they meet the operational constraint boundaries. If overruns occur, the optimization objective is to minimize the deviation of the control target. Optimization algorithms such as sequential quadratic programming are used to iteratively adjust the control quantity, and the simulation is repeated until all constraints are met. This ensures that the final executable control commands are not only technically feasible but also always within the safe operating envelope of the power grid, significantly improving the safety and robustness of the control.

[0037] Step S4: Compare the response data with the estimated state of the digital twin in two directions to generate a residual compensation signal and update the parameters used to correct the mapping bias in the updated digital mirror model to obtain a digital mirror model with corrected parameters.

[0038] Furthermore, the digital mirror model includes a runtime layer, which dynamically updates the system state through a recursive state estimation algorithm; Step S4 can be implemented through the following preferred embodiments, including steps S41-S44: S41: Extract state quantities directly related to the executable control instructions from the estimated state of the digital twin as predicted values, extract actual measurement data from the response data as actual values, and calculate the state prediction error based on each predicted value and actual value; S42: Obtain the actual control quantity of each terminal device in the power system when executing executable control instructions as the execution value, extract the control target value of each terminal device from the executable control instructions as the instruction value, and calculate the instruction execution error based on each execution value and instruction value; S43: Combine the state prediction errors and instruction execution errors to obtain a comprehensive error vector; S44: Generate a residual compensation signal based on the comprehensive error vector, and update the weight coefficients of the physical constraints in the first neural network, the output bias parameters in the second neural network, and the parameters of the recursive state estimation algorithm based on the residual compensation signal to obtain a digital mirror model with corrected parameters.

[0039] Traditional digital twin systems often focus only on state estimation errors, neglecting the feedback of deviations in instruction execution, leading to a gradual disconnect between the model and the physical system after long-term operation. This embodiment constructs a two-way error perception mechanism through steps S41–S44: on the one hand, it calculates the state prediction error, reflecting the modeling deviation in the uplink mapping; on the other hand, it calculates the instruction execution error, capturing the actual execution deviation in the downlink inversion. The combined error vector formed by these two calculations comprehensively represents the degree of mapping mismatch between the physical and information spaces. Based on this, the system synchronously updates the physical constraint weight coefficients of the first neural network (adjusting physical reliability), the output bias parameters of the second neural network (correcting behavioral prediction deviations), and the noise covariance of the recursive state estimation algorithm (optimizing state tracking sensitivity), achieving multi-level, multi-dimensional adaptive correction. This allows the digital mirror model to continuously approximate the real system during continuous operation, forming a truly "more accurate with use" closed-loop evolution capability.

[0040] In practical implementation, this application will be further described as follows: The first step is to acquire data and build a digital twin model (i.e., a digital mirror model).

[0041] Here, multi-source heterogeneous data includes: Electrical measurement data such as voltage, current, and power collected by the SCADA system (sampling period 1-5 seconds), synchronous phasor data collected by the PMU (sampling frequency 50-100 frames / second), and user electricity consumption data collected by the smart meter (sampling period 15 minutes-1 hour). Temperature, irradiance, and wind speed data collected by meteorological monitoring stations (sampling period 15 minutes to 1 hour), as well as electricity market transaction data and historical user response behavior data; The equipment parameters (static) include: generator rated capacity, synchronous reactance, transient reactance, inertial time constant, transformer rated capacity, short-circuit impedance, turns ratio, transmission line impedance, admittance, energy storage system rated capacity, charge and discharge efficiency, initial SOC value, distributed power inverter capacity and control mode, and the physical topology connection relationship of each terminal device.

[0042] As an optimal approach, when acquiring multi-source heterogeneous data, data completion for monitoring blind spots is also included: (1) Obtain measured data of the measurement coverage area in the new power system. The measurement coverage area refers to the nodes and branches equipped with SCADA measurement terminals or PMU measurement devices. The measured data shall include at least the voltage amplitude, voltage phase angle, injected active power, injected reactive power of each measurement node, and power flow data of each measurement branch.

[0043] (2) Input the measured data into the physical constraint generative adversarial network. The physical constraint generative adversarial network consists of a generator and a discriminator. The output of the generator is connected to the input of the discriminator, and the output of the discriminator is connected to the feedback input of the generator. The generator adopts a multi-layer fully connected neural network. The input is a random noise vector and part of the observation data of the measurement coverage area. The output is virtual measurement data of the monitoring blind area. The discriminator adopts a multi-layer fully connected neural network. The input is real data or virtual data output by the generator. The output is the probability that the data is real data.

[0044] The generator's loss function incorporates voltage drop constraints and power balance constraints. The voltage drop constraint is based on Kirchhoff's voltage law, and the power balance constraint is based on Kirchhoff's current law. The generator's total loss function... ,in To combat the losses, For physical loss, These are the physical constraint weighting coefficients.

[0045] Physical loss The specific calculation method is as follows: the virtual measurement data of the blind zone output by the generator is combined with the actual measurement data of the measurement coverage area to form a complete full network data snapshot.

[0046] On this data snapshot, calculate the voltage drop equation residuals for each branch: ; This formula appears as a physical constraint term in the generator loss function of a Physically Constrained Generative Adversarial Network (PC-GAN). It is used to constrain the virtual measurement data of the blind zone generated by the generator to satisfy the physical laws of the power grid. The parameters in the formula are defined as follows: in The node number at the beginning of the branch road. Number the node at the end of the branch road. Indicates the connection node With nodes Branch lines (lines or transformers). For nodes voltage amplitude, For nodes The voltage amplitude; branch road The voltage drop calculation residual is the deviation between the theoretical voltage drop and the actual voltage difference. for and The branch resistance value, for a line, is taken from the resistance per unit length of the line multiplied by the length; for a transformer, it is the short-circuit resistance referred to a certain side. For the node Flow to Node The effective power, measured in watts (W) or per-unit (pu), is expressed as the power flowing out of the node. Positive direction; branch road The reactance value is expressed in ohms (Ω). For a line, this value is taken as the reactance per unit length of the line multiplied by the length. For a transformer, this value is the short-circuit reactance referred to a certain side (per unit value). For the node Flow to Node Reactive power, measured in var (Var) or per-unit (pu), is expressed as the power flowing out of the node. Positive direction; In this invention, this formula does not participate in the control of the actual physical system, but serves as a hard constraint term for the "physical loss" of the physical constraint generative adversarial network. That is, the virtual blind zone data generated by the generator is substituted into this formula, and if the calculated... If the deviation from the theoretical voltage drop is too large, the generated data is determined to be inconsistent with physical laws, thereby penalizing the generator to ensure that its output virtual measurement data strictly follows the circuit principle.

[0047] Calculate the power balance equation residuals for each node: ; ; Physical loss .

[0048] The training of the physical constraint generative adversarial network adopts an alternating optimization strategy: the generator parameters are fixed, and the discriminator is trained to maximize the discrimination accuracy; the discriminator parameters are fixed, and the generator is trained so that its output data can both deceive the discriminator and satisfy the physical constraints. The two stages are alternated until the network converges.

[0049] The generator in the trained physical constraint generative adversarial network generates virtual measurement data for the monitoring blind zone. The virtual measurement data is then merged with the actual measurement data to form a complete multi-source heterogeneous dataset.

[0050] Next, when constructing the digital twin model, as a preferred option, this application constructs a digital mirror model system based on device parameters, which includes a device digital mirror layer, a running digital mirror layer, and an interactive digital mirror layer; The device digital mirror layer, the running digital mirror layer, and the interactive digital mirror layer are interconnected through a unified data bus. The output of the device digital mirror layer is connected to the input of the running digital mirror layer, the output of the running digital mirror layer is connected to the input of the interactive digital mirror layer, and the output of the interactive digital mirror layer is connected to the correction input of the running digital mirror layer, forming a three-layer progressive interactive loop.

[0051] I. The device digital mirror layer is a device-level model built based on the device's static parameters, used to provide physical constraint boundary conditions for running the digital mirror layer.

[0052] Taking a synchronous generator as an example, the digital mirror layer of the equipment contains at least a fifth-order practical model of the synchronous generator, including the rotor motion equation, stator voltage equation, rotor winding electromagnetic transient equation, and transfer function models of the excitation system and speed regulation system.

[0053] Taking energy storage systems as an example, the equipment digital mirror layer should at least include the equivalent circuit model of the energy storage battery and the average switching cycle model of the energy storage converter. Taking power transmission and distribution networks as an example, the equipment digital mirror layer should at least include the steady-state model of the power grid based on the node admittance matrix and the dynamic thermal model of each branch.

[0054] The physical constraints provided by the device digital mirror layer include power balance constraints, Kirchhoff's voltage and current laws constraints, upper and lower limits of device output constraints, and voltage / frequency operating limit constraints.

[0055] II. The running digital mirror layer constructs a dynamic running digital twin model of the device and system based on real-time measurement data, which is used to provide the running state context for the interactive digital mirror layer.

[0056] The digital mirror layer receives real-time data from SCADA, PMU, and smart meters. It then uses an extended Kalman filter (a weighted recursive state estimation algorithm) to dynamically track and correct the internal state variables of the device's digital mirror layer online. Taking a synchronous generator as an example, the digital mirror layer uses the generator mechanism model in the device digital mirror layer as the state prediction equation, and the measured terminal voltage and current of the PMU as the measurement update input. The EKF recursively estimates the generator's internal state variables that cannot be directly measured, such as the power angle, speed, and transient electromotive force at the current moment. Taking energy storage systems as an example, the digital mirror layer uses the battery equivalent circuit model in the device's digital mirror layer as the state prediction equation, and the terminal voltage and charging / discharging current measured by SCADA as the measurement update input. The internal states such as SOC, SOH and polarization voltage are tracked online through EKF. Taking the power transmission and distribution network as an example, the digital mirror layer operates based on the voltage amplitude, voltage phase angle, and injected active and reactive power measurement data of each measurement node. It uses the weighted least squares method to identify the network topology and estimate the state, and outputs the optimal estimated values ​​of voltage amplitude and phase angle of each node in the entire network.

[0057] The output of the digital mirror layer is not limited to the state estimate at the current moment, but also includes the state prediction at future moments and its confidence interval obtained by extrapolating from the state prediction equation, which is used to support subsequent situational awareness and forward decision-making.

[0058] Third, the interactive digital mirror layer constructs a digital twin model of multi-entity interactive behavior based on multi-entity interactive data. The multi-entity (i.e. market behavior entities) includes distributed photovoltaic owners, distributed wind power owners, energy storage system operators, electric vehicle aggregators, and flexible load users. The multi-entity interactive data includes historical bidding data, winning bid data, settlement price data, historical records of each entity's response to control instructions, and static attribute data of each entity.

[0059] The interactive digital mirror layer extracts response behavior feature vectors for each entity based on the aforementioned multi-entity interaction data. These vectors include response capacity features, response speed features, performance reliability features, economic sensitivity features, and time pattern features. A long short-term memory network-attention mechanism is then used to perform time-series modeling of these features, outputting a predicted response behavior model for each entity under the current operating conditions. The input to this model is the expected adjustment amount, incentive price, and execution time window of the candidate control command. The output is the probability of the entity accepting the command, the expected actual response amount, and the confidence interval of the response deviation. Furthermore, the digital twin model of the interactive digital mirror layer is also used to generate a behavior prediction deviation signal. This signal is fed back to the operational digital mirror layer to correct load prediction deviations within the operational digital mirror layer.

[0060] Secondly, after obtaining multi-source heterogeneous data, spatiotemporal alignment processing is performed to generate standardized fused feature data. This spatiotemporal alignment processing includes temporal dimension alignment and spatial dimension alignment.

[0061] Time dimension alignment is achieved in the following way: using a unified timestamp across the entire network as a benchmark, an elastic time window interpolation algorithm is used to align data with different sampling frequencies. The elastic time window interpolation algorithm dynamically encrypts interpolation nodes when the system is disturbed, and maintains the standard interpolation step size when the system is in a steady state.

[0062] Specifically, based on the IEEE 1588 precise time protocol, nanosecond-level time synchronization is achieved across the entire network, and a unified format timestamp tag and quality identifier are added to each frame of multi-source heterogeneous data.

[0063] Then, the system frequency change rate and node voltage amplitude change rate are monitored in real time. When the absolute value of the frequency change rate exceeds the preset threshold or the node voltage amplitude change rate exceeds the preset threshold, it is determined that the system has a disturbance, and the interpolation step size is automatically dynamically reduced from the steady-state value to the transient value.

[0064] During steady-state operation of the system, a larger interpolation step size is used to reduce computational redundancy. For SCADA data, cubic spline interpolation is used to map it to the standard time grid. For meteorological data, piecewise cubic Hermite interpolation is used. When the system is disturbed, the interpolation step size is automatically and dynamically reduced to the transient value, and the encrypted interpolation is maintained during the duration of the disturbance. For PMU data, its original sampled value is directly used as the value on the standard time grid without additional interpolation.

[0065] Spatial dimension alignment is achieved in the following way: a spatial mapping matrix is ​​generated based on the power grid topology model, and data of different spatial granularities are mapped to the standard spatial coordinate system through the spatial mapping matrix. The dimension of the spatial mapping matrix is ​​determined by the number of power grid nodes, and the elements in the spatial mapping matrix are calculated by the impedance weight between each measurement point and the standard spatial coordinate.

[0066] Specifically, based on the physical topology connections in the static parameters of the equipment, a node-branch association model of the distribution network is constructed. Based on the impedance parameters of each branch, a node admittance matrix Y is established, and then a spatial mapping matrix M is constructed. The spatial mapping matrix M is an n_s×n_m dimensional matrix, where n_s is the number of standard spatial nodes and n_m is the number of measurement points. The matrix element M_ij represents the weight coefficient of the data of the j-th measurement point to the i-th standard spatial node. The weight coefficient is determined according to the impedance amplitude between the measurement point and the standard node. The closer the electrical distance, the greater the weight. And it satisfies that for each standard node, the sum of the weights of all the measurement points it receives is 1.

[0067] For example, when mapping the voltage amplitude data of measurement point A at the beginning of a feeder and measurement point B at the end of the feeder to the standard spatial node, voltage drop compensation is performed according to the line impedance ratio between the measurement point and the standard node. For active power and reactive power data, the power transmission loss between each measurement point and the standard node is used for calculation.

[0068] After completing the time dimension alignment and spatial dimension alignment, the aligned data is normalized and standardized to generate standardized fused feature data. Each standardized fused feature data includes a timestamp label, spatial coordinate label, data type label, and fused feature vector.

[0069] Next, based on the standardized fusion feature data generated in the previous step, an uplink mapping from physical space to information space is performed. This can be achieved through a dedicated uplink mapping module during deployment, which includes a splitting unit, a first mapping network, a second mapping network, and a dynamic collaborative correction unit.

[0070] Specifically, the first step is to divide the flow according to the strength of physical laws: The shunting unit categorizes standardized fused feature data into strong physical law feature data and weak physical law feature data according to the strength of physical laws. Data such as voltage, current, node injected power, branch power flow, and system frequency are classified as strong physical law feature data because they strictly obey Kirchhoff's laws, power flow equations, and generator swing equations. Data such as ambient temperature, light intensity, wind speed, and user behavior characteristics are classified as weak physical law feature data because they lack precise physical equation descriptions.

[0071] A first mapping network embedding physical constraints is used to process data with strong physical laws: the first output of the splitting unit is connected to the input of the first mapping network. The splitting unit sends the data with strong physical laws to the first mapping network, which performs mapping processing on the data with embedded physical constraints. The first mapping network is a mapper based on a physical information neural network. The network structure of the physical information neural network adopts a multi-layer fully connected neural network with the Swish activation function, a network depth of 8 layers, and 128 neurons in each layer. The loss function of the physical information neural network consists of a weighted sum of a data loss term and a physical loss term. The total loss function is: ; in The data loss term represents the mean square error between the output value of the physical information neural network mapping and the actual measurement value. This is the physical loss term, representing the degree to which the mapped output value violates the laws of physics; These are the physical constraint weighting coefficients.

[0072] The loss function of the physical information neural network embeds the power flow equations and generator swing equations. The physical loss term L_physics is embedded in the residuals of the power flow equations. For each node i in the power grid, the active power imbalance is... and reactive power imbalance The calculation expression is: ; ; in and The active power and reactive power are injected into the power source at node i, respectively. and Let represent the active power and reactive power absorbed by the load at node i, respectively. and Let be the voltage magnitudes at nodes i and j, respectively. The voltage phase angle difference between node i and node j and Let represent the mutual conductance and mutual susceptance between nodes i and j in the nodal admittance matrix, respectively, and N(i) be the set of nodes directly connected to node i. The physical loss term... .

[0073] By using the residuals of the aforementioned power flow equations as soft constraints for the physical information neural network, the mapping output of the physical information neural network tends to satisfy Kirchhoff's laws under any operating condition.

[0074] A data-driven second mapping network is used to process weak physical law feature data: the second output of the splitting unit is connected to the input of the second mapping network. The splitting unit feeds the weak physical law feature data to the second mapping network, which performs data-driven mapping processing on the weak physical law feature data. The second mapping network is a mapper based on a deep residual network. The network structure of the deep residual network contains several residual blocks. Each residual block contains two fully connected layers and skip connections. The skip connections directly add the input of the residual block to the output of the residual block. The activation function is ReLU. The network depth is 6 residual blocks, and the number of neurons in each layer is 64. The deep residual network learns the nonlinear mapping relationship between weak physical law feature data and information space states in a purely data-driven manner.

[0075] The outputs of the first and second mapping networks are fused using a dynamic collaborative correction step: The outputs of the first mapping network and the second mapping network are connected to the input of the dynamic collaborative correction unit. The first mapping result output by the first mapping network and the second mapping result output by the second mapping network are input to the dynamic collaborative correction unit. The dynamic collaborative correction unit integrates the outputs of the first mapping network and the second mapping network to obtain high-precision digital twin state data in the information space.

[0076] The dynamic collaborative correction stage includes a residual monitoring gating unit and an online extreme learning machine. The first input of the residual monitoring gating unit is connected to the output of the first mapping network, the second input of the residual monitoring gating unit is connected to the output of the second mapping network, and the third input of the residual monitoring gating unit is used to receive actual measurement data.

[0077] The residual monitoring gating unit calculates in real time the first mapping deviation between the first mapping result output by the first mapping network and the actual measurement data, and the second mapping deviation between the second mapping result output by the second mapping network and the actual measurement data. The output of the residual monitoring gating unit is connected to the input of the online extreme learning machine, and the first mapping deviation and the second mapping deviation are sent to the online extreme learning machine.

[0078] The online extreme learning machine fits the residual signal based on the first mapping bias and the second mapping bias. The online extreme learning machine is a single hidden layer feedforward neural network. Its hidden layer weights and biases are randomly generated and then fixed. The output weights are updated online through the recursive least squares method. The standardized fused feature data at the current time is used as the input of the online extreme learning machine, and the mapping bias is used as the target output of the online extreme learning machine. The online extreme learning machine fits a residual estimate.

[0079] The first feedback output of the online extreme learning machine is connected to the boundary condition correction input of the first mapping network, and the residual signal is backpropagated to the first mapping network and the second mapping network to realize dynamic collaborative correction between the first mapping network and the second mapping network. The residual estimate is fed back to the first mapping network and the second mapping network according to a preset allocation ratio. The ratio allocated to the first mapping network is α, and the ratio allocated to the second mapping network is (1-α). The residual signal allocated to the first mapping network is converted into the adjustment amount of the weight coefficient λ of the physical loss term of the physical information neural network. The residual signal allocated to the second mapping network is used to fine-tune the output bias of the last fully connected layer of the deep residual network.

[0080] Through the above-mentioned physical law strong and weak current separation processing and dynamic collaborative correction fusion, high-precision digital twin state data in the information space is finally obtained. The high-precision digital twin state data includes at least the voltage amplitude and phase angle of each node in the entire network, the active and reactive power flow of each branch, the power angle and speed of each generator, the SOC of each energy storage system, the frequency of the entire network and its rate of change, and the predictive response behavior characteristics of each market actor.

[0081] The digital mirror model system is updated based on high-precision digital twin status data: electrical quantities in the high-precision digital twin status data are written into the operational digital mirror layer; multi-agent behavior prediction results are written into the interactive digital mirror layer; and equipment aging status updates are written into the equipment digital mirror layer.

[0082] Next, the control target instruction is obtained, and the downlink inversion processing from information space to physical space is performed.

[0083] Specifically, the system acquires control target instructions from the information space. These instructions originate from a global situational awareness and optimization decision-making system deployed in the cloud and are transmitted in a structured data format, including target type fields, target value fields, target area fields, time scale fields, and priority fields.

[0084] Based on the updated digital mirror model system, the downlink inversion process from the information space to the physical space of the control target command execution is performed. The downlink inversion process includes the following sub-steps: (1) Decomposition of control target instructions.

[0085] The control target command is input into the command semantic parsing engine. The command semantic parsing engine connects to the updated digital mirror model system. Based on the static parameters and current operating status of each terminal device in the device digital mirror layer, the command semantic parsing engine decomposes the control target command into device-level control commands for each terminal device.

[0086] Taking the control target of "raising the overall voltage of a certain feeder by 1%" as an example, the decomposition process of the instruction semantic parsing engine is as follows: First, extract the node-branch topology of the feeder and the voltage-reactive power sensitivity matrix of each node from the equipment digital mirror layer. Second, based on the voltage-reactive power sensitivity matrix, use the least squares method to solve the reactive power injection increment required by each reactive power source node. Third, introduce the multi-agent behavior constraints of the interactive digital mirror layer, query the predicted response behavior model of the subject to which each reactive power source device belongs, evaluate the acceptance probability and expected response deviation of each subject to the instruction, and for subjects with an acceptance probability lower than a preset threshold, transfer the reactive power increment part allocated to the subject with a higher acceptance probability. After the transfer, verify again whether the allocation of each subject is within the rated capacity constraint of its equipment digital mirror layer. Fourth, generate equipment-level physical control instructions.

[0087] (2) Virtual simulation and deduction.

[0088] Device-level control commands are used as physical control commands. These physical control commands are then substituted back into the running digital mirror layer of the updated digital mirror model system. Virtual simulation is performed in the running digital mirror layer to generate simulation results. The control quantities of each device in the physical control commands are used as forced injection quantities and substituted into the mechanism model of the device digital mirror layer. The Newton-Raphson power flow calculation method is used to perform steady-state power flow simulation. The simulation results include the voltage amplitude, voltage phase angle, power flow of each branch, and total network loss of the entire network after the command is executed.

[0089] (3) Feasibility verification and correction.

[0090] The simulation results are input into the feasibility verifier, which is connected to the device digital mirror layer. The feasibility verifier reads the operational constraint boundaries of each terminal device from the device digital mirror layer and determines whether the simulation results exceed the operational constraint boundaries.

[0091] Feasibility verification includes constraint checks in the following three dimensions: (1) Voltage constraints: Whether the voltage amplitude of each node is within the allowable range after the deduction; (2) Equipment constraints: Whether the output of each device is within its rated capacity limit and whether the energy storage SOC is within a reasonable range; (3) Frequency constraints: If active power adjustment is involved, calculate whether the steady-state frequency is within the allowable range.

[0092] When the simulation results show that the physical control command exceeds the operational constraint boundary, the command correction algorithm is automatically triggered to correct the physical control command, generate the corrected physical control command, and substitute the corrected physical control command back to the running digital mirror layer for a second simulation until the simulation result meets the operational constraint boundary. The physical control command that meets the operational constraint boundary is then output as the execution control command.

[0093] The instruction correction algorithm takes minimizing the deviation of the control target as the objective function and uses equipment constraints and system constraints as constraints. It uses a sequential quadratic programming method to iteratively adjust the control quantities of each device. If a feasible solution that satisfies all constraints is found within the preset maximum number of iterations, the solution is used as the corrected physical control instruction output. If no feasible solution can be found, a "instruction not feasible" signal is returned to the cloud.

[0094] Once the verification is successful or the correction is completed, an execution control command is output and sent to the corresponding terminal device for execution via the communication network.

[0095] Finally, the actual response data is obtained, and bidirectional closed-loop error comparison and online updates are performed.

[0096] Acquire the actual response data of the executed control commands in the actual physical space. The actual response data includes: the actual control value executed by the terminal device, the actual execution timestamp, the execution result status code, and the actual system status measured by SCADA / PMU after execution.

[0097] The actual response data is compared with the high-precision digital twin state data using a two-way closed-loop error comparison, which includes two independent error comparison links: Link 1, State Prediction Error Comparison: Extract the state quantities directly related to the execution of control commands from the high-precision digital twin state data as predicted values, and use the actual system state measured by SCADA / PMU after execution as the actual values.

[0098] Calculate the state error vector between the predicted value and the actual value.

[0099] Link 2, Instruction Execution Error Comparison: Extract the control target values ​​of each device in the execution control instructions as instruction values, and use the actual control values ​​executed by the terminal devices as execution values.

[0100] Calculate the instruction error vector between the instruction value and the executed value.

[0101] The state error vector and the command error vector are concatenated into a comprehensive error vector. A residual compensation signal is generated based on the comparison results, and the recursive least squares method is used to process the comprehensive error vector. Online identification is performed to extract the systematic bias components. The recursive formula for the recursive least squares method is as follows: ; ; ; Where t is the discrete-time index, K(t) is the gain vector, and P(t) is the covariance matrix. Let μ be the input feature vector, and μ be the forgetting factor. To identify the system deviation parameter vector, It is mapped to the residual compensation signal C_comp.

[0102] The mapping correction parameters of the dynamic collaborative correction stage are updated online using the residual compensation signal, and the residual compensation signal C_comp is distributed according to the following path: (1) Feed the physical deviation components in C_comp back to the first mapping network to adjust the key physical parameters in the boundary conditions of the physical information neural network; (2) Feed the behavioral bias component in C_comp back to the second mapping network as a new training sample input to the online extreme learning machine and update the output weight matrix of the online extreme learning machine; (3) Feed back the integrated bias component in C_comp to the running digital mirror layer to correct the model noise covariance matrix and measurement noise covariance matrix of the extended Kalman filter.

[0103] Finally, the specific embodiments of this application are supplemented as follows: The experimental scenario was set as a feeder of an actual 10kV distribution network, including 12 nodes, 3 distributed photovoltaic inverters, 1 energy storage converter, and several load nodes. The experiment was divided into four operating conditions: Operating condition 1 was the design condition; Operating condition 2 was the sudden irradiation condition; Operating condition 3 was the sudden load increase condition; and Operating condition 4 was the combined extreme condition (the sudden irradiation and the sudden load increase occurred simultaneously, and the actual parameters of the equipment deviated from the nameplate values ​​by 15%-20%).

[0104] Under the above four operating conditions, the mapping accuracy of the following three mapping methods was tested respectively (with the root mean square error (RMSE) of voltage amplitude mapping as the evaluation index): Method A is a pure mechanism model method; Method B is a pure data-driven method; Method C is the method of this invention (including dynamic collaborative correction).

[0105] The experimental results are shown in the table below: Experimental results show that under the design conditions, the mapping accuracy of the three methods is not significantly different. However, in scenarios deviating from the design conditions, such as conditions two, three, and four, the pure mechanistic model method suffers a sharp increase in error due to parameter mismatch (RMSE up to 0.089 pu), while the pure data-driven method suffers a decrease in reliability due to the training data not covering extreme conditions (RMSE up to 0.071 pu). In contrast, the method of this invention maintains high mapping accuracy in all four conditions (RMSE stable within the range of 0.008-0.015 pu) through dynamic collaborative correction of physical law strength-based processing, PINN mapping embedded with physical constraints, and data-driven ResNet mapping.

[0106] Furthermore, this embodiment also verifies the technical effectiveness of the bidirectional closed-loop error comparison and online update mechanism of the present invention through simulation experiments.

[0107] The experimental setup is as follows: On the distribution network feeder model, a continuous 24-hour operation scenario is simulated, including the daily variation of light intensity, the daily variation curve of load, and two random disturbance events. The operation scenario is divided into two groups: the first group (control group) disables the online update function of this application; the second group (experimental group) enables all functions. Both groups of experiments use the uplink mapping method of this invention to record the voltage amplitude mapping RMSE of the two groups of experiments at each moment within 24 hours, and sample once every 5 minutes.

[0108] Control group: mean RMSE 0.032 pu, maximum RMSE 0.071 pu (occurred after the irradiation mutation at 15:00), standard deviation of RMSE 0.019 pu.

[0109] Experimental group: RMSE mean 0.011pu, RMSE maximum 0.018pu (occurred at the initial moment of the disturbance, and quickly recovered to the mean level within 2 minutes), RMSE standard deviation 0.004pu.

[0110] Experimental results show that after enabling the bidirectional closed-loop error comparison and online update function, the mean, maximum, and standard deviation of the mapping accuracy are significantly better than those of the control group. In particular, after sudden changes in irradiation and load, the experimental group updates the mapping correction parameters of the dynamic collaborative correction link online through the residual compensation signal, enabling the mapping accuracy to quickly recover to the steady-state level after the disturbance (recovery time is about 2 minutes). In contrast, the control group, lacking an online update mechanism, has a persistently low mapping accuracy after the disturbance. The above experimental results verify the significant technical effect of the bidirectional closed-loop error comparison and online update mechanism of this invention in improving the adaptability and robustness of the digital twin system.

[0111] In summary, compared with the prior art, the above embodiments of this application have the following beneficial effects: Real-time acquisition of multi-source heterogeneous data and equipment parameters of the power system, and construction of a digital mirror model based on the equipment parameters, provides a digital foundation with physical structural constraints for subsequent mapping and control, ensuring the interpretability of state estimation within the physical boundaries of the equipment; furthermore, by dividing multi-source heterogeneous data into strongly regular data and weakly regular data according to the strength of physical laws, and mapping them respectively using a first neural network embedded with physical constraints and a data-driven second neural network, the electrical measurement data that obeys the power grid physical equations is constrained by physical laws and maintains consistency during the mapping process. Simultaneously, non-electrical data representing the behavior of environmental and market actors adapts to uncertainty through a flexible data-driven approach, thus balancing the physical reliability and data adaptability of the modeling; after collaborative correction, the digital twin estimated state is obtained and the digital mirror model is updated. Dynamic complementary correction is achieved by fusing the deviation information of the two types of mapping outputs, suppressing the deviation of a single modeling paradigm in operating condition or sample deviation. The system avoids performance degradation when the conditions are insufficient. Simultaneously, by acquiring control target commands and performing virtual simulations based on the updated digital twin model, the high-level control intentions can be pre-enacted in a high-fidelity digital space, generating executable control commands before being sent to the physical system. This avoids unverified commands directly impacting the power grid and causing over-limit or instability. Furthermore, by comparing response data with the estimated state of the digital twin in both directions, residual compensation signals are generated and the parameters in the model used to correct mapping deviations are updated. This achieves joint perception and feedback correction of state prediction deviations and command execution deviations, enabling the digital twin model to continuously correct its mapping deviations based on operational feedback and maintain state consistency over long-term operation. These features work synergistically to construct a complete link: a high-fidelity uplink mapping from physical to information, a feasible downlink inversion from information to physical, and a two-way closed-loop feedback correction. This solves the problems of insufficient control accuracy and unexecutable commands caused by the lack of a two-way closed-loop mapping mechanism between physical and information spaces in existing digital twin technologies.

[0112] Example 2: Please refer to Figure 2 Based on the same inventive concept, the present invention discloses a power system digital twin bidirectional mapping construction system, comprising: a data acquisition module M1, a first update module M2, a response module M3, and a second update module M4; The data acquisition module M1 is used to collect multi-source heterogeneous data and equipment parameters of the power system in real time, and to construct a digital mirror model based on the equipment parameters. The first update module M2 is used to divide the multi-source heterogeneous data into strong regularity data and weak regularity data according to the strength of physical laws. The first neural network with embedded physical constraints and the second neural network driven by data are used to map the strong regularity data and weak regularity data respectively. Then, the digital twin estimation state is obtained through collaborative correction and the digital mirror model is updated to obtain the updated digital mirror model. Furthermore, the first update module M2 includes: a spatiotemporal processing unit and a splitting unit; The spatiotemporal processing unit is used to perform spatiotemporal alignment processing on the multi-source heterogeneous data to generate standardized fused feature data. The current splitting unit is used to treat electrical measurement data that obey the power grid physical equations in the standardized fusion feature data as strong regularity data, and non-electrical measurement data that characterize environmental factors and market behavior as weak regularity data in the standardized fusion feature data.

[0113] In this preferred embodiment, spatiotemporal alignment processing of multi-source heterogeneous data is performed to generate standardized fused feature data, eliminating the dimensional mismatch in time sampling frequency and spatial measurement location of data from the source, grid, load, and storage sides, providing a consistent input for subsequent strong and weak current distribution according to physical laws. Furthermore, electrical measurement data that obeys the physical equations of the power grid in the standardized fused feature data are regarded as strong regularity data, and non-electrical measurement parts that characterize environmental factors and market behavior are regarded as weak regularity data, clarifying the technical basis of the current distribution criterion, giving subsequent differentiated modeling a clear physical boundary, and improving the rationality and feasibility of the mapping strategy.

[0114] Furthermore, the first neural network is a multilayer fully connected neural network trained with a loss function constructed by weighting data loss terms and physical loss terms; the second neural network is a deep residual network containing skip connections; wherein, the data loss term is an error measure between the mapped output value and the actual measured value; the physical loss term is characterized by calculating the sum of the squares of the active power imbalance and reactive power imbalance in the power flow equation of the power grid.

[0115] In this preferred embodiment, the data loss term ensures fitting accuracy by measuring the error between the mapped output value and the actual measured value, while the physical loss term quantifies the degree of violation of Kirchhoff's laws by calculating the sum of the squares of the active power imbalance and reactive power imbalance in the power grid power flow equation. The weighted synergy of the two terms enables the network output to approximate the measured data while forcibly satisfying the basic physical constraints of the power grid. The second neural network is a deep residual network containing skip connections. The residual structure alleviates the gradient vanishing problem and improves the ability to represent complex nonlinear relationships in weakly regular data, thereby enhancing the robustness and generalization performance of the overall mapping model.

[0116] Furthermore, the first update module M2 also includes: a mapping unit, a deviation unit, a fitting unit, and a compensation fusion unit; The mapping unit is used to input the strong regularity data into the first neural network to calculate the first state component representing the electrical state of the power grid, and to input the weak regularity data into the second neural network to calculate the second state component representing the environmental and market behavior state. The deviation unit is used to calculate the first deviation between the first state component and the corresponding electrical measurement value in the standardized fusion feature data, and the second deviation between the second state component and the corresponding environmental and behavioral measurement value in the standardized fusion feature data. The fitting unit is used to obtain the residual signal by using an online extreme learning machine and recursive least squares method based on the first deviation, the second deviation and the standardized fusion feature data. The compensation fusion unit is used to compensate and fuse the first state component and the second state component according to the residual signal, and output the digital twin estimated state.

[0117] In this preferred embodiment, strong regularity data is input into a first neural network to calculate a first state component representing the electrical state of the power grid, and weak regularity data is input into a second neural network to calculate a second state component representing the environmental and market behavior states. This achieves decoupled modeling and independent output of the two types of states. Furthermore, by calculating the first deviation between the first state component and the corresponding electrical measurement value, and the second deviation between the second state component and the corresponding environmental and behavioral measurement values, the local errors of the two models are quantified respectively. Based on the two types of deviations and standardized fused feature data, an online extreme learning machine is used and a recursive least squares method is applied to fit the residual signal, enabling the correction process to have real-time performance and lightweight parameters. Finally, the two state components are compensated and fused according to the residual signal, and a digital twin estimated state is output. This achieves dynamic collaborative fusion of the two heterogeneous states under error-driven conditions, improving the accuracy and consistency of the overall state estimation.

[0118] The response module M3 is used to acquire the control target instruction, perform virtual simulation and deduction of the control target instruction based on the updated digital mirror model, generate executable control instructions and send them to the power system for execution, and obtain response data. Furthermore, the response module M3 includes: a decomposition unit, a query unit, and a correction unit; The decomposition unit is used to decompose the control target instruction into preliminary control instructions for each terminal device based on the control target instruction and the static parameters and current operating status of each terminal device in the digital mirror model. The query unit is used to perform a prediction query on each corresponding preliminary control command based on the preset response behavior prediction model of each terminal device to which the market behavior entity belongs in the digital mirror model, so as to obtain the acceptance probability and expected response deviation of each market behavior entity to the preliminary control command. The correction unit is used to correct each of the preliminary control commands based on the acceptance probability and expected response deviation, and generate each device-level control command as the executable control command.

[0119] In this preferred embodiment, the control target instruction is decomposed into preliminary control instructions based on the static parameters and current operating status of each terminal device in the digital mirror model, ensuring that the instruction decomposition conforms to the physical capability boundaries of the devices. Furthermore, the preliminary control instructions are predicted and queried based on the preset response behavior prediction model of the market actors to which each terminal device belongs, obtaining the acceptance probability and expected response deviation of each actor, explicitly incorporating the uncertainty of multi-agent behavior into the instruction generation process. Then, the preliminary control instructions are corrected based on the acceptance probability and expected response deviation to generate device-level control instructions as executable control instructions, making the final issued instructions closer to the actual execution intentions and capabilities, thus improving the success rate and reliability of the control instructions.

[0120] Furthermore, the response module M3 also includes: a deduction unit and an adjustment unit; The deduction unit is used to substitute each of the device-level control commands back into the digital mirror model to perform steady-state power flow deduction and generate the system steady-state state. The adjustment unit is used to read the operational constraint boundaries of each terminal device from the digital mirror model, determine whether the steady-state state of the system meets the operational constraint boundaries; if not, with the minimum deviation of the control target as the objective and the operational constraint boundaries as the constraint conditions, the control quantities of each device-level control instruction are iteratively adjusted using an optimization algorithm, and the adjusted control quantities are substituted back into the digital mirror model for secondary deduction until the steady-state state of the system meets the operational constraint boundaries, and the adjusted device-level control instruction is output as the final executable control instruction.

[0121] In this preferred embodiment, control commands at each device level are substituted back into the digital mirror model for steady-state power flow simulation to generate a system steady-state state and predict the grid operation conditions after command execution. Furthermore, the operational constraint boundaries of each terminal device are read from the digital mirror model, and it is determined whether the system steady-state state meets these boundaries, thus performing a preliminary verification of the command's feasibility. If not, the control quantity of each device-level control command is iteratively adjusted using an optimization algorithm with the goal of minimizing the deviation of the control target and the operational constraint boundaries as constraints. The adjusted control quantity is then substituted back into the model for a second simulation until the constraints are met, outputting the final executable control command. This ensures that the issued commands are strictly executable in the physical system and do not trigger grid safety limits, enhancing the safety and robustness of the control.

[0122] The second update module M4 is used to perform a bidirectional error comparison between the response data and the estimated state of the digital twin, generate a residual compensation signal, and update the parameters used to correct the mapping deviation in the updated digital mirror model to obtain a digital mirror model with corrected parameters.

[0123] Furthermore, the digital mirror model includes a runtime layer, which dynamically updates the system state through a recursive state estimation algorithm; The second update module M4 includes: a prediction error calculation unit, an execution error calculation unit, a splicing unit, and a parameter update unit; The prediction error calculation unit is used to extract state quantities directly related to the executable control instructions from the digital twin estimated state as predicted values, extract actual measurement data from the response data as actual values, and calculate the state prediction error based on each predicted value and actual value. The execution error calculation unit is used to obtain the actual control quantity of each terminal device in the power system when executing executable control instructions as the execution value, extract the control target value of each terminal device from the executable control instructions as the instruction value, and calculate the instruction execution error based on each execution value and instruction value. The splicing unit is used to splice the state prediction errors and instruction execution errors to obtain a comprehensive error vector. The parameter update unit is used to generate a residual compensation signal based on the comprehensive error vector, and update the weight coefficients of the physical constraints in the first neural network, the output bias parameters in the second neural network, and the parameters of the recursive state estimation algorithm based on the residual compensation signal to obtain a digital mirror model with corrected parameters.

[0124] In this preferred embodiment, state variables directly related to executable control commands are extracted from the digital twin estimated state as predicted values, and actual measurement data are extracted from the response data as actual values. The state prediction error is calculated to reflect the uplink mapping deviation. Simultaneously, the actual control variables executed by the terminal device are obtained as execution values, and control target values ​​are extracted from the executable control commands as command values. The command execution error is calculated to reflect the downlink execution deviation. The two types of errors are then concatenated to obtain a comprehensive error vector, which comprehensively characterizes the overall mismatch degree of the bidirectional mapping. Finally, a residual compensation signal is generated based on the comprehensive error vector, and the physical constraint weight coefficients in the first neural network, the output bias parameters in the second neural network, and the parameters of the recursive state estimation algorithm are updated accordingly. This achieves multi-level synchronous correction of model parameters and improves the adaptive accuracy and stability of the system under long-term operation.

[0125] In summary, compared with existing technologies, the embodiments of this application have the following beneficial effects: Real-time acquisition of multi-source heterogeneous data and equipment parameters of the power system, and construction of a digital mirror model based on the equipment parameters, provides a digital foundation with physical structural constraints for subsequent mapping and control, ensuring the interpretability of state estimation within the physical boundaries of the equipment; furthermore, by dividing multi-source heterogeneous data into strongly regular and weakly regular data according to the strength of physical laws, and mapping them respectively using a first neural network embedded with physical constraints and a data-driven second neural network, the electrical measurement data that obeys the power grid physical equations is constrained by physical laws and maintains consistency during the mapping process. Simultaneously, non-electrical data representing the behavior of environmental and market actors adapts to uncertainty through a flexible data-driven approach, thus balancing the physical reliability and data adaptability of the model; further, through collaborative correction, a digital twin estimated state is obtained and the digital mirror model is updated. Dynamic complementary correction is achieved by fusing the deviation information of the two types of mapping outputs, suppressing the deviation of a single modeling paradigm due to operating condition shifts or sample inconsistencies. The system effectively mitigates performance degradation over time. Simultaneously, by acquiring control target commands and performing virtual simulations based on the updated digital twin model, the high-level control intentions can be pre-enacted in a high-fidelity digital space, generating executable control commands before being sent to the physical system. This avoids unverified commands directly impacting the power grid and causing over-limit or instability. Furthermore, by comparing response data with the estimated state from the digital twin in both directions, residual compensation signals are generated and the parameters in the model used to correct mapping deviations are updated. This achieves joint perception and feedback correction of state prediction deviations and command execution deviations, enabling the digital twin model to continuously correct its mapping deviations based on operational feedback and maintain state consistency over long-term operation. These features work synergistically to construct a complete link: a high-fidelity uplink mapping from physical to information, a feasible downlink inversion from information to physical, and a two-way closed-loop feedback correction. This solves the problems of insufficient control accuracy and unexecutable commands caused by the lack of a two-way closed-loop mapping mechanism between physical and information spaces in existing digital twin technologies.

[0126] Example 3: This invention also provides a computer program product, including a computer program or instructions, capable of running on a computing device or stored in any available medium. When the computer program product is run on at least one computing device, it causes the at least one computing device to execute any of the power system digital twin bidirectional mapping construction methods of this invention.

[0127] Example 4: This invention also provides a computer-readable storage medium storing at least one executable instruction. When the executable instruction is executed on a power system digital twin bidirectional mapping construction system, it causes the power system digital twin bidirectional mapping construction system to perform one of the power system digital twin bidirectional mapping construction methods described in any of the above method embodiments.

[0128] The specific working processes of each module described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. The division of modules is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules may be combined or integrated into another system.

[0129] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for constructing a bidirectional mapping for a digital twin of a power system, characterized in that, include: Real-time acquisition of multi-source heterogeneous data and equipment parameters of the power system, and construction of a digital mirror model based on the equipment parameters; The multi-source heterogeneous data is divided into strong-law data and weak-law data according to the strength of physical laws. The strong-law data and weak-law data are mapped by a first neural network embedded with physical constraints and a second data-driven neural network, respectively. The digital twin estimation state is obtained through collaborative correction and the digital mirror model is updated to obtain the updated digital mirror model. The control target instruction is obtained, and the control target instruction is virtually simulated and deduced based on the updated digital mirror model to generate executable control instructions and send them to the power system for execution, thereby obtaining response data. The response data is compared with the estimated state of the digital twin in two directions to generate a residual compensation signal and update the parameters used to correct the mapping deviation in the updated digital mirror model, thus obtaining a digital mirror model with corrected parameters.

2. The method for constructing a bidirectional digital twin mapping for a power system as described in claim 1, characterized in that, The step of dividing the multi-source heterogeneous data into strongly regular data and weakly regular data according to the strength of physical laws includes: The multi-source heterogeneous data is subjected to spatiotemporal alignment processing to generate standardized fused feature data; Electrical measurement data that conform to the power grid physical equations in the standardized fusion feature data are considered as strong regularity data, while non-electrical measurement data that characterize environmental factors and the behavior of market participants are considered as weak regularity data.

3. The method for constructing a bidirectional digital twin mapping for a power system as described in claim 1, characterized in that, The first neural network is a multilayer fully connected neural network trained with a loss function constructed by weighting data loss terms and physical loss terms; the second neural network is a deep residual network containing skip connections; wherein, the data loss term is an error measure between the mapped output value and the actual measured value; the physical loss term is characterized by calculating the sum of the squares of the active power imbalance and reactive power imbalance in the power flow equation of the power grid.

4. The method for constructing a bidirectional digital twin mapping for a power system as described in claim 2, characterized in that, The process of mapping the strong-pattern data and weak-pattern data using a first neural network with preset embedded physical constraints and a second data-driven neural network, respectively, and then obtaining the digital twin estimated state through collaborative correction, includes: The strong regularity data is input into the first neural network to calculate the first state component representing the electrical state of the power grid, and the weak regularity data is input into the second neural network to calculate the second state component representing the environmental and market behavior state. Calculate the first deviation between the first state component and the corresponding electrical measurement value in the standardized fusion feature data, and the second deviation between the second state component and the corresponding environmental and behavioral measurement values ​​in the standardized fusion feature data; Based on the first bias, the second bias, and the standardized fusion feature data, the residual signal is obtained by using an online extreme learning machine and fitting it using the recursive least squares method. The first state component and the second state component are compensated and fused according to the residual signal to output the estimated state of the digital twin.

5. The method for constructing a bidirectional digital twin mapping for a power system as described in claim 1, characterized in that, The virtual simulation and deduction of the control target command based on the updated digital mirror model generates executable control commands, including: Based on the control target instruction, and combined with the static parameters and current operating status of each terminal device in the digital mirror model, the control target instruction is decomposed into preliminary control instructions for each terminal device. Based on the preset response behavior prediction model of each terminal device in the digital mirror model, the corresponding preliminary control instructions are predicted and queried to obtain the acceptance probability and expected response deviation of each market entity to the preliminary control instructions. Based on the acceptance probability and expected response deviation, the preliminary control instructions are modified to generate device-level control instructions, which are then used as executable control instructions.

6. The method for constructing a bidirectional digital twin mapping for a power system as described in claim 5, characterized in that, After generating the control commands for each device level, the process includes: Each of the device-level control commands is substituted back into the digital mirror model to perform steady-state power flow deduction and generate the system steady-state state. The operating constraint boundaries of each terminal device are read from the digital mirror model. It is determined whether the steady-state state of the system satisfies the operating constraint boundaries. If not, the control quantity of each device-level control instruction is iteratively adjusted using an optimization algorithm with the goal of minimizing the deviation of the control target and the operating constraint boundaries as constraints. The adjusted control quantities are then substituted back into the digital mirror model for secondary deduction until the steady-state state of the system satisfies the operating constraint boundaries. Finally, the adjusted device-level control instruction is output as the final executable control instruction.

7. The method for constructing a bidirectional digital twin mapping for a power system as described in claim 1, characterized in that, The digital mirror model includes a runtime layer, which dynamically updates the system state through a recursive state estimation algorithm. The step of performing a bidirectional error comparison between the response data and the estimated digital twin state, generating a residual compensation signal, and updating the parameters used to correct the mapping bias in the updated digital mirror model to obtain a parameter-corrected digital mirror model includes: The state quantities directly related to the executable control instructions are extracted from the estimated state of the digital twin as predicted values, and the actual measurement data are extracted from the response data as actual values. Based on each predicted value and actual value, the state prediction error is calculated. The actual control quantity of each terminal device in the power system when executing executable control instructions is obtained as the execution value. The control target value of each terminal device is extracted from the executable control instructions as the instruction value. Based on the execution value and the instruction value, the instruction execution error is calculated. By concatenating the state prediction errors and instruction execution errors, a comprehensive error vector is obtained. Based on the comprehensive error vector, a residual compensation signal is generated, and based on the residual compensation signal, the weight coefficients of the physical constraints in the first neural network, the output bias parameters in the second neural network, and the parameters of the recursive state estimation algorithm are updated to obtain a digital mirror model with corrected parameters.

8. A power system digital twin bidirectional mapping construction system, characterized in that, include: Data acquisition module, first update module, response module, and second update module; The data acquisition module is used to collect multi-source heterogeneous data and equipment parameters of the power system in real time, and to construct a digital mirror model based on the equipment parameters. The first update module is used to divide the multi-source heterogeneous data into strong regularity data and weak regularity data according to the strength of physical laws, and use a preset first neural network embedded with physical constraints and a data-driven second neural network to map the strong regularity data and weak regularity data respectively, and then obtain the digital twin estimation state through collaborative correction and update the digital mirror model to obtain the updated digital mirror model. The response module is used to acquire the control target instruction, perform virtual simulation and deduction of the control target instruction based on the updated digital mirror model, generate executable control instructions and send them to the power system for execution, and obtain response data. The second update module is used to perform a bidirectional error comparison between the response data and the estimated state of the digital twin, generate a residual compensation signal, and update the parameters used to correct the mapping deviation in the updated digital mirror model to obtain a digital mirror model with corrected parameters.

9. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed, they implement a method for constructing a bidirectional digital twin mapping of a power system as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements a method for constructing a bidirectional digital twin mapping of a power system as described in any one of claims 1-7.

Images