Self-adaptive double-layer simulation method and system based on intelligent optimization and residual correction

Abstract

The invention discloses a self-adaptive double-layer simulation method and system based on intelligent optimization and residual correction, and belongs to the technical field of system simulation. The method comprises the following steps: receiving a simulation task and intelligently deciding a simulation strategy; when the simulation strategy decision is double-layer simulation, global scanning and automatic optimization are carried out through first-layer rapid simulation, and a key area is identified; performing refined simulation on the key area by fusing a physical information neural network through second-layer precise simulation; and finally fusing two layers of results. In addition, the system is provided with an outer ring residual error correction link, and model parameters are corrected online and model drift is resisted by comparing a simulation result with real data. Through double-layer cooperation and closed-loop correction, the contradiction between precision and efficiency in the simulation process is effectively solved, the self-adaptive capacity and long-term reliability of simulation are improved, and the method is particularly suitable for full-life-cycle simulation of complex systems such as a ship power system.

Description

Technical Field

[0001] This invention belongs to the field of system simulation technology, specifically relating to a two-layer simulation framework that combines rapid simulation and accurate simulation, and a simulation method and system for ship power systems that integrates physical information neural network (PINN) and automatic residual correction. Background Technology

[0002] Marine propulsion systems are the core component of marine power plants, and their simulation technology is crucial for system design, fault diagnosis, and operational optimization. Traditional system simulation methods mainly employ a single simulation mode, either based on rapid simulation of simplified models (such as linear models or empirical formulas) or on accurate simulation based on high-fidelity models (such as computational fluid dynamics simulation).

[0003] The existing technical solutions are mainly as follows: The first method is a dynamic simulation model development approach for heat exchangers based on Modelica. This method achieves dynamic simulation through parameter initialization and calculation modules (including calculation units such as the main condensation zone and heat exchanger), but it lacks an adaptive switching mechanism between fast and accurate simulation, making it difficult to balance the requirements of real-time performance and accuracy.

[0004] The second approach involves developing dynamic simulation models for heat exchangers with multiple structures and pressure nodes. This method performs simulations by defining structural parameters and operating conditions, but it does not incorporate intelligent algorithms for automatic parameter optimization, nor does it integrate physical constraints with data-driven simulation.

[0005] In summary, the existing technology has the following main drawbacks: (1) Conflict between real-time performance and accuracy: A single simulation mode cannot simultaneously meet the needs of real-time control and fine analysis. High-precision models (such as multi-physics coupled models that include fluid, heat and structure) take too long to calculate and cannot meet the needs of design iteration or real-time state prediction; while fast models are not accurate enough and may lead to design defects or inaccurate predictions.

[0006] (2) Lack of adaptive switching mechanism: Existing technologies require manual selection of simulation modes and cannot automatically switch simulation strategies according to scenarios (such as fault diagnosis and optimized operation). Especially in the rapid simulation layer, it is necessary to frequently target different operating conditions (such as full-speed navigation, low-speed cruising, and emergency reversing), and a single model is unlikely to maintain optimal performance under all operating conditions. Engineers need to select models based on experience, which is cumbersome and not optimal.

[0007] (3) Deviation between model and physical system (model drift): The simulation model and the real physical system will deviate due to factors such as equipment aging and environmental changes (e.g., noise, sea conditions, etc.), i.e. model drift. Traditional simulation does not consider the cumulative error caused by model drift and lacks an online correction mechanism, which causes the simulation results to gradually deviate from the true value over time.

[0008] Therefore, there is an urgent need for a simulation method and system that can adaptively switch, intelligently optimize, and correct errors online. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide an adaptive two-layer simulation method and system to solve the problems of difficulty in balancing simulation accuracy and efficiency and model selection relying on experience in the prior art.

[0010] To achieve the above objectives, the present invention adopts the following technical solution.

[0011] An adaptive two-layer simulation method based on intelligent optimization and residual correction includes the following steps: S100, Task Reception and Decision-Making: Receives simulation tasks, parses task constraints, and automatically decides on simulation strategies, including two-layer simulation. S200, Adaptive Switching and Two-Layer Simulation: Based on the simulation strategy decided in step S100, the corresponding simulation process is automatically activated. When the simulation strategy decision is two-layer simulation, steps S210 to S240 are executed, wherein... S210, First-level fast simulation: Build and run the first-level fast simulation, select the best fast model from multiple candidate fast models for global fast simulation, and obtain preliminary results; S220, Key Identification: Based on preset criteria, key areas or key time periods are identified from preliminary results; S230, Second-layer precise simulation: For key areas or key time periods, construct or call the second-layer precise simulation, use a high-precision model to perform local fine simulation, and obtain local high-precision results; S240, Result Fusion: The preliminary results from step S210 are fused with the local high-precision results from step S230 to generate a preliminary simulation report.

[0012] In some embodiments, the method further includes the following steps: S300, Residual Correction: Based on the residuals between the real system data and the preliminary simulation report, the simulation model parameters are corrected online, and the model library is updated.

[0013] In some embodiments, step S300 includes: S310, when sensor data from the actual ship propulsion system is available, use it as the true value of the reference system; S320, calculate the residuals between the corresponding physical quantities in the preliminary simulation report and the true values ​​of the reference system; S330, based on the residuals, automatically fine-tunes the parameters of the model in step S210 and / or step S230 through a correction algorithm; S340 updates the corrected model parameters to the model library.

[0014] In some embodiments, step S300 is executed periodically, or is triggered when new real system data is received.

[0015] In some embodiments, in step S100, the task constraints include at least one of scenario type, accuracy requirements, real-time requirements, and computing resource budget; automatic decision-making is achieved through a predefined decision rule base or a machine learning classifier.

[0016] In some embodiments, the method includes an automatic switching mechanism that triggers the operation mechanism based on task constraints.

[0017] In some embodiments, the simulation strategy further includes single fast simulation and / or single accurate simulation. The automatic switching mechanism operation mechanism includes: if the scenario type is fault diagnosis, the simulation strategy decision is two-layer simulation or single accurate simulation; if the scenario type is real-time control, the simulation strategy decision is single fast simulation; if the uncertainty of single fast simulation or first-layer fast simulation exceeds a preset accuracy threshold, the simulation strategy decision is single accurate simulation; if the system's required response time is lower than a preset real-time threshold, the simulation strategy decision is single fast simulation.

[0018] In some embodiments, step S210 includes: S211, calls multiple candidate fast models from a fast simulation algorithm set containing simplified models optimized for different operating conditions; S212 allows multiple candidate fast models to be run in parallel or in rapid serial execution within a reduced computational domain or time window. S213, automatically selects the optimal fast model based on preset fast evaluation criteria; S214 uses the optimal fast model to perform fast global and / or full-time simulations and obtain preliminary results.

[0019] In some embodiments, step S230 includes: S231, For key areas or key time periods, the high-precision model adopts a physical information neural network, which uses its control equations as constraints to embed the loss function. S232, the data provided by the first layer of rapid simulation in step S210 at the boundary of the key area or the start and end points of the time period are used as the boundary conditions or initial conditions of the physical information neural network. S233, train and / or fine-tune the physical information neural network to perform high-precision, meshless, fine-grained simulation of key areas or key time periods, and obtain local high-precision results.

[0020] The present invention also provides an adaptive two-layer simulation system, comprising: The task management and decision-making module is used to receive simulation tasks and automatically decide on simulation strategies. The adaptive switching module, connected to the task management and decision-making module, is used to automatically activate the corresponding simulation process based on the simulation strategy of the decision. A dual-layer simulation engine, connected to an adaptive switching module, is activated when the simulation strategy decision is dual-layer simulation. The dual-layer simulation engine includes a first-layer fast simulation module, a key identification module, a second-layer accurate simulation module, and a result fusion module. The first-level rapid simulation module is used to perform rapid simulation and automatic optimization. The key identification module is used to identify key areas or key time periods; The second-layer precise simulation module is used to perform precise simulations in key areas or key time periods. The results fusion module is used to fuse fast and accurate simulation results; as well as The residual correction module is used to calculate residuals and update simulation model parameters online. The model library, connected to the two-layer simulation engine and residual correction module, is used to store and manage simulation models and their parameters.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention achieves a collaborative mechanism between the first-layer fast simulation (automatic algorithm set optimization) and the second-layer accurate simulation (PINN fusion) through a two-layer simulation framework. It realizes adaptive switching logic triggered by multiple conditions (scene, accuracy, real-time performance) through an automatic switching mechanism. It realizes an online parameter correction method and residual correction network method based on reference system error through residual correction. This solves the contradiction between real-time performance and accuracy, improves simulation reliability, and realizes full automation, intelligence and adaptability of the simulation process.

[0022] (1) Improved accuracy: Through physical information neural network and residual correction mechanism, simulation error is significantly reduced, which can be reduced to 5% or even below 3% (10% for traditional methods).

[0023] (2) Balance between efficiency and accuracy: By adopting the strategy of "global fast scanning + local precise focusing" and introducing automatic optimization in the first layer of fast simulation, we can ensure that the near-optimal fast model is used under any working condition. The overall efficiency far exceeds that of traditional methods, while ensuring the high accuracy of key parts.

[0024] (3) Flexibility: Adapt to different working conditions through an automatic switching mechanism to avoid “over-simulation” or “under-simulation”.

[0025] (4) Engineering practicality: The unique residual correction process enables the model to resist model drift, solves the problem of model drift with changes in time and operating conditions, enables the simulation system to learn from real data and update itself, becoming a continuously evolving "digital twin", significantly improving the long-term reliability and stability of simulation, and is suitable for the whole life cycle management of ship power systems.

[0026] (5) Intelligence: The algorithm set automatically optimizes and uses PINN physical constraints, taking into account both data-driven and model reliability, and realizes full automation, intelligence and adaptability of the simulation process. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall process of the adaptive switching mechanism of the present invention; Figure 2 This is a flowchart illustrating one embodiment of the two-layer simulation process of the present invention; Figure 3 This is a flowchart illustrating an embodiment of the first-level rapid simulation; Figure 4 This is a flowchart illustrating an embodiment of the second-layer accurate simulation. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0029] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0030] like Figure 1 and Figure 2 As shown in some embodiments of this application, the present invention provides an adaptive two-layer simulation method based on intelligent optimization and residual correction, including step S100 task reception and decision-making, step S200 adaptive switching and two-layer simulation, and step S300 residual correction, as detailed below: S100, Task Reception and Decision Making.

[0031] The system receives simulation tasks and analyzes their scenario type (e.g., design optimization, fault diagnosis, real-time prediction), accuracy requirements, real-time requirements, and computational resource budget. Based on these constraints, it automatically determines the simulation strategy for this task using a predefined decision rule base or a lightweight machine learning classifier: a single fast simulation, a single accurate simulation, or initiating a two-layer simulation process.

[0032] S200, adaptive switching and two-layer simulation.

[0033] Based on the simulation strategy decided in step S100, the corresponding simulation process is automatically activated. When the simulation strategy decision is a two-layer simulation, steps S210 to S240 are executed. S210, First-level fast simulation: Build and run the first-level fast simulation, select the best fast model from multiple candidate fast models for global fast simulation, and obtain preliminary results; S220, Key Identification: Based on preset criteria, key areas or key time periods are identified from preliminary results; S230, Second-layer precise simulation: For key areas or key time periods, construct or call the second-layer precise simulation, use a high-precision model to perform local fine simulation, and obtain local high-precision results; S240, Result Fusion: The preliminary results of step S210 are fused with the local high-precision results of step S230 (usually the corresponding part of the preliminary results of step S210 is replaced by the local high-precision results of step S230) to generate a preliminary simulation report.

[0034] S300, residual correction.

[0035] Based on the residuals between real system data and the preliminary simulation report, the simulation model parameters are corrected online, and the model library is updated.

[0036] like Figure 3 As shown, in some embodiments, step S210 includes: S211 calls multiple candidate fast models from a fast simulation algorithm set containing simplified models optimized for different working conditions. Simplified models include linear models based on transfer functions, reduced-order models based on data, and lightweight neural network proxy models. S212 allows multiple candidate fast models to be run in parallel or in rapid serial execution within a reduced computational domain or time window. S213, based on preset fast evaluation criteria, automatically selects the optimal fast model. Fast evaluation criteria include fitting to a small amount of high-precision data, calculation speed, numerical stability, etc. S214 uses the optimal fast model to perform fast global and / or full-time simulations and obtain preliminary results.

[0037] In some embodiments, step S220 analyzes the preliminary results obtained in step S214 and automatically identifies key areas (such as water pump valves, etc.) or key time periods (such as the moment the host starts) that need to be focused on based on preset key criteria (such as: physical quantity gradient exceeding the threshold, system efficiency reaching an inflection point, vibration and noise index exceeding the standard).

[0038] like Figure 4 As shown, in some embodiments, step S230 includes: S231, for the identified key areas or key time periods, construct or call a second-layer high-precision simulation model. This model is preferably a Physical Information Neural Network (PINN), with its control equations (such as the dynamic equations of the transmission system and the thermodynamic equations) used as constraint embedding loss functions; S232, the data provided by the first layer of rapid simulation in step S210 at the boundary of the key area or the start and end points of the time period are used as the boundary conditions or initial conditions of the physical information neural network. S233, train and / or fine-tune the physical information neural network to perform high-precision, meshless, fine-grained simulation of key areas or key time periods, and obtain local high-precision results.

[0039] In some embodiments, step S300 includes: S310, when sensor data of a real ship power system is available (such as bench test data or real ship monitoring data), it shall be used as the true value of the reference system. S320, calculate the residuals between the corresponding physical quantities (such as output shaft torque and cooling water temperature) in the preliminary simulation report and the true values ​​of the reference system; S330, based on the residuals, automatically fine-tunes the parameters of the model in step S210 and / or step S230 through correction algorithms (such as online correction based on the error model, data assimilation based on Kalman filtering, or inversion identification of unknown parameters in the PINN model); S340 updates the corrected model parameters to the model library for the next simulation task, thereby achieving synchronous evolution of the simulation model and the physical system and maintaining its long-term accuracy.

[0040] In some embodiments, step S300 is executed periodically, or is triggered when new real system data is received.

[0041] In some embodiments, the adaptive two-layer simulation method based on intelligent optimization and residual correction of the present invention is equipped with an automatic switching mechanism, which triggers the operation mechanism according to task constraints.

[0042] In some embodiments, the automatic switching mechanism operation mechanism includes: Scenario-driven: If the scenario type is fault diagnosis (such as leakage), the simulation strategy decision is either two-layer simulation or single precise simulation. For example, if the condenser is blocked, the automatic switching mechanism switches to single precise simulation, and PINN locates the fault point, improving the fault detection accuracy and saving computing resources. If the scenario type is real-time control, the simulation strategy decision is single fast simulation. Accuracy threshold: If the uncertainty of a single fast simulation or the first-level fast simulation exceeds the preset accuracy threshold (e.g., uncertainty > 10%), the simulation strategy decision is to use a single accurate simulation. Real-time constraint: If the required response time of the system is lower than the preset real-time threshold (e.g., response time < 1 second), the simulation strategy decision is to use a single fast simulation. For example, if the response time is < 0.5 seconds during normal navigation, a single fast simulation will dominate.

[0043] In some embodiments, the present invention also provides an adaptive two-layer simulation system, comprising: The task management and decision-making module is used to receive simulation tasks and automatically decide on simulation strategies. The adaptive switching module, connected to the task management and decision-making module, is used to automatically activate the corresponding simulation process based on the simulation strategy of the decision. A dual-layer simulation engine, connected to an adaptive switching module, is activated when the simulation strategy decision is dual-layer simulation. The dual-layer simulation engine includes a first-layer fast simulation module, a key identification module, a second-layer accurate simulation module, and a result fusion module. The first-level rapid simulation module is used to perform rapid simulation and automatic optimization. The key identification module is used to identify key areas or key time periods; The second-layer precise simulation module is used to perform precise simulations in key areas or key time periods. The results fusion module is used to fuse fast and accurate simulation results; as well as The residual correction module is used to calculate residuals and update simulation model parameters online. The model library, connected to the two-layer simulation engine and residual correction module, is used to store and manage simulation models and their parameters.

[0044] For ease of description, the above system is described by dividing it into various modules based on their functions. Of course, when implementing this specification, the functions of each module can be implemented in one or more software and / or hardware components.

[0045] In some embodiments, the function of the first-layer rapid simulation module is to perform rapid calculations of the thermodynamic processes of the power system and output key parameters (such as pressure and temperature). Its core design is as follows: (1) Algorithm set integration: includes intelligent algorithms such as genetic algorithm and particle swarm optimization, which automatically seek optimization based on input objectives (such as maximizing efficiency and minimizing energy consumption); (2) Simplify the physical model: Based on the reduced-order model (ROM) or equivalent circuit model, reduce the amount of calculation. For example, the heat exchanger model adopts the lumped parameter method instead of the distributed parameter method. (3) Input / output: Input is the system boundary conditions (such as steam flow rate and temperature); output is the optimized operating parameters (such as valve opening and pump speed).

[0046] In some embodiments, the function of the second-layer accurate simulation module is to be activated in scenarios requiring high precision, and to perform fine simulation by integrating PINN (Physical Information Neural Network). Its core design is as follows: (1) PINN integration: Physical constraints such as the Navier-Stokes equation and the law of conservation of energy are added as loss functions to neural network training to ensure that the output conforms to physical laws; (2) Multi-pressure node modeling: For the multi-structural characteristics of the power system (such as shell and tube heat exchangers, direct contact mixing condensers, etc.), a multi-pressure node model is established to calculate the shell-side / tube-side parameters. (3) Dynamic data-driven: Correct model parameters (such as heat transfer coefficient) through real-time sensor data.

[0047] In some embodiments, the residual correction module functions to compare the residuals of the simulation output with those of the reference system (physical or high-fidelity model) and automatically correct the model parameters. The implementation methods include the following: 1. Implementation Method 1: (1) Residual calculation: e(t)=|Y_sim(t)-Y_real(t)|, where Y is a key parameter (such as pressure); (2) Correction algorithm: Use Kalman filtering or least squares method to update the model coefficients (such as heat exchanger resistance coefficient) online. (3) Closed-loop feedback: A correction is performed every 10 simulation cycles to prevent error accumulation.

[0048] 2. Implementation Method Two: Residual correction can also be achieved by using a residual correction network, which introduces residual blocks at the microscopic level.

[0049] This patent application provides the following embodiments.

[0050] Example 1: This example uses a marine main engine cooling system as an example to demonstrate the complete process of this application in a fault diagnosis scenario, as follows: Step S100, Task Reception and Decision. The system receives a simulation task to "diagnose the cause of abnormal cooling system temperature in a certain type of ship's main engine during cruise." The task constraints are as follows: scenario type is "fault diagnosis," accuracy requirement is high (error ≤ 5%), real-time requirement is medium (response time ≤ 10 minutes), and computational resource budget is medium. Based on the above requirements, the system automatically decides that the simulation strategy for this task is to "start a two-layer simulation process."

[0051] Step S200, adaptive switching and two-layer simulation, execute steps S210 to S240 as follows: Step S210, First-layer fast simulation. The system first calls three candidate models from the fast simulation algorithm set: "one-dimensional heat transfer simplified model, neural network-based surrogate model, and equivalent circuit thermal model". Then, it runs the three models in parallel within a narrowed time window (data from the most recent 30 minutes). The evaluation criteria are "fitting degree with historical high-precision data, calculation speed, and numerical stability". The conclusion is that "the neural network surrogate model is the optimal fast model". Finally, the global fast simulation is executed, and the preliminary results of the host cooling system temperature distribution are output.

[0052] Step S220, Key Identification. Based on the preset criterion "temperature gradient exceeds the threshold (>10°C / m) and system efficiency reaches an inflection point", the above preliminary results are analyzed to identify "the high-temperature water outlet section of the main unit" as the key area and "within 5 minutes after a sudden increase in the main unit load" as the key time period.

[0053] Step S230, second-layer accurate simulation. First, construct the following Physical Information Neural Network (PINN) model: the governing equations are "unsteady-state heat conduction equation + convection heat transfer equation", the boundary conditions are provided by the first-layer fast simulation, and the initial conditions are taken from the simulation results of the first layer. Then, train PINN to perform a refined simulation of the high-temperature water outlet section, and finally output the local high-precision results of the local temperature field, thermal stress distribution, and potential blockage location.

[0054] S240, Results Fusion. The first-layer rapid simulation of the global temperature distribution is fused with the second-layer accurate simulation of the local high-precision temperature field to generate a preliminary simulation report, which points out the risk of local blockage in the high-temperature water outlet section.

[0055] S300, Residual Correction. The system first acquires sensor data (temperature, flow rate) from the propagating ship, then calculates the residual between the simulation results and the measured values. Next, it uses the Kalman filter method to correct the heat transfer coefficient in PINN and updates the parameters of the corresponding model in the model library. The corrected model is then used for subsequent simulations to improve long-term prediction accuracy.

[0056] Example 2: This example demonstrates how the system adaptively selects a single fast simulation strategy in a real-time control scenario, as detailed below: Step S100, Task Reception and Decision-Making. The system receives a simulation task for "real-time prediction of host torque for controlling the fuel injection system." The task constraints are as follows: scenario type is "real-time control," accuracy requirement is medium (error ≤ 8%), real-time requirement is high (response time ≤ 0.5 seconds), and computing resources are limited. Based on these requirements, the system automatically decides that the simulation strategy for this task is "start a single fast simulation process."

[0057] Step S200: Adaptive switching. The system skips the second layer of precise simulation and directly uses the first layer of fast simulation. It calls the pre-trained optimal fast model (lightweight neural network) to output torque prediction results, which are directly used for control command generation.

[0058] S300, Residual Correction (periodic execution). The system uses ship data to perform online corrections to the fast model every 24 hours and updates the model weights using the least squares method to ensure that the model maintains accuracy during long-term operation.

[0059] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. The present invention is described with reference to flowchart illustrations and / or block diagrams of methods, systems (apparatus), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0060] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An adaptive double-layer simulation method based on intelligent optimization and residual correction, characterized in that, The method comprises the following steps: S100, task receiving and decision making: receiving a simulation task, parsing task constraints, and automatically deciding a simulation strategy, the simulation strategy comprising double-layer simulation; S200, adaptive switching and double-layer simulation: automatically enabling a corresponding simulation process according to the simulation strategy decided in step S100, and when the simulation strategy decision is double-layer simulation, performing steps S210 to S240, wherein, S210, first-layer fast simulation: constructing and running a first-layer fast simulation, selecting an optimal fast model from a plurality of candidate fast models for global fast simulation, and obtaining preliminary results; S220, key identification: identifying a key area or a key time period from the preliminary results based on preset criteria; S230, second-layer accurate simulation: constructing or calling a second-layer accurate simulation for the key area or the key time period, performing local fine simulation by using a high-precision model, and obtaining local high-precision results; S240, result fusion: fusing the preliminary results of step S210 and the local high-precision results of step S230 to generate a preliminary simulation report.

2. The method of claim 1, wherein, The method further comprises the following steps: S300, residual correction: based on residuals between real system data and the preliminary simulation report, correcting simulation model parameters online, and updating a model library.

3. The method of claim 2, wherein, The step S300 comprises: S310, when sensor data of a real ship power system is available, taking the sensor data as reference system true values; S320, calculating residuals between corresponding physical quantities in the preliminary simulation report and the reference system true values; S330, based on the residuals, automatically fine-tuning parameters of models in steps S210 and / or S230 by a correction algorithm; S340, updating the corrected model parameters to the model library.

4. The method of claim 2, wherein, The step S300 is periodically performed or triggered for execution when new real system data is received.

5. The method of claim 1, wherein, In the step S100, the task constraints comprise at least one of a scene type, a precision requirement, a real-time requirement, and a computing resource budget; and the automatic decision making is realized by a pre-defined decision rule library or a machine learning classifier.

6. The method of claim 5, wherein, The method is provided with an automatic switching mechanism, and the automatic switching mechanism triggers a running mechanism according to the task constraints.

7. The method of claim 6, wherein, The simulation strategy further comprises single fast simulation and / or single accurate simulation, and the automatic switching mechanism running mechanism comprises: if the scene type is fault diagnosis, the simulation strategy decision is double-layer simulation or single accurate simulation; if the scene type is real-time control, the simulation strategy decision is single fast simulation; if the uncertainty of single fast simulation or the first-layer fast simulation exceeds a preset precision threshold, the simulation strategy decision is single accurate simulation; and if a response time required by the system is lower than a preset real-time threshold, the simulation strategy decision is single fast simulation.

8. The method of claim 1, wherein, The step S210 comprises: S211, calling a plurality of candidate fast models from a fast simulation algorithm set comprising simplified models optimized for different working conditions; S212, running the plurality of candidate fast models in parallel or fast serial in a narrowed computing domain or time window. S213, automatically selecting the optimal fast model according to the preset fast evaluation criteria; S214, using the optimal fast model to perform global and / or full-time fast simulation and obtaining preliminary results.

9. The method of claim 1, wherein, The step S230 includes: S231, for the key area or key period, the high-precision model adopts a physical information neural network, which embeds its control equation as a constraint into a loss function; S232, providing data provided by the first layer fast simulation in step S210 at the key area boundary or time period start and end point as the boundary condition or initial condition of the physical information neural network; S233, training and / or fine-tuning the physical information neural network to perform high-precision, meshless fine simulation on the key area or key period to obtain local high-precision results.

10. An adaptive two-layer simulation system implementing the method of any one of claims 1-9, characterized by Comprise: Task management and decision module, for receiving simulation tasks and automatically deciding simulation strategies; Adaptive switching module, connected with the task management and decision module, for automatically enabling corresponding simulation processes according to the decided simulation strategies; Double-layer simulation engine, connected with the adaptive switching module, the adaptive switching module starts the double-layer simulation engine when the simulation strategy is decided as double-layer simulation, the double-layer simulation engine includes a first layer fast simulation module, a key identification module, a second layer accurate simulation module, and a result fusion module, wherein, The first layer fast simulation module is used for performing fast simulation and automatic optimization; The key identification module is used for identifying the key area or key period; The second layer accurate simulation module is used for performing accurate simulation on the key area or key period; The result fusion module is used for fusing fast and accurate simulation results; And The residual correction module is used for calculating residual and updating simulation model parameters online; The model library, connected with the double-layer simulation engine and the residual correction module, is used for storing and managing simulation models and their parameters.

Images