Generator set multiple quantitative analysis system and method based on thermodynamic digital twinning and artificial intelligence

Abstract

The invention discloses a generator set multiple quantitative analysis system and method based on thermodynamic digital twinning and artificial intelligence, and the system comprises a thermodynamic digital twinning module which is used for simulating the thermal state of a generator set in real time according to boundary conditions; the expert model module is used for calculating unit performance and virtual power station performance according to the actual parameters; the AI analysis optimization module is used for calling intermediate parameters of the digital twin and expert model to perform quantitative analysis and diagnosis, generating an optimization strategy and verifying the strategy; and the credibility guarantee module is used for guaranteeing the interpretability and reliability of the system. According to the method, an AI model with digital twin and expert models as data and knowledge basis and physical rules as constraints is constructed, closed-loop management of intelligent analysis, diagnosis and operation optimization of the generator set is realized, and the problems of insufficient modeling precision, insufficient strategy reliability, insufficient credibility, unbalanced data privacy and generalization ability and the like in the prior art are remarkably solved.

Description

Technical Field

[0001] This invention belongs to the field of generator set optimization control technology, specifically relating to a generator set multi-quantization analysis system and method based on thermodynamic digital twin and artificial intelligence. Background Technology

[0002] Against the backdrop of the deepening implementation of the "dual carbon" goals and the accelerated transformation of the energy structure, generator sets, as the core energy supply carriers of the power system, are directly related to the improvement of energy utilization efficiency, the reduction of carbon emission intensity, and the safe supply of the power system in terms of their operating efficiency, safety and stability and low carbon level. Therefore, the optimization of generator set operation has become a key lever for the high-quality development of the industry.

[0003] Currently, the mainstream solutions for generator set analysis and optimization fall into three categories: First, mechanistic models are built based on the first principles of thermodynamics, using conservation equations to characterize the unit's thermodynamic processes and performance indicators. However, these models have limited adaptability to complex nonlinear dynamic processes and multi-temporal scale characteristics. Second, pure data-driven AI models are trained based on historical operating data for state fitting and strategy generation. However, these models lack physical mechanism support and are prone to producing invalid results that violate the laws of thermodynamics. Third, digital twins, expert rule bases, and AI algorithms are applied independently. Basic functions are achieved through manual data transmission or simple interface interaction. Optimization strategies are mostly generated using traditional control algorithms or unconstrained AI algorithms, with feasibility only verified through simple rules. Some solutions attempt to introduce data fusion or model correction techniques, but they lack systematic quality assessment, dynamic calibration, and cross-module collaboration mechanisms, making it difficult to form closed-loop optimization capabilities.

[0004] Existing technologies attempt to integrate digital twins and AI models into generator set analysis systems. These systems simulate generator state using digital twins and combine AI models to assist in performance calculations and strategy generation. However, these technologies still have significant drawbacks and fail to meet practical application requirements: limited modeling accuracy and high prediction errors; insufficient strategy reliability, with optimization strategy generation lacking rigorous thermodynamic constraint embedding and closed-loop verification mechanisms; insufficient system credibility, with AI decision-making lacking interpretable analysis and uncertainty assessment; and difficulty in balancing data privacy protection and model generalization ability. These are the shortcomings of existing technologies.

[0005] In view of this, it is very necessary to provide a multi-quantization analysis system and method for generator sets based on thermodynamic digital twins and artificial intelligence to solve the above-mentioned defects in the prior art. Summary of the Invention

[0006] The purpose of this invention is to address the technical shortcomings of existing technologies, such as limited modeling accuracy, high prediction errors, insufficient strategy reliability, lack of rigorous thermodynamic constraint embedding and closed-loop verification mechanisms in strategy generation, lack of system credibility, lack of interpretability analysis and uncertainty assessment in AI decision-making, and difficulty in balancing data privacy protection and model generalization ability. This invention provides a system and method for designing a multi-quantitative analysis of generator sets based on thermodynamic digital twins and artificial intelligence to solve the aforementioned technical problems.

[0007] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a multi-quantitative analysis system for generator sets based on thermodynamic digital twins and artificial intelligence, including a thermodynamic digital twin module, an expert model module, an AI analysis and optimization module, and a reliability assurance module; each module communicates through a unified data bus and API interface to form a closed-loop optimization system; The thermodynamic digital twin module, in which: The collected multi-source heterogeneous data undergoes quality assessment and adaptive weighted fusion processing to remove outlier data and output standardized data. Based on thermodynamic laws and conservation equations, a differential algebraic equation model of the unit's thermodynamic process is constructed, and thermodynamically sensed reduced-order modeling is used to simplify the model complexity. The thermodynamic state data output by the model simulation is compared with the actual measured data, and the model parameters are dynamically adjusted through a parameter estimation method based on Bayesian inference. The model is then optimized according to a graded calibration strategy to output high-fidelity unit thermodynamic state data. The expert model module, in which: A hybrid modeling framework is constructed that integrates first-principles models and AI models through residual connections. Based on the fundamental laws of thermodynamics and actual operating parameters of generator sets, and combined with high-fidelity thermodynamic state data of the units output by the thermodynamic digital twin module, quantitative analysis of the performance of individual units and the overall performance of the virtual power station is carried out. The AI ​​analysis and optimization module, in which: By calling the unit's thermal state data from the thermodynamic digital twin module and the performance analysis data from the expert model module, anomaly identification and root cause analysis are completed through deep neural networks and knowledge graph technology; operation optimization strategies are generated using physical constraint reinforcement learning algorithms; and the strategies are verified with the help of structure-preserving neural networks and related constraint loss functions, forming a closed-loop management, and outputting verified optimization strategies to provide decision support for the operation optimization of generator units and virtual power plants. The trustworthiness assurance module, in which: By visualizing and interpreting the AI ​​decision-making logic, the transparency of decision-making is improved. Uncertainty assessments are conducted on the output results of each module, and confidence interval estimates are provided to clarify the reliable range of the results. At the same time, a specific architecture is adopted to achieve multi-site model collaborative optimization while preserving the privacy of sensitive operational data, taking into account both data security and model generalization ability, and providing trust support for high-risk decision-making scenarios and large-scale applications of the system.

[0008] Furthermore, the thermodynamic digital twin module is used to simulate the thermal state of the unit according to the boundary conditions and output high-fidelity thermal state data of the unit. It includes a multi-source data fusion unit, a dynamic simulation unit and a real-time calibration unit. The multi-source data fusion unit is specifically: Multi-source heterogeneous data from the generator set's DCS system, SIS system, and sensor system are collected, and data quality is assessed. The data quality assessment includes integrity checks, physical consistency verification, and timeliness assessment. The integrity check calculates an integrity score (C) based on the proportion of missing data. The physical consistency verification obtains a consistency score (CS) through testing against fundamental thermodynamic laws. The timeliness assessment is achieved through timestamp analysis and delay calculation, yielding a timeliness score (T). The overall data quality is quantified using a data quality assessment formula, which is:

[0009] in, For quality score, For completeness score, For consistency score, Score for timeliness; The quality score is collaboratively calculated with the preset inherent confidence metric value of the data source (preset based on the data source's accuracy level, operational stability, and other characteristics). A softmax normalization algorithm is used to normalize the fusion calculation result in the direction with dimension 0, generating dynamic fusion weights for each data source. This weight allocation follows an adaptive rule that "the higher the quality score and the stronger the inherent confidence, the larger the weight ratio," ensuring that high-quality data dominates the fusion result.

[0010] Based on the generated dynamic fusion weights, the effective multi-source heterogeneous data that has passed quality assessment is weighted and summed. Data dimension alignment and numerical integration are achieved through tensor operations. Finally, standardized fusion data with abnormal interference removed and consistency and accuracy are output, along with the corresponding dynamic fusion weights. This provides boundary conditions and initial parameters that conform to the actual operating conditions of the unit for the subsequent dynamic simulation unit.

[0011] The dynamic simulation unit is specifically: The system receives standardized fusion data output from the multi-source data fusion unit and uses it as boundary conditions and initial parameters for the simulation model, providing basic data support for the simulation of the thermal process and ensuring that the simulation input is consistent with the actual operating conditions of the unit. Based on the first and second laws of thermodynamics, and combined with the mass conservation equation and momentum conservation equation, a differential-algebraic equation model of the unit's thermodynamic process is constructed. The dynamic characteristics of the unit's thermodynamic system are described by differential equations, and the static constraint relationships between the system's parameters are described by algebraic equations. The mass conservation equation is expressed as follows:

[0012] in, Let be the local rate of change of fluid density with time t; Let be the rate of change of fluid mass due to transport. For fluid density, This is the fluid velocity vector.

[0013] The first law of thermodynamics, also known as the law of conservation of energy, is applied to thermodynamic systems and is expressed as follows:

[0014] in, For fluid density, For isobaric specific heat capacity, For fluid temperature, The divergence of heat conduction flux; The expression for the conservation of momentum is:

[0015] in, For fluid density, This refers to fluid pressure. For fluid velocity vector, For dynamic viscosity, The Laplace operator for velocity; Thermodynamically-aware reduced-order modeling technique is adopted. By using a sparse autoencoder, key feature variables that significantly affect the thermodynamic process are extracted from multiple original variables contained in the differential algebraic equation model. Redundant variables are eliminated, reducing the computational complexity of the model without sacrificing the core simulation accuracy. Run the optimized differential-algebraic equation model to simulate the thermodynamic operation of the unit under the current boundary conditions and output preliminary thermodynamic state simulation data.

[0016] The real-time calibration unit is specifically: The system receives preliminary thermodynamic state simulation data output by the dynamic simulation unit, and simultaneously receives actual measurement data collected by the generator set's DCS system and SIS system; it matches the preliminary thermodynamic state simulation data with the actual measurement data one by one, and calculates the parameter deviation between the two. Constrained by the laws of thermodynamics, a parameter estimation method based on Bayesian inference is adopted, combined with extended Kalman filtering to dynamically adjust the key parameters of the model. In this process, the process noise matrix is ​​a diagonal matrix [0.01, 0.05, 0.02, ...], and the observation noise matrix is ​​a diagonal matrix [0.02, 0.03, 0.01, ...]. Perform the operation according to the graded calibration strategy: high frequency calibration (calibration cycle of 1 minute) for key operating parameters such as main steam temperature and pressure, medium frequency calibration (calibration cycle of 5 minutes) for performance parameters such as heat rate and efficiency, and low frequency calibration (calibration cycle of 30 minutes) for equipment status parameters corresponding to wear and scaling. The final output is high-fidelity thermal status data of the unit.

[0017] Furthermore, the expert model module includes a unit performance calculation unit, a virtual power plant performance evaluation unit, and a hybrid modeling strategy unit; The hybrid modeling strategy unit is specifically as follows: Based on the first and second laws of thermodynamics and the principles of combustion dynamics, a first-principles model is constructed to accurately characterize the core physical mechanisms of energy conversion, heat transfer and pollutant generation in the unit, ensuring the physical consistency of the modeling process. At the same time, a deep neural network structure is used to build an AI model. By learning the residuals between the historical operating data of the generator set and the calculation results of the first-principles model, the calculation errors caused by simplification assumptions and nonlinear characteristics of the system are compensated in the physical modeling process. The core design logic of the hybrid modeling framework is as follows: a first-principles model is used for subsystems with clear mechanisms, while an AI model is used for subsystems with complex mechanisms or those that are difficult to model accurately. The two models work together through a residual connection architecture. The first-principles model outputs preliminary performance calculation results, while the AI ​​model learns the error between the first-principles model and the actual system and corrects the preliminary calculation results through residuals. Finally, a hybrid modeling framework with both physical interpretability and computational accuracy is formed, providing core modeling support for the unit performance calculation unit.

[0018] The unit performance calculation unit is specifically as follows: Based on the hybrid modeling framework provided by the hybrid modeling strategy unit, the high-fidelity thermodynamic state data of the generator set output by the thermodynamic digital twin module is received. Combined with the actual operating parameters of the generator set (such as fuel characteristics, steam parameters, load conditions, etc.), the core performance data of the generator set is calculated using the fundamental laws of thermodynamics (the first law of thermodynamics and the second law of thermodynamics) as the core calculation basis. This includes key performance indicators such as heat rate, operating efficiency, coal consumption, and pollutant (such as NOx) emission concentration.

[0019] The virtual power plant performance evaluation unit is specifically as follows: Based on the core performance data of each unit output by the unit performance calculation unit, and receiving grid dispatch instructions, market price signals, and equipment constraints, an overall performance evaluation of the virtual power plant is conducted. During the evaluation process, the calculation of peak-shaving capacity of each unit and the overall operational economic analysis of the virtual power plant are used as core evaluation dimensions. The evaluation results are quantified using a comprehensive performance scoring formula, which is: Overall performance score = 0.6 × peak shaving capacity + 0.4 × operating economy Among them, peak-shaving capacity has a weight of 0.6, and operational economy has a weight of 0.4; The final output is an overall performance evaluation conclusion of the virtual power plant, providing core decision-making basis for the AI ​​analysis and optimization module to generate operation optimization strategies.

[0020] Furthermore, the AI ​​analysis and optimization module includes a quantitative analysis and diagnosis unit, a strategy generation unit, and a strategy verification unit; The quantitative analysis and diagnostic unit specifically comprises: Based on the high-fidelity unit thermal state data output by the thermodynamic digital twin module and the performance analysis data (including core indicators such as heat rate, efficiency, coal consumption, and pollutant emission concentration) output by the expert model module, multi-dimensional quantitative analysis is carried out using deep neural network and knowledge graph technology. By capturing feature patterns of data deviating from normal operating range through deep neural networks, and combining the generator set equipment association logic and operating rules integrated with knowledge graphs, operational deviation anomalies are identified and their root causes are traced. Based on the time-series change data of performance indicators, the performance evolution law within the normal operating life cycle of the equipment is learned through deep neural networks, capturing the degradation characteristics of gradual performance decline. The correlation between degradation characteristics and equipment operating status is verified with the help of knowledge graphs, realizing accurate detection and root cause location of equipment degradation anomalies. Finally, the anomaly diagnosis conclusions and root cause analysis results are output, providing core decision-making basis for the strategy generation unit to formulate targeted operation optimization strategies.

[0021] The strategy generation unit is specifically: A policy generation model is constructed using a physical constraint reinforcement learning algorithm. The total number of training rounds is set, and operations such as policy generation, constraint verification, and parameter optimization are performed cyclically in each round to ensure that the model gradually converges to the optimal state through multiple iterations.

[0022] In each round of training, based on the policy network of the current reinforcement learning model, the current operating status data of the generating unit and the virtual power station are input (including high-fidelity thermodynamic state data output by the thermodynamic digital twin module, performance data output by the expert model module, and abnormal diagnosis conclusions and root cause analysis results output by the quantitative analysis and diagnosis unit) to generate operation optimization strategies for the generating unit and the virtual power station (in terms of load allocation, combustion optimization, equipment maintenance and adjustment, etc.); at the same time, exploration noise is introduced to ensure the diversity of strategy search during training and avoid the model getting trapped in local optima.

[0023] Furthermore, the physical constraint reinforcement learning algorithm embeds the fundamental laws of thermodynamics as constraints into the reward function and loss function to achieve constraint verification; the constraints include energy conservation constraints and entropy increase inequality constraints; during the modeling process, the constraint weight λ decays from 1.0 to 0.1 according to a preset rule, while the data weight increases from 0.1 to 1.0, achieving dynamic coordination between constraints and data.

[0024] Furthermore, the energy conservation constraint is based on the first law of thermodynamics. During the constraint verification phase, the difference between the total input energy and the total output energy of the system after the strategy execution is calculated to ensure that the constraint error is satisfied.

[0025] in, Input total energy into the system, The system outputs total energy; Furthermore, the entropy increase inequality constraint, based on the second law of thermodynamics, constructs a negative entropy production penalty mechanism through the ReLU function, defining the entropy increase constraint penalty term as:

[0026] in, The entropy change of the system after the strategy is executed, when When <0 (i.e., negative entropy production occurs, violating the second law of thermodynamics), the penalty term outputs a non-zero value; otherwise, it outputs a zero value, thus effectively constraining negative entropy production. The verification results of the two constraints are combined into a physical constraint violation quantification value; Based on the quantification of physical constraints, the reward and loss functions are constructed: On the one hand, the constraint-based reward is calculated, with the following formula:

[0027] Where r is the basic reward. The constraint weights are the penalty terms for violating physical constraints; Construct the policy update loss function, the formula is:

[0028] in, To use PPO as the base loss, the influence of enhanced constraints on policy updates is strengthened, and policy network optimization is completed through a parameter update mechanism. The policy verification unit is specifically: The system receives the operation optimization strategy output by the strategy generation unit, and simultaneously collects the basic thermal state data of the unit output by the thermodynamic digital twin module and the performance benchmark data output by the expert model module to construct a verification input dataset to ensure the consistency between the simulation scenario and the actual operating conditions of the unit. The pre-configured structure-preserving neural network is activated. This network contains four hidden layers, each with 128 neurons. The input layer dimension matches the feature dimension of the validation input dataset, and the output layer dimension corresponds to the system state parameters after policy execution. An adaptive Swish activation function is used, with the expression:

[0029] in, =1, which is a trainable parameter used to enhance the network's ability to fit the state of complex nonlinear systems; The structure maintains that the constraint loss function of the neural network includes energy conservation constraints and entropy increase inequality constraints, which are consistent with the physical constraints of the policy generation unit. The running optimization strategy in the verification input dataset is used as the input of the structure-preserving neural network. Through forward propagation calculation, the system state parameters of the generator set and virtual power station after the strategy is executed are digitally simulated to achieve accurate mapping of the optimization strategy execution effect. The forward propagation logic follows the basic rules of neural network signal transmission and completes data transformation and state output based on the preset network layers and activation functions. The core execution flow of forward propagation is as follows: first, the data of each hidden layer is processed by the adaptive Swish activation function, and then physical constraints are embedded in the output layer to complete the final output; Based on the output of the neural network and the verification of physical constraints, the total verification loss value is calculated. If the loss value meets the preset compliance requirements, the optimization strategy is determined to comply with the basic laws of thermodynamics and the requirements for safe operation of the equipment. The optimization strategy is then output to the generator set / DCS system for execution, and the verification result is fed back to the thermodynamic digital twin module for model parameter updates. If the loss value does not meet the preset compliance requirements, the optimization strategy is determined to have the risk of violating physical laws. A rejection instruction and details of the constraint violation are output to the strategy generation unit, triggering the regeneration of the optimization strategy.

[0030] Furthermore, the trustworthiness assurance module includes an explainable AI technology unit, an uncertainty quantification unit, and a data privacy protection unit; The explainable AI technology unit specifically includes: Using the anomaly diagnosis conclusions, optimization strategies, and corresponding related data output by the AI ​​analysis and optimization module as the analysis objects, the formation basis of the anomaly diagnosis conclusions and the generation path of the optimization strategies are logically decomposed through a locally interpretable model to locate the key influencing factors of decision-making; the mechanism of action of each input feature and decision result is traced by feature correlation analysis technology to clarify the feature weights and correlation logic; through the complete process of "feature importance analysis → decision path tracing → natural language interpretation generation", the AI ​​decision-making logic is visualized and interpreted, improving the transparency of decision-making.

[0031] The uncertainty quantification unit is specifically: An uncertainty quantification model is constructed based on a Bayesian machine learning framework. The analysis objects include the unit thermal state simulation results output by the thermodynamic digital twin module, the unit and virtual power plant performance calculation results output by the expert model module, the anomaly diagnosis conclusions, prediction results, and operation optimization strategy related evaluation data output by the AI ​​analysis and optimization module. The error sources, data fluctuation range, and model uncertainty of various results are systematically analyzed through Bayesian inference mechanism. The confidence intervals corresponding to each result are calculated and output, clarifying the reliable value range of the results under different confidence levels, and intuitively presenting the degree of uncertainty of the output data.

[0032] The data privacy protection unit specifically comprises: A multi-site collaborative optimization mechanism is constructed using a federated learning architecture, with local sensitive operating data of each generator unit and the collaborative training requirements of multi-site models as the core inputs. Through the distributed training mode of federated learning, each generator unit can participate in the collaborative training of cross-site models only through parameter sharing and gradient aggregation, while retaining the original sensitive data locally and not transmitting or leaking privacy information. This avoids the privacy leakage risks faced by sensitive data during transmission and centralized storage. At the same time, the advantages of multi-site data are integrated to jointly optimize the model, effectively improving the model's adaptability to different unit operating conditions and regional operating scenarios, and enhancing the model's generalization performance.

[0033] Secondly, the present invention also provides a multi-quantitative analysis method for generator sets based on thermodynamic digital twins and artificial intelligence, which communicates through a unified data bus and API interface; Step S1, the thermal state simulation and data generation step, in which: The collected multi-source heterogeneous data undergoes quality assessment and adaptive weighted fusion processing to remove outlier data and output standardized data. Based on thermodynamic laws and conservation equations, a differential algebraic equation model of the unit's thermodynamic process is constructed, and thermodynamically sensed reduced-order modeling is used to simplify the model complexity. The thermodynamic state data output by the model simulation is compared with the actual measured data, and the model parameters are dynamically adjusted through a parameter estimation method based on Bayesian inference. The model is then optimized according to a graded calibration strategy to output high-fidelity unit thermodynamic state data. Step S2, the step of quantitative analysis of unit and virtual power plant performance, in which: A hybrid modeling framework is constructed that integrates first-principles models and AI models through residual connections. Based on the fundamental laws of thermodynamics and actual operating parameters of generator sets, and combined with high-fidelity thermodynamic state data of the units, quantitative analysis is conducted on the performance of individual units and the overall performance of the virtual power station. Step S3, the step of generating and verifying the anomaly diagnosis and operation optimization strategy, in which: By calling upon the thermal state data and performance analysis data of the generating units, anomaly identification and root cause analysis are completed through deep neural networks and knowledge graph technology; physical constraint reinforcement learning algorithms are used to generate operation optimization strategies; and the strategies are verified with the help of structure-preserving neural networks and related constraint loss functions, forming a closed-loop management and outputting verified optimization strategies to provide decision support for the operation optimization of generating units and virtual power plants. Step S4, the steps for ensuring system trustworthiness and data security, in which: By visualizing and interpreting the AI ​​decision-making logic, the transparency of decision-making is improved. Uncertainty assessments are conducted on the output results of each module, and confidence interval estimates are provided to clarify the reliable range of the results. At the same time, a specific architecture is adopted to achieve multi-site model collaborative optimization while preserving the privacy of sensitive operational data, taking into account both data security and model generalization ability, and providing trust support for high-risk decision-making scenarios and large-scale applications of the system.

[0034] Furthermore, step S1 is used to simulate the thermal state of the unit based on boundary conditions and output high-fidelity thermal state data of the unit; specifically, it also includes the following steps: Step S101, the multi-source data fusion step, in which: Multi-source heterogeneous data from the generator set's DCS system, SIS system, and sensor system are collected, and data quality is assessed. The data quality assessment includes integrity checks, physical consistency verification, and timeliness assessment. The integrity check calculates an integrity score (C) based on the proportion of missing data. The physical consistency verification obtains a consistency score (CS) through testing against fundamental thermodynamic laws. The timeliness assessment is achieved through timestamp analysis and delay calculation, yielding a timeliness score (T). The overall data quality is quantified using a data quality assessment formula, which is:

[0035] in, For quality score, For completeness score, For consistency score, Score for timeliness; The quality score is collaboratively calculated with the preset inherent confidence metric value of the data source (preset based on the data source's accuracy level, operational stability, and other characteristics). A softmax normalization algorithm is used to normalize the fusion calculation result in the direction with dimension 0, generating dynamic fusion weights for each data source. This weight allocation follows an adaptive rule that "the higher the quality score and the stronger the inherent confidence, the larger the weight ratio," ensuring that high-quality data dominates the fusion result.

[0036] Based on the generated dynamic fusion weights, the effective multi-source heterogeneous data that has passed quality assessment is weighted and summed. Data dimension alignment and numerical integration are achieved through tensor operations. Finally, standardized fusion data with abnormal interference removed and consistency and accuracy are output, along with the corresponding dynamic fusion weights. This provides boundary conditions and initial parameters that conform to the actual operating conditions of the unit for the subsequent dynamic simulation unit.

[0037] Step S102, the dynamic simulation step, in which: The system receives standardized fusion data and uses it as boundary conditions and initial parameters for the simulation model, providing basic data support for the simulation of thermal processes and ensuring that the simulation input is consistent with the actual operating conditions of the unit. Based on the first and second laws of thermodynamics, and combined with the mass conservation equation and momentum conservation equation, a differential-algebraic equation model of the unit's thermodynamic process is constructed. The dynamic characteristics of the unit's thermodynamic system are described by differential equations, and the static constraint relationships between the system's parameters are described by algebraic equations. The mass conservation equation is expressed as follows:

[0038] in, Let be the local rate of change of fluid density with time t; Let be the rate of change of fluid mass due to transport. For fluid density, This is the fluid velocity vector.

[0039] The first law of thermodynamics, also known as the law of conservation of energy, is applied to thermodynamic systems and is expressed as follows:

[0040] in, For fluid density, For isobaric specific heat capacity, For fluid temperature, The divergence of heat conduction flux; The expression for the conservation of momentum is:

[0041] in, For fluid density, This refers to fluid pressure. For fluid velocity vector, For dynamic viscosity, The Laplace operator for velocity; Thermodynamically-aware reduced-order modeling technique is adopted. By using a sparse autoencoder, key feature variables that significantly affect the thermodynamic process are extracted from multiple original variables contained in the differential algebraic equation model. Redundant variables are eliminated, reducing the computational complexity of the model without sacrificing the core simulation accuracy. Run the optimized differential-algebraic equation model to simulate the thermodynamic operation of the unit under the current boundary conditions and output preliminary thermodynamic state simulation data.

[0042] Step S103, the real-time calibration step, in which: Receive preliminary thermodynamic state simulation data, and simultaneously receive actual measurement data collected by the generator set's DCS system and SIS system; match the preliminary thermodynamic state simulation data with the actual measurement data one by one, and calculate the parameter deviation between the two; Constrained by the laws of thermodynamics, a parameter estimation method based on Bayesian inference is adopted, combined with extended Kalman filtering to dynamically adjust the key parameters of the model. In this process, the process noise matrix is ​​a diagonal matrix [0.01, 0.05, 0.02, ...], and the observation noise matrix is ​​a diagonal matrix [0.02, 0.03, 0.01, ...]. Perform the operation according to the graded calibration strategy: high frequency calibration (calibration cycle of 1 minute) for key operating parameters such as main steam temperature and pressure, medium frequency calibration (calibration cycle of 5 minutes) for performance parameters such as heat rate and efficiency, and low frequency calibration (calibration cycle of 30 minutes) for equipment status parameters corresponding to wear and scaling. The final output is high-fidelity thermal status data of the unit.

[0043] Furthermore, step S2 specifically includes the following steps: Step S201, the hybrid modeling step, in which: Based on the first and second laws of thermodynamics and the principles of combustion dynamics, a first-principles model is constructed to accurately characterize the core physical mechanisms of energy conversion, heat transfer and pollutant generation in the unit, ensuring the physical consistency of the modeling process. At the same time, a deep neural network structure is used to build an AI model. By learning the residuals between the historical operating data of the generator set and the calculation results of the first-principles model, the calculation errors caused by simplification assumptions and nonlinear characteristics of the system are compensated in the physical modeling process. The core design logic of the hybrid modeling framework is as follows: a first-principles model is used for subsystems with clear mechanisms, and an AI model is used for subsystems with complex mechanisms or difficult to model accurately. The two models work together through a residual connection architecture. The first-principles model outputs preliminary performance calculation results, and the AI ​​model learns the error between the first-principles model and the actual system. The preliminary calculation results are corrected through residuals, and finally a hybrid modeling framework with both physical interpretability and computational accuracy is formed. Step S202, the step of unit performance calculation, in which: Based on the hybrid modeling framework, high-fidelity thermodynamic state data of the generator unit is received and combined with the actual operating parameters of the generator unit (such as fuel characteristics, steam parameters, load conditions, etc.). Using the fundamental laws of thermodynamics (the first law of thermodynamics and the second law of thermodynamics) as the core calculation basis, the core performance data of the generator unit is calculated, including key performance indicators such as heat rate, operating efficiency, coal consumption, and pollutant (such as NOx) emission concentration.

[0044] Step S203, the virtual power plant performance evaluation step, in which: Based on the core performance data of each generating unit, and receiving grid dispatch instructions, market price signals, and equipment constraints, an overall performance evaluation of the virtual power plant is conducted. During the evaluation process, the calculation of peak-shaving capacity of each generating unit and the overall operational economic analysis of the virtual power plant are used as core evaluation dimensions. The evaluation results are quantified using a comprehensive performance scoring formula, which is as follows: Overall performance score = 0.6 × peak shaving capacity + 0.4 × operating economy Among them, peak-shaving capacity has a weight of 0.6, and operational economy has a weight of 0.4; The final output is the overall performance evaluation conclusion of the virtual power plant; Furthermore, step S3 specifically includes the following steps: Step S301, the quantitative analysis and diagnosis step, in which: Based on high-fidelity unit thermal state data and performance analysis data (including core indicators such as heat rate, efficiency, coal consumption, and pollutant emission concentration), multi-dimensional quantitative analysis is carried out using deep neural network and knowledge graph technology; By capturing feature patterns of data deviating from normal operating range through deep neural networks, and combining the generator set equipment association logic and operating rules integrated by knowledge graphs, operational deviation anomalies are identified and their root causes are traced. Based on the time-series change data of performance indicators, the performance evolution law within the normal operating life cycle of the equipment is learned through deep neural networks, capturing the degradation characteristics of gradual performance decline. The correlation between degradation characteristics and equipment operating status is verified with the help of knowledge graphs, realizing accurate detection and root cause location of equipment degradation anomalies. Finally, anomaly diagnosis conclusions and root cause analysis results are output. Step S302, the strategy generation step, in which: A strategy generation model is constructed using a physical constraint reinforcement learning algorithm. High-fidelity unit thermal state data, unit and virtual power plant performance data, as well as anomaly diagnosis conclusions and root cause analysis results are used as joint data inputs to generate operation optimization strategies (including load allocation, combustion optimization, equipment maintenance and adjustment, etc.) for generator units and virtual power plants. Furthermore, the physical constraint reinforcement learning algorithm embeds the fundamental laws of thermodynamics as constraints into the reward function, specifically including energy conservation constraints and entropy increase inequality constraints; during the modeling process, the constraint weight λ decays from 1.0 to 0.1 according to a preset rule, while the data weight increases from 0.1 to 1.0, realizing dynamic coordination between constraints and data; Furthermore, the energy conservation constraint is based on the first law of thermodynamics, calculating the difference between the total input energy and the total output energy of the system after the strategy is executed, ensuring that the constraint error is satisfied:

[0045] in, Input total energy into the system, The system outputs total energy; Furthermore, the entropy increase inequality constraint, based on the second law of thermodynamics, constructs a negative entropy production penalty mechanism through the ReLU function, defining the entropy increase constraint penalty term as:

[0046] in, The entropy change of the system after the strategy is executed, when When <0 (i.e., negative entropy production occurs, violating the second law of thermodynamics), the penalty term outputs a non-zero value; otherwise, it outputs a zero value, thus effectively constraining negative entropy production. The core execution loop of the physical constraint reinforcement learning algorithm is: episode iteration → policy execution → physical constraint verification → constraint reward calculation → policy update. This loop enables the linkage between constraint verification and policy optimization.

[0047] Step S303, the policy verification step, in which: The system receives operational optimization strategies and collects basic thermal state data and performance benchmark data of the unit to construct a verification input dataset, ensuring the consistency between the simulated scenario and the actual operating conditions of the unit. The pre-configured structure-preserving neural network is activated. This network contains four hidden layers, each with 128 neurons. The input layer dimension matches the feature dimension of the validation input dataset, and the output layer dimension corresponds to the system state parameters after policy execution. An adaptive Swish activation function is used, with the expression:

[0048] in, =1, which is a trainable parameter used to enhance the network's ability to fit the state of complex nonlinear systems; The structure maintains that the constraint loss function of the neural network includes energy conservation constraints and entropy increase inequality constraints, which are consistent with the physical constraints of the policy generation unit. The running optimization strategy in the verification input dataset is used as the input of the structure-preserving neural network. Through forward propagation calculation, the system state parameters of the generator set and virtual power station after the strategy is executed are digitally simulated to achieve accurate mapping of the optimization strategy execution effect. The forward propagation logic follows the basic rules of neural network signal transmission and completes data transformation and state output based on the preset network layers and activation functions. The core execution flow of forward propagation is as follows: first, the data of each hidden layer is processed by the adaptive Swish activation function, and then physical constraints are embedded in the output layer to complete the final output; Based on the output of the neural network and the verification of physical constraints, the total verification loss value is calculated. If the loss value meets the preset compliance requirements, the optimization strategy is determined to comply with the basic laws of thermodynamics and the requirements for safe operation of the equipment. The optimization strategy is then output to the generator set / DCS system for execution, and the verification result is fed back to the thermodynamic digital twin module for model parameter updates. If the loss value does not meet the preset compliance requirements, the optimization strategy is determined to have the risk of violating physical laws. A rejection instruction and details of the constraint violation are output to the strategy generation unit, triggering the regeneration of the optimization strategy.

[0049] Furthermore, step S4 specifically includes: Step S401, Explaining the steps of AI technology, in which: Taking anomaly diagnosis conclusions, operational optimization strategies, and corresponding related data as the analysis objects, this study uses a locally interpretable model to logically decompose the formation basis of anomaly diagnosis conclusions and the generation path of optimization strategies, thereby identifying key influencing factors in decision-making. It also leverages feature correlation analysis technology to trace the interaction mechanism between each input feature and the decision result, clarifying feature weights and their correlation logic. Through a complete process of "feature importance analysis → decision path tracing → natural language interpretation generation," the study achieves a visual interpretation of AI decision-making logic, enhancing decision transparency.

[0050] Step S402, the uncertainty quantification step, in which: An uncertainty quantification model is constructed based on a Bayesian machine learning framework. The analysis objects include the results of unit thermal state simulation, the performance calculation results of the unit and virtual power station, the anomaly diagnosis conclusions, the prediction results, and the relevant evaluation data of operation optimization strategies. The model systematically analyzes the error sources, data fluctuation range, and model uncertainty of various results through Bayesian inference mechanism. The model calculates and outputs the confidence intervals corresponding to each result, clarifies the reliable value range of the results under different confidence levels, and intuitively presents the degree of uncertainty of the output data.

[0051] Step S403, the data privacy protection step, in which: A multi-site collaborative optimization mechanism is constructed using a federated learning architecture, with local sensitive operating data of each generator unit and the collaborative training requirements of multi-site models as the core inputs. Through the distributed training mode of federated learning, each generator unit can participate in the collaborative training of cross-site models only through parameter sharing and gradient aggregation, while retaining the original sensitive data locally and not transmitting or leaking privacy information. This avoids the privacy leakage risks faced by sensitive data during transmission and centralized storage. At the same time, the advantages of multi-site data are integrated to jointly optimize the model, effectively improving the model's adaptability to different unit operating conditions and regional operating scenarios, and enhancing the model's generalization performance.

[0052] The beneficial effects of this invention are as follows: A unified data bus and API interface enable module communication; the thermodynamic digital twin module, through multi-source data quality assessment and fusion, thermodynamic process modeling and order reduction, and real-time calibration using Bayesian inference, outputs high-fidelity unit thermodynamic state data; the expert model module employs a hybrid modeling framework combining first-principles calculations and AI residual connections to complete the calculation of core unit performance and the overall evaluation of the virtual power plant; the AI ​​analysis and optimization module, relying on deep neural networks, physical constraint reinforcement learning, and structure-preserving neural networks, achieves a closed loop of anomaly diagnosis, strategy generation, and verification; and the credibility assurance module, through interpretable AI, Bayesian uncertainty quantification, and federated learning, balances decision transparency, result reliability, and data privacy. This solution significantly improves prediction accuracy, enhances the reliability of optimization strategies, and improves the economic efficiency of unit operation, providing a complete technical solution for the digital and intelligent transformation of the power generation industry.

[0053] Furthermore, the design principle of this invention is reliable, the structure is simple, and it has a very wide range of application prospects.

[0054] Therefore, it is evident that the present invention has outstanding substantive features and significant progress compared with the prior art, and the beneficial effects of its implementation are also obvious. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0056] Figure 1 This embodiment provides an overall architecture diagram of a generator set multi-quantization analysis system based on thermodynamic digital twins and artificial intelligence.

[0057] Figure 2 This is a schematic diagram of a thermodynamic digital twin module. Detailed Implementation

[0058] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The following embodiments are explanations of the present invention, but the present invention is not limited to the following implementation methods.

[0059] Example 1: This embodiment provides a multi-quantization analysis system for generator sets based on thermodynamic digital twins and artificial intelligence, such as... Figure 1 As shown, it includes a thermodynamic digital twin module, an expert model module, an AI analysis and optimization module, and a reliability assurance module; each module communicates with an API interface through a unified data bus to form a closed-loop optimization system; The thermodynamic digital twin module, in which: The collected multi-source heterogeneous data undergoes quality assessment and adaptive weighted fusion processing to remove outlier data and output standardized data. Based on thermodynamic laws and conservation equations, a differential algebraic equation model of the unit's thermodynamic process is constructed, and thermodynamically sensed reduced-order modeling is used to simplify the model complexity. The thermodynamic state data output by the model simulation is compared with the actual measured data, and the model parameters are dynamically adjusted through a parameter estimation method based on Bayesian inference. The model is then optimized according to a graded calibration strategy to output high-fidelity unit thermodynamic state data. The expert model module, in which: A hybrid modeling framework is constructed that integrates first-principles models and AI models through residual connections. Based on the fundamental laws of thermodynamics and actual operating parameters of generator sets, and combined with high-fidelity thermodynamic state data of the units output by the thermodynamic digital twin module, quantitative analysis of the performance of individual units and the overall performance of the virtual power station is carried out. The AI ​​analysis and optimization module, in which: By calling the unit's thermal state data from the thermodynamic digital twin module and the performance analysis data from the expert model module, anomaly identification and root cause analysis are completed through deep neural networks and knowledge graph technology; operation optimization strategies are generated using physical constraint reinforcement learning algorithms; and the strategies are verified with the help of structure-preserving neural networks and related constraint loss functions, forming a closed-loop management, and outputting verified optimization strategies to provide decision support for the operation optimization of generator units and virtual power plants. The trustworthiness assurance module, in which: By visualizing and interpreting the AI ​​decision-making logic, the transparency of decision-making is improved. Uncertainty assessments are conducted on the output results of each module, and confidence interval estimates are provided to clarify the reliable range of the results. At the same time, a specific architecture is adopted to achieve multi-site model collaborative optimization while preserving the privacy of sensitive operational data, taking into account both data security and model generalization ability, and providing trust support for high-risk decision-making scenarios and large-scale applications of the system.

[0060] Furthermore, the thermodynamic digital twin module, such as Figure 2 As shown, it is used to simulate the thermal state of the unit according to the boundary conditions and output high-fidelity thermal state data of the unit. It includes a multi-source data fusion unit, a dynamic simulation unit and a real-time calibration unit. The multi-source data fusion unit is specifically: Multi-source heterogeneous data from the generator set's DCS system, SIS system, and sensor system are collected, and data quality is assessed. The data quality assessment includes integrity checks, physical consistency verification, and timeliness assessment. The integrity check calculates an integrity score (C) based on the proportion of missing data. The physical consistency verification obtains a consistency score (CS) through testing against fundamental thermodynamic laws. The timeliness assessment is achieved through timestamp analysis and delay calculation, yielding a timeliness score (T). The overall data quality is quantified using a data quality assessment formula, which is:

[0061] in, For quality score, For completeness score, For consistency score, Score for timeliness; The quality score is collaboratively calculated with the preset inherent confidence metric value of the data source (preset based on the data source's accuracy level, operational stability, and other characteristics). A softmax normalization algorithm is used to normalize the fusion calculation result in the direction with dimension 0, generating dynamic fusion weights for each data source. This weight allocation follows an adaptive rule that "the higher the quality score and the stronger the inherent confidence, the larger the weight ratio," ensuring that high-quality data dominates the fusion result.

[0062] Based on the generated dynamic fusion weights, the effective multi-source heterogeneous data that has passed quality assessment is weighted and summed. Data dimension alignment and numerical integration are achieved through tensor operations. Finally, standardized fusion data with abnormal interference removed and consistency and accuracy are output, along with the corresponding dynamic fusion weights. This provides boundary conditions and initial parameters that conform to the actual operating conditions of the unit for the subsequent dynamic simulation unit.

[0063] The dynamic simulation unit is specifically: The system receives standardized fusion data output from the multi-source data fusion unit and uses it as boundary conditions and initial parameters for the simulation model, providing basic data support for the simulation of the thermal process and ensuring that the simulation input is consistent with the actual operating conditions of the unit. Based on the first and second laws of thermodynamics, and combined with the mass conservation equation and momentum conservation equation, a differential-algebraic equation model of the unit's thermodynamic process is constructed. The dynamic characteristics of the unit's thermodynamic system are described by differential equations, and the static constraint relationships between the system's parameters are described by algebraic equations. The mass conservation equation is expressed as follows:

[0064] in, Let be the local rate of change of fluid density with time t; Let be the rate of change of fluid mass due to transport. For fluid density, This is the fluid velocity vector.

[0065] The first law of thermodynamics, also known as the law of conservation of energy, is applied to thermodynamic systems and is expressed as follows:

[0066] in, For fluid density, For isobaric specific heat capacity, For fluid temperature, The divergence of heat conduction flux; The expression for the conservation of momentum is:

[0067] in, For fluid density, This refers to fluid pressure. For fluid velocity vector, For dynamic viscosity, The Laplace operator for velocity; Thermodynamically-aware reduced-order modeling technique is adopted. By using a sparse autoencoder, key feature variables that significantly affect the thermodynamic process are extracted from multiple original variables contained in the differential algebraic equation model. Redundant variables are eliminated, reducing the computational complexity of the model without sacrificing the core simulation accuracy. Run the optimized differential-algebraic equation model to simulate the thermodynamic operation of the unit under the current boundary conditions and output preliminary thermodynamic state simulation data.

[0068] The real-time calibration unit is specifically: The system receives preliminary thermodynamic state simulation data output by the dynamic simulation unit, and simultaneously receives actual measurement data collected by the generator set's DCS system and SIS system; it matches the preliminary thermodynamic state simulation data with the actual measurement data one by one, and calculates the parameter deviation between the two. Constrained by the laws of thermodynamics, a parameter estimation method based on Bayesian inference is adopted, combined with extended Kalman filtering to dynamically adjust the key parameters of the model. In this process, the process noise matrix is ​​a diagonal matrix [0.01, 0.05, 0.02, ...], and the observation noise matrix is ​​a diagonal matrix [0.02, 0.03, 0.01, ...]. Perform the operation according to the graded calibration strategy: high frequency calibration (calibration cycle of 1 minute) for key operating parameters such as main steam temperature and pressure, medium frequency calibration (calibration cycle of 5 minutes) for performance parameters such as heat rate and efficiency, and low frequency calibration (calibration cycle of 30 minutes) for equipment status parameters corresponding to wear and scaling. The final output is high-fidelity thermal status data of the unit.

[0069] Furthermore, the expert model module includes a unit performance calculation unit, a virtual power plant performance evaluation unit, and a hybrid modeling strategy unit; The hybrid modeling strategy unit is specifically as follows: Based on the first and second laws of thermodynamics and the principles of combustion dynamics, a first-principles model is constructed to accurately characterize the core physical mechanisms of energy conversion, heat transfer and pollutant generation in the unit, ensuring the physical consistency of the modeling process. At the same time, a deep neural network structure is used to build an AI model. By learning the residuals between the historical operating data of the generator set and the calculation results of the first-principles model, the calculation errors caused by simplification assumptions and nonlinear characteristics of the system are compensated in the physical modeling process. The core design logic of the hybrid modeling framework is as follows: a first-principles model is used for subsystems with clear mechanisms, while an AI model is used for subsystems with complex mechanisms or those that are difficult to model accurately. The two models work together through a residual connection architecture. The first-principles model outputs preliminary performance calculation results, while the AI ​​model learns the error between the first-principles model and the actual system and corrects the preliminary calculation results through residuals. Finally, a hybrid modeling framework with both physical interpretability and computational accuracy is formed, providing core modeling support for the unit performance calculation unit.

[0070] The unit performance calculation unit is specifically as follows: Based on the hybrid modeling framework provided by the hybrid modeling strategy unit, the high-fidelity thermodynamic state data of the generator set output by the thermodynamic digital twin module is received. Combined with the actual operating parameters of the generator set (such as fuel characteristics, steam parameters, load conditions, etc.), the core performance data of the generator set is calculated using the fundamental laws of thermodynamics (the first law of thermodynamics and the second law of thermodynamics) as the core calculation basis. This includes key performance indicators such as heat rate, operating efficiency, coal consumption, and pollutant (such as NOx) emission concentration.

[0071] The virtual power plant performance evaluation unit is specifically as follows: Based on the core performance data of each unit output by the unit performance calculation unit, and receiving grid dispatch instructions, market price signals, and equipment constraints, an overall performance evaluation of the virtual power plant is conducted. During the evaluation process, the calculation of peak-shaving capacity of each unit and the overall operational economic analysis of the virtual power plant are used as core evaluation dimensions. The evaluation results are quantified using a comprehensive performance scoring formula, which is: Overall performance score = 0.6 × peak shaving capacity + 0.4 × operating economy Among them, peak-shaving capacity has a weight of 0.6, and operational economy has a weight of 0.4; The final output is an overall performance evaluation conclusion of the virtual power plant, providing core decision-making basis for the AI ​​analysis and optimization module to generate operation optimization strategies.

[0072] Furthermore, the AI ​​analysis and optimization module includes a quantitative analysis and diagnosis unit, a strategy generation unit, and a strategy verification unit; The quantitative analysis and diagnostic unit specifically comprises: Based on the high-fidelity unit thermal state data output by the thermodynamic digital twin module and the performance analysis data (including core indicators such as heat rate, efficiency, coal consumption, and pollutant emission concentration) output by the expert model module, multi-dimensional quantitative analysis is carried out using deep neural network and knowledge graph technology. By capturing feature patterns of data deviating from normal operating range through deep neural networks, and combining the generator set equipment association logic and operating rules integrated with knowledge graphs, operational deviation anomalies are identified and their root causes are traced. Based on the time-series change data of performance indicators, the performance evolution law within the normal operating life cycle of the equipment is learned through deep neural networks, capturing the degradation characteristics of gradual performance decline. The correlation between degradation characteristics and equipment operating status is verified with the help of knowledge graphs, realizing accurate detection and root cause location of equipment degradation anomalies. Finally, the anomaly diagnosis conclusions and root cause analysis results are output, providing core decision-making basis for the strategy generation unit to formulate targeted operation optimization strategies.

[0073] The strategy generation unit is specifically: A policy generation model is constructed using a physical constraint reinforcement learning algorithm. The total number of training rounds is set, and operations such as policy generation, constraint verification, and parameter optimization are performed cyclically in each round to ensure that the model gradually converges to the optimal state through multiple iterations.

[0074] In each round of training, based on the policy network of the current reinforcement learning model, the current operating status data of the generating unit and the virtual power station are input (including high-fidelity thermodynamic state data output by the thermodynamic digital twin module, performance data output by the expert model module, and abnormal diagnosis conclusions and root cause analysis results output by the quantitative analysis and diagnosis unit) to generate operation optimization strategies for the generating unit and the virtual power station (in terms of load allocation, combustion optimization, equipment maintenance and adjustment, etc.); at the same time, exploration noise is introduced to ensure the diversity of strategy search during training and avoid the model getting trapped in local optima.

[0075] Furthermore, the physical constraint reinforcement learning algorithm embeds the fundamental laws of thermodynamics as constraints into the reward function and loss function to achieve constraint verification; the constraints include energy conservation constraints and entropy increase inequality constraints; during the modeling process, the constraint weight λ decays from 1.0 to 0.1 according to a preset rule, while the data weight increases from 0.1 to 1.0, achieving dynamic coordination between constraints and data.

[0076] Furthermore, the energy conservation constraint, based on the first law of thermodynamics, calculates the difference between the total input energy and the total output energy of the system after the strategy is executed during the constraint verification phase, ensuring that the constraint error is satisfied.

[0077] in, Input total energy into the system, The system outputs total energy; Furthermore, the entropy increase inequality constraint, based on the second law of thermodynamics, constructs a negative entropy production penalty mechanism through the ReLU function, defining the entropy increase constraint penalty term as:

[0078] in, The entropy change of the system after the strategy is executed, when When <0 (i.e., negative entropy production occurs, violating the second law of thermodynamics), the penalty term outputs a non-zero value; otherwise, it outputs a zero value, thus effectively constraining negative entropy production. The verification results of the two constraints are combined into a physical constraint violation quantification value; Based on the quantification of physical constraints, the reward and loss functions are constructed: On the one hand, the constraint-based reward is calculated, with the following formula:

[0079] Where r is the basic reward. The constraint weights are the penalty terms for violating physical constraints; Construct the policy update loss function, the formula is:

[0080] in, To use PPO as the base loss, the influence of enhanced constraints on policy updates is strengthened, and policy network optimization is completed through a parameter update mechanism. The policy verification unit is specifically: The system receives the operation optimization strategy output by the strategy generation unit, and simultaneously collects the basic thermal state data of the unit output by the thermodynamic digital twin module and the performance benchmark data output by the expert model module to construct a verification input dataset to ensure the consistency between the simulation scenario and the actual operating conditions of the unit. The pre-configured structure-preserving neural network is activated. This network contains four hidden layers, each with 128 neurons. The input layer dimension matches the feature dimension of the validation input dataset, and the output layer dimension corresponds to the system state parameters after policy execution. An adaptive Swish activation function is used, with the expression:

[0081] in, =1, which is a trainable parameter used to enhance the network's ability to fit the state of complex nonlinear systems; The structure maintains that the constraint loss function of the neural network includes energy conservation constraints and entropy increase inequality constraints, which are consistent with the physical constraints of the policy generation unit. The running optimization strategy in the verification input dataset is used as the input of the structure-preserving neural network. Through forward propagation calculation, the system state parameters of the generator set and virtual power station after the strategy is executed are digitally simulated to achieve accurate mapping of the optimization strategy execution effect. The forward propagation logic follows the basic rules of neural network signal transmission and completes data transformation and state output based on the preset network layers and activation functions. The core execution flow of forward propagation is as follows: first, the data of each hidden layer is processed by the adaptive Swish activation function, and then physical constraints are embedded in the output layer to complete the final output; Based on the output of the neural network and the verification of physical constraints, the total verification loss value is calculated. If the loss value meets the preset compliance requirements, the optimization strategy is determined to comply with the basic laws of thermodynamics and the requirements for safe operation of the equipment. The optimization strategy is then output to the generator set / DCS system for execution, and the verification result is fed back to the thermodynamic digital twin module for model parameter updates. If the loss value does not meet the preset compliance requirements, the optimization strategy is determined to have the risk of violating physical laws. A rejection instruction and details of the constraint violation are output to the strategy generation unit, triggering the regeneration of the optimization strategy.

[0082] Furthermore, the trustworthiness assurance module includes an explainable AI technology unit, an uncertainty quantification unit, and a data privacy protection unit; The explainable AI technology unit specifically includes: Using the anomaly diagnosis conclusions, optimization strategies, and corresponding related data output by the AI ​​analysis and optimization module as the analysis objects, the formation basis of the anomaly diagnosis conclusions and the generation path of the optimization strategies are logically decomposed through a locally interpretable model to locate the key influencing factors of decision-making; the mechanism of action of each input feature and decision result is traced by feature correlation analysis technology to clarify the feature weights and correlation logic; through the complete process of "feature importance analysis → decision path tracing → natural language interpretation generation", the AI ​​decision-making logic is visualized and interpreted, improving the transparency of decision-making.

[0083] The uncertainty quantification unit is specifically: An uncertainty quantification model is constructed based on a Bayesian machine learning framework. The analysis objects include the unit thermal state simulation results output by the thermodynamic digital twin module, the unit and virtual power plant performance calculation results output by the expert model module, the anomaly diagnosis conclusions, prediction results, and operation optimization strategy related evaluation data output by the AI ​​analysis and optimization module. The error sources, data fluctuation range, and model uncertainty of various results are systematically analyzed through Bayesian inference mechanism. The confidence intervals corresponding to each result are calculated and output, clarifying the reliable value range of the results under different confidence levels, and intuitively presenting the degree of uncertainty of the output data.

[0084] The data privacy protection unit specifically comprises: A multi-site collaborative optimization mechanism is constructed using a federated learning architecture, with local sensitive operating data of each generator unit and the collaborative training requirements of multi-site models as the core inputs. Through the distributed training mode of federated learning, each generator unit can participate in the collaborative training of cross-site models only through parameter sharing and gradient aggregation, while retaining the original sensitive data locally and not transmitting or leaking privacy information. This avoids the privacy leakage risks faced by sensitive data during transmission and centralized storage. At the same time, the advantages of multi-site data are integrated to jointly optimize the model, effectively improving the model's adaptability to different unit operating conditions and regional operating scenarios, and enhancing the model's generalization performance.

[0085] Example 2: This invention also provides a multi-quantization analysis method for generator sets based on thermodynamic digital twins and artificial intelligence, which communicates through a unified data bus and API interface; Step S1, the thermal state simulation and data generation step, in which: The collected multi-source heterogeneous data undergoes quality assessment and adaptive weighted fusion processing to remove outlier data and output standardized data. Based on thermodynamic laws and conservation equations, a differential algebraic equation model of the unit's thermodynamic process is constructed, and thermodynamically sensed reduced-order modeling is used to simplify the model complexity. The thermodynamic state data output by the model simulation is compared with the actual measured data, and the model parameters are dynamically adjusted through a parameter estimation method based on Bayesian inference. The model is then optimized according to a graded calibration strategy to output high-fidelity unit thermodynamic state data. Step S2, the step of quantitative analysis of unit and virtual power plant performance, in which: A hybrid modeling framework is constructed that integrates first-principles models and AI models through residual connections. Based on the fundamental laws of thermodynamics and actual operating parameters of generator sets, and combined with high-fidelity thermodynamic state data of the units, quantitative analysis is conducted on the performance of individual units and the overall performance of the virtual power station. Step S3, the step of generating and verifying the anomaly diagnosis and operation optimization strategy, in which: By calling upon the thermal state data and performance analysis data of the generating units, anomaly identification and root cause analysis are completed through deep neural networks and knowledge graph technology; physical constraint reinforcement learning algorithms are used to generate operation optimization strategies; and the strategies are verified with the help of structure-preserving neural networks and related constraint loss functions, forming a closed-loop management and outputting verified optimization strategies to provide decision support for the operation optimization of generating units and virtual power plants. Step S4, the steps for ensuring system trustworthiness and data security, in which: By visualizing and interpreting the AI ​​decision-making logic, the transparency of decision-making is improved. Uncertainty assessments are conducted on the output results of each module, and confidence interval estimates are provided to clarify the reliable range of the results. At the same time, a specific architecture is adopted to achieve multi-site model collaborative optimization while preserving the privacy of sensitive operational data, taking into account both data security and model generalization ability, and providing trust support for high-risk decision-making scenarios and large-scale applications of the system.

[0086] Furthermore, step S1 is used to simulate the thermal state of the unit based on boundary conditions and output high-fidelity thermal state data of the unit; specifically, it also includes the following steps: Step S101, the multi-source data fusion step, in which: Multi-source heterogeneous data from the generator set's DCS system, SIS system, and sensor system are collected, and data quality is assessed. The data quality assessment includes integrity checks, physical consistency verification, and timeliness assessment. The integrity check calculates an integrity score (C) based on the proportion of missing data. The physical consistency verification obtains a consistency score (CS) through testing against fundamental thermodynamic laws. The timeliness assessment is achieved through timestamp analysis and delay calculation, yielding a timeliness score (T). The overall data quality is quantified using a data quality assessment formula, which is:

[0087] in, For quality score, For completeness score, For consistency score, Score for timeliness; The quality score is co-calculated with the pre-defined confidence metric of the data source. The softmax normalization algorithm is used to normalize the fusion calculation result in the direction with dimension 0, generating dynamic fusion weights for each data source. Based on the generated dynamic fusion weights, the effective multi-source heterogeneous data that has been quality assessed is weighted and summed. Tensor operations are used to achieve data dimension alignment and numerical integration, and finally, standardized fusion data and corresponding dynamic fusion weights are output. Step S102, the dynamic simulation step, in which: The system receives standardized fusion data and uses it as boundary conditions and initial parameters for the simulation model, providing basic data support for the simulation of thermal processes and ensuring that the simulation input is consistent with the actual operating conditions of the unit. Based on the first and second laws of thermodynamics, and combined with the mass conservation equation and momentum conservation equation, a differential-algebraic equation model of the unit's thermodynamic process is constructed. The dynamic characteristics of the unit's thermodynamic system are described by differential equations, and the static constraint relationships between the system's parameters are described by algebraic equations. The mass conservation equation is expressed as follows:

[0088] in, Let be the local rate of change of fluid density with time t; Let be the rate of change of fluid mass due to transport. For fluid density, This is the fluid velocity vector.

[0089] The first law of thermodynamics, also known as the law of conservation of energy, is applied to thermodynamic systems and is expressed as follows:

[0090] in, For fluid density, For isobaric specific heat capacity, For fluid temperature, The divergence of heat conduction flux; The expression for the conservation of momentum is:

[0091] in, For fluid density, This refers to fluid pressure. For fluid velocity vector, For dynamic viscosity, The Laplace operator for velocity; Thermodynamically-aware reduced-order modeling technique is adopted. By using a sparse autoencoder, key feature variables that significantly affect the thermodynamic process are extracted from multiple original variables contained in the differential algebraic equation model. Redundant variables are eliminated, reducing the computational complexity of the model without sacrificing the core simulation accuracy. Run the optimized differential-algebraic equation model to simulate the thermodynamic operation of the unit under the current boundary conditions and output preliminary thermodynamic state simulation data.

[0092] Step S103, the real-time calibration step, in which: Receive preliminary thermodynamic state simulation data, and simultaneously receive actual measurement data collected by the generator set's DCS system and SIS system; match the preliminary thermodynamic state simulation data with the actual measurement data one by one, and calculate the parameter deviation between the two; Constrained by the laws of thermodynamics, a parameter estimation method based on Bayesian inference is adopted, combined with extended Kalman filtering to dynamically adjust the key parameters of the model. In this process, the process noise matrix is ​​a diagonal matrix [0.01, 0.05, 0.02, ...], and the observation noise matrix is ​​a diagonal matrix [0.02, 0.03, 0.01, ...]. Perform the operation according to the graded calibration strategy: high frequency calibration (calibration cycle of 1 minute) for key operating parameters such as main steam temperature and pressure, medium frequency calibration (calibration cycle of 5 minutes) for performance parameters such as heat rate and efficiency, and low frequency calibration (calibration cycle of 30 minutes) for equipment status parameters corresponding to wear and scaling. The final output is high-fidelity thermal status data of the unit.

[0093] Furthermore, step S2 specifically includes the following steps: Step S201, the hybrid modeling step, in which: Based on the first and second laws of thermodynamics and the principles of combustion dynamics, a first-principles model is constructed to accurately characterize the core physical mechanisms of energy conversion, heat transfer and pollutant generation in the unit, ensuring the physical consistency of the modeling process. At the same time, a deep neural network structure is used to build an AI model. By learning the residuals between the historical operating data of the generator set and the calculation results of the first-principles model, the calculation errors caused by simplification assumptions and nonlinear characteristics of the system are compensated in the physical modeling process. The core design logic of the hybrid modeling framework is as follows: a first-principles model is used for subsystems with clear mechanisms, and an AI model is used for subsystems with complex mechanisms or difficult to model accurately. The two models work together through a residual connection architecture. The first-principles model outputs preliminary performance calculation results, and the AI ​​model learns the error between the first-principles model and the actual system. The preliminary calculation results are corrected through residuals, and finally a hybrid modeling framework with both physical interpretability and computational accuracy is formed. Step S202, the step of unit performance calculation, in which: Based on the hybrid modeling framework, high-fidelity thermodynamic state data of the generator unit is received and combined with the actual operating parameters of the generator unit (such as fuel characteristics, steam parameters, load conditions, etc.). Using the fundamental laws of thermodynamics (the first law of thermodynamics and the second law of thermodynamics) as the core calculation basis, the core performance data of the generator unit is calculated, including key performance indicators such as heat rate, operating efficiency, coal consumption, and pollutant (such as NOx) emission concentration.

[0094] Step S203, the virtual power plant performance evaluation step, in which: Based on the core performance data of each generating unit, and receiving grid dispatch instructions, market price signals, and equipment constraints, an overall performance evaluation of the virtual power plant is conducted. During the evaluation process, the calculation of peak-shaving capacity of each generating unit and the overall operational economic analysis of the virtual power plant are used as core evaluation dimensions. The evaluation results are quantified using a comprehensive performance scoring formula, which is as follows: Overall performance score = 0.6 × peak shaving capacity + 0.4 × operating economy Among them, peak-shaving capacity has a weight of 0.6, and operational economy has a weight of 0.4; The final output is the overall performance evaluation conclusion of the virtual power plant; Furthermore, step S3 specifically includes the following steps: Step S301, the quantitative analysis and diagnosis step, in which: Based on high-fidelity unit thermal state data and performance analysis data (including core indicators such as heat rate, efficiency, coal consumption, and pollutant emission concentration), multi-dimensional quantitative analysis is carried out using deep neural network and knowledge graph technology; By capturing feature patterns of data deviating from normal operating range through deep neural networks, and combining the generator set equipment association logic and operating rules integrated by knowledge graphs, operational deviation anomalies are identified and their root causes are traced. Based on the time-series change data of performance indicators, the performance evolution law within the normal operating life cycle of the equipment is learned through deep neural networks, capturing the degradation characteristics of gradual performance decline. The correlation between degradation characteristics and equipment operating status is verified with the help of knowledge graphs, realizing accurate detection and root cause location of equipment degradation anomalies. Finally, anomaly diagnosis conclusions and root cause analysis results are output. Step S302, the strategy generation step, in which: A strategy generation model is constructed using a physical constraint reinforcement learning algorithm. High-fidelity unit thermal state data, unit and virtual power plant performance data, as well as anomaly diagnosis conclusions and root cause analysis results are used as joint data inputs to generate operation optimization strategies (including load allocation, combustion optimization, equipment maintenance and adjustment, etc.) for generator units and virtual power plants. Furthermore, the physical constraint reinforcement learning algorithm embeds the fundamental laws of thermodynamics as constraints into the reward function, specifically including energy conservation constraints and entropy increase inequality constraints; during the modeling process, the constraint weight λ decays from 1.0 to 0.1 according to a preset rule, while the data weight increases from 0.1 to 1.0, realizing dynamic coordination between constraints and data; Furthermore, the energy conservation constraint is based on the first law of thermodynamics, calculating the difference between the total input energy and the total output energy of the system after the strategy is executed, ensuring that the constraint error is satisfied:

[0095] in, Input total energy into the system, The system outputs total energy; Furthermore, the entropy increase inequality constraint, based on the second law of thermodynamics, constructs a negative entropy production penalty mechanism through the ReLU function, defining the entropy increase constraint penalty term as:

[0096] in, The entropy change of the system after the strategy is executed, when When <0 (i.e., negative entropy production occurs, violating the second law of thermodynamics), the penalty term outputs a non-zero value; otherwise, it outputs a zero value, thus effectively constraining negative entropy production. The core execution loop of the physical constraint reinforcement learning algorithm is: episode iteration → policy execution → physical constraint verification → constraint reward calculation → policy update. This loop enables the linkage between constraint verification and policy optimization.

[0097] Step S303, the policy verification step, in which: The system receives operational optimization strategies and collects basic thermal state data and performance benchmark data of the unit to construct a verification input dataset, ensuring the consistency between the simulated scenario and the actual operating conditions of the unit. The pre-configured structure-preserving neural network is activated. This network contains four hidden layers, each with 128 neurons. The input layer dimension matches the feature dimension of the validation input dataset, and the output layer dimension corresponds to the system state parameters after policy execution. An adaptive Swish activation function is used, with the expression:

[0098] in, =1, which is a trainable parameter used to enhance the network's ability to fit the state of complex nonlinear systems; The structure maintains that the constraint loss function of the neural network includes energy conservation constraints and entropy increase inequality constraints, which are consistent with the physical constraints of the policy generation unit. The running optimization strategy in the verification input dataset is used as the input of the structure-preserving neural network. Through forward propagation calculation, the system state parameters of the generator set and virtual power station after the strategy is executed are digitally simulated to achieve accurate mapping of the optimization strategy execution effect. The forward propagation logic follows the basic rules of neural network signal transmission and completes data transformation and state output based on the preset network layers and activation functions. The core execution flow of forward propagation is as follows: first, the data of each hidden layer is processed by the adaptive Swish activation function, and then physical constraints are embedded in the output layer to complete the final output; Based on the output of the neural network and the verification of physical constraints, the total verification loss value is calculated. If the loss value meets the preset compliance requirements, the optimization strategy is determined to comply with the basic laws of thermodynamics and the requirements for safe operation of the equipment. The optimization strategy is then output to the generator set / DCS system for execution, and the verification result is fed back to the thermodynamic digital twin module for model parameter updates. If the loss value does not meet the preset compliance requirements, the optimization strategy is determined to have the risk of violating physical laws. A rejection instruction and details of the constraint violation are output to the strategy generation unit, triggering the regeneration of the optimization strategy.

[0099] Furthermore, step S4 specifically includes: Step S401, Explaining the steps of AI technology, in which: Taking anomaly diagnosis conclusions, operational optimization strategies, and corresponding related data as the analysis objects, this study uses a locally interpretable model to logically decompose the formation basis of anomaly diagnosis conclusions and the generation path of optimization strategies, thereby identifying key influencing factors in decision-making. It also leverages feature correlation analysis technology to trace the interaction mechanism between each input feature and the decision result, clarifying feature weights and their correlation logic. Through a complete process of "feature importance analysis → decision path tracing → natural language interpretation generation," the study achieves a visual interpretation of AI decision-making logic, enhancing decision transparency.

[0100] Step S402, the uncertainty quantification step, in which: An uncertainty quantification model is constructed based on a Bayesian machine learning framework. The analysis objects include the results of unit thermal state simulation, the performance calculation results of the unit and virtual power station, the anomaly diagnosis conclusions, the prediction results, and the relevant evaluation data of operation optimization strategies. The model systematically analyzes the error sources, data fluctuation range, and model uncertainty of various results through Bayesian inference mechanism. The model calculates and outputs the confidence intervals corresponding to each result, clarifies the reliable value range of the results under different confidence levels, and intuitively presents the degree of uncertainty of the output data.

[0101] Step S403, the data privacy protection step, in which: A multi-site collaborative optimization mechanism is constructed using a federated learning architecture, with local sensitive operating data of each generator unit and the collaborative training requirements of multi-site models as the core inputs. Through the distributed training mode of federated learning, each generator unit can participate in the collaborative training of cross-site models only through parameter sharing and gradient aggregation, while retaining the original sensitive data locally and not transmitting or leaking privacy information. This avoids the privacy leakage risks faced by sensitive data during transmission and centralized storage. At the same time, the advantages of multi-site data are integrated to jointly optimize the model, effectively improving the model's adaptability to different unit operating conditions and regional operating scenarios, and enhancing the model's generalization performance.

[0102] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The methods disclosed in the embodiments are described simply because they correspond to the systems disclosed in the embodiments; relevant details can be found in the method section.

[0103] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0104] In the embodiments provided by this invention, it should be understood that the disclosed systems, methods, and approaches can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, or other forms.

[0105] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0106] In addition, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit.

[0107] Similarly, in the various embodiments of the present invention, each processing unit can be integrated into a functional module, or each processing unit can exist physically, or two or more processing units can be integrated into a functional module.

[0108] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0109] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0110] The above-disclosed embodiments are merely preferred embodiments of the present invention, but the present invention is not limited thereto. Any non-creative variations that can be conceived by those skilled in the art, as well as any improvements and modifications made without departing from the principles of the present invention, should fall within the protection scope of the present invention.

Claims

1. A multi-quantitative analysis system for generator sets based on thermodynamic digital twins and artificial intelligence, characterized in that, include: Thermodynamic digital twin module, in which: The collected multi-source heterogeneous data is subjected to quality assessment and adaptive weighted fusion processing to output standardized data; Based on the laws of thermodynamics and conservation equations, a differential algebraic equation model of the unit's thermodynamic process is constructed. Thermodynamically sensed reduced-order modeling is used to simplify the model complexity. The thermodynamic state data output by the model simulation is compared with the actual measured data. The model parameters are dynamically adjusted by a parameter estimation method based on Bayesian inference and optimized according to a graded calibration strategy to output high-fidelity unit thermodynamic state data. The expert model module contains: A hybrid modeling framework is constructed that integrates first-principles models and AI models through residual connections. Based on the fundamental laws of thermodynamics and actual operating parameters of generator sets, and combined with high-fidelity thermodynamic state data of the units output by the thermodynamic digital twin module, quantitative analysis of the performance of individual units and the overall performance of the virtual power station is carried out. The AI ​​analysis and optimization module contains: The system utilizes unit thermal state data from the thermodynamic digital twin module and performance analysis data from the expert model module, employing deep neural networks and knowledge graph technology to perform anomaly identification and root cause analysis. It then generates operational optimization strategies using physical constraint reinforcement learning algorithms, verifies these strategies with structure-preserving neural networks and constraint loss functions, and outputs validated optimization strategies. The trustworthiness assurance module contains: The AI ​​decision-making logic is visualized and interpreted, and the uncertainty of the output results of each module is assessed and confidence intervals are estimated. A federated learning architecture is adopted to achieve collaborative optimization of multi-site models.

2. The system according to claim 1, characterized in that, The thermodynamic digital twin module includes a multi-source data fusion unit, a dynamic simulation unit, and a real-time calibration unit; The multi-source data fusion unit is specifically: Collect multi-source heterogeneous data from generator set DCS system, SIS system, and sensor system, and conduct data quality assessment on the multi-source heterogeneous data; The dynamic simulation unit is specifically: The system receives standardized fused data output from the multi-source data fusion unit and uses it as the boundary conditions and initial parameters of the simulation model. Based on the first and second laws of thermodynamics, and combined with the mass conservation equation and momentum conservation equation, a differential algebraic equation model of the unit's thermodynamic process is constructed. Thermodynamically-aware reduced-order modeling technique is adopted. By using a sparse autoencoder, key feature variables that have a significant impact on the thermodynamic process are extracted from multiple original variables contained in the differential algebraic equation model. Redundant variables are eliminated, reducing the computational complexity of the model. Run the reduced-order optimized differential-algebraic equation model to simulate the thermodynamic operation of the unit under the current boundary conditions and output preliminary thermodynamic state simulation data; The real-time calibration unit is specifically: The system receives preliminary thermodynamic state simulation data output by the dynamic simulation unit, and simultaneously receives actual measurement data collected by the generator set's DCS system and SIS system; it matches the preliminary thermodynamic state simulation data with the actual measurement data one by one, and calculates the parameter deviation between the two. Constrained by the laws of thermodynamics, a parameter estimation method based on Bayesian inference is adopted, combined with extended Kalman filtering to dynamically adjust the key parameters of the model. The operation is performed according to a hierarchical calibration strategy: high-frequency calibration of key operating parameters, medium-frequency calibration of performance parameters, and low-frequency calibration of equipment status parameters; finally, high-fidelity unit thermal state data are output.

3. The system according to claim 2, characterized in that, In the multi-source data fusion unit, data quality assessment includes integrity check, physical consistency verification, and timeliness assessment; The integrity check is performed by calculating the proportion of missing data to obtain an integrity score; the physical consistency verification is performed by testing against the fundamental laws of thermodynamics to obtain a consistency score; the timeliness assessment is performed by timestamp analysis and delay calculation to obtain a timeliness score; and the overall data quality is quantified according to the data quality assessment formula, which is: in, For quality score, For completeness score, For consistency score, Score for timeliness; The quality score is co-calculated with the pre-defined confidence metric of the data source. The softmax normalization algorithm is used to normalize the fusion calculation result in the direction with dimension 0, generating dynamic fusion weights for each data source. Based on the generated dynamic fusion weights, the effective multi-source heterogeneous data that has passed the quality assessment is weighted and summed. Tensor operations are used to achieve data dimension alignment and numerical integration, and finally, standardized fusion data and corresponding dynamic fusion weights are output.

4. The system according to claim 1, characterized in that, The expert model module includes a unit performance calculation unit, a virtual power plant performance evaluation unit, and a hybrid modeling strategy unit; The hybrid modeling strategy unit is specifically as follows: Based on the first and second laws of thermodynamics and the principles of combustion dynamics, a first-principles model is constructed; an AI model is built using a deep neural network structure, and the calculation error in the physical modeling process is compensated by learning the residual between the historical operating data of the generator set and the calculation results of the first-principles model. First-principles models are used for subsystems with well-defined mechanisms, while AI models are used for subsystems with complex mechanisms or those that are difficult to model accurately. The two models work together through a residual connection architecture. The first-principles model outputs preliminary performance calculation results, the AI ​​model learns the error between the first-principles model and the actual system, corrects the preliminary calculation results through residuals, and finally forms a hybrid modeling framework; The unit performance calculation unit is specifically as follows: Based on the hybrid modeling framework provided by the hybrid modeling strategy unit, the high-fidelity thermodynamic state data of the generator set output by the thermodynamic digital twin module is received, and the core performance data of the generator set is calculated based on the basic laws of thermodynamics, combined with the actual operating parameters of the generator set. The virtual power plant performance evaluation unit is specifically as follows: Based on the core performance data of each unit output by the unit performance calculation unit, the system receives grid dispatch instructions, market price signals, and equipment constraints to conduct an overall performance evaluation of the virtual power plant. During the evaluation process, the calculation of the peak-shaving capacity of each unit and the overall economic analysis of the virtual power plant are taken as core evaluation dimensions, and the evaluation results are quantified through a comprehensive performance scoring formula. Finally, the overall performance evaluation conclusion of the virtual power plant is output.

5. The system according to claim 1, characterized in that, The AI ​​analysis and optimization module includes a quantitative analysis and diagnosis unit, a strategy generation unit, and a strategy verification unit. The quantitative analysis and diagnostic unit specifically comprises: Based on high-fidelity unit thermal state data and performance analysis data, deep neural networks are used to capture feature patterns that deviate from the normal operating range. Combined with the generator set equipment association logic and operating rules integrated by knowledge graph, operational deviation anomalies are identified and their root causes are traced. Based on the time-series variation data of performance indicators, the system learns the performance evolution pattern during the normal operation life cycle of the equipment through deep neural networks, captures the degradation characteristics of gradual performance decline, verifies the correlation between degradation characteristics and equipment operating status with the help of knowledge graphs, and locates the accurate detection and root cause of equipment degradation anomalies; finally, it outputs anomaly diagnosis conclusions and root cause analysis results. The strategy generation unit is specifically: A strategy generation model is constructed using a physical constraint reinforcement learning algorithm. High-fidelity unit thermal state data, unit and virtual power plant performance data, as well as anomaly diagnosis conclusions and root cause analysis results are used as data inputs to generate operation optimization strategies for generator units and virtual power plants. The policy verification unit is specifically: The system receives the operation optimization strategy output by the strategy generation unit, and simultaneously collects high-fidelity unit thermal state data and performance data to construct a verification input dataset. The operation optimization strategy from the verification input dataset is input into a pre-set structure-preserving neural network. Through forward propagation calculation, the system state parameters of the generator unit and virtual power station after the strategy is executed are digitally simulated. Based on the output of the structure-preserving neural network and the physical constraint verification, the total verification loss value is calculated. If the loss value meets the preset compliance requirements, the optimization strategy is output. If the conditions are not met, a rejection instruction and details of the constraint violation are output to the policy generation unit, triggering policy regeneration.

6. The system according to claim 5, characterized in that, The physical constraint reinforcement learning algorithm embeds the fundamental laws of thermodynamics as constraints into the reward function, specifically including energy conservation constraints and entropy increase inequality constraints. During the modeling process, the constraint weight λ decays from 1.0 to 0.1 according to a preset rule, while the data weight increases from 0.1 to 1.0, achieving dynamic coordination between constraints and data. The energy conservation constraint is based on the first law of thermodynamics. It calculates the difference between the total input energy and the total output energy of the system after the strategy is executed, ensuring that the constraint error is satisfied. in, Input total energy into the system, The system outputs total energy; The entropy increase inequality constraint, based on the second law of thermodynamics, constructs a negative entropy production penalty mechanism using the ReLU function, defining the entropy increase constraint penalty term as follows: in, The entropy change of the system after the strategy is executed, when When the value is less than 0, the penalty term outputs a non-zero value; otherwise, it outputs a zero value, thus effectively constraining the production of negative entropy.

7. The system according to claim 5, characterized in that, The structure-preserving neural network includes 4 hidden layers, each with 128 neurons. The input layer dimension matches the feature dimension of the validation input dataset, and the output layer dimension corresponds to the system state parameter dimension after the policy is executed. The adaptive Swish activation function is used, and its expression is: in, =1, indicating a trainable parameter; The structure maintains that the neural network's constraint loss function includes energy conservation constraints and entropy increase inequality constraints.

8. The system according to claim 1, characterized in that, The credibility assurance module includes an explainable AI technology unit, an uncertainty quantification unit, and a data privacy protection unit; The explainable AI technology unit specifically includes: Taking the anomaly diagnosis conclusions, optimization strategies, and corresponding related data output by the AI ​​analysis and optimization module as the analysis objects, the formation basis of the anomaly diagnosis conclusions and the generation path of the optimization strategies are logically decomposed through a locally interpretable model to locate the key influencing factors of decision-making; and the mechanism of action of each input feature and decision result is traced by feature correlation analysis technology to clarify the feature weights and correlation logic. The uncertainty quantification unit is specifically: An uncertainty quantification model is constructed based on a Bayesian machine learning framework. The analysis objects are the unit thermal state simulation results output by the thermodynamic digital twin module, the unit and virtual power plant performance calculation results output by the expert model module, the anomaly diagnosis conclusions, prediction results, and operation optimization strategy related evaluation data output by the AI ​​analysis and optimization module. The error sources, data fluctuation range, and model uncertainty of various results are systematically analyzed through Bayesian inference mechanism; the confidence intervals corresponding to each result are calculated and output. The data privacy protection unit specifically comprises: A multi-site collaborative optimization mechanism is constructed using a federated learning architecture, with sensitive local operating data of each generator unit and the collaborative training requirements of multi-site models as the core inputs. Through the distributed training mode of federated learning, each generator unit can participate in the collaborative training of cross-site models only through parameter sharing, gradient aggregation, etc., while retaining the original sensitive data locally and not transmitting or leaking privacy information. This avoids the privacy leakage risk faced by sensitive data during transmission and centralized storage. The advantages of multi-site data are integrated to jointly optimize the model.

9. A multi-quantitative analysis system method for generator sets based on thermodynamic digital twins and artificial intelligence, characterized in that, Includes the following steps: S1, the steps of thermodynamic state simulation and data generation, in which: The collected multi-source heterogeneous data is subjected to quality assessment and adaptive weighted fusion processing to output standardized data; Based on the laws of thermodynamics and conservation equations, a differential algebraic equation model of the unit's thermodynamic process is constructed. Thermodynamically sensed reduced-order modeling is used to simplify the model complexity. The thermodynamic state data output by the model simulation is compared with the actual measured data. The model parameters are dynamically adjusted by a parameter estimation method based on Bayesian inference and optimized according to a graded calibration strategy to output high-fidelity unit thermodynamic state data. S2, the steps of quantitative analysis of unit and virtual power plant performance, in which: A hybrid modeling framework is constructed that integrates first-principles models and AI models through residual connections. Based on the fundamental laws of thermodynamics and actual operating parameters of generator sets, and combined with high-fidelity thermodynamic state data of the units output by the thermodynamic digital twin module, quantitative analysis of the performance of individual units and the overall performance of the virtual power station is carried out. S3, the steps for anomaly diagnosis and operational optimization strategy generation and verification, in which: By calling upon the unit's thermal state data and performance analysis data, anomaly identification and root cause analysis are completed through deep neural networks and knowledge graph technology; an operational optimization strategy is generated using a physical constraint reinforcement learning algorithm; the strategy is verified using a structure-preserving neural network and related constraint loss functions, and the verified optimization strategy is output. S4. Steps for ensuring system trustworthiness and data security, in which: The AI ​​decision-making logic is visualized and interpreted, and the uncertainty of the output results of each module is assessed and confidence intervals are estimated. A specific architecture is used to achieve collaborative optimization of multi-site models.

10. The method according to claim 9, characterized in that, Step S1 specifically includes: Receive preliminary thermodynamic state simulation data, and simultaneously receive actual measurement data collected by the generator set's DCS system and SIS system; match the preliminary thermodynamic state simulation data with the actual measurement data one by one, and calculate the parameter deviation between the two; Constrained by the laws of thermodynamics, a parameter estimation method based on Bayesian inference is adopted, combined with extended Kalman filtering to dynamically adjust the key parameters of the model. In this process, the process noise matrix is ​​a diagonal matrix [0.01, 0.05, 0.02, ...], and the observation noise matrix is ​​a diagonal matrix [0.02, 0.03, 0.01, ...]. The operation is performed according to the graded calibration strategy: high frequency calibration of key operating parameters, cycle 1 minute; medium frequency calibration of performance parameters, cycle 5 minutes; low frequency calibration of equipment status parameters, cycle 30 minutes; and finally output high-fidelity unit thermal status data.

Images