CN-121997753-A - Multiple quantitative analysis system and method for generator set based on thermodynamic digital twin and artificial intelligence

CN121997753ACN 121997753 ACN121997753 ACN 121997753ACN-121997753-A

Abstract

The invention discloses a multiple quantitative analysis system and method for a generator set based on thermodynamic digital twin and artificial intelligence, wherein the multiple quantitative analysis system comprises a thermodynamic digital twin module, an expert model module, an AI analysis optimization module and a credibility guarantee module, wherein the thermodynamic digital twin module is used for simulating the thermodynamic state of the generator set in real time according to boundary conditions, the expert model module is used for calculating the performance of the generator set and the performance of a virtual power station according to actual parameters, the AI analysis optimization module is used for calling intermediate parameters of the digital twin and the expert model to conduct quantitative analysis, diagnosis and generate an optimization strategy and verify the strategy, and the credibility guarantee module is used for guaranteeing the interpretability and the reliability of the system. The method constructs an AI model taking digital twin and expert models as data and knowledge bases and taking physical rules as constraints, realizes closed-loop management of intelligent analysis, diagnosis and operation optimization of the generator set, and remarkably improves the problems of insufficient modeling accuracy, insufficient strategy reliability, insufficient credibility, unbalanced data privacy and generalization capability and the like in the prior art.

Inventors

WANG JIAN
CAO YUANFU
Cui Haodi
LI YAPENG
Zhang Tingbing
MA YONGFENG
SUN HAITAO
XUE YAQING
YANG DONGPING
NING ZHIXING
PENG YIFU

Assignees

青岛华丰伟业电力科技工程有限公司

Dates

Publication Date: 20260508
Application Date: 20260128

Claims (10)

1. A thermodynamic digital twin and artificial intelligence-based multiple quantitative analysis system for a generator set, comprising: A thermodynamic digital twin module in which: The method comprises the steps of collecting multisource heterogeneous data, carrying out quality evaluation and self-adaptive weighted fusion processing on the multisource heterogeneous data, outputting standardized data, constructing a differential algebraic equation model of a unit thermodynamic process based on thermodynamic laws and conservation equations, simplifying model complexity by adopting reduced modeling of thermodynamic perception, comparing thermodynamic state data output by model simulation with actual measurement data, dynamically adjusting model parameters by a parameter estimation method based on Bayesian reasoning, optimizing according to a hierarchical calibration strategy, and outputting high-fidelity unit thermodynamic state data; An expert model module in which: Constructing a hybrid modeling framework of cooperative work of a first sexual principle model and an AI model through residual connection, and carrying out quantitative analysis on unit monomer performance and overall performance of a virtual power station based on thermodynamic basic law and actual operation parameters of the generator set by combining thermodynamic state data of the high-fidelity unit output by a thermodynamic digital twin module; AI analysis optimization module, wherein: Invoking the thermodynamic state data of the unit of the thermodynamic digital twin module and the performance analysis data of the expert model module, completing anomaly identification and root cause analysis through a deep neural network and a knowledge graph technology, generating an operation optimization strategy by using a physical constraint reinforcement learning algorithm, maintaining the neural network and constraint loss function verification strategy by means of a structure, and outputting the verified optimization strategy; A credibility assurance module, wherein: And performing visual reading on the AI decision logic, performing uncertainty evaluation on output results of all modules, providing confidence interval estimation, and realizing multi-site model collaborative optimization by adopting a federal learning architecture.
2. The system of claim 1, wherein the thermodynamic digital twin module comprises a multi-source data fusion unit, a dynamic simulation unit, and a real-time calibration unit; the multi-source data fusion unit specifically comprises: collecting multi-source heterogeneous data from a generator set DCS system, an SIS system and a sensor system, and performing data quality evaluation on the multi-source heterogeneous data; the dynamic simulation unit specifically comprises: based on the first law of thermodynamics, the second law of thermodynamics, and the combination of the conservation of mass equation and conservation of momentum equation, a differential algebraic equation model of the thermodynamic process of the unit is constructed; The thermodynamic perception reduced order modeling technology is adopted, key characteristic variables which have obvious influence on a thermodynamic process are extracted from a plurality of original variables contained in a differential algebra equation model through a sparse self-encoder, redundant variables are removed, and the calculation complexity of the model is reduced; Operating the differential algebraic equation model after the reduced order optimization, simulating the thermodynamic operation state of the unit under the current boundary condition, and outputting preliminary thermodynamic state simulation data; The real-time calibration unit specifically comprises: Receiving preliminary thermodynamic state simulation data output by a dynamic simulation unit, and simultaneously receiving actual measurement data acquired by a generator set DCS system and an SIS system; the thermodynamic law is used as constraint, a parameter estimation method based on Bayesian reasoning is adopted, the extended Kalman filtering is combined to dynamically adjust key parameters of the model, the operation is performed according to a hierarchical calibration strategy, namely, key operation parameters are calibrated in high frequency, performance parameters are calibrated in medium frequency, equipment state parameters are calibrated in low frequency, and finally, thermodynamic state data of the high-fidelity unit are output.
3. The system of claim 2, wherein in the multi-source data fusion unit, data quality assessment includes integrity checking, physical consistency verification, and timeliness assessment; The integrity check is carried out by calculating the proportion of missing data to obtain an integrity score, the physical consistency verification is carried out by checking the thermodynamic basic law to obtain a consistency score, the timeliness assessment is carried out by time stamp analysis and delay calculation to obtain an timeliness score, and the overall quality of the data is quantified according to a data quality assessment formula, wherein the data quality assessment formula is as follows: Wherein, the For the quality score to be a quality score, As a result of the score of the integrity, In order to be a consistency score, Is a timeliness score; The method comprises the steps of carrying out collaborative calculation on quality scores and preset inherent confidence quantification values of data sources, carrying out normalization processing on fusion calculation results in the direction of 0 dimension by adopting a softmax normalization algorithm to generate dynamic fusion weights corresponding to the data sources, carrying out weighted summation calculation on effective multi-source heterogeneous data subjected to quality evaluation based on the generated dynamic fusion weights, realizing data dimension alignment and numerical integration through tensor operation, and finally outputting standardized fusion data and corresponding dynamic fusion weights.
4. The system according to claim 1, wherein the expert model module comprises a unit performance calculation unit, a virtual power station performance evaluation unit, a hybrid modeling strategy unit; the hybrid modeling strategy unit specifically comprises the following components: Constructing an AI model by adopting a deep neural network structure, and compensating calculation errors in a physical modeling process by learning residual errors of historical operation data of a generator set and calculation results of the first principle model; The first principle model outputs a preliminary performance calculation result, the AI model learns errors between the first principle model and an actual system, and the preliminary calculation result is corrected through residual errors to finally form a mixed modeling frame; The unit performance calculating unit specifically comprises: based on a mixed modeling framework provided by a mixed modeling strategy unit, receiving thermodynamic state data of the high-fidelity unit output by the thermodynamic digital twin module, and calculating core performance data of the generator unit based on a thermodynamic basic law by combining actual operation parameters of the generator unit; The virtual power station performance evaluation unit specifically comprises: based on the core performance data of each unit output by the unit performance calculation unit, receiving a power grid dispatching instruction, market price signals and equipment constraint conditions, carrying out overall performance evaluation of the virtual power plant, taking peak shaving capacity calculation of each unit and overall operation economy analysis of the virtual power plant as core evaluation dimensions in the evaluation process, quantifying the evaluation results through a comprehensive performance scoring formula, and finally outputting the overall performance evaluation conclusion of the virtual power plant.
5. The system of claim 1, wherein the AI analysis optimization module comprises a quantitative analysis and diagnosis unit, a policy generation unit, and a policy verification unit; the quantitative analysis and diagnosis unit specifically comprises: Based on the time sequence change data of the performance index, the degradation characteristic of gradual degradation of the performance is captured through the performance evolution rule in the normal operation life cycle of the deep neural network learning equipment, the relevance of the degradation characteristic and the equipment operation state is verified by means of the knowledge graph, the accurate detection and root cause of the equipment degradation abnormality are positioned, and finally the abnormality diagnosis conclusion and root cause analysis result is output; the strategy generation unit specifically comprises: Constructing a strategy generation model by adopting a physical constraint reinforcement learning algorithm, and generating an operation optimization strategy for the generator set and the virtual power station by taking high-fidelity unit thermodynamic state data, unit and virtual power station performance data, as well as abnormality diagnosis results and root cause analysis results as data input; the policy verification unit specifically comprises: The method comprises the steps of receiving an operation optimization strategy output by a strategy generation unit, collecting thermodynamic state data and performance data of a high-fidelity unit at the same time, constructing a verification input data set, inputting the operation optimization strategy in the verification input data set into a preset structure-keeping neural network, performing digital simulation on system state parameters of a generator set and a virtual power station after the strategy is executed through forward propagation calculation of the network, calculating a verification total loss value based on an output result of the structure-keeping neural network and a physical constraint verification condition, outputting the optimization strategy if the loss value meets a preset compliance requirement, and outputting a refuting instruction and constraint violation details to the strategy generation unit if the loss value does not meet the preset compliance requirement, and triggering strategy regeneration.
6. The system of claim 5, wherein the physical constraint reinforcement learning algorithm embeds a thermodynamic basic law as a constraint condition into a reward function, and specifically comprises an energy conservation constraint and an entropy increase inequality constraint, wherein in the modeling process, a constraint weight lambda is attenuated to 0.1 from 1.0 according to a preset rule, and a data weight is increased to 1.0 from 0.1, so that dynamic coordination of the constraint and the data is realized; According to the first law of thermodynamics, the energy conservation constraint calculates the difference value between the total input energy and the total output energy of the system after strategy execution, and ensures that constraint errors meet the following conditions: Wherein, the The total energy input to the system is provided, Outputting total energy for the system; The entropy increase inequality constraint is characterized in that a negative entropy production penalty mechanism is constructed through a ReLU function according to a second law of thermodynamics, and an entropy increase constraint penalty term is defined as follows: Wherein, the For entropy change value of system after policy execution, when When <0, the penalty term outputs a non-zero value, otherwise, a zero value is output, and the effective constraint on the negative entropy production is realized.
7. The system of claim 5, wherein the structure preserving neural network comprises 4 hidden layers, each layer is provided with 128 neurons, an input layer dimension matches a feature dimension of the verification input data set, and an output layer dimension corresponds to a system state parameter dimension after policy execution; the adaptive Swish activation function is adopted, and the expression is as follows: Wherein, the =1, A trainable parameter; the constraint loss function of the structure-preserving neural network includes an energy conservation constraint and an entropy increase inequality constraint.
8. The system of claim 1, wherein the trust assurance module comprises an interpretable AI technology unit, an uncertainty quantization unit, and a data privacy protection unit; The interpretable AI technical unit comprises the following concrete steps: Taking the abnormal diagnosis conclusion, the operation optimization strategy and the corresponding associated data output by the AI analysis optimization module as analysis objects, logically disassembling the formation basis of the abnormal diagnosis conclusion and the generation path of the optimization strategy through a local interpretable model, and positioning key influence factors of decision; the uncertainty quantization unit specifically comprises: Based on a Bayesian machine learning framework, an uncertainty quantization model is constructed, and an error source, a data fluctuation range and model uncertainty of various results are systematically analyzed through a Bayesian reasoning mechanism by taking a unit thermodynamic state simulation result output by a thermodynamic digital twin module, a unit and virtual power station performance calculation result output by an expert model module, an abnormal diagnosis conclusion output by an AI analysis optimization module, a prediction result and operation optimization strategy related evaluation data as analysis objects; the data privacy protection unit specifically comprises: The method comprises the steps of constructing a multi-station collaborative optimization mechanism by adopting a federal learning architecture, taking local sensitive operation data of each generator set and the multi-station model collaborative training requirement as core input, enabling each generator set to participate in the collaborative training of a cross-station model only by means of parameter sharing, gradient aggregation and the like on the premise that original sensitive data is locally reserved, privacy information is not transmitted outwards or leaked through a federal learning distributed training mode, avoiding privacy leakage risks of sensitive data in the transmission and centralized storage processes, and integrating multi-station data advantages to perform joint optimization on the model.
9. The multiple quantitative analysis system method of the generator set based on the thermodynamic digital twin and artificial intelligence is characterized by comprising the following steps: s1, simulating a thermodynamic state and generating data, wherein in the step: The method comprises the steps of collecting multisource heterogeneous data, carrying out quality evaluation and self-adaptive weighted fusion processing on the multisource heterogeneous data, outputting standardized data, constructing a differential algebraic equation model of a unit thermodynamic process based on thermodynamic laws and conservation equations, simplifying model complexity by adopting reduced modeling of thermodynamic perception, comparing thermodynamic state data output by model simulation with actual measurement data, dynamically adjusting model parameters by a parameter estimation method based on Bayesian reasoning, optimizing according to a hierarchical calibration strategy, and outputting high-fidelity unit thermodynamic state data; s2, quantitatively analyzing the performance of the unit and the virtual power station, wherein the steps are as follows: Constructing a hybrid modeling framework of cooperative work of a first sexual principle model and an AI model through residual connection, and carrying out quantitative analysis on unit monomer performance and overall performance of a virtual power station based on thermodynamic basic law and actual operation parameters of the generator set by combining thermodynamic state data of the high-fidelity unit output by a thermodynamic digital twin module; s3, generating and verifying an abnormality diagnosis and operation optimization strategy, wherein in the step: The method comprises the steps of generating a running optimization strategy by a physical constraint reinforcement learning algorithm, verifying the strategy by means of a structure-preserving neural network and related constraint loss functions, and outputting the verified optimization strategy; s4, a step of system credibility and data security assurance, wherein in the step: and performing visual interpretation on the AI decision logic, performing uncertainty evaluation on output results of all modules, providing confidence interval estimation, and adopting a specific architecture to realize multi-site model collaborative optimization.
10. The method according to claim 9, wherein in step S1, the method further comprises: Receiving preliminary thermodynamic state simulation data, and simultaneously receiving actual measurement data collected by a DCS system and an SIS system of the generator set; Taking a thermodynamic law as a constraint, adopting a parameter estimation method based on Bayesian reasoning, and dynamically adjusting key parameters of a model by combining with an extended Kalman filter, wherein a process noise matrix used in the process is a diagonal matrix [0.01,0.05,0.02, the term. ], and an observation noise matrix is a diagonal matrix [0.02,0.03,0.01, the term. ]; the operation is executed according to a grading calibration strategy, wherein the period is 1 minute, the period is 5 minutes, the period is 30 minutes, and the thermodynamic state data of the high-fidelity unit is finally output.

Description

Multiple quantitative analysis system and method for generator set based on thermodynamic digital twin and artificial intelligence Technical Field The invention belongs to the technical field of optimal control of generator sets, and particularly relates to a thermodynamic digital twin and artificial intelligence-based multiple quantitative analysis system and method for a generator set. Background Under the background of deep propulsion of a 'double-carbon' target and acceleration transformation of an energy structure, the generator set is used as a core energy supply carrier of an electric power system, and the operation efficiency, the safety stability and the low-carbon level of the generator set are directly related to the improvement of the energy utilization efficiency, the reduction of the carbon emission intensity and the safety supply of the electric power system, so that the operation optimization of the generator set becomes a key grip for high-quality development of the industry. At present, the mainstream solution of the analysis and optimization of the generator set is mainly divided into three types, namely, a mechanism model is built based on a thermodynamic first principle, a thermodynamic process and performance indexes of the generator set are described through a conservation equation, but the adaptation capability to a complex nonlinear dynamic process and multi-time-space scale characteristics is limited, a pure data driving AI model is trained by means of historical operation data and used for state fitting and strategy generation, but no physical mechanism support is needed, invalid results against the thermodynamic law are easily generated, a digital twin, expert rule library and AI algorithm are independently applied, basic functions are realized through manual data transmission or simple interface interaction, an optimization strategy is mostly generated by adopting a traditional control algorithm or an unconstrained AI algorithm, and feasibility is verified only through simple rules. Some schemes attempt to introduce data fusion or model correction technology, but lack quality evaluation, dynamic calibration and cross-module collaboration mechanisms of the system, so that closed loop optimization capability is difficult to form. The prior art is a generator set analysis system which tries to integrate a digital twin and an AI model, and the state of the generator set is simulated through the digital twin, and the AI model is combined to assist performance calculation and strategy generation. However, the technology still has significant defects, and is difficult to meet the practical application requirements, such as limited modeling precision, higher prediction error, insufficient strategy reliability, lack of strict thermodynamic constraint embedding and closed-loop verification mechanism for optimizing strategy generation, lack of system reliability, lack of interpretability analysis and uncertainty assessment for AI decision making, and difficulty in balancing data privacy protection and model generalization capability, which is the defect of the prior art. In view of the above, it is desirable to provide a system and a method for multiple quantitative analysis of a generator set based on thermodynamic digital twinning and artificial intelligence, so as to solve the above-mentioned drawbacks in the prior art. Disclosure of Invention The invention aims to provide a multiple quantitative analysis system and method for a generator set based on thermodynamic digital twin and artificial intelligence, which are designed to solve the technical problems of limited modeling precision, high prediction error, insufficient strategy reliability, lack of strict thermodynamic constraint embedding and closed loop verification mechanism for optimizing strategy generation, lack of system reliability, lack of interpretability analysis and uncertainty evaluation for AI decision, and difficulty in balancing data privacy protection and model generalization capability in the prior art. In order to achieve the above purpose, the present invention provides the following technical solutions: the invention provides a multiple quantitative analysis system of a generator set based on thermodynamic digital twin and artificial intelligence, which comprises a thermodynamic digital twin module, an expert model module, an AI analysis optimization module and a credibility guarantee module, wherein each module is communicated with an API interface through a unified data bus to form a closed-loop optimization system; the thermodynamic digital twin module, in which: The method comprises the steps of collecting multisource heterogeneous data, carrying out quality evaluation and self-adaptive weighted fusion treatment, removing abnormal data and outputting standardized data, constructing a differential algebraic equation model of a thermodynamic process of a unit based on thermodynamic laws and conservation equatio