Agent-driven intelligent AI depth prediction method and system for power system

Abstract

The invention relates to the technical field of power equipment control, in particular to an Agent-driven intelligent AI depth prediction method and system for a power system, and provides the following scheme that real-time operation data of a gas turbine is collected through a sensor, historical data and a preset mechanism model are combined, an intelligent AI model is constructed, and key parameters which are difficult to directly measure are predicted. And through deviation analysis, abnormity and deviation sources in the operation state are identified, reinforcement learning and a particle swarm optimization algorithm are utilized to dynamically update model parameters, optimization control parameters are generated, and the operation state of the gas turbine is adjusted. According to the method, the difficult-to-measure parameters can be accurately and reversely deduced from the easy-to-measure data, the operation efficiency and stability are continuously optimized through closed-loop control, the dynamic adjustment requirement under the complex working condition is met, and the economical efficiency and reliability of the gas turbine are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment control, and particularly to an Agent-driven intelligent AI deep prediction method and system for a power system. Background Art

[0002] Existing gas turbine operation control methods mainly rely on static mechanism models and empirical parameters for prediction and regulation. However, due to the nonlinear dynamic changes during the operation of gas turbines, especially the key parameters (such as turbine inlet temperature, compression ratio, gas mixture ratio, etc.) are difficult to directly measure, traditional methods are difficult to achieve high-precision prediction and real-time optimization under dynamic operating conditions. In addition, the identification and regulation of abnormal deviations in the existing technology are usually lagging, and it is difficult to adapt to complex operating condition changes, which limits the further improvement of gas turbine performance.

[0003] The existing technology has the problems proposed in this background art: it is difficult to accurately and real-time infer the key difficult-to-measure parameters of gas turbine operation through easily measurable parameters, and it is impossible to achieve efficient optimization and precise control under dynamic operating conditions. To solve the above problems, this application designs an Agent-driven intelligent AI deep prediction method and system for a power system. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide, in view of the deficiencies of the existing technology, an Agent-driven intelligent AI deep prediction method and system for a power system. By collecting the real-time operation data of a gas turbine through sensors, combining historical data and a preset mechanism model, an intelligent AI model is constructed to predict the key parameters (such as turbine inlet temperature, compression ratio, and gas mixture ratio) that are difficult to directly measure. Through deviation analysis, the abnormal conditions and deviation sources in the operation state are identified, and the model parameters are dynamically updated using reinforcement learning and particle swarm optimization algorithms to generate optimized control parameters and adjust the operation state of the gas turbine. The present invention can accurately infer difficult-to-measure parameters from easily measurable data, continuously optimize the operation efficiency and stability through closed-loop control, adapt to the dynamic adjustment requirements under complex operating conditions, and improve the economy and reliability of gas turbines.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] An Agent-driven intelligent AI deep prediction method for a power system, applied to a gas turbine, the method comprising:

[0007] Collect gas turbine data through sensors, quantitatively analyze the gas turbine data, and calculate an ideal operation data set;

[0008] Input the gas turbine data and a preset mechanism model into a preset intelligent AI model, and output a field operation data set through the intelligent AI model;

[0009] Based on the DCS time series data mining technology, pattern recognition and anomaly detection are carried out on the gas turbine, and control parameters are calculated according to the deviation analysis between the on-site operation data set and the ideal operation data set.

[0010] The method further includes:

[0011] Extract the characteristic parameters of the physical process of the gas turbine operation, analyze and obtain the influencing factors of the characteristic parameters of each component, and construct a mechanism model, where the mechanism model includes the flow compression process of the compressor, the thermal conversion process of the combustor, and the mechanical energy output process of the turbine.

[0012] Quantitatively analyze the gas turbine data to calculate the ideal operation data set, including:

[0013] Clean the gas turbine operation data collected by the sensor;

[0014] Obtain the historical operation data of the gas turbine, and divide the historical operation data into operation states under different load conditions through cluster analysis;

[0015] According to the mechanism model, calculate the optimal operation data corresponding to the operation state under each load condition;

[0016] According to the thermodynamics and energy balance equations, calculate the ideal operation parameters corresponding to the optimal operation data under each working condition, where the ideal operation parameters include compression ratio, gas mixing ratio, and turbine inlet temperature.

[0017] The intelligent AI model includes a first AI model, which is used to model the probability distribution of the compressor inlet and outlet pressures according to the Gaussian mixture model, and output the operating compression ratio by combining the distribution characteristics of the gas turbine operation data and the mechanism model.

[0018] The intelligent AI model further includes a second AI model, which is used to perform time series modeling on the gas turbine operation data, capture the non-linear characteristics and time-dependent relationships in the combustion process of the operation data, calculate the change trend of the gas mixing ratio, and output the operating gas mixing ratio.

[0019] The intelligent AI model further includes a third AI model, which is used to reverse infer the inlet temperature of the turbine, including:

[0020] An input layer, which is used to receive the gas turbine operation data and normalize the operation data, where the operation data also includes the gas turbine historical operation data;

[0021] The modeling layer is used to make a preliminary prediction of the turbine inlet temperature through an ensemble learning algorithm, calculate the preliminary prediction result, and model the normalized operation data based on a graph neural network to construct a graph structure with the operating components and characteristic parameters of the gas turbine as nodes;

[0022] Propagate according to the preliminary prediction result in the graph structure, adjust the inconsistency between the mechanism model and the operation data, and output the feature representation;

[0023] The prediction layer is used to fuse the historical operation data and feature representation of the gas turbine through temporal embedding to generate a temporal feature vector, process the temporal feature vector through a Transformer network, calculate the temperature prediction value, calculate the temporal dependence relationship of the temperature through a self-attention mechanism, calculate the temperature change trend according to the temporal dependence relationship, and correct the temperature prediction value according to the temperature change trend, and output the operating turbine inlet temperature.

[0024] The modeling layer includes an ensemble learning sub-layer, a graph neural network sub-layer and a fusion sub-layer;

[0025] The ensemble learning sub-layer is used to integrate multiple machine learning models to generate multiple preliminary learning sub-models. Each sub-model selects operation data for training according to the mechanism model, outputs the predicted temperature, calculates the weighted weight according to the hit rate of each sub-model in historical predictions, and calculates the preliminary prediction result according to the weighted weight and the predicted temperature;

[0026] The graph neural network sub-layer is used to construct a graph structure with the operating components and characteristic parameters of the gas turbine as nodes according to the operation data, physical mechanism and correlation relationship between components of the gas turbine. Among them, each graph node represents a component or characteristic parameter, and each edge represents the physical dependence relationship between nodes;

[0027] Propagate the preliminary prediction result in the graph structure through an information propagation mechanism. Each node adjusts the preliminary prediction result according to the environmental context it is in during the propagation process, and locally adjusts the graph structure through the feature information between nodes;

[0028] The fusion sub-layer is used to propagate the preliminary prediction result and the adjustment output by the graph structure, update the internal parameters according to the adjustment result, and calculate the feature representation.

[0029] And according to the deviation analysis of the on-site operation data set and the ideal operation data set, the calculated control parameters include:

[0030] Quantitatively analyze the deviation between the on-site operation data set and the ideal operation data set, and calculate the deviation value of the gas turbine operation parameters;

[0031] Output control parameters according to the deviation value, where the reinforcement learning algorithm is applied to output the control parameters, and the Q-learning algorithm is used in the reinforcement learning algorithm to optimize the turbine control parameters through an interactive learning process. The Q-learning algorithm includes state, action, and reward;

[0032] The state represents the operating data and deviation value of the gas turbine, including the on-site operating data set, the ideal operating data set, and the deviation value;

[0033] The action represents the decision made by the Q-learning algorithm in the current state, and the decision is used to adjust the control parameters of the gas turbine;

[0034] The reward measures the operating performance after the adjustment of the gas turbine control parameters. The reward is updated according to the adjusted deviation value and operating state. The reward includes error reward, stability reward, and efficiency reward;

[0035] Calculate the reward function according to the reward, and use the reward function to guide the Q-learning algorithm to update the learning strategy after each adjustment, and output the gas turbine control parameters.

[0036] The method further includes updating the parameters of the intelligent AI model according to the deviation value of the gas turbine operating parameters. The update includes:

[0037] Set the first optimization goal, which is to maximize the parameter inversion accuracy of the intelligent AI model according to the deviation value;

[0038] Define the space of the parameters in the intelligent AI model to obtain a set of space parameters, where the parameters include the learning rate, weight parameters, and time series window of each sub-model in the model;

[0039] Calculate the optimization objective function through the particle swarm optimization algorithm. The optimization objective function is constructed according to the deviation value of the gas turbine operating data, the deviation between the optimized operating parameters and the ideal operating parameters, and the improvement effect of the overall performance of the gas turbine;

[0040] Calculate the difference between the real-time operating data of the gas turbine, the optimized control parameters, and the output of the target deviation. Iteratively update the set of space parameters according to the difference and the optimization objective function to generate the optimal parameter solution of the intelligent AI model.

[0041] An Agent-driven intelligent AI deep prediction system for power systems. The system includes a data acquisition module, a data analysis module, and an optimization control module;

[0042] The data acquisition module is used to collect the operating data of the gas turbine and the historical data set of the gas turbine operation;

[0043] The data analysis module is used to quantitatively analyze the gas turbine data, calculate the ideal operation data set, input the gas turbine data and a preset mechanism model into a preset intelligent AI model, and output the on-site operation data set through the intelligent AI model;

[0044] The optimization control module is used to perform pattern recognition and anomaly detection on the gas turbine according to the DCS timing data mining technology, and calculate control parameters based on the deviation analysis between the on-site operation data set and the ideal operation data set.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] By combining the mechanism model with the intelligent AI model, the present invention realizes the accurate inverse derivation of the difficult-to-measure key parameters (such as turbine inlet temperature, compression ratio, etc.) of the gas turbine operation from the easily measurable parameters, and optimizes and dynamically adjusts the control parameters through deviation analysis and reinforcement learning to form a closed-loop control system. It can adapt to complex dynamic working conditions in real time, and improve the operation efficiency, stability and economy of the gas turbine. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Other features, objects and advantages of the present invention will become more apparent by reading the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0048] Figure 1 It is a schematic flow chart of the Agent-driven intelligent AI depth prediction method for the power system in Embodiment 1 of the present invention;

[0049] Figure 2 It is a schematic diagram of the power system intelligent agent structure in Embodiment 1 of the present invention;

[0050] Figure 3 It is a schematic diagram of the gas turbine monitoring in Embodiment 1 of the present invention;

[0051] Figure 4 It is a structural diagram of the intelligent AI model in Embodiment 1 of the present invention;

[0052] Figure 5 It is a structural diagram of the third AI model in Embodiment 1 of the present invention;

[0053] Figure 6 It is a structural diagram of the modeling layer in Embodiment 1 of the present invention;

[0054] Figure 7 It is a schematic flow chart of a method for updating the parameters of the intelligent AI model in Embodiment 1 of the present invention;

[0055] Figure 8 It is a module diagram of the Agent-driven intelligent AI depth prediction system for the power system in Embodiment 2 of the present invention. Detailed implementation manners

[0056] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0057] Embodiment 1

[0058] Please refer to Figure 1 , an embodiment provided by the present invention: an Agent-driven intelligent AI deep prediction method for a power system, and the specific steps of the method are as follows:

[0059] S1: Collect gas turbine data through sensors, quantitatively analyze the gas turbine data, and calculate an ideal operation data set;

[0060] In this step, the data collection of the gas turbine is completed by sensors deployed in the compressor, combustion chamber, and turbine, specifically including real-time data such as compressor outlet pressure, combustion chamber fuel injection volume, turbine exhaust temperature, air flow rate, and environmental conditions (such as environmental temperature and atmospheric pressure). After the data is collected, these data are subjected to cleaning, normalization, and time alignment processing to ensure the integrity and consistency of the data. Using historical operation data, the historical operation data is divided into operation states under different load conditions through cluster analysis, and the ideal operation parameters of each load condition, including compression ratio, gas mixture ratio, and turbine inlet temperature, are calculated in combination with a mechanism model. An ideal operation data set covering different operation conditions of the gas turbine is generated, providing a standardized reference for subsequent deviation analysis and control optimization.

[0061] S2: Input the gas turbine data and a preset mechanism model into a preset intelligent AI model, and output a field operation data set through the intelligent AI model;

[0062] In this step, the intelligent AI model is composed of multiple sub-models, including a first AI model based on a Gaussian mixture model, a second AI model based on time series modeling, and a third AI model for back-calculating the turbine inlet temperature. The gas turbine data collected in real time is processed by the mechanism model and then input into the intelligent AI model for in-depth learning and modeling analysis. Among them, the first AI model is used to predict the compression ratio, the second AI model is used to capture the dynamic change trend of the gas mixture ratio, and the third AI model combines a graph neural network and a time series feature embedding method to back-calculate the turbine inlet temperature. Through the multi-layer structure of the model, it is possible to accurately predict difficult-to-measure parameters based on easily measurable data, generate a field operation data set, and thus comprehensively characterize the operation state of the gas turbine.

[0063] S3: Based on the DCS time series data mining technology, perform pattern recognition and anomaly detection on the gas turbine, and calculate the control parameters according to the deviation analysis between the on-site operation data set and the ideal operation data set.

[0064] In this step, through quantitative analysis of the deviation values between the on-site operation data set and the ideal operation data set, identify the abnormal patterns in the operation of the gas turbine, and calculate the contribution degree of each component (such as the compressor, combustion chamber, and turbine) to the overall deviation. According to the deviation values, optimize the control parameters through reinforcement learning (Q-learning), and dynamically adjust the compressor inlet pressure, combustion chamber fuel injection amount, and turbine load distribution ratio to maximize the operation efficiency of the gas turbine and reduce the deviation values. Through the closed-loop optimization method, not only can it quickly respond to abnormal situations under dynamic working conditions, but also can achieve the adaptive optimization of the model through continuous iteration, thereby significantly improving the operation stability and economy of the gas turbine.

[0065] In this embodiment, the Agent refers to the power system intelligent agent. Input the collected gas turbine data into the power system intelligent agent. The power system intelligent agent includes a data processing and classification layer and multiple data processing modules. The data processing module includes an intelligent AI model. The data processing and classification layer analyzes the gas turbine data and inputs it into the corresponding intelligent AI model of the power system.

[0066] Please refer to Figure 2 , the schematic diagram of the structure of the power system intelligent agent in the embodiment of the present invention. In this embodiment, the power system intelligent agent is an integrated intelligent control system. Its main function is to receive different data and specific requirements from the power system, analyze them, and output the processing results. The power system intelligent agent includes a data processing and classification layer and multiple data processing modules. Each data processing module is configured with different types of models, and can process the input data according to specific task requirements. The task of the data processing and classification layer in this architecture is to analyze the input data of the model, and according to the requirements of different tasks, input the data into the corresponding data processing module to ensure that the tasks can be effectively processed and accurate processing results can be obtained.

[0067] Specifically, as the core component, the power system intelligent agent undertakes the data analysis and task scheduling functions of the entire system, playing a control role like a brain. Its main task is to receive the real-time operation data from different sensors and data sources in the power system, and make intelligent decisions according to specific task requirements. Through the built-in data processing and classification layer and multiple data processing modules, the intelligent agent selects the most suitable processing method under different tasks, and transmits the input data to the most suitable data processing module to ensure that each task can be efficiently processed and accurate results can be obtained.

[0068] Furthermore, the data processing classification layer of the intelligent agent, as a key part bridging different data processing modules, can quickly identify the task type and transfer data to the corresponding data processing modules according to the task requirements. For example, in the data monitoring task of the gas turbine in the power system, the intelligent agent will automatically judge the task type based on the real-time data and transfer the data related to the operating state of the gas turbine to the data processing module configured with the intelligent AI model, so as to deduce the data that is difficult to measure.

[0069] Please refer to Figure 3 , the schematic diagram of gas turbine monitoring in the embodiment of the present invention. Specifically, in practical applications, the monitoring and control of the operating state of the gas turbine rely on these key parameters such as compression ratio, gas mixture ratio, and turbine inlet temperature. However, these key parameters are usually difficult to obtain directly due to the limitations of high temperature, high pressure, or complex working conditions. For example, due to the high-temperature and high-pressure environment, traditional sensors are difficult to work stably for a long time at the turbine inlet temperature; the calculation of the gas mixture ratio requires accurate monitoring of the air and fuel ratio inside the combustion chamber, but direct measurement is usually affected by the combustion chamber structure and dynamic changes, resulting in unstable or difficult-to-obtain data. Therefore, the reliability and accuracy of the gas turbine operation monitoring system are insufficient, which in turn affects the dynamic optimization ability of the gas turbine under complex working conditions.

[0070] Specifically, by combining the easily measurable operating parameters of the gas turbine (such as compressor outlet pressure, air flow rate, fuel injection amount, and turbine exhaust temperature, etc.), the key parameters that are difficult to measure directly are inversely deduced based on the intelligent AI model, thus effectively solving the above problems and further realizing the comprehensive characterization and efficient control of the operating state of the gas turbine. For example, during high-load operation, if the traditional method directly relies on measuring the turbine inlet temperature, data inaccuracy or even sensor damage may occur due to the limit tolerance of the sensor. However, by using the easily measurable parameters such as turbine exhaust temperature and fuel injection amount through the method of this embodiment, the turbine inlet temperature is deduced through the intelligent AI model, which not only avoids the risk of sensor damage but also improves the stability and accuracy of the data.

[0071] The specific steps of the preset mechanism model construction process are as follows:

[0072] Extract the characteristic parameters of the physical process of the gas turbine operation;

[0073] Specifically, the physical process of gas turbine operation involves air compression by the compressor, thermal conversion in the combustion chamber, and mechanical energy output by the turbine. According to the physical process, the specific corresponding characteristic parameters can be known. The compressor characteristic parameters include outlet pressure, air flow rate, and outlet temperature; the combustion chamber characteristic parameters include fuel injection amount, gas mixture ratio, and combustion product concentration (such as CO2, NOx, etc.); the turbine characteristic parameters include inlet temperature, exhaust temperature, and power output.

[0074] Specifically, these characteristic parameters directly reflect the operating states of various components and have a significant impact on the overall performance of the gas turbine. By comprehensively extracting these parameters, sufficient basic data can be provided for subsequent modeling and analysis, ensuring the reliability and coverage of the model.

[0075] Calculate the thermodynamic characteristics, statistical characteristics, and time-frequency characteristics of the characteristic parameters, where the characteristic parameters include compressor characteristic parameters, combustor characteristic parameters, and turbine characteristic parameters;

[0076] Specifically, the compressor outlet pressure and air flow rate are used to calculate the compression ratio and specific work. The fuel injection amount and combustion product concentration in the combustor are combined with thermodynamic equations to calculate the combustion efficiency. The turbine exhaust temperature is combined with the enthalpy drop model to deduce the energy conversion efficiency. In this way, the physical operating state of the gas turbine is converted into characteristic variables with practical physical meanings, while enhancing the adaptability of the model to different operating conditions.

[0077] Process the thermodynamic characteristics, statistical characteristics, and time-frequency characteristics through correlation analysis to extract core variables;

[0078] Specifically, by calculating the correlation coefficient matrix between characteristic variables, high-correlation and low-correlation variables are identified. For example, analyze the correlation between the fuel injection amount and combustion efficiency, the correlation between the compressor outlet pressure and air flow rate, and the correlation between the turbine exhaust temperature and energy conversion efficiency. Among the high-correlation variables, further extract the most representative core variables, such as the compression ratio, gas mixture ratio, and turbine inlet temperature, through the principal component analysis method. Reduce the data dimension, improve the calculation efficiency of the model, and at the same time ensure the accurate description of the physical state by the model.

[0079] Construct a mechanism model that describes the operating states of various components inside the gas turbine according to the physical mechanism and the core variables.

[0080] Specifically, for the compressor part, a flow compression model is established through the energy equation and aerodynamic characteristic curves; for the combustor part, a combustion kinetics model is constructed based on the correlation between the fuel-air mixture ratio and the thermal conversion efficiency; for the turbine part, an output power model is constructed by combining the enthalpy drop equation and mechanical efficiency. The mechanism model is optimized by polynomial fitting and piecewise linear regression methods to enable it to maintain a high accuracy under multiple operating conditions. The purpose of constructing the mechanism model is to provide physical constraints for the intelligent AI model and make up for the overfitting problem that may occur in the AI model in complex industrial scenarios. For example, under the condition of switching between partial load and full load, the output of the mechanism model can be used as prior knowledge for the AI model to guide the parameter optimization and prediction adjustment of the AI model.

[0081] The specific steps of S1 are as follows:

[0082] S1.1: Clean the operation data of the gas turbine collected by the sensors;

[0083] Specifically, to ensure the reliability and accuracy of the data, the collected data is first subjected to a cleaning operation. The specific steps of cleaning include: removing outliers, such as instantaneous noise data caused by sensor failures or network jitters; interpolating missing values, using a time series interpolation algorithm to interpolate the missing data according to the trends of the previous and subsequent time points; normalizing the collected data of different dimensions, standardizing the data into a dimensionless form to meet the input requirements of subsequent models. Eliminate the influence of data noise and outliers on model training and analysis results.

[0084] S1.2: Obtain the historical operation data of the gas turbine, and divide the historical operation data into operation states under different load conditions through cluster analysis;

[0085] In this embodiment, the recorded data of the long-term operation of the gas turbine is extracted from the historical database, and these data cover the multi-dimensional operation parameters of the gas turbine under different load conditions. To effectively distinguish different operating conditions, a cluster analysis method is used to classify the historical data. Specifically, the K-means clustering algorithm is used to divide the historical operation data into multiple typical load states, including low load, medium load, and high load states, according to the characteristic vectors of the data (such as compressor outlet pressure, gas mixture ratio, and exhaust temperature). Through cluster analysis, the operation rules under each load state can be accurately identified. The reason for doing this is that the operation characteristics and the variation rules of key parameters of the gas turbine under different load conditions are significantly different, and directly analyzing the unclassified data will lead to distorted results. Through cluster analysis, an independent reference standard can be established for each operating condition, improving the pertinence and accuracy of the subsequent calculation of ideal operation parameters.

[0086] S1.3: Calculate the optimal operation data for the operation state corresponding to each load condition according to the mechanism model;

[0087] In this embodiment, an independent mechanism model is established for each load state. Combining the principles of thermodynamics and the energy conservation of gas turbines, the optimal operating data under each load state is calculated. For example, in the low load state, the mechanism model calculates the optimal values of air flow rate, fuel injection quantity, and turbine exhaust gas temperature through the flow compression ratio of the compressor, the gas mixing efficiency of the combustor, and the enthalpy drop relationship of the turbine. In the medium and high load states, the rated power output of the gas turbine is considered, and the corresponding optimal parameters are calculated in combination with the dynamic efficiency curves of the compressor and the turbine. Through this load condition classification and optimal operating parameter calculation, a clear reference target can be established for each load state, thereby providing an accurate ideal operating data set for deviation analysis under the dynamic operating conditions of the gas turbine. By adapting to the operating characteristics of different load conditions of the gas turbine, the applicability and accuracy of subsequent deviation analysis are improved.

[0088] S1.4: According to the thermodynamics and energy balance equations, calculate the ideal operating parameters corresponding to the optimal operating data under each condition, where the ideal operating parameters include compression ratio, gas mixing ratio, and turbine inlet temperature.

[0089] In this embodiment, through the principles of thermodynamics and energy balance, the optimal operating data under each condition is further converted into ideal operating parameters. Specifically, first, calculate the compression ratio according to the ratio of the compressor outlet pressure to the inlet pressure, which directly reflects the air compression efficiency of the compressor; then, in combination with the air flow rate and fuel injection quantity, calculate the gas mixing ratio through the gas mixing model of the combustor, which reflects the combustion efficiency and fuel utilization rate of the combustor; finally, based on the gas mixing ratio and the combustor outlet pressure, combined with the enthalpy drop formula and the combustion heat loss model, estimate the inlet temperature of the turbine. The calculation of these ideal operating parameters is based on the physical operating mechanism model of the gas turbine, and its advantage is that it can provide accurate reference values for deviation analysis of on-site operating data. In scenarios where it is difficult to directly measure key parameters such as high temperature and high pressure, the ideal values of key parameters can be deduced through easily measurable data, thus solving the technical problem of relying on direct measurement in traditional methods and laying a foundation for the intelligent optimization control of gas turbines.

[0090] Please refer to Figure 4 , the intelligent AI model structure diagram of the embodiment of the present invention. The intelligent AI model includes a first AI model, a second AI model, and a third AI model. Each model respectively models and predicts the key characteristic parameters (compression ratio, gas mixing ratio, and turbine inlet temperature) of the gas turbine operation, and each undertakes different functions.

[0091] It should be noted that in this embodiment, since three parameters that cannot be directly measured are calculated, the intelligent AI model is used to process the known and easily measurable parameters, so as to inversely deduce the parameters that cannot be directly measured. It can be understood that in the specific model processing process, the specific values of the unknown quantities are inversely deduced based on the known quantities. In the application of neural networks, learning the non-linear mapping relationship between the input data and the output target, and deriving the unknown data (difficult-to-measure operating parameters) based on the current known data (easily measurable operating parameters) is a prediction process. Specifically, in a broad sense, prediction not only includes inferences about the future, but also includes the derivation of information that cannot be directly observed currently. Therefore, the neural network model in this embodiment can also be replaced by a common prediction task model.

[0092] In this embodiment, the first AI model models the probability distributions of the compressor inlet and outlet pressures through a Gaussian mixture model (GMM) for inversely deducing the compression ratio during the operation of the gas turbine. The GMM utilizes the pressure change laws in the historical data of the gas turbine to construct a statistical distribution model of the compressor operation, and combines the real-time operation data for dynamic adjustment to output an accurate inversely deduced value of the compression ratio.

[0093] Specifically, the compression ratio is a key parameter of the compressor operation state, directly affecting the overall efficiency and combustion performance of the gas turbine. However, since the compression ratio is the ratio of the inlet and outlet pressures, its change is greatly affected by the dynamic fluctuations of the operating conditions. Traditional methods only rely on static formulas to calculate the compression ratio and cannot adapt to the complex changes under dynamic conditions. Through GMM modeling, the statistical laws of historical data and real-time data can be fully utilized to construct a compression ratio calculation method that not only has global generalization ability but also can dynamically adapt to changes in operating conditions. First, a superposition model of multiple Gaussian distributions is established using the historical operation data of the gas turbine (such as the distribution of the compressor inlet and outlet pressures under different load conditions), and each Gaussian distribution represents the pressure characteristics of an operating mode. In the modeling process, the GMM adopts the expectation-maximization (EM) algorithm to iteratively optimize the mean, variance, and weight parameters of each Gaussian distribution to make it highly match the distribution characteristics of the historical data. When the gas turbine is actually operating, the real-time collected inlet pressure and outlet pressure data will be used as inputs to dynamically adjust the parameters of the GMM through maximum likelihood estimation, so as to accurately inversely deduce the compression ratio value under the current operating conditions.

[0094] Furthermore, statistical means are used to capture the non-linear characteristics and complex correlation relationships in the compressor operation, and reliable inversely deduced results can be obtained even in the case of large data noise. This provides basic support for the optimization of the gas turbine operation efficiency and the dynamic regulation of the combustion chamber air supply conditions.

[0095] In this embodiment, the second AI model uses BiLSTM (Bidirectional Long Short-Term Memory Network) to perform time series modeling on the fuel-air ratio, capturing the non-linear characteristics and time-dependent relationships in the combustion chamber operation data. The fuel-air ratio is affected by various factors such as fuel injection quantity, air flow rate, and combustion chamber temperature, and its dynamic changes have strong temporal characteristics. The BiLSTM model can consider the influence of both historical and future time points simultaneously, and accurately model the dynamic change trend of the fuel-air ratio through a bidirectional learning structure.

[0096] Specifically, the fuel-air ratio is an important influencing factor for the combustion chamber efficiency, and its accurate prediction is crucial for combustion optimization and fuel utilization. However, traditional methods often ignore the long-term dependence relationships in the time series when dealing with the prediction of the fuel-air ratio, resulting in insufficient modeling ability for the dynamic characteristics of the combustion chamber. By introducing the BiLSTM model, it is possible to capture the potential non-linear and time-dependent laws during the operation of the combustion chamber, improving the accuracy and robustness of the fuel-air ratio prediction. During modeling, BiLSTM first normalizes the real-time operation data of the combustion chamber, mapping multi-dimensional input data (such as the time series of fuel injection quantity and air flow rate) to a unified feature space. Then, BiLSTM uses its bidirectional learning structure, where the forward network is responsible for capturing the influence of historical time points on the fuel-air ratio, and the backward network is responsible for analyzing the possible change trends of future time points. Through this bidirectional modeling method, BiLSTM can effectively capture the dynamic characteristics of the combustion chamber operation and generate accurate prediction results of the fuel-air ratio.

[0097] Furthermore, in practical applications, by real-time inputting easily measurable parameters such as fuel injection quantity and air flow rate, the model can output the predicted value of the fuel-air ratio under the current operating state. This modeling method not only improves the dynamic control ability of the combustion efficiency but also provides guarantee for the stable operation of the gas turbine under high load conditions.

[0098] Please refer to Figure 5 , the structural diagram of the third AI model in the embodiment of the present invention. The design of the third AI model in this embodiment aims to optimize the prediction accuracy of the turbine. Through a multi-level deep learning algorithm, combined with thermal characteristics and temperature change trends, it realizes efficient prediction and control of temperature. The input layer normalizes the real-time data from different sources to provide consistent input for the model. The modeling layer combines the ensemble learning algorithm and the graph neural network, improving the accuracy and stability of temperature prediction through multi-model integration and information sharing between nodes. The prediction layer, through the combination of temporal embedding and the Transformer network, can make full use of historical data and graph structure features, deeply explore temporal laws, and improve the accuracy and generalization ability of temperature prediction.

[0099] The third AI model specifically includes:

[0100] An input layer for receiving gas turbine operation data and normalizing the operation data, where the operation data further includes historical operation data of the gas turbine;

[0101] Normalization is to eliminate the dimensional differences between different data sources, enabling all data to be processed within the same scale, thereby avoiding improper impacts of certain parameters on the model during the training process. For example, current loads usually vary within a large range, while temperature and external environment data usually have a smaller fluctuation range. Normalization can prevent this difference from causing the model to be biased towards certain input features.

[0102] A modeling layer for initially predicting the turbine inlet temperature through an ensemble learning algorithm, calculating the initial prediction result, and modeling the normalized operation data based on a graph neural network to construct a graph structure with gas turbine operation components and characteristic parameters as nodes;

[0103] Propagate according to the initial prediction result in the graph structure, adjust the inconsistency between the mechanism model and the operation data, and output a feature representation;

[0104] Furthermore, the modeling layer first uses an ensemble learning algorithm to initially predict the temperature. The ensemble learning algorithm can effectively reduce the impact of the prediction error of a single model by weighted averaging the prediction results of multiple different models, improving the accuracy and robustness of the prediction result. At the same time, the modeling layer also models the normalized data through a graph neural network (GNN). The graph neural network processes complex multi-dimensional data by establishing a graph structure and utilizing the spatial and temporal relationships between sensors, and shares data and prediction information between nodes through an information propagation mechanism.

[0105] Specifically, through the integration of multiple models, the prediction results of each sub-model are weighted and fused to ensure that the final prediction result is more reliable. For example, among multiple machine learning algorithms, some models may be more accurate in predicting the temperature change trend, while other models may be better at capturing the impact of gas changes. Ensemble learning can effectively integrate these advantages into the final result. Through the graph neural network, the spatial and temporal dependence relationships between sensor data can be captured, providing a more accurate prediction. For data with a large number of sensors and complex environmental variables in the power system, the deep connections therein can be effectively mined.

[0106] A prediction layer, which is used to fuse the historical operation data and feature representation of the gas turbine through temporal embedding to generate a temporal feature vector, process the temporal feature vector through a Transformer network, calculate the temperature prediction value, calculate the temporal dependence of the temperature through a self-attention mechanism, calculate the temperature change trend according to the temporal dependence, correct the temperature prediction value according to the temperature change trend, and output the operating turbine inlet temperature.

[0107] Specifically, the Transformer network is a model based on the self-attention mechanism that can process long-sequence data and is particularly suitable for temporal prediction tasks. In temperature prediction, the Transformer can automatically learn the temporal dependence according to the historical trend of temperature change, so as to calculate the future temperature change trend. Through the self-attention mechanism, the model can automatically assign different weights to each data point of the temperature at different time points, ensuring that important historical information can receive more attention in the temporal data.

[0108] Adopting the Transformer network can improve the modeling ability of the temperature change trend. Especially in complex temporal data, it can effectively capture the non-linear change law of the temperature. At the same time, the self-attention mechanism can enhance the interpretability of the model, making it clear which historical time points have a greater impact on the current prediction during the temperature prediction process.

[0109] Please refer to Figure 6 , the structure diagram of the modeling layer in the embodiment of the present invention. The modeling layer includes an ensemble learning sub-layer, a graph neural network sub-layer, and a fusion sub-layer;

[0110] The ensemble learning sub-layer is used to integrate multiple machine learning models to generate multiple preliminary learning sub-models. Each sub-model selects real-time data for training according to the mechanism model, outputs the predicted temperature, calculates the weighted weight according to the hit rate of each sub-model in historical predictions, and calculates the preliminary prediction result according to the weighted weight and the predicted temperature;

[0111] Specifically, the core idea of the ensemble learning sub-layer is to improve the accuracy and stability of prediction through multi-model integration. In this embodiment, the ensemble learning sub-layer combines multiple different types of machine learning models, including decision trees, support vector machines, and random forests, to generate multiple preliminary learning sub-models. These sub-models are respectively used to capture different characteristic parameters and non-linear relationships in the operation of the gas turbine.

[0112] Furthermore, each sub-model selects real-time data for training according to the mechanism model of the gas turbine. For example, the decision tree model gives priority to processing the relationship between the compressor outlet pressure and the turbine inlet temperature with clear segmentation characteristics; the support vector machine model is suitable for processing high-dimensional non-linear relationships, such as the dynamic impact of the gas mixing ratio change on the turbine inlet temperature; the random forest model comprehensively learns the overall parameters by processing diverse input data (such as fuel injection volume and air flow rate).

[0113] Furthermore, during the data training process, the sub-models will extract the characteristic laws under different load conditions by using the historical operation data set of the gas turbine. The mechanism model provides the basic thermodynamic and energy balance relationships (such as the relationship between the compressor efficiency and the temperature change), ensuring that the training process conforms to the physical operation logic of the gas turbine.

[0114] Furthermore, after each sub-model is trained, it outputs the predicted turbine inlet temperature value according to the input real-time data. To evaluate the prediction reliability of each sub-model, the ensemble learning sub-layer uses the historical hit rate index as the basis for weight allocation. Specifically, the higher the matching degree between the predicted value and the true value of the sub-model under the past operating conditions, the greater its weighted weight in the current prediction.

[0115] Furthermore, the ensemble learning sub-layer performs weighted fusion on the prediction results of each sub-model through a weight calculation formula to generate a preliminary prediction result. This weighted fusion strategy not only ensures the overall accuracy of the model output, but also can be supplemented and corrected by other models when the performance of a single model is unstable, improving the stability of the prediction.

[0116] The graph neural network sub-layer is used to construct a graph structure with the operating components and characteristic parameters of the gas turbine as nodes according to the operating data, physical mechanism and the correlation relationship between components of the gas turbine. Among them, each graph node represents a component or a characteristic parameter, and each edge represents the physical dependence relationship between the nodes;

[0117] Specifically, the core components of the gas turbine include a compressor, a combustion chamber and a turbine. These components form a complex physical coupling relationship through the transfer of air flow, heat energy and mechanical energy. In this embodiment, the nodes not only represent the core components of the gas turbine, but also include its key characteristic parameters, such as the outlet pressure of the compressor, the fuel injection volume of the combustion chamber and the exhaust temperature of the turbine. The construction of this graph structure can comprehensively characterize the operating state of the gas turbine.

[0118] Furthermore, in order to capture the dynamic changes in the operating state of the gas turbine, the edges in the graph structure are designed to reflect the physical dependencies between nodes, such as the air flow coupling between the compressor outlet pressure and the combustor inlet conditions, the thermodynamic correlation between the combustor outlet temperature and the turbine inlet conditions, etc. The weights of these edges are jointly determined by the mechanism model of the gas turbine and the real-time collected data, and are dynamically adjusted through the mapping function of the physical model. For example, the edge weight of the air flow relationship can be calculated from the flow equation and the compression ratio, while the edge weight of the heat energy transfer is derived from the energy balance equation. This design of dynamic weights ensures that the graph structure can adapt to the operating changes of the gas turbine in real time.

[0119] Propagate the preliminary prediction results in the graph structure through the information propagation mechanism. Each node in the graph not only obtains its own input features, but also adjusts its output features based on the information of adjacent nodes. Each node adjusts the preliminary prediction results according to the environmental context it is in during the propagation process, and locally adjusts the graph structure through the feature information between nodes;

[0120] Specifically, the initial value of the feature vector of each node comes from the real-time collected data, such as the sensor measurement value of the compressor outlet pressure. Then, calculate the updated value of the node by combining the feature information of adjacent nodes and the edge weights. For example, during the propagation process, the outlet pressure feature of the compressor node will be combined with the inlet conditions of the combustor node, and the feature is updated through weighted summation and a non-linear activation function (such as the ReLU function). This propagation mechanism can effectively capture the coupling relationship between the components of the gas turbine, making the prediction results more in line with the actual operating state.

[0121] Furthermore, in order to improve the adaptability of the model, in this embodiment, the feature representation between nodes is optimized through the local adjustment mechanism. When there is a deviation between the preliminary prediction result and the physical mechanism or actual data, the local adjustment mechanism will make corrections according to the context information of specific nodes. For example, when the preliminary predicted value of the turbine inlet temperature deviates from the actual value, the adjustment mechanism will give priority to the nodes related to the turbine inlet conditions (such as the combustor outlet temperature and the compressor outlet pressure), and reduce the prediction error through re-propagation and weighted optimization. This process can optimize the prediction accuracy of the overall model through local constraints.

[0122] Specifically, each node receives information through adjacent nodes. The weight of each edge determines the intensity and influence of information propagation. In the context of sensor data, the message passing between nodes is to transmit data with weights through the edges. For example, a temperature sensor may update its temperature value based on the data from a current load sensor and an external environment sensor. This process is achieved by aggregating the information of neighbor nodes, that is, the node calculates a new state according to the data of neighbor nodes. This mechanism can capture the complex dependencies between different sensor data, especially the temporal and spatial correlations of temperature and load changes. This update method can effectively capture the changing trend of temperature and quickly respond to sudden abnormal situations. In addition, the information propagation between nodes will also be dynamically adjusted according to environmental changes. For example, some sensors may be more sensitive under specific environmental conditions, while other sensors need to be adjusted under different load conditions. After each information propagation, the graph structure itself will also be adjusted to optimize the transmission of information flow. For example, some nodes may be more dependent on the information of other nodes under specific environmental conditions. At this time, the edge weights between nodes in the graph structure will be adjusted to reflect this new dependency.

[0123] Furthermore, the initial prediction results come from the ensemble learning sub-layer, which provides the initial output data for the graph neural network. By propagating these initial prediction results in the graph structure, the dependency relationships between nodes in the graph can be utilized to optimize the prediction results.

[0124] The fusion sub-layer is used to propagate the initial prediction results and the adjustments output by the graph structure, update the internal parameters according to the adjustment results, and calculate the feature representation.

[0125] Internal parameters refer to the adjustable weights and bias terms of each computing unit in the fusion sub-layer. These parameters are continuously adjusted through optimization methods such as gradient descent during the training process to improve the prediction accuracy of the model.

[0126] Specifically, in each round of training, the gradient of the loss function with respect to each internal parameter is calculated through the backpropagation algorithm, and the weights and biases are adjusted using the optimization algorithm. During the backpropagation process, the gradient of the loss function will be gradually transmitted to the fusion sub-layer, and each weight is updated by calculating the gradient, so that the fused result is more accurate and further optimizes the performance of the model.

[0127] By optimizing and updating the internal parameters, the fusion sub-layer can gradually adjust its weights in each training process, so that the final fused feature representation is more accurate and can better serve downstream tasks. This collaborative optimization mechanism effectively improves the performance of the model and avoids the limitations of simple weighting.

[0128] The calculation formula of the feature representation is as follows:

[0129]

[0130] Among them, H i (t + 1) represents the updated feature representation of node i, t represents the t-th propagation, σ represents the activation function, N(i) represents the set of neighboring nodes of node i, j represents a single node in the set of neighboring nodes, A ij represents the adjacency matrix between node i and node j, D i represents the degree matrix of node i, D j represents the degree matrix of node j, represents matrix multiplication, H j (t) represents the feature representation of node j in the t-th propagation, y o represents the adjusted preliminary prediction result, W represents the internal parameter matrix, H i (t) represents the feature representation of node i in the t-th propagation.

[0131] The specific steps of S3 are as follows:

[0132] S3.1: By quantitatively analyzing the deviation between the on-site operation dataset and the ideal operation dataset, calculate the deviation values of the gas turbine operation parameters;

[0133] Specifically, by comparing the on-site operation dataset with the ideal operation dataset, calculate the deviation values of the key parameters of the gas turbine item by item. For example, the compression ratio deviation, the gas mixture ratio deviation, and the turbine inlet temperature deviation.

[0134] Furthermore, in the process of calculating the deviation values, statistical analysis methods are used to process the real-time data, including denoising, normalization, and time alignment, to ensure the accuracy and consistency of the deviation analysis. By dynamically monitoring and statistically analyzing the deviation values, capture the change trends of the key parameters of the gas turbine under different operating states, and convert the deviation situation of the on-site operation data into specific inputs that can be used for optimization and control, thereby providing a basis for optimizing the control parameters.

[0135] S3.2: According to the deviation values, output control parameters through a reinforcement learning algorithm. Among them, the reinforcement learning algorithm applies the Q-learning algorithm, and optimizes the turbine control parameters through an interactive learning process. The Q-learning algorithm includes states, actions, and rewards;

[0136] The state represents the gas turbine operation data and deviation values, including the on-site operation dataset, the ideal operation dataset, and the deviation values;

[0137] The action represents the decision made by the Q-learning algorithm in the current state, and this decision is used to adjust the control parameters of the gas turbine;

[0138] The reward measures the operating performance after the adjustment of the gas turbine control parameters. The reward is updated according to the adjusted deviation value and operating state, and the reward includes error reward, stability reward, and efficiency reward;

[0139] In this embodiment, the application of the Q-learning algorithm is based on the deviation value and operating state of the gas turbine, and three core elements of state, action, and reward are designed. The state consists of the operating data and deviation value of the gas turbine, including the on-site operating data set (such as the compressor outlet pressure, fuel injection quantity, etc.), the ideal operating data set (such as the reference value calculated by the mechanism model), and the above deviation value. These states characterize the operating state and performance deviation of the gas turbine at the current moment, providing input for optimal control.

[0140] Furthermore, the action represents the adjustment decision of the gas turbine control parameters in the current state, including adjusting the compressor inlet pressure set value, fuel injection quantity ratio, turbine load distribution, etc. Through reinforcement learning, the Q-learning algorithm selects the optimal action in different states to gradually reduce the deviation value.

[0141] Furthermore, the reward is a quantitative evaluation of the control effect, and three dimensions of error reward, stability reward, and efficiency reward are designed. The error reward is used to quantify the change in the deviation value after the adjustment of the control parameters. The smaller the deviation, the higher the reward; the stability reward is used to evaluate the fluctuation degree of the operating data after the adjustment. The more stable the data, the higher the reward; the efficiency reward is used to measure the improvement effect of the gas turbine operating efficiency. The higher the efficiency, the higher the reward. These rewards ensure that the algorithm optimization not only focuses on minimizing the deviation but also considers the stability and economy of the operation.

[0142] Furthermore, the Q-learning algorithm gradually optimizes the control parameters through an iterative learning strategy, enabling the gas turbine to respond to deviations in real time and adjust the operating state under dynamic working conditions. Reinforcement learning can explore and learn the optimal control strategy under complex working conditions without predefined fixed control rules, enhancing the adaptive ability of the system.

[0143] S3.3: Calculate the reward function according to the reward, and use the reward function to guide the Q-learning algorithm to update the learning strategy after each adjustment, and output the gas turbine control parameters.

[0144] In this embodiment, the design of the reward function comprehensively considers the magnitude of the deviation value, the stability of the operating state, and the improvement of overall efficiency. Specifically, the reward function combines these dimensions into a unified metric by calculating the error reward, stability reward, and efficiency reward in real time, which is used to measure the effect of the current control action. After each adjustment of the control parameters by the system, the error reward is calculated by comparing the change in the deviation value before and after the adjustment.

[0145] Furthermore, to ensure the convergence and real-time performance of the optimization process, a mechanism for dynamically adjusting the learning rate is adopted, enabling the Q-learning algorithm to quickly converge to the optimal control strategy under different operating conditions. At the same time, the result of the reward function is input into the update process of the Q-learning algorithm through backpropagation to guide the algorithm to select a better control strategy.

[0146] Furthermore, the output gas turbine control parameters include the compressor inlet pressure, fuel injection quantity ratio, and turbine load distribution, etc. The dynamic adjustment of these parameters during the operation of the gas turbine can effectively reduce the deviation value and improve the operating efficiency. For example, under high-load conditions, the system realizes the precise control of the turbine inlet temperature by increasing the fuel injection quantity and optimizing the compressor inlet pressure, avoiding problems such as overheating or a decrease in operating efficiency.

[0147] Please refer to Figure 7 , the flowchart of a method for updating the parameters of an intelligent AI model according to an embodiment of the present invention updates the parameters of the intelligent AI model according to the deviation value of the gas turbine operating parameters. The specific steps of the method are as follows:

[0148] S401: Set the first optimization goal, where the first optimization goal is to maximize the parameter inverse inference accuracy of the intelligent AI model according to the deviation value;

[0149] In this embodiment, the set first optimization goal is to maximize the inverse inference accuracy of the intelligent AI model according to the deviation value of the gas turbine operating parameters. To achieve this goal, first, the source of the deviation value is analyzed, specifically including the comparison deviation between the on-site operation data and the ideal operation data (such as the deviation of the compression ratio, gas mixing ratio, and turbine inlet temperature), and the influence weight of the deviation on the overall performance of the gas turbine (such as thermal efficiency, output power). These deviation values directly reflect the adaptability of the model to the actual working conditions. Therefore, the optimization goal is formulated by minimizing the deviation value and improving the prediction accuracy of key parameters. Through this setting, it is ensured that the intelligent AI model can not only dynamically adapt to different working conditions but also further optimize the operating efficiency of the gas turbine.

[0150] S402: Define the space of the parameters in the intelligent AI model to obtain the space parameter set;

[0151] Among them, the parameters include the learning rate, weight parameters, and time series window of each sub-model in the model;

[0152] In this embodiment, the learning rate controls the update speed of the model parameters, the weight parameters determine the influence ratio of different data features, and the time series window is used to capture the time correlation of the operation data. In actual operation, by restricting the ranges of these parameters, a multi-dimensional parameter space is constructed to ensure the effectiveness and stability of the search process. In particular, the range of the time series window needs to be dynamically adjusted according to the time span of the historical data to ensure that the model can capture the change trend in the critical time period.

[0153] Specifically, by constructing the parameter space, the optimization process of the model is transformed into a parameter search problem, avoiding the problem of relying on empirical settings for parameter adjustment in traditional methods. This makes the parameter adjustment more flexible and scientific, and at the same time enhances the generalization ability of the model.

[0154] S403: Calculate the optimization objective function through the particle swarm optimization algorithm;

[0155] The optimization objective function is constructed based on the deviation value of the gas turbine operation data, the deviation between the optimized operation parameters and the ideal operation parameters, and the improvement effect of the overall performance of the gas turbine;

[0156] In this embodiment, each particle represents a parameter combination of an intelligent AI model, and its position and velocity change in the multi-dimensional parameter space with the goal of minimizing the deviation value and improving the overall performance. The position of the particle is determined by the objective function jointly constructed by the deviation value of the gas turbine operation data (such as the deviation of the compression ratio and the turbine inlet temperature), the deviation between the optimized operation parameters and the ideal operation parameters, and the overall performance of the gas turbine (such as the improvement of the thermal efficiency). The particle swarm uses the information of the historical best position and the swarm best position during the search process to dynamically adjust the search path, thus ensuring the efficiency and accuracy of the search process.

[0157] Specifically, through the particle swarm optimization algorithm, the defect that the traditional gradient optimization method is prone to fall into local optimum is effectively avoided, and at the same time, the convergence speed of parameter optimization is accelerated, ensuring that the optimization result has global optimality in the complex parameter space.

[0158] S404: Calculate the difference between the real-time operation data of the gas turbine, the optimized control parameters, and the output of the target deviation;

[0159] In this embodiment, by introducing the dynamic difference analysis method, the main sources of deviation are decomposed. For example, the compression ratio may be affected by the inlet air flow rate, and the gas mixing ratio may be affected by the change in the fuel injection amount, so as to distribute the overall deviation to each key component. Through such difference calculation, a clear direction can be provided for subsequent parameter optimization.

[0160] S405: Iteratively update the set of spatial parameters according to the difference and the optimized objective function to generate the optimal parameter solution of the intelligent AI model.

[0161] In this embodiment, according to the calculation results of the difference value and the optimized objective function, the parameter space is iteratively updated through a particle swarm optimization algorithm, and finally the optimal parameter solution of the intelligent AI model is generated. The update process combines the optimal deviation results and the historical optimal solutions in each iteration to dynamically adjust the learning rate, weights, and time series window of the model. In each iteration, by updating the model parameters in real time, the model can more accurately deduce the key difficult-to-measure parameters (such as turbine inlet temperature), and at the same time further reduce the deviation value.

[0162] Embodiment 2

[0163] Please refer to Figure 8 , the present invention provides an embodiment: an Agent-driven intelligent AI deep prediction system for a power system, the system includes a data acquisition module, a data analysis module, and an optimization control module;

[0164] The data acquisition module is used to collect the operating data of the gas turbine and the historical data set of the gas turbine operation;

[0165] The data analysis module is used to quantitatively analyze the gas turbine data, calculate the ideal operating data set, input the gas turbine data and a preset mechanism model into a preset intelligent AI model, and output the on-site operating data set through the intelligent AI model;

[0166] The optimization control module is used to perform pattern recognition and anomaly detection on the gas turbine according to the DCS time series data mining technology, and calculate the control parameters according to the deviation analysis between the on-site operating data set and the ideal operating data set.

[0167] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. An Agent-driven intelligent AI deep prediction method for power systems, applied to gas turbines, characterized in that, The method includes: Collecting gas turbine data through sensors, quantitatively analyzing the gas turbine data, and calculating an ideal operation data set; Inputting the gas turbine data and a preset mechanism model into a preset intelligent AI model, and outputting an on-site operation data set through the intelligent AI model; Performing pattern recognition and anomaly detection on the gas turbine according to the DCS time series data mining technology, and calculating control parameters based on the deviation analysis between the on-site operation data set and the ideal operation data set.

2. The Agent-driven intelligent AI deep prediction method for power systems according to claim 1, wherein The method further includes: Extracting characteristic parameters of the physical process of the gas turbine operation, analyzing and obtaining influencing factors of the characteristic parameters of each component, and constructing a mechanism model, where the mechanism model includes the flow compression process of the compressor, the thermal conversion process of the combustor, and the mechanical energy output process of the turbine.

3. The Agent-driven intelligent AI deep prediction method for power systems according to claim 1, characterized in that Quantitatively analyzing the gas turbine data and calculating an ideal operation data set includes: Cleaning the gas turbine operation data collected through sensors; Obtaining the historical operation data of the gas turbine, and dividing the historical operation data into operation states under different load conditions through cluster analysis; Calculating the optimal operation data for the operation state corresponding to each load condition according to the mechanism model; Calculating ideal operation parameters corresponding to the optimal operation data for each working condition according to the thermodynamics and energy balance equations, where the ideal operation parameters include the compression ratio, the gas mixing ratio, and the turbine inlet temperature.

4. The Agent-driven intelligent AI deep prediction method for a power system according to claim 1, characterized in that, The intelligent AI model includes a first AI model, which is used to model the probability distribution of the compressor inlet and outlet pressures according to the Gaussian mixture model, and output the operation compression ratio by combining the distribution characteristics of the gas turbine operation data and the mechanism model.

5. The Agent-driven intelligent AI deep prediction method for power systems according to claim 1, wherein The intelligent AI model further includes a second AI model, which is used to perform time series modeling on the gas turbine operation data, capture the non-linear characteristics and time-dependent relationships in the combustion process of the operation data, calculate the change trend of the gas mixing ratio, and output the operation gas mixing ratio.

6. The Agent-driven intelligent AI deep prediction method for a power system according to claim 2, wherein The intelligent AI model further includes a third AI model, which is used to reverse-infer the inlet temperature of the turbine, including: An input layer, which is used to receive the gas turbine operation data and perform normalization processing on the operation data, where the operation data further includes the historical operation data of the gas turbine; A modeling layer, which is used to perform a preliminary prediction on the turbine inlet temperature through an ensemble learning algorithm, calculate a preliminary prediction result, and model the normalized operation data according to the graph neural network to construct a graph structure with the gas turbine operation components and characteristic parameters as nodes; Propagating according to the preliminary prediction result in the graph structure, adjusting the inconsistency between the mechanism model and the operation data, and outputting a feature representation; A prediction layer, which is used to fuse the historical operation data of the gas turbine and the feature representation through time series embedding to generate a time series feature vector, process the time series feature vector through a Transformer network, calculate a temperature prediction value, calculate the time series dependence of the temperature through a self-attention mechanism, correct the temperature prediction value according to the time series dependence, and output the operation turbine inlet temperature.

7. The Agent-driven intelligent AI depth prediction method for a power system according to claim 6, wherein The modeling layer includes an integrated learning sub-layer, a graph neural network sub-layer, and a fusion sub-layer; The integrated learning sub-layer is used to integrate multiple machine learning models to generate multiple preliminary learning sub-models. Each sub-model selects operating data for training according to the mechanism model, outputs the predicted temperature, calculates the weighted weights according to the hit rate of each sub-model in historical predictions, and calculates the preliminary prediction result according to the weighted weights and the predicted temperature; The graph neural network sub-layer is used to construct a graph structure with the operating components and characteristic parameters of the gas turbine as nodes according to the operating data, physical mechanism, and correlation relationship between components of the gas turbine. Among them, each graph node represents a component or characteristic parameter, and each edge represents the physical dependence relationship between nodes; The preliminary prediction result is propagated in the graph structure through the information propagation mechanism. Each node adjusts the preliminary prediction result according to the environmental context it is in during the propagation process, and locally adjusts the graph structure through the feature information between nodes; The fusion sub-layer is used to propagate the preliminary prediction result and the adjustment output by the graph structure, update the internal parameters according to the adjustment result, and calculate the feature representation.

8. The Agent-driven intelligent AI depth prediction method for a power system according to claim 1, wherein The calculation of the control parameters according to the deviation analysis between the on-site operation data set and the ideal operation data set includes: Quantitatively analyze the deviation between the on-site operation data set and the ideal operation data set to calculate the deviation value of the gas turbine operation parameters; According to the deviation value, output the control parameters through the reinforcement learning algorithm. Among them, the reinforcement learning algorithm applies the Q-learning algorithm, and optimizes the gas turbine control parameters through an interactive learning process. The Q-learning algorithm includes state, action, and reward; The state represents the gas turbine operation data and the deviation value, including the on-site operation data set, the ideal operation data set, and the deviation value; The action represents the decision made by the Q-learning algorithm in the current state, and the decision is used to adjust the control parameters of the gas turbine; The reward measures the operation performance after the adjustment of the gas turbine control parameters, and updates the reward according to the adjusted deviation value and operation state. The reward includes error reward, stability reward, and efficiency reward; Calculate the reward function according to the reward, and guide the Q-learning algorithm to update the learning strategy after each adjustment through the reward function, and output the gas turbine control parameters.

9. The Agent-driven intelligent AI deep prediction method for power systems according to claim 8, wherein The method further includes updating the parameters of the intelligent AI model according to the deviation value of the gas turbine operation parameters. The update includes: Set the first optimization goal, which is to maximize the parameter inversion accuracy of the intelligent AI model according to the deviation value; Define the space of the parameters in the intelligent AI model to obtain the space parameter set, where the parameters include the learning rate, weight parameters, and time series window of each sub-model in the model; Calculate the optimization objective function through the particle swarm optimization algorithm. The optimization objective function is constructed according to the deviation value of the gas turbine operation data, the deviation between the optimized operation parameters and the ideal operation parameters, and the improvement effect of the overall performance of the gas turbine; Calculate the difference between the real-time operation data of the gas turbine, the optimized control parameters and the output of the target deviation, and iteratively update the spatial parameter set according to the difference and the optimized objective function to generate the optimal parameter solution of the intelligent AI model.

10. An Agent-driven intelligent AI depth prediction system for a power system, implemented based on the Agent-driven intelligent AI depth prediction method according to any one of claims 1-9, characterized in that, The system includes a data acquisition module, a data analysis module and an optimization control module; The data acquisition module is used to collect the operation data of the gas turbine and the historical data set of the gas turbine operation; The data analysis module is used to quantitatively analyze the gas turbine data, calculate the ideal operation data set, input the gas turbine data and a preset mechanism model into a preset intelligent AI model, and output the on-site operation data set through the intelligent AI model; The optimization control module is used to perform pattern recognition and anomaly detection on the gas turbine according to the DCS time series data mining technology, and calculate the control parameters according to the deviation analysis between the on-site operation data set and the ideal operation data set.

Images