Method for online simulation of operating conditions for gas power plants

Abstract

The application discloses an operation condition online simulation method for a gas power station, and relates to the field of operation simulation, and comprises the following steps: step 1: obtaining operation index setting values of each operation equipment of the gas power station under different operation modes and corresponding real-time index feedback data through a preset acquisition component, and constructing an operation condition simulation simulation framework of the gas power station based on the operation index setting values and the index feedback data; step 2: continuously extracting state characteristics in a preset sampling period based on the index feedback data acquired by each acquisition component in real time, and generating a state characteristic sequence corresponding to each acquisition component; the method can sensitively capture the deviation trend of a certain parameter relative to the overall correlation mode of the system, can discover early fault signs of those whose amplitude does not exceed a threshold value but whose mode has appeared abnormal, and realizes intelligent distinction of equipment abnormality and sensor failure.

Description

Technical Field

[0001] This invention relates to the field of operation simulation technology, specifically to an online simulation method for the operating conditions of gas-fired power plants. Background Technology

[0002] Gas-fired power plants, as clean and efficient peak-shaving and distributed energy sources, play an increasingly important role in modern power systems. Their operation involves a complex thermodynamic cycle comprised of gas turbines, waste heat boilers, steam turbines, generators, and numerous auxiliary systems, with highly variable and interconnected operating conditions. With the widespread adoption of the Industrial Internet of Things (IIoT) and big data technologies, power plants are equipped with a massive number of sensors and data acquisition components, continuously generating multi-dimensional time-series data reflecting equipment status, such as temperature, pressure, flow rate, and vibration. This provides the possibility for in-depth data-driven insights into the power plant's operational status. Online simulation technology has become a key tool for optimizing operation and preventing risks. It aims to construct a virtual model synchronously mapped to the physical power plant, enabling real-time reproduction of actual operating conditions, future projections, and anomaly predictions.

[0003] Traditional simulation models are mostly built based on design parameters or historical steady-state data, primarily used for offline design and training. They struggle to integrate deeply with real-time, changing operational data, resulting in simulation results that fail to accurately reflect the current dynamic operating conditions of the power plant. Early warnings are delayed, and the ability to identify early, gradually changing, or correlated anomalies is insufficient. Furthermore, they are prone to generating numerous invalid alarms due to normal fluctuations in operating conditions, interfering with operational judgment. When data anomalies are detected, alarms are typically triggered, but it's difficult to quickly distinguish between actual equipment failures and sensor malfunctions. In the latter case, the system lacks an effective real-time data restoration and compensation mechanism, causing downstream simulations to run on erroneous data, rendering them meaningless and unable to dynamically demonstrate the impact of faults and post-recovery conditions, hindering rapid decision-making by operators. Summary of the Invention

[0004] (a) Technical problems to be solved In view of the above-mentioned shortcomings of the existing technology, the present invention provides an online simulation method for the operating conditions of gas-fired power plants, which can effectively solve the problems of the existing technology.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: This invention discloses an online simulation method for the operating conditions of a gas-fired power plant, comprising the following steps: Step 1: Obtain the operating index settings and corresponding real-time index feedback data of each operating device in different working modes of the gas power plant through the preset acquisition components, and construct the operating condition simulation framework of the gas power plant based on the operating index settings and index feedback data. Step 2: Based on the real-time indicator feedback data acquired by each acquisition component, continuously extract state features within a preset sampling period to generate a state feature sequence corresponding to each acquisition component; collect the state feature sequences of all acquisition components to establish a global state pool. Step 3: Train the state pool using deep learning algorithms to establish a data acquisition correlation analysis model to characterize the correlation and matching relationship between the state feature sequences of each acquisition component. The model identifies the abnormal fluctuation trend of each state feature sequence in the overall state pool. Step 4: Based on the identified abnormal fluctuation trend, determine whether to trigger a risk alarm according to the preset fluctuation threshold. If the alarm conditions are met, generate and output a risk alarm signal. Step 5: Based on the abnormal fluctuation trend, the acquisition attributes of each acquisition component, and the preset performance indicators of the acquisition components, calculate the fault risk coefficient of the corresponding acquisition component through the preset fault identification model; if the fault risk coefficient exceeds the set threshold, it is determined to be a fault of the acquisition component, and the abnormal data is restored to standard data based on the fault identification results in order to restore the correct indicator feedback data. Step 6: Simultaneously submit the risk alarm results generated in Step 4 and the standard data restored in Step 5 to the operating condition simulation framework for batch-by-batch operating condition simulation demonstrations, so as to show the power plant operating conditions after different risk states or fault recovery online.

[0006] Furthermore, the different operating modes in step 1 include the start-up mode, rated load operation mode, variable load regulation mode, and shutdown mode of the gas power plant; the operating equipment includes gas turbines, waste heat boilers, steam turbines, and generators; the operating index settings include temperature, pressure, flow rate, speed, and power settings, and the index feedback data are the actual measured values ​​of the corresponding operating indexes.

[0007] Furthermore, the process of constructing the working condition simulation framework in step 1 includes the following steps: Step 11: Through the acquisition component, synchronously acquire the set values ​​of operating indicators under different working modes and the corresponding historical and real-time indicator feedback data from the distributed control system and historical database of the gas power plant. The data covers at least one complete power plant start-up, shutdown and load regulation cycle. Step 12: Clean, align and normalize the obtained indicator feedback data, and divide the processed data into multiple independent datasets corresponding to startup, rated operation, variable load and shutdown modes according to the preset working mode discrimination rules. Step 13: Based on the partitioned dataset, establish a simulation sub-model for each working mode that combines theoretical expected values ​​with real-time calculated correction values. The simulation sub-model simulates and outputs the expected equipment state and system response based on the input index settings. Step 14: Integrate the simulation sub-models of each working mode, the real-time data access interface, and the working condition demonstration engine into a preset software platform to form the working condition simulation framework.

[0008] Furthermore, the extraction process of the state feature sequence in step 2 is as follows: For each acquisition component, the continuous index feedback data stream acquired within the preset sampling period is cleaned and aligned. The sliding time window technique is used to extract multiple subsequences of equal length from the data stream; For each subsequence, time-domain feature extraction and frequency-domain feature extraction are performed in parallel: Temporal feature extraction includes calculating the statistical features of the subsequence, which at least include the mean, standard deviation, skewness, kurtosis, and normalized peak factor. Frequency domain feature extraction includes performing a fast Fourier transform on the subsequence to obtain its spectrum, and extracting the main frequency component and its amplitude, spectral entropy, and energy proportion within a specific frequency band. The time-domain features and frequency-domain features extracted within each time window are combined into a feature vector and arranged in chronological order of the time windows to form a sequence of state features that continuously evolve within the preset period of the acquisition component.

[0009] Furthermore, the process of constructing the data collection correlation analysis model in step 3 includes: The state feature sequences of all acquisition components in the state pool are aligned and standardized according to a unified timestamp to construct a multi-dimensional time series input matrix; The deep learning algorithm adopts an encoder and decoder architecture with long short-term memory network units as the core. During the training phase, the model is trained unsupervised using state pool data under historical normal working conditions. The training objective is to learn the dynamic correlation pattern between state feature sequences through the encoder and reconstruct the input sequence through the decoder. The encoder is used to learn and compress the high-dimensional spatiotemporal features of the state feature sequences of each acquisition component, and the decoder is used to reconstruct or predict the normal state of each sequence based on the encoded features. After training, the hidden layer state and internal weight matrix of the model represent the complex nonlinear correlation and matching relationship between the state sequences of each acquisition component. Real-time state pool data is input into the trained model, and the model outputs the corresponding reconstructed sequence. By calculating the residual between the real-time sequence and the reconstructed sequence, and combining the distribution statistics of the residual in the state pool, a quantitative criterion for identifying abnormal fluctuation trends is generated.

[0010] Furthermore, the preset fluctuation threshold in step 4 adopts a dynamic threshold setting method, which adaptively calculates the alarm threshold of each state characteristic sequence based on the current working mode of the gas power plant and historical normal operation data of the same period; when the amplitude or duration of the abnormal fluctuation trend exceeds its corresponding dynamic threshold, a risk alarm is triggered.

[0011] Furthermore, the preset process for the fault identification model in step 5 is as follows: Collect sample data marked with different fault types and normal states of acquisition components during historical operation. The sample data includes abnormal fluctuation patterns of state characteristic sequences, attribute information of acquisition components, and corresponding performance index data. A neural network model with multiple inputs and a single output is constructed. Its input layer receives the abnormal fluctuation trend feature vector, the attribute encoding vector of the collected components, and the performance index vector, respectively. The core of the model includes feature extraction, feature fusion based on the attention mechanism, and fault risk regression. The model is trained using the prepared sample data, and the model parameters are adjusted by optimizing the algorithm to minimize the error between the fault risk coefficient output by the model and the true fault label of the sample. The diagnostic accuracy and generalization ability of the model are verified using an independent test dataset. After verification, the trained fault identification model is integrated into the preset local software.

[0012] Furthermore, in step 5, during the operation phase, the fault identification model analyzes the pattern of abnormal fluctuation trends to determine whether it is a sudden, drifting, or periodic anomaly. It combines the attribute information of the acquisition component, such as component type, installation location, measurement principle, and maintenance history; and calculates the fault risk coefficient by assigning attention weights to the preset acquisition component performance index library. The fault risk coefficient is used to quantify the probability of the acquisition component failing.

[0013] Furthermore, the standard data restoration process in step 5 is as follows: When the fault identification model determines that a specific acquisition component is faulty, based on the association matching relationship established by the acquisition association analysis model, multiple normal acquisition components that are associated with the faulty acquisition component more than a preset association degree threshold are selected from the state pool. Based on the real-time status feature sequences of multiple normal acquisition components, a pre-trained regression completion network is used to estimate the standard indicator data that the faulty acquisition component should have output during the fault period, which conforms to the overall operation logic of the power plant, so as to replace or correct the abnormal data caused by the fault and complete the data restoration.

[0014] Furthermore, in the batch operation condition demonstration process in step 6, the operation condition simulation framework automatically matches the corresponding simulation scenario library based on the input risk alarm level and the restored data; each batch of demonstrations includes visualization of operation condition parameters, highlighting of anomaly location, deduction of fault impact, and comparison of restored data with original data to support operators in making online decisions.

[0015] (III) Beneficial Effects Compared with the known prior art, the technical solution provided by this invention has the following beneficial effects: 1. By constructing a global state pool and applying a data acquisition correlation analysis model to mine the complex spatiotemporal correlation matching relationships between the state feature sequences, a system-level normal operation mode benchmark was established. This enables anomaly detection to no longer be limited to the exceeding of limits of a single parameter, but to keenly capture the deviation trend of a parameter relative to the overall correlation mode of the system. It can detect early fault signs that have not exceeded the threshold but have already shown abnormality in the mode earlier, improve the foresight and accuracy of risk warning, and reduce the false alarm rate caused by isolated judgment.

[0016] 2. By comprehensively considering abnormal patterns, component attributes, and performance indicators, the fault identification model is used to quantitatively calculate the fault risk coefficient, realizing the intelligent distinction between equipment abnormalities and sensor faults. Once the fault is determined to be a data acquisition component, the system can immediately estimate and restore the standard data that the faulty component should have in real time based on the data of other strongly correlated normal components in the correlation analysis model through regression completion network. This effectively ensures the reliability of the data input to subsequent simulation stages and lays a data foundation for high-quality simulation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0018] Fig. 1 This is a schematic diagram of the overall process of the present invention; Fig. 2 This is a flowchart illustrating the construction process of the working condition simulation framework in this invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] The present invention will be further described below with reference to embodiments.

[0021] The online simulation method for the operating conditions of a gas-fired power plant in this embodiment, such as... Figs. 1-2 As shown, it includes the following steps: Step 1: Acquire the operating index settings and corresponding real-time index feedback data of each operating device in different working modes of the gas-fired power plant through preset acquisition components. Construct a working condition simulation framework for the gas-fired power plant based on the operating index settings and index feedback data. Different working modes include the gas-fired power plant's start-up mode, rated load operation mode, variable load regulation mode, and shutdown mode. The operating devices include gas turbines, waste heat boilers, steam turbines, and generators. The operating index settings include temperature, pressure, flow rate, speed, and power setpoints. The index feedback data are the actual measured values ​​of the corresponding operating indicators.

[0022] Step 2: Based on the real-time indicator feedback data acquired by each acquisition component, continuously extract state features within a preset sampling period to generate a state feature sequence corresponding to each acquisition component; collect the state feature sequences of all acquisition components to establish a global state pool; the process of extracting the state feature sequence is as follows: For each acquisition component, the continuous index feedback data stream acquired within the preset sampling period is cleaned and aligned. The sliding time window technique is used to extract multiple subsequences of equal length from the data stream; For each subsequence, time-domain feature extraction and frequency-domain feature extraction are performed in parallel: Temporal feature extraction includes calculating the statistical features of the subsequence, which at least include the mean, standard deviation, skewness, kurtosis, and normalized peak factor. Frequency domain feature extraction includes performing a fast Fourier transform on the subsequence to obtain its spectrum, and extracting the main frequency component and its amplitude, spectral entropy, and energy proportion within a specific frequency band. The time-domain features and frequency-domain features extracted within each time window are combined into a feature vector and arranged in chronological order of the time windows to form a sequence of state features that continuously evolve within the preset period of the acquisition component. Data quality is ensured through data cleaning and alignment. The dynamic and periodic patterns of equipment operation are fully captured by utilizing time-domain statistical features and frequency-domain features. The sliding window technique forms a continuous sequence, which enhances the time-series sensitivity of state trend analysis. The combination of multi-dimensional feature vectors improves the detection accuracy and efficiency of abnormal fluctuations, providing a real-time data foundation for subsequent correlation analysis and fault diagnosis, thereby supporting the accuracy and response speed of the online simulation system.

[0023] Step 3: Train the state pool using deep learning algorithms to establish a data acquisition correlation analysis model, which characterizes the correlation and matching relationships between the state feature sequences of each acquisition component. The model identifies abnormal fluctuation trends of each state feature sequence in the overall state pool. The construction process of the data acquisition correlation analysis model includes: The state feature sequences of all acquisition components in the state pool are aligned and standardized according to a unified timestamp to construct a multi-dimensional time series input matrix; The deep learning algorithm adopts an encoder and decoder architecture with long short-term memory network units as the core. During the training phase, the model is trained unsupervised using state pool data under historical normal working conditions. The training objective is to learn the dynamic correlation pattern between state feature sequences through the encoder and reconstruct the input sequence through the decoder. The encoder is used to learn and compress the high-dimensional spatiotemporal features of the state feature sequences of each acquisition component, and the decoder is used to reconstruct or predict the normal state of each sequence based on the encoded features. After training, the hidden layer state and internal weight matrix of the model represent the complex nonlinear correlation and matching relationship between the state sequences of each acquisition component. Real-time state pool data is input into the trained model, and the model outputs the corresponding reconstructed sequence. By calculating the residual between the real-time sequence and the reconstructed sequence, and combining the distribution statistics of the residual in the state pool, a quantitative criterion for identifying abnormal fluctuation trends is generated. By using deep learning to automatically learn and quantify the dynamic correlations between various devices and parameters from the overall operation data of the power plant, it can not only more sensitively capture potential fault precursors where a single parameter has not yet shown obvious abnormalities, but also effectively distinguish whether the abnormality is caused by the actual operation of the equipment or by the failure of a certain data acquisition component itself. This improves the accuracy and timeliness of early warnings and provides key technical support for proactive predictive intelligent operation and maintenance.

[0024] Step 4: Based on the identified abnormal fluctuation trend, determine whether to trigger a risk alarm according to the preset fluctuation threshold. If the alarm conditions are met, generate and output a risk alarm signal. The preset fluctuation threshold adopts a dynamic threshold setting method, which adaptively calculates the alarm threshold of each state characteristic sequence based on the current working mode of the gas power plant and historical normal operation data of the same period. When the amplitude or duration of the abnormal fluctuation trend exceeds its corresponding dynamic threshold, a risk alarm is triggered.

[0025] Step 5: Based on the abnormal fluctuation trend, the acquisition attributes of each acquisition component, and the preset performance indicators of the acquisition components, calculate the fault risk coefficient of the corresponding acquisition component using a preset fault identification model; if the fault risk coefficient exceeds a set threshold, it is determined that the acquisition component is faulty, and the abnormal data is restored to standard data based on the fault identification results to recover the correct indicator feedback data; the preset process of the fault identification model is as follows: Collect sample data marked with different fault types and normal states of acquisition components during historical operation. The sample data includes abnormal fluctuation patterns of state characteristic sequences, attribute information of acquisition components, and corresponding performance index data. A neural network model with multiple inputs and a single output is constructed. Its input layer receives the abnormal fluctuation trend feature vector, the attribute encoding vector of the collected components, and the performance index vector, respectively. The core of the model includes feature extraction, feature fusion based on the attention mechanism, and fault risk regression. The model is trained using the prepared sample data, and the model parameters are adjusted by optimizing the algorithm to minimize the error between the fault risk coefficient output by the model and the true fault label of the sample. The diagnostic accuracy and generalization ability of the model are verified using an independent test dataset. After verification, the trained fault identification model is integrated into the preset local software.

[0026] During the operation phase, the fault identification model analyzes the pattern of abnormal fluctuation trends to determine whether it is a sudden, drifting, or periodic anomaly. It combines the attribute information of the acquisition component, such as component type, installation location, measurement principle, and maintenance history. By referring to a preset performance index library of acquisition components, it calculates the fault risk coefficient through attention weight allocation. The fault risk coefficient is used to quantify the probability of the acquisition component failing.

[0027] The process of restoring standard data is as follows: When the fault identification model determines that a specific acquisition component is faulty, based on the association matching relationship established by the acquisition association analysis model, multiple normal acquisition components that are associated with the faulty acquisition component more than a preset association degree threshold are selected from the state pool. Based on the real-time status feature sequences of multiple normal acquisition components, a pre-trained regression completion network is used to estimate the standard indicator data that the faulty acquisition component should have output during the fault period and that conforms to the overall operation logic of the power plant, so as to replace or correct the abnormal data caused by the fault and complete the data restoration. The construction process of the regression completion network is based on historical normal operation data of the gas-fired power plant, and specifically includes the following steps: The training dataset is constructed by extracting the state feature sequences of all collected components during fault-free operation from a long-term historical database. The data is then cleaned and aligned to ensure that the timestamps of each component's sequence are synchronized. An encoder-decoder architecture is adopted, in which the encoder is a multi-layer long short-term memory network used to learn high-order spatiotemporal correlations between multi-source normal state feature sequences; the decoder is a fully connected layer responsible for mapping the correlated features to the standard output sequence of the target acquisition component. Using a specific data acquisition component as a simulated target, the data from that component is randomly occluded for a portion of the time period during training. The network is trained to predict the standard sequence for the occluded time period using sequences from other related components as input. The loss function employs mean squared error, and the influence of strongly correlated components is amplified by weights provided by the correlation analysis model. The dataset is divided using cross-validation, and overfitting is prevented by an early stopping strategy. The final network should be able to stably output a high-precision completed sequence that matches the true value after inputting the sequence of related components. The regression completion network learns the implicit operational relationships between multiple components through the encoder and achieves reliable estimation of data under fault conditions with the help of the decoder. This construction process enables the network to restore local missing or abnormal data from the overall system operating state, providing high-quality data support for online simulation and enhancing the robustness of the system and the continuity of simulation under component failure.

[0028] Step 6: Simultaneously submit the risk alarm results generated in Step 4 and the restored standard data from Step 5 to the operating condition simulation framework for batch-by-batch operating condition simulation demonstrations. This will allow for online display of the power plant's operating conditions under different risk states or after fault recovery. During the batch-by-batch operating condition demonstrations, the operating condition simulation framework will automatically match the corresponding simulation scenario library based on the input risk alarm level and the restored data. Each batch of demonstrations includes visualization of operating condition parameters, highlighting of anomaly locations, fault impact simulation, and a comparison between the restored data and the original data to support operators in making online decisions.

[0029] Compared with existing technologies, this embodiment deeply integrates real-time monitoring, anomaly diagnosis, and online simulation to construct a closed-loop system from data acquisition to operational condition demonstration. It not only realizes fault alarms but also proactively demonstrates the impact of faults and recovery status through simulation. It introduces a deep learning model to identify associated anomalies and performs fault determination and data restoration of the acquired components based on multi-source information fusion, thereby improving the system's reliability and data integrity under component failures. It adopts a dynamic threshold alarm and a batch operational condition demonstration mechanism, enabling the system to adapt to different operating modes. It also intuitively supports operators' decision-making through visual simulation, enhancing the real-time performance and operability of operational condition management.

[0030] At other levels, this embodiment also provides a process for constructing a working condition simulation framework, such as... Fig. 2 As shown, it includes the following steps: Step 11: Through the acquisition component, synchronously acquire the set values ​​of operating indicators under different working modes and the corresponding historical and real-time indicator feedback data from the distributed control system and historical database of the gas power plant. The data covers at least one complete power plant start-up, shutdown and load regulation cycle. Step 12: Clean, align and normalize the obtained indicator feedback data, and divide the processed data into multiple independent datasets corresponding to startup, rated operation, variable load and shutdown modes according to the preset working mode discrimination rules. Step 13: Based on the partitioned dataset, establish a simulation sub-model for each working mode, combining theoretical expected values ​​and real-time calculated correction values. The simulation sub-model simulates and outputs the expected equipment state and system response based on the input indicator settings. The simulation sub-model calculates the theoretical expected value of the equipment state based on the input operating indicator settings. At the same time, the data-driven model built through a deep neural network learns the system response deviation patterns caused by equipment characteristics, environmental factors, etc. in actual operation based on a large amount of historical matching data, and outputs the real-time calculated correction value. During simulation, the simulation sub-model calls these two parts in parallel, integrates the theoretical expected value and the real-time calculated correction value of the data-driven model based on the current input, and finally outputs the simulation result. Step 14: Integrate the simulation sub-models of each working mode, the real-time data access interface, and the working condition demonstration engine into a preset software platform to form the working condition simulation framework. This framework has the functions of receiving real-time data, calling the corresponding model to perform simulation calculations, and outputting visual demonstration results. This operating condition simulation framework can autonomously identify abnormal fluctuations in operating status and potential equipment failures, and dynamically correct simulation parameters and logic accordingly. It deeply integrates online monitoring, fault recovery, and decision-making simulation, which not only improves the accuracy and timeliness of status perception, but also enhances the operators' ability to predict complex operating conditions and their emergency response capabilities, thereby providing digital support for the safe, efficient, and optimized operation of gas-fired power plants.

[0031] In summary, this invention achieves full-process online and automated operation from data acquisition to simulation verification, enabling real-time and continuous monitoring and simulation of power plant operation status, greatly improving the timeliness and proactivity of operating condition management. By constructing a data acquisition correlation analysis model through deep learning, it can not only identify anomalies in individual devices, but also capture hidden correlation anomalies and trend drifts from the perspective of the entire system, thereby improving the accuracy of fault warning and early detection capability. By designing a dual-path mechanism for handling anomalies, it not only ensures immediate alarms for major risks, but also distinguishes between equipment failures and data anomalies through intelligent diagnosis, and reliably restores abnormal data, thus ensuring the data quality of subsequent simulation analysis and the reliability of decision-making basis. This invention integrates alarm information and restoration data into a working condition simulation framework for multi-scenario demonstrations, providing efficient support for operation optimization and emergency plan formulation, and enhancing the overall reliability, safety and economy of gas power plant operation.

[0032] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for online simulation of operating conditions of a gas-fired power plant, characterized in that, Includes the following steps: Step 1: Obtain the operating index settings and corresponding real-time index feedback data of each operating device in different working modes of the gas power plant through the preset acquisition components, and construct the operating condition simulation framework of the gas power plant based on the operating index settings and index feedback data. Step 2: Based on the real-time indicator feedback data acquired by each acquisition component, continuously extract state features within a preset sampling period to generate a state feature sequence corresponding to each acquisition component; collect the state feature sequences of all acquisition components to establish a global state pool. Step 3: Train the state pool using deep learning algorithms to establish a data collection and correlation analysis model. The model identifies the abnormal fluctuation trends of each state feature sequence in the overall state pool. Step 4: Based on the identified abnormal fluctuation trend, determine whether to trigger a risk alarm according to the preset fluctuation threshold. If the alarm conditions are met, generate and output a risk alarm signal. Step 5: Based on the abnormal fluctuation trend, the acquisition attributes of each acquisition component, and the preset performance indicators of the acquisition components, calculate the fault risk coefficient of the corresponding acquisition component through the preset fault identification model. If the fault risk coefficient exceeds the set threshold, it is determined to be a fault in the data acquisition component, and the abnormal data is restored to standard data based on the fault identification results. Step 6: Simultaneously submit the risk alarm results generated in Step 4 and the standard data restored in Step 5 to the working condition simulation framework for batch working condition simulation demonstrations.

2. The online simulation method for operating conditions of a gas-fired power plant according to claim 1, characterized in that, The different operating modes in step 1 include the start-up mode, rated load operation mode, variable load regulation mode, and shutdown mode of the gas power plant; the operating equipment includes gas turbines, waste heat boilers, steam turbines, and generators; the operating index settings include temperature, pressure, flow rate, speed, and power settings, and the index feedback data are the actual measured values ​​of the corresponding operating indexes.

3. The online simulation method for operating conditions of a gas-fired power plant according to claim 1, characterized in that, The process of constructing the working condition simulation framework in step 1 includes the following steps: Step 11: Through the acquisition component, synchronously acquire the set values ​​of operating indicators under different working modes and the corresponding historical and real-time indicator feedback data from the distributed control system and historical database of the gas power plant; Step 12: Clean, align and normalize the obtained indicator feedback data, and divide the processed data into multiple independent datasets corresponding to startup, rated operation, variable load and shutdown modes according to the preset working mode discrimination rules. Step 13: Based on the partitioned dataset, establish a simulation sub-model for each working mode that combines theoretical expected values ​​with real-time calculated correction values. The simulation sub-model simulates and outputs the expected equipment state and system response based on the input index settings. Step 14: Integrate the simulation sub-models of each working mode, the real-time data access interface, and the working condition demonstration engine into a preset software platform to form the working condition simulation framework.

4. The online simulation method for operating conditions of a gas-fired power plant according to claim 1, characterized in that, The extraction process of the state feature sequence in step 2 is as follows: For each acquisition component, the continuous index feedback data stream acquired within the preset sampling period is cleaned and aligned. The sliding time window technique is used to extract multiple subsequences of equal length from the data stream; For each subsequence, time-domain feature extraction and frequency-domain feature extraction are performed in parallel: Temporal feature extraction includes calculating the statistical features of the subsequence; Frequency domain feature extraction includes performing a fast Fourier transform on the subsequence to obtain its spectrum; The time-domain features and frequency-domain features extracted within each time window are combined into a feature vector and arranged in chronological order of the time windows to form a sequence of state features that continuously evolve within the preset period of the acquisition component.

5. The online simulation method for operating conditions of a gas-fired power plant according to claim 1, characterized in that, The process of constructing the data collection correlation analysis model in step 3 includes: The state feature sequences of all acquisition components in the state pool are aligned and standardized according to a unified timestamp to construct a multi-dimensional time series input matrix; An encoder-decoder architecture with long short-term memory network units as the core is adopted as the deep learning algorithm. During the training phase, the model is trained unsupervised using state pool data under historical normal working conditions. The training objective is to learn the dynamic correlation pattern between state feature sequences through the encoder and reconstruct the input sequence through the decoder. Real-time state pool data is input into the trained model, and the model outputs the corresponding reconstructed sequence. By calculating the residual between the real-time sequence and the reconstructed sequence, and combining the distribution statistics of the residual in the state pool, a quantitative criterion for identifying abnormal fluctuation trends is generated.

6. The online simulation method for operating conditions of a gas-fired power plant according to claim 1, characterized in that, The preset fluctuation threshold in step 4 adopts a dynamic threshold setting method, which adaptively calculates the alarm threshold of each state characteristic sequence based on the current working mode of the gas power plant and the historical normal operation data of the same period. A risk alarm is triggered when the amplitude or duration of an abnormal fluctuation trend exceeds its corresponding dynamic threshold.

7. The online simulation method for operating conditions of a gas-fired power plant according to claim 1, characterized in that, The preset process for the fault identification model in step 5 is as follows: Collect sample data marked with different fault types and normal states of acquisition components during historical operation. The sample data includes abnormal fluctuation patterns of state characteristic sequences, attribute information of acquisition components, and corresponding performance index data. Construct a neural network model with multiple inputs and a single output, where the input layer receives the abnormal fluctuation trend feature vector, the attribute encoding vector of the acquisition component, and the performance index vector, respectively. The model is trained using the prepared sample data, and the model parameters are adjusted by optimizing the algorithm to minimize the error between the fault risk coefficient output by the model and the true fault label of the sample. The diagnostic accuracy and generalization ability of the model are verified using an independent test dataset. After verification, the trained fault identification model is integrated into the preset local software.

8. The online simulation method for operating conditions of a gas-fired power plant according to claim 1, characterized in that, In step 5, the fault identification model analyzes the pattern of abnormal fluctuation trends during the operation phase, and determines whether it is a sudden, drifting, or periodic anomaly, combined with the attribute information of the acquisition component, such as component type, installation location, measurement principle, and maintenance history. The failure risk coefficient is calculated by comparing it with a pre-set database of performance indicators for the data acquisition components and allocating attention weights.

9. The online simulation method for operating conditions of a gas-fired power plant according to claim 1, characterized in that, The standard data restoration process in step 5 is as follows: When the fault identification model determines that a specific acquisition component is faulty, based on the association matching relationship established by the acquisition association analysis model, multiple normal acquisition components that are associated with the faulty acquisition component more than a preset association degree threshold are selected from the state pool. Based on the real-time status feature sequences of multiple normal acquisition components, a pre-trained regression completion network is used to estimate the standard indicator data that the faulty acquisition component should have output during the fault period, which conforms to the overall operation logic of the power plant, so as to replace or correct the abnormal data caused by the fault and complete the data restoration.

10. The online simulation method for operating conditions of a gas-fired power plant according to claim 1, characterized in that, In the batch operation condition demonstration process in step 6, the operation condition simulation framework automatically matches the corresponding simulation scenario library according to the input risk alarm level and the restored data; each batch of demonstrations includes operation condition parameter visualization, anomaly location highlighting, fault impact deduction, and comparison display of restored data and original data.

Images