Gas turbine hot-hanging intelligent prediction and solution method and system based on big data

Abstract

The present disclosure provides a kind of big data-based gas turbine hot suspension intelligent prediction and solution method, system, by constructing high-frequency multidimensional data acquisition module, real-time obtains gas turbine operating parameter, and the original data is rejected exception, complete missing data etc.Pretreatment, and then based on historical normal operation data, fault precursor data and fault data, construct dynamic optimization history database, adopt the machine learning algorithm of random forest and LSTM neural network fusion training obtains intelligent prediction model, in the process of unit operation, the parameters after real-time pretreatment are input model to output fault occurrence probability, time window and precursor level, and compare and verify with the same kind of working condition data in history database in real time, according to the deviation threshold and double verification mechanism triggers graded intervention measures, while the intervention effect is fed back to the model for incremental learning and dynamic optimization, so as to realize the advance prediction, accurate grading response and active intervention of gas turbine hot suspension fault.

Description

Technical Field

[0001] This invention relates to the field of gas turbine power generation, and in particular to a smart prediction and solution method and system for gas turbine thermal suspension based on big data. Background Technology

[0002] Gas turbines, as highly efficient power machines, are widely used in power generation, aerospace, industrial drives, and many other fields. The stability of their startup process directly determines the reliability, economy, and continuity of the unit's operation. During the startup and acceleration process of a gas turbine, thermal runaway is an occasional but highly dangerous fault. This fault not only leads to turbine startup failure and direct economic losses, but can also cause irreversible damage to the unit's thermal path components due to problems such as a sudden rise in turbine exhaust temperature, significantly reducing the unit's service life and long-term operational economy.

[0003] The typical characteristics of thermal overhang failure are: during the startup phase, the gas turbine speed stops increasing or even shows a downward trend. Even with continuous increases in fuel supply, the speed cannot be pushed up further. At the same time, the turbine exhaust temperature will rapidly climb to the unit's set exhaust temperature limit. If not intervened in time, it can easily lead to more serious equipment failure. Currently, most existing gas turbine control systems adopt a passive fault handling mechanism, that is, they only prevent equipment damage by cutting off fuel or emergency shutdown after thermal overhang has occurred. This makes it impossible to predict the failure risk in advance and intervene proactively, resulting in delayed fault response and difficulty in guaranteeing a successful startup.

[0004] With the rapid development of big data technology and artificial intelligence machine learning algorithms, intelligent monitoring and fault prediction of industrial equipment has become an industry trend. Gas turbines generate massive amounts of operational data during long-term operation, containing patterns and early warning signs of thermal suspension faults. However, current technologies have not effectively mined and utilized this data, failing to predict thermal suspension faults in advance. Therefore, there is an urgent need for an intelligent prediction and solution for gas turbine thermal suspension faults that integrates big data and machine learning technologies. This solution would overcome the limitations of traditional monitoring and prediction methods, enabling accurate prediction and proactive intervention of thermal suspension faults, and ensuring stable startup and long-term reliable operation of gas turbines. Summary of the Invention

[0005] To address the aforementioned technical problems, embodiments of this application provide a big data-based intelligent prediction and solution for gas turbine thermal suspension, including: Step S1: Collect various operating parameter data during the operation of the gas turbine, and preprocess the collected raw data to obtain the preprocessed current operating parameter data; Step S2: Construct a historical database of gas turbine operating parameters. The historical database includes a normal operation dataset, a thermal suspension failure precursor dataset, and a thermal suspension failure dataset. The database is dynamically updated and backed up. Step S3: Based on the data in the historical database, a fusion algorithm is used to train a gas turbine thermal suspension intelligent prediction model, and the model is optimized and verified. Step S4: Input the preprocessed current operating parameter data into the intelligent prediction model to obtain the thermal suspension failure prediction result, and compare and verify the current operating parameter data with the data of the same working condition in the historical database in real time. Based on the judgment result of the comparison deviation value and the preset threshold, trigger the corresponding response measures. Step S5: Implement corresponding graded intervention measures according to the level of fault precursors, and collect parameter data after intervention and feed it back to the intelligent prediction model for dynamic self-learning and model optimization.

[0006] In conjunction with the first aspect, the various types of operating parameter data mentioned in step S1 include at least: fuel parameters, atmospheric parameters, exhaust temperature, gas turbine speed, valve opening, natural gas calorific value, guide vane opening, pressure difference, anti-surge valve opening, and operating current.

[0007] In conjunction with the first aspect, the preprocessing described in step S1 includes: Using the 3σ criterion combined with the Grubbs criterion, and setting the confidence level to 99%, outlier data that exceeds the reasonable range is automatically identified and removed. Linear interpolation was used to complete the missing data; The Z-score standardization method was used to convert all parameter data into standardized data with a mean of 0 and a standard deviation of 1. The wavelet transform algorithm is adopted, and the db4 wavelet basis is selected. The standardized data is subjected to multi-level wavelet decomposition to extract low-frequency components and remove high-frequency components.

[0008] In conjunction with the first aspect, the historical database mentioned in step S2: The normal operation dataset is archived according to the unit's operating conditions, which include start-up cleaning, ignition, speed increase, and stable operation. The thermal suspension failure precursor dataset contains operating parameter data from 1 second to 600 seconds before the occurrence of a thermal suspension failure, and marks the changing trends of core parameters in the failure precursor stage; The thermal suspension failure dataset contains all parameter data at the moment of occurrence of a thermal suspension failure and during the duration of the failure.

[0009] In conjunction with the first aspect, the fusion algorithm in step S3 is a fusion algorithm combining the random forest algorithm and the LSTM neural network algorithm, specifically including: The feature importance evaluation function of the random forest algorithm is used to rank the feature importance of the preprocessed full-dimensional running parameters and select the core feature parameters as the main input of the model. The data in the historical database is divided into training, validation and test sets according to the proportions, and the model is trained accordingly. The hyperparameters of the model were optimized using a grid search method combined with a Bayesian optimization algorithm. The stability of the model was verified using k-fold cross-validation. The optimized model was finally validated using test set data.

[0010] In conjunction with the first aspect, the core characteristic parameters include: gas turbine speed, exhaust temperature, fuel valve opening, natural gas calorific value, IGV opening, compressor inlet filter pressure difference, compressor inlet temperature, anti-surge valve opening, and SFC current; the remaining parameters serve as auxiliary inputs; the model output of the thermal suspension fault prediction results includes the probability of fault occurrence, the time when the fault may occur, and the level of fault precursors.

[0011] In conjunction with the first aspect, step S4 specifically includes: Based on the current operating conditions, environmental conditions and fuel quality of the unit, normal operating data and fault precursor data under similar conditions are selected from the historical database. The Euclidean distance algorithm is used to calculate the deviation between the current operating parameter data and historical data of the same type. A preset comparison deviation threshold is set. When the comparison deviation value does not exceed the threshold, the unit is judged to be operating normally. When the comparison deviation value exceeds the threshold, it is judged that there is a risk of thermal suspension failure, and corresponding intervention measures are triggered according to the level of failure precursor.

[0012] In conjunction with the first aspect, step S4 also includes a dual verification mechanism: When the probability of failure output by the model is greater than or equal to the first preset value and the comparison deviation value exceeds the threshold, intervention measures are triggered immediately. When the probability of failure output by the model is between the second preset value and the first preset value and the comparison deviation exceeds the threshold, an early warning is issued and continuous monitoring is performed. If the parameters do not improve within a preset time and the deviation continues to increase, intervention measures are triggered.

[0013] In conjunction with the first aspect, the tiered intervention measures described in step S5 include: The gas turbine control system switches to the backup temperature control line, raising the exhaust temperature limit to the safe upper limit; At the same time, the anti-surge valve control logic is adjusted to close the anti-surge valve and reduce the speed, thereby increasing the amount of air entering the gas turbine.

[0014] A second aspect of this disclosure provides a gas turbine thermal suspension prediction and solution system based on big data, comprising: The data acquisition module is used to collect various operating parameter data during the operation of the gas turbine. The preprocessing module is used to preprocess the collected raw data to obtain the preprocessed current operating parameter data; The historical database module is used to store historical data of gas turbine operating parameters, including normal operation data, thermal suspension failure precursor data, and thermal suspension failure data. The machine learning prediction model module is trained based on data from a historical database and is used to output thermal suspension failure prediction results based on current operating parameter data. The comparison and verification module is used to compare and verify the current operating parameter data with the same operating condition data in the historical database in real time, and trigger a response based on the judgment result of the comparison deviation value and the preset threshold. The graded intervention module is used to execute corresponding graded intervention measures based on the level of fault precursors. The feedback optimization module is used to collect parameter data after intervention and feed it back to the machine learning prediction model module for dynamic self-learning and model optimization.

[0015] Beneficial Effects: This disclosure provides a big data-based intelligent prediction and solution system for gas turbine thermal suspension. It acquires gas turbine operating parameters in real time by constructing a high-frequency, multi-dimensional data acquisition module. The raw data undergoes refined preprocessing, including anomaly removal, missing data completion, standardization, and noise reduction. A dynamically optimized historical database is then built based on historical normal operation data, fault precursor data, and fault data. An intelligent prediction model is trained using a machine learning algorithm that combines random forest and LSTM neural networks. During unit operation, the preprocessed parameters are input into the model to output the fault occurrence probability, time window, and precursor level. The model is then compared and verified in real time with similar operating conditions in the historical database. Based on deviation thresholds and a dual verification mechanism, graded intervention measures are triggered. The intervention effect is fed back to the model for incremental learning and dynamic optimization. This achieves early prediction, precise graded response, and proactive intervention for gas turbine thermal suspension faults, avoiding over-intervention or delayed intervention, and significantly improving fault avoidance capabilities and the safety and reliability of unit operation. Attached Figure Description

[0016] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0017] Figure 1 This is a flowchart illustrating a big data-based intelligent prediction and solution for gas turbine thermal suspension according to an embodiment of this application; Figure 2 This is a structural diagram of a gas turbine thermal suspension intelligent prediction and solution system based on big data, according to an embodiment of this application. Detailed Implementation

[0018] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0019] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0020] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0021] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0022] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0023] like Figure 1 The flowchart shown is a big data-based intelligent prediction and solution for gas turbine thermal suspension according to an embodiment of the application, including: Step S1: Collect various operating parameter data during the operation of the gas turbine, and preprocess the collected raw data to obtain the preprocessed current operating parameter data; The various operating parameter data mentioned in step S1 include at least: fuel parameters, atmospheric parameters, exhaust temperature, gas turbine speed, valve opening, natural gas calorific value, guide vane opening, pressure difference, anti-surge valve opening, and operating current.

[0024] The preprocessing described in step S1 includes: Using the 3σ criterion combined with the Grubbs criterion, and setting the confidence level to 99%, outlier data that exceeds the reasonable range is automatically identified and removed. Linear interpolation was used to complete the missing data; The Z-score standardization method was used to convert all parameter data into standardized data with a mean of 0 and a standard deviation of 1. The wavelet transform algorithm is adopted, and the db4 wavelet basis is selected. The standardized data is subjected to multi-level wavelet decomposition to extract low-frequency components and remove high-frequency components.

[0025] This embodiment takes a 300MW heavy-duty gas turbine as the application object and adopts the intelligent prediction and solution for gas turbine thermal suspension based on big data provided in this disclosure. The specific implementation of step S1 is as follows: First, the collected operating parameter data includes at least the following: fuel parameters such as fuel flow rate, fuel pressure, and fuel temperature; atmospheric parameters such as atmospheric pressure and atmospheric humidity; exhaust temperature; gas turbine speed; fuel valve opening; natural gas calorific value; IGV opening (guide vane opening); compressor inlet filter pressure difference; anti-surge valve opening; and SFC operating current. Exhaust temperature is monitored using thermocouples arranged at multiple points, with the average value taken as the monitoring value. The data acquisition frequency is set to 10Hz, and data is transmitted to the edge computing unit via TCP / IP protocol, ensuring a transmission delay of ≤100ms.

[0026] After data collection, the raw data undergoes fine preprocessing, as detailed below: The system employs a combination of the 3σ criterion and the Grubbs criterion, with a confidence level of 99%, to automatically identify and remove outlier data that exceeds reasonable ranges. For example, the normal fluctuation range for exhaust temperature is 200-600℃. If the exhaust temperature is 850℃ at a certain moment, the system determines that this value exceeds the 3σ range, verifies it as an outlier using the Grubbs criterion, and automatically removes it. Simultaneously, it records the time of the anomaly (e.g., 10:23:45 on [Date]) and the parameter type. The same method is used to handle other outlier data such as a sudden drop in fuel flow to 0 (not in a shutdown state) or a sudden increase in engine speed exceeding 3000 r / min.

[0027] For missing data caused by transmission interruptions or sensor malfunctions, linear interpolation is used to complete the data. For example, if the calorific value data of natural gas is missing at a certain moment, five consecutive valid data points before and after that moment are selected, and the missing value is calculated using a linear interpolation formula. After completion, the data error is controlled within ±1%.

[0028] The Z-score standardization method is used to convert all parameter data into standardized data with a mean of 0 and a standard deviation of 1, eliminating dimensional differences. For example, fuel flow rate (unit: Nm³ / h) and exhaust temperature (unit: °C) are both converted into dimensionless data after standardization.

[0029] The wavelet transform algorithm is adopted, and the db4 wavelet basis is selected. The standardized data is decomposed into three levels of wavelet decomposition to extract low-frequency components that characterize the precursors of hot suspension failure (such as slow increase in exhaust temperature and decrease in speed increase rate). High-frequency components (noise caused by environmental interference and equipment vibration) are removed. Finally, standardized, noise-free, continuous and complete preprocessed data are obtained and stored in a temporary data buffer for subsequent model training and real-time prediction.

[0030] Step S2: Construct a historical database of gas turbine operating parameters. The historical database includes a normal operation dataset, a thermal suspension failure precursor dataset, and a thermal suspension failure dataset. The database is dynamically updated and backed up. The historical database mentioned in step S2: The normal operation dataset is archived according to the unit's operating conditions, which include start-up cleaning, ignition, speed increase, and stable operation. The thermal suspension failure precursor dataset contains operating parameter data from 1 second to 600 seconds before the occurrence of a thermal suspension failure, and marks the changing trends of core parameters in the failure precursor stage; The thermal suspension failure dataset contains all parameter data at the moment of occurrence of a thermal suspension failure and during the duration of the failure.

[0031] Based on the data preprocessing completed in step S1, a historical database of gas turbine operating parameters is constructed. This embodiment uses a MySQL relational database, deployed on an industrial server. The database is divided into three main data modules, which are interconnected and work together to support each other. The specific construction process is as follows: Normal operating parameter data for this type of gas turbine were collected under different environmental conditions (atmospheric pressure 80-120 kPa, atmospheric humidity 20%-100%, atmospheric temperature -10-40℃). The data were categorized and archived according to the unit's operating conditions, including four conditions: start-up purging, ignition, speed increase, and stable operation. The sample size for each condition was no less than 10,000 sets. For example, under the start-up speed increase condition (speed 0-3000 r / min), 12,000 sets of normal data were collected, with corresponding environmental conditions (e.g., atmospheric pressure 101 kPa, humidity 60%, temperature 25℃) and fuel parameters (e.g., natural gas calorific value 35 MJ / m³) labeled to ensure the representativeness of the samples.

[0032] Operating parameter data for the gas turbine were collected from 1 second to 600 seconds before each thermal suspension failure (12 times in this embodiment). The trends in core parameters during the pre-failure phase were highlighted. For example, 300 seconds before a certain thermal suspension failure occurred, the exhaust temperature slowly increased from 450℃ to 520℃, and the rate of increase in engine speed decreased from 5 r / s to 2 r / s. The time of occurrence of the precursor (300 seconds before the failure), the corresponding environmental conditions (atmospheric pressure 101 kPa, atmospheric humidity 60%), and fuel parameters (natural gas calorific value 35 MJ / m³) were also noted to help the model capture pre-failure characteristics.

[0033] Real-time parameter data were collected during the 12 hot suspension failures, including all 15 types of parameter data at the moment of failure and during the failure process. For example, at the moment of a hot suspension failure, the engine speed stopped at 2500 r / min and began to decrease, the exhaust temperature rapidly rose to 620℃ (exceeding the unit's set exhaust temperature limit of 600℃), the fuel flow rate continued to increase but the engine speed did not increase. All parameter data for this process were completely recorded and used to train the model's fault identification capability.

[0034] A dynamic database update and optimization mechanism is established, with automatic updates scheduled every Sunday at 2:00 AM to replenish normal operation data and newly added fault case data for the week. Data classification labels are also optimized to ensure accuracy. A data backup mechanism is established, employing a combination of redundant and local backups: local backups are performed daily, storing data on the server's local disk and external hard drives; redundant backups are performed weekly, synchronizing data to a remote backup server to prevent data loss and ensure long-term database stability. The core function of this historical database is to provide massive amounts of sample data for training machine learning prediction models, and to provide a reference standard for real-time comparison and verification of prediction results, achieving a closed loop of "real-time prediction - historical comparison - accurate judgment".

[0035] Step S3: Based on the data in the historical database, a fusion algorithm is used to train a gas turbine thermal suspension intelligent prediction model, and the model is optimized and verified. The fusion algorithm in step S3 is a combination of the random forest algorithm and the LSTM neural network algorithm, specifically including: The feature importance evaluation function of the random forest algorithm is used to rank the feature importance of the preprocessed full-dimensional running parameters and select the core feature parameters as the main input of the model. The data in the historical database is divided into training, validation and test sets according to the proportions, and the model is trained accordingly. The hyperparameters of the model were optimized using a grid search method combined with a Bayesian optimization algorithm. The stability of the model was verified using k-fold cross-validation. The optimized model was finally validated using test set data.

[0036] The core characteristic parameters include: gas turbine speed, exhaust temperature, fuel valve opening, natural gas calorific value, IGV opening, compressor inlet filter pressure difference, compressor inlet temperature, anti-surge valve opening, and SFC current; the remaining parameters are used as auxiliary inputs; the model output thermal suspension fault prediction results include the probability of fault occurrence, the time when the fault may occur, and the level of fault precursors.

[0037] This embodiment, considering the nonlinear and multi-factor synergistic influence characteristics of thermal suspension failure in this type of gas turbine, employs a fusion algorithm combining random forest and LSTM neural network algorithms to construct an intelligent prediction model for gas turbine thermal suspension. Model training, optimization, and validation are all performed on an industrial server using Python programming language and based on the Scikit-learn and TensorFlow frameworks. The specific implementation process is as follows: The feature importance assessment function of the random forest algorithm was used to rank the feature importance of 15 categories of preprocessed full-dimensional operating parameters. A feature importance threshold of 0.05 was set to select the core feature parameters most strongly correlated with thermal suspension faults. The final determined core feature parameters include: turbine speed, exhaust temperature, fuel valve opening, natural gas calorific value, IGV opening, compressor inlet filter differential pressure, compressor inlet temperature, anti-surge valve opening, and SFC current, totaling 9 categories of parameters as the model's primary inputs. The remaining 6 categories of parameters (fuel flow rate, fuel pressure, fuel temperature, atmospheric pressure, atmospheric humidity, and compressor inlet filter differential pressure—note: the original text already lists differential pressure as core, so auxiliary parameters should be adjusted to others, such as fuel flow rate, fuel pressure, fuel temperature, atmospheric pressure, and atmospheric humidity) are used as auxiliary inputs. This reduces the computational load of model training, improves training efficiency, retains key fault-related features, and avoids overfitting caused by feature redundancy.

[0038] The preprocessed data in the historical database was divided into training, validation, and test sets in a 7:2:1 ratio. The training set was used for model parameter fitting, the validation set for hyperparameter tuning, and the test set for final model performance verification. During training, nine core feature parameters and six auxiliary parameters were selected as model inputs. The probability of hot suspension failure (0-100%), the time frame for failure occurrence (0-600s, i.e., the failure may occur within the next 0-600s), and the level of early warning signs (1-3, where level 1 is a slight early warning sign, level 2 is a moderate early warning sign, and level 3 is a severe early warning sign) were used as the model output targets. The number of training iterations was set to 1000, the learning rate was set to 0.001, and the Adam optimizer was used for model parameter optimization.

[0039] A grid search algorithm combined with Bayesian optimization was used to iteratively fine-tune the core hyperparameters of the model. After optimization, the random forest had 100 decision trees with a depth of 15 layers; the LSTM neural network had 3 hidden layers with 128, 64, and 32 neurons respectively, and a dropout coefficient of 0.3; at the same time, the model's regularization parameter was set, with an L2 regularization coefficient of 0.001, to avoid overfitting or underfitting and ensure the model's generalization ability.

[0040] A 5-fold cross-validation method was employed, dividing the training set into five subsets. Four subsets were selected sequentially as training data, and one subset as validation data, repeating this process five times. The model prediction accuracy was calculated for each validation iteration, and the average of the five validation results was taken as the cross-validation accuracy. In this embodiment, the cross-validation accuracy exceeded 90%, ensuring the model's stability.

[0041] The optimized model was finally validated using test set data. Fifteen types of parameters were input into the test set, and the model's output of fault occurrence probability, occurrence time, and precursor level were compared with the actual data labeled in the test set. The model's prediction accuracy, fault occurrence time prediction error, and precursor level determination accuracy were calculated. Validation showed that the model's prediction accuracy reached 92%, the fault occurrence time prediction error was within ±50 seconds, and the precursor level determination accuracy was 88%. Finally, a stable, reliable, and high-precision intelligent prediction model for gas turbine thermal suspension was formed, deployed on an industrial server, and linked in real-time with the data acquisition module and historical database.

[0042] Step S4: Input the preprocessed current operating parameter data into the intelligent prediction model to obtain the thermal suspension failure prediction result, and compare and verify the current operating parameter data with the data of the same working condition in the historical database in real time. Based on the judgment result of the comparison deviation value and the preset threshold, trigger the corresponding response measures. Step S4 specifically includes: Based on the current operating conditions, environmental conditions and fuel quality of the unit, normal operating data and fault precursor data under similar conditions are selected from the historical database. The Euclidean distance algorithm is used to calculate the deviation between the current operating parameter data and historical data of the same type. A preset comparison deviation threshold is set. When the comparison deviation value does not exceed the threshold, the unit is judged to be operating normally. When the comparison deviation value exceeds the threshold, it is judged that there is a risk of thermal suspension failure, and corresponding intervention measures are triggered according to the level of failure precursor.

[0043] Step S4 also includes a dual verification mechanism: When the probability of failure output by the model is greater than or equal to the first preset value and the comparison deviation value exceeds the threshold, intervention measures are triggered immediately. When the probability of failure output by the model is between the second preset value and the first preset value and the comparison deviation exceeds the threshold, an early warning is issued and continuous monitoring is performed. If the parameters do not improve within a preset time and the deviation continues to increase, intervention measures are triggered.

[0044] During the start-up and normal operation of the gas turbine, the intelligent prediction model is linked with the historical database in real time, requiring no manual intervention. In this embodiment, the specific implementation process of step S4 is as follows: The data acquisition module collects 15 types of parameter data of the unit's current operation in real time at a frequency of 10Hz, and quickly transmits them to the edge computing unit. Following the preprocessing process described in step S1 (outlier removal, missing value completion, standardization and noise reduction), the preprocessed current operating parameter data is obtained within 100ms.

[0045] The preprocessed current operating parameter data (including 9 types of core feature parameters and 6 types of auxiliary parameters) is input in real time into the intelligent prediction model trained and optimized in step S3. The model completes feature extraction, computational analysis, and outputs the hot suspension fault prediction results within 50ms, specifically including: fault occurrence probability, possible fault time, fault precursor level, and core parameter anomaly alerts. For example, at a certain moment, the model outputs: fault occurrence probability 85%, possible fault time 120s later, fault precursor level 3, and core parameter anomaly alerts "exhaust temperature continues to rise, speed increase rate decreases".

[0046] The model automatically retrieves relevant data from the historical database for real-time comparison and verification. The comparison logic is as follows: based on the current operating conditions of the unit (e.g., during startup and acceleration phase, speed 2200 r / min), environmental conditions (atmospheric pressure 101 kPa, atmospheric humidity 55%, atmospheric temperature 25℃), and fuel quality (natural gas calorific value 34.8 MJ / m³), normal operating data and fault precursor data under similar conditions are filtered from the historical database. The Euclidean distance algorithm is used to calculate the deviation between the feature vector of the current operating parameter data and the feature vector of similar historical data.

[0047] A preset comparison deviation threshold is established, which is dynamically calibrated based on the current gas turbine model, operating conditions, and historical fault data. In this embodiment, the deviation threshold is set to 0.8. When the comparison deviation value does not exceed 0.8, the unit is determined to be operating normally, and the model continues to monitor in real time. When the comparison deviation value exceeds 0.8, the unit is determined to have a risk of thermal suspension failure, and corresponding intervention measures are triggered according to the failure precursor level (level 1-3). Simultaneously, the current prediction results, comparison data, and judgment results are written into the historical database for subsequent incremental training and optimization of the model.

[0048] To avoid false positives or false negatives, this embodiment sets up a dual verification mechanism, specifically: When the probability of failure output by the model is ≥80% (first preset value) and the comparison deviation value exceeds 0.8, intervention measures are triggered immediately. When the model outputs a fault occurrence probability between 60% and 80% (the second preset value is 60%, and the first preset value is 80%), and the comparison deviation exceeds 0.8, the control system issues a warning (such as an audible and visual alarm) while continuously monitoring parameter changes. If the parameters do not improve within 10 seconds and the deviation continues to increase, intervention measures are triggered; if the parameters return to normal and the deviation falls back to within 0.8, the warning is canceled, and only relevant data is recorded.

[0049] Through the aforementioned dual verification mechanism, "accurate prediction and tiered response" are achieved, avoiding excessive or delayed intervention. For example, during a certain operation, the model output a fault probability of 75% (within the 60%-80% range), with a deviation value of 0.85. The system issued an early warning and continued monitoring. During the monitoring period, the exhaust temperature gradually decreased and the speed returned to normal. After 6 seconds, the deviation value dropped back to 0.75, and the system automatically canceled the warning without triggering intervention, thus avoiding unnecessary unit adjustments.

[0050] Step S5: Implement corresponding graded intervention measures according to the level of fault precursors, and collect parameter data after intervention and feed it back to the intelligent prediction model for dynamic self-learning and model optimization.

[0051] The tiered intervention measures described in step S5 include: The gas turbine control system switches to the backup temperature control line, raising the exhaust temperature limit to the safe upper limit; At the same time, the anti-surge valve control logic is adjusted to close the anti-surge valve and reduce the speed, thereby increasing the amount of air entering the gas turbine.

[0052] Based on the fault precursor level output in step S4 and the dual verification results, this embodiment formulates a graded intervention strategy. All intervention measures are automatically triggered and executed by the gas turbine control system, while the intervention effect is fed back in real time, forming a closed loop of "intervention-feedback-optimization". The specific implementation is as follows: This embodiment implements corresponding intervention measures for different levels of fault precursors. It should be noted that, based on the actual operating characteristics of this type of gas turbine, Level 1 (minor precursor), Level 2 (moderate precursor), and Level 3 (severe precursor) all employ a unified core intervention strategy, but the triggering timing and response speed can be dynamically adjusted according to the precursor level. Specific intervention measures include: Switching to the backup temperature control line: The gas turbine control system quickly switches to the backup temperature control line. Compared to the basic temperature control line, this backup temperature control line increases the exhaust temperature limit from 600℃ to 630℃ (safe upper limit), avoiding unit speed-up obstacles due to overly strict temperature control, while ensuring that the exhaust temperature does not exceed the safe range (≤630℃).

[0053] Adjust the anti-surge valve control logic: Simultaneously adjust the anti-surge valve control logic, reduce the anti-surge valve closing speed from 2800 r / min to 2600 r / min, so that the anti-surge valve enters the closing process earlier, increases the amount of air entering the gas turbine body, thereby reducing the turbine exhaust temperature and curbing the trend of hot suspension.

[0054] For example, when the model outputs a fault precursor level of 3, a fault occurrence probability of 85%, and a comparison deviation value of 1.2 (exceeding the threshold of 0.8) in step S4, the gas turbine control system completes the switching of the backup temperature control line and the adjustment of the anti-surge valve control logic within 50ms. Subsequently, the system monitors key parameters such as exhaust temperature and speed in real time to observe the intervention effect. Actual test data shows that within 30 seconds after the intervention, the exhaust temperature decreased from 620℃ to 590℃, and the speed recovered from 2500r / min to the normal acceleration rate, effectively curbing the thermal suspension trend.

[0055] For Level 1 or Level 2 precursors, the intervention can be delayed through the double verification mechanism in step S4 (e.g., issue an early warning first, monitor continuously for 10 seconds, and trigger the intervention only if the parameters do not improve). However, once triggered, the intervention will be performed in the same way.

[0056] After the intervention, the predictive model dynamically learns based on the post-intervention parameter data, adjusting the internal weight and threshold parameters to optimize prediction accuracy and intervention logic. The specific process is as follows: Intervention data collection: After the intervention measures are implemented, the data collection module continuously collects operating parameter data for at least 600 seconds after the intervention, including the change curves of core parameters such as exhaust temperature, speed, fuel valve opening, and anti-surge valve opening.

[0057] Effectiveness evaluation: The edge computing unit automatically evaluates the intervention effect and calculates key indicators (such as the decrease in exhaust temperature, the speed recovery rate, and the intervention response time). For example, in the case above, the exhaust temperature decreased by 30°C within 30 seconds, and the speed recovered to the normal acceleration curve, indicating that the intervention was effective.

[0058] Incremental training: The intervention case and complete parameter data before and after the intervention were entered into the historical database as incremental training samples. At the same time, the model's recognition weights for fault precursors under similar operating conditions were adjusted. For example, in this case, the feature combination of "exhaust temperature 620℃, speed 2500r / min" was marked as a valid precursor, and the model will correspondingly increase the sensitivity of such features in subsequent predictions.

[0059] Threshold optimization: Based on feedback from multiple interventions, dynamically calibrate the comparison deviation threshold and the preset probability value for dual verification in step S4. For example, if the exhaust temperature drops rapidly after multiple consecutive interventions without over-intervention, maintain the current threshold; if intervention lag occurs, appropriately reduce the deviation threshold or probability trigger value.

[0060] Through the complete closed loop of "prediction-intervention-feedback-optimization" described above, the recognition accuracy and intervention accuracy of the intelligent prediction model are continuously improved. In this embodiment, within 6 months of continuous operation, a total of 12 early warnings were triggered, 5 actual interventions were carried out, and 5 hot suspension failures were successfully avoided, with no missed warnings and an intervention success rate of 100%. This significantly improved the safety and reliability of the gas turbine during the start-up and acceleration phases and during variable load operation.

[0061] like Figure 2 As shown, this disclosure presents a gas turbine thermal suspension prediction and solution system based on big data, comprising: The data acquisition module 210 is used to collect various operating parameter data during the operation of the gas turbine. This module consists of an industrial-grade PLC data acquisition terminal, a high-precision sensor group, an Ethernet transmission unit, and an edge computing unit. The sensor group includes 15 types of parameters: fuel flow sensor, fuel pressure sensor, fuel temperature sensor, atmospheric pressure sensor, atmospheric humidity sensor, exhaust temperature sensor (multiple points arranged for average averaging), gas turbine speed sensor, fuel valve opening sensor, natural gas calorific value detector, IGV opening sensor, compressor inlet filter differential pressure sensor, compressor inlet temperature sensor, anti-surge valve opening sensor, and SFC operating current sensor. The data acquisition frequency is set to 10Hz, and data is transmitted to the edge computing unit via TCP / IP protocol with a transmission delay ≤100ms.

[0062] The preprocessing module 220 is used to preprocess the collected raw data to obtain preprocessed current operating parameter data. This module receives the raw data transmitted by the data acquisition module 210 and performs the following preprocessing operations: outlier data is removed using the 3σ criterion combined with the Grubbs criterion (confidence level 99%); missing data is completed using linear interpolation (e.g., taking 5 valid points before and after the missing value); all parameters are converted into dimensionless data with a mean of 0 and a standard deviation of 1 using the Z-score standardization method; low-frequency effective features are extracted and high-frequency noise is removed using the wavelet transform algorithm (db4 wavelet basis, 3-level decomposition), finally obtaining standardized, noise-free, continuous and complete preprocessed data.

[0063] Historical database module 230 stores historical operating parameter data of the gas turbine, including normal operation data, thermal suspension failure precursor data, and thermal suspension failure data. This module uses a MySQL relational database and is deployed on an industrial server. The database contains three main datasets: a normal operation dataset (categorized by start-up cleaning, ignition, speed increase, and stable operation conditions, with ≥10,000 samples per condition), a thermal suspension failure precursor dataset (containing operating data from 1 second to 600 seconds before 12 thermal suspension failures, annotating the changing trends of core parameters), and a thermal suspension failure dataset (containing all parameters at the moment of failure and throughout its duration). The database is set to automatically update every Sunday at 2:00 AM and employs a combination of local backup (once daily) and redundant backup (once weekly) to prevent data loss.

[0064] Machine learning prediction model module 240, trained on data from a historical database, is used to output thermal suspension fault prediction results based on current operating parameter data. This module is trained on data from historical database module 230 using an algorithm that combines random forest and LSTM neural network. First, by ranking the importance of features through random forest, nine core feature parameters (gas turbine speed, exhaust temperature, fuel valve opening, natural gas calorific value, IGV opening, compressor inlet filter pressure difference, compressor inlet temperature, anti-surge valve opening, and SFC current) are selected as the main inputs, and the remaining six are auxiliary inputs. During training, the training, validation, and test sets were divided in a 7:2:1 ratio, with 1000 iterations and a learning rate of 0.001. The Adam optimizer was used. Hyperparameters were determined through grid search combined with Bayesian optimization (random forest with 100 trees and 15 layers, LSTM with 3 hidden layers and 128 / 64 / 32 neurons, dropout 0.3, and L2 regularization coefficient 0.001). After 5-fold cross-validation (accuracy > 90%) and test set validation (prediction accuracy 92%), the model was deployed on an industrial server, outputting the probability of failure (0-100%), the time it may take for a failure to occur (0-600s), and the level of fault precursors (1-3).

[0065] The comparison and verification module 250 is used to compare and verify the current operating parameter data with similar operating condition data in the historical database in real time, and trigger a response based on the judgment result of the comparison deviation value and the preset threshold. This module is linked in real time with the machine learning prediction model module 240. Based on the current operating conditions of the unit (such as the speed of 2200 r / min during the start-up and acceleration phase), environmental conditions (atmospheric pressure 101 kPa, humidity 55%, temperature 25℃), and fuel quality (calorific value 34.8 MJ / m³), it filters normal operating data and fault precursor data under similar operating conditions from the historical database module 230, and uses the Euclidean distance algorithm to calculate the deviation value between the current parameter feature vector and the historical data. The preset deviation threshold is 0.8. When the deviation value is ≤0.8, the unit is judged to be normal; when the deviation value is >0.8, a risk of thermal suspension failure is judged. A dual verification mechanism is set up: if the probability of failure output by the model is ≥80% and the deviation value is >0.8, intervention is triggered immediately; if the probability is between 60% and 80% and the deviation value is >0.8, an audible and visual warning is issued and monitoring continues for 10 seconds. If the parameters do not improve and the deviation value continues to increase, intervention is triggered; otherwise, the warning is canceled.

[0066] The graded intervention module 260 is used to execute corresponding graded intervention measures according to the level of fault precursors. This module receives the trigger signal from the comparison and verification module 250 and executes graded intervention measures according to the level of fault precursors. The intervention measures include: switching the gas turbine control system to the backup temperature control line, increasing the exhaust temperature limit from 600℃ to 630℃ (safe upper limit); and synchronously adjusting the anti-surge valve control logic, reducing the anti-surge valve closing speed from 2800 r / min to 2600 r / min to increase the amount of air entering the gas turbine. All interventions are completed automatically within 50ms. For example, when the model outputs a precursor level of 3, a fault probability of 85%, and a deviation value of 1.2, the system executes the above interventions, and the measured exhaust temperature drops from 620℃ to 590℃ within 30s, and the speed returns to normal.

[0067] The feedback optimization module 270 is used to collect parameter data after intervention and feed it back to the machine learning prediction model module for dynamic self-learning and model optimization. This module collects parameter data within at least 600 seconds after intervention, evaluates the intervention effect (such as the decrease in exhaust temperature and the recovery rate of engine speed), and enters the intervention case and complete data before and after intervention into the historical database module 230 as incremental training samples. It also adjusts the weight parameters of the machine learning prediction model module 240 and the deviation threshold of the comparison and verification module 250. Through the "prediction-intervention-feedback-optimization" closed loop, the system has avoided hot suspension failures a total of 5 times within 6 months of continuous operation, with an intervention success rate of 100%.

[0068] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A smart prediction and solution for gas turbine thermal suspension based on big data, characterized in that, include: Step S1: Collect various operating parameter data during the operation of the gas turbine, and preprocess the collected raw data to obtain the preprocessed current operating parameter data; Step S2: Construct a historical database of gas turbine operating parameters. The historical database includes a normal operation dataset, a thermal suspension failure precursor dataset, and a thermal suspension failure dataset. The database is dynamically updated and backed up. Step S3: Based on the data in the historical database, a fusion algorithm is used to train a gas turbine thermal suspension intelligent prediction model, and the model is optimized and verified. Step S4: Input the preprocessed current operating parameter data into the intelligent prediction model to obtain the thermal suspension failure prediction result, and compare and verify the current operating parameter data with the data of the same working condition in the historical database in real time. Based on the judgment result of the comparison deviation value and the preset threshold, trigger the corresponding response measures. Step S5: Implement corresponding graded intervention measures according to the level of fault precursors, and collect parameter data after intervention and feed it back to the intelligent prediction model for dynamic self-learning and model optimization.

2. The method according to claim 1, characterized in that, The various operating parameter data mentioned in step S1 include at least: fuel parameters, atmospheric parameters, exhaust temperature, gas turbine speed, valve opening, natural gas calorific value, guide vane opening, pressure difference, anti-surge valve opening, and operating current.

3. The method according to claim 1, characterized in that, The preprocessing described in step S1 includes: Using the 3σ criterion combined with the Grubbs criterion, and setting the confidence level to 99%, outlier data that exceeds the reasonable range is automatically identified and removed. Linear interpolation was used to complete the missing data; The Z-score standardization method was used to convert all parameter data into standardized data with a mean of 0 and a standard deviation of 1. The wavelet transform algorithm is adopted, and the db4 wavelet basis is selected. The standardized data is subjected to multi-level wavelet decomposition to extract low-frequency components and remove high-frequency components.

4. The method according to claim 1, characterized in that, The historical database mentioned in step S2: The normal operation dataset is archived according to the unit's operating conditions, which include start-up cleaning, ignition, speed increase, and stable operation. The thermal suspension failure precursor dataset contains operating parameter data from 1 second to 600 seconds before the occurrence of a thermal suspension failure, and marks the changing trends of core parameters in the failure precursor stage; The thermal suspension failure dataset contains all parameter data at the moment of occurrence of a thermal suspension failure and during the duration of the failure.

5. The method according to claim 1, characterized in that, The fusion algorithm in step S3 is a combination of the random forest algorithm and the LSTM neural network algorithm, specifically including: The feature importance evaluation function of the random forest algorithm is used to rank the feature importance of the preprocessed full-dimensional running parameters and select the core feature parameters as the main input of the model. The data in the historical database is divided into training, validation and test sets according to the proportions, and the model is trained accordingly. The hyperparameters of the model were optimized using a grid search method combined with a Bayesian optimization algorithm. The stability of the model was verified using k-fold cross-validation. The optimized model was finally validated using test set data.

6. The method according to claim 5, characterized in that, The core characteristic parameters include: gas turbine speed, exhaust temperature, fuel valve opening, natural gas calorific value, IGV opening, compressor inlet filter pressure difference, compressor inlet temperature, anti-surge valve opening, and SFC current; the remaining parameters are used as auxiliary inputs; the model output thermal suspension fault prediction results include the probability of fault occurrence, the time when the fault may occur, and the level of fault precursors.

7. The method according to claim 1, characterized in that, Step S4 specifically includes: Based on the current operating conditions, environmental conditions and fuel quality of the unit, normal operating data and fault precursor data under similar conditions are selected from the historical database. The Euclidean distance algorithm is used to calculate the deviation between the current operating parameter data and historical data of the same type. A preset comparison deviation threshold is set. When the comparison deviation value does not exceed the threshold, the unit is judged to be operating normally. When the comparison deviation value exceeds the threshold, it is judged that there is a risk of thermal suspension failure, and corresponding intervention measures are triggered according to the level of failure precursor.

8. The method according to claim 7, characterized in that, Step S4 also includes a dual verification mechanism: When the probability of failure output by the model is greater than or equal to the first preset value and the comparison deviation value exceeds the threshold, intervention measures are triggered immediately. When the probability of failure output by the model is between the second preset value and the first preset value and the comparison deviation exceeds the threshold, an early warning is issued and continuous monitoring is performed. If the parameters do not improve within a preset time and the deviation continues to increase, intervention measures are triggered.

9. The method according to claim 1, characterized in that, The tiered intervention measures described in step S5 include: The gas turbine control system switches to the backup temperature control line, raising the exhaust temperature limit to the safe upper limit; At the same time, the anti-surge valve control logic is adjusted to close the anti-surge valve and reduce the speed, thereby increasing the amount of air entering the gas turbine.

10. A gas turbine thermal suspension prediction and solution system based on big data, characterized in that, include: The data acquisition module is used to collect various operating parameter data during the operation of the gas turbine. The preprocessing module is used to preprocess the collected raw data to obtain the preprocessed current operating parameter data; The historical database module is used to store historical data of gas turbine operating parameters, including normal operation data, thermal suspension failure precursor data, and thermal suspension failure data. The machine learning prediction model module is trained based on data from a historical database and is used to output thermal suspension failure prediction results based on current operating parameter data. The comparison and verification module is used to compare and verify the current operating parameter data with the same operating condition data in the historical database in real time, and trigger a response based on the judgment result of the comparison deviation value and the preset threshold. The graded intervention module is used to execute corresponding graded intervention measures based on the level of fault precursors. The feedback optimization module is used to collect parameter data after intervention and feed it back to the machine learning prediction model module for dynamic self-learning and model optimization.

Images