Performance prediction method of gas turbine secondary air system based on transformer

Abstract

The present invention relates to the technical field of gas turbine performance prediction, and discloses a transformer-based gas turbine secondary air system performance prediction method, which comprises obtaining and preprocessing historical operating data of a healthy gas turbine secondary air system under steady-state conditions, extracting key feature parameters, and time-series-producing key feature sequences; constructing and training a first Transformer model using the key feature sequences, and predicting the turbine outer cylinder chamber pressure and compressor exhaust pressure in combination with real-time operating data; constructing a residual time series by comparing measured values ​​with predicted values, and predicting future residuals using a second Transformer model, thereby quantitatively evaluating the degree of system performance degradation and generating a health assessment report; this method can improve the performance prediction accuracy and maintenance efficiency of the gas turbine secondary air system, ensure stable system operation, reduce maintenance costs, and improve energy utilization efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of gas turbine performance prediction, and more particularly to a transformer-based gas turbine secondary air system performance prediction method. Background Art

[0002] As a critical component of modern industry, gas turbines rely on the stable support of their secondary air systems (SASs). This system distributes compressed air to ensure turbine cooling, bearing sealing, and other functions, ensuring long-term, stable operation in high-temperature, high-pressure environments. However, with increasing operating time, degradation of the secondary air system's performance directly impacts the gas turbine's overall efficiency and component life. Therefore, real-time monitoring and accurate prediction of secondary air system performance are crucial.

[0003] In the prior art, Chinese patent application CN113187592A discloses a secondary air system monitoring method, apparatus, device, and storage medium. This method determines whether a system fault exists by obtaining a baseline flow difference value and auxiliary calculation value under different secondary air pump conditions. However, this method can only determine whether a system fault exists and cannot provide early warning of gradual degradation of system performance. Specifically, this method determines whether a system fault exists by setting preset stability and effectiveness conditions. However, this method lacks the ability to predict system performance degradation trends and cannot provide advance maintenance recommendations to operation and maintenance personnel.

[0004] A Chinese patent application, publication number CN117328978A, provides a flow estimation method for a secondary air system. This method primarily analyzes the system's instantaneous flow or characteristics by iteratively calculating mass flow. However, this method primarily focuses on analyzing instantaneous flow or system characteristics and fails to effectively support long-term performance evaluation. Specifically, this method analyzes system characteristics by estimating the mass flow of the secondary air injection valve. However, this method lacks the ability to predict long-term performance trends and cannot provide comprehensive decision support for the long-term operation and maintenance of gas turbines.

[0005] Existing technical solutions primarily focus on real-time monitoring and fault diagnosis, lacking intelligent prediction of secondary air system performance degradation trends. These solutions mostly rely on static analysis and short-term fault diagnosis, failing to fully utilize the extensive historical data accumulated during gas turbine operation. Furthermore, data-driven approaches fail to fully leverage the advantages of deep learning models to capture the dynamic interactions between multi-dimensional features. Summary of the Invention

[0006] To overcome the aforementioned shortcomings of the existing technology, the present invention provides a Transformer-based method for predicting gas turbine secondary air system performance. This method accurately predicts system performance through in-depth analysis of historical and real-time operating data from the gas turbine secondary air system. This method effectively improves prediction accuracy, promptly identifies signs of system performance degradation, and provides comprehensive health assessment reports to maintenance personnel, helping to rationalize maintenance planning, reduce maintenance costs, and ensure safe and stable gas turbine operation.

[0007] The present invention is applicable to various scenarios where gas turbines are used, such as power plants and power supply systems in industrial production. In these scenarios, stable operation of the gas turbine is crucial. This method can monitor secondary air system performance in real time, identifying potential problems in advance and ensuring efficient and stable operation of the gas turbine, thereby avoiding production interruptions and economic losses caused by system failures.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] The Transformer-based gas turbine secondary air system performance prediction method includes:

[0010] Obtaining historical operating data containing timestamps of a healthy gas turbine secondary air system under steady-state conditions and preprocessing the historical operating data; extracting key feature parameters from the preprocessed historical operating data, and time-sequencing the key feature parameters according to the timestamps to obtain a key feature sequence;

[0011] Based on the key feature sequence, the first Transformer model is constructed and trained; the real-time operating data sequence containing timestamps of the current gas turbine secondary air system is obtained, and the turbine outer cylinder cavity pressure prediction value time series is obtained based on the real-time operating data sequence and the trained first Transformer model. and compressor exhaust pressure prediction time series

[0012] Obtain the current gas turbine turbine outer cylinder pressure measured value time series P cyl and the measured value time series of compressor exhaust pressure P comp , according to P cyl 、P comp 、 and Get the current residual time series of the performance ratio; Based on the current residual time series and the pre-built second Transformer model, predict the residual time series of the next Y time points according to Quantitatively evaluate the performance degradation of the current gas turbine secondary air system and generate a health assessment report.

[0013] Furthermore, the key characteristic parameters include the compressor inlet total temperature T in , compressor inlet wall static pressure P in , compressor outlet pressure P out , compressor outlet temperature T out and turbine cooling valve opening V cool ;

[0014] The key feature sequence is obtained as follows:

[0015] The total temperature T at the compressor inlet is calculated based on the timestamp. in Perform time series processing to obtain the compressor inlet total temperature time series T in-seq ; Compressor inlet wall static pressure P according to timestamp in Perform time series processing to obtain the compressor inlet wall static pressure time series P in-seq ; Compressor outlet pressure P according to timestamp out Perform time series processing to obtain the compressor outlet pressure time series P out-seq ; Compressor outlet temperature T according to timestamp out Perform time series processing to obtain the compressor outlet temperature time series T out-seq ; According to the timestamp, the turbine cooling valve opening V cool Perform time series processing to obtain the turbine cooling valve opening time series V cool-seq ;

[0016] T in-seq 、P in-seq 、P out-seq 、T out-seq and V cool-seq Align by timestamp and build key feature sequence TS seq =(TS seq,1 ,TS seq,2 ,…,TS seq,n ), where TS seq,k For TS seq The kth element in , n is TS seq The total number of elements in , 1≤k≤n.

[0017] Furthermore, the TS seq Each element in is a sextuple:

[0018] TS seq,k =(t k ,T in,k ,P in,k ,P out,k ,T out,k ,V cool,k );

[0019] Among them, t k For TS seq The timestamp of the kth element in T in,k is the timestamp t k The total temperature at the compressor inlet under in,k is the timestamp t k The static pressure on the compressor inlet wall under out,k is the timestamp t k The compressor outlet pressure at out,k is the timestamp t k The compressor outlet temperature under cool,k is the timestamp t k Turbine cooling valve opening below.

[0020] Furthermore, the first Transformer model includes an input embedding layer, a position encoding layer, a self-attention mechanism, a multi-head attention mechanism, a feedforward neural network, and an output layer;

[0021] The constructing and training of the first Transformer model includes:

[0022] From the key feature sequence TS seq Each element of TS seq,k Extract the model input feature vector TS in,k =(T in,k ,P in,k ,P out,k ,T out,k ,V cool,k );

[0023] Input the model into the feature vector TS in,k Input to the input embedding layer for encoding to obtain the timestamp t k The embedded feature representation E k ;

[0024] The timestamp t is obtained through the position encoding layer k The position encoding vector P under k ;

[0025] The position encoding vector P k Add to the embedded feature representation E k , get the timestamp t k The model input feature representation Z k ;

[0026] According to the model input feature representation Z k Get the input sequence Z={Z1,Z2,…,Z n}.

[0027] Furthermore, constructing and training the first Transformer model further includes:

[0028] Through the self-attention mechanism in the first Transformer model, the correlation between the model input feature representations corresponding to different timestamps in the input sequence Z is calculated to obtain a one-dimensional feature correlation matrix;

[0029] The multi-head attention mechanism in the first Transformer model uses n2 attention heads to parallelly calculate the model input feature representation Z k The relationship between each input sub-feature in is used to obtain the multi-head attention weight matrix;

[0030] The single-dimensional feature correlation matrix and the multi-head attention weight matrix are nonlinearly transformed through a feedforward neural network, and the output layer maps the output of the feedforward neural network to the turbine outer cylinder pressure and the compressor exhaust pressure.

[0031] Furthermore, the position encoding vector P k =[P k,0 ,P k,1 ,P k,2 ,P k,3 ,…,P k,2d-1 ], where d represents the dimension of the position encoding vector, Represents the position encoding vector at the jth * The value of the dimension, j * is the dimension index in the position encoding vector, 0≤j * ≤2d-1; the values ​​in even dimensions are calculated using the sine function, and the values ​​in odd dimensions are calculated using the cosine function.

[0032] Furthermore, the model input feature representation Z k Contains five input sub-features, namely timestamp t k The total inlet temperature characteristic T' in,k , timestamp t k The static pressure characteristics of the inlet wall under in,k , timestamp t k Outlet pressure characteristic P' out,k , timestamp t k The outlet temperature characteristic T' out,k and timestamp t k Turbine cooling valve opening characteristic V' under cool,k ;

[0033] The n2 attention heads include at least head 1 head1 = (T' in,k 、P' out,k ), head 2 head2 = (V' cool,k 、P' in,k) and head 3head3=(V' cool,k , T' out,k ).

[0034] Furthermore, the time series of the predicted value of the turbine outer cylinder pressure is obtained based on the real-time running data sequence and the trained first Transformer model. and compressor exhaust pressure prediction time series include:

[0035] Extracting real-time compressor inlet total temperature from real-time operating data series Real-time compressor inlet wall static pressure Real-time compressor outlet pressure Real-time compressor outlet temperature and real-time turbine cooling valve opening

[0036] Respectively and Perform time series processing to obtain the real-time compressor inlet total temperature sequence Real-time compressor inlet wall static pressure sequence Real-time compressor outlet pressure series Real-time compressor outlet temperature series and real-time turbine cooling valve opening sequence

[0037] according to And the first Transformer model, get the time series of the turbine outer cylinder pressure prediction value and compressor exhaust pressure prediction time series

[0038] Furthermore, the basis And the first Transformer model, get the time series of the turbine outer cylinder pressure prediction value and compressor exhaust pressure prediction time series include:

[0039] Will and According to the timestamp alignment, construct the real-time key feature vector set TN seq =(TN seq,1 ,TN seq,2 ,…,TN seq,m ), where TN seq,j is the real-time key feature vector at the jth time point, m is the total number of time points, 1≤j≤m; in, is the real-time compressor inlet total temperature at the jth time point, is the real-time static pressure of the compressor inlet wall at the jth time point, is the real-time compressor outlet pressure at the jth time point, is the real-time compressor outlet temperature at the jth time point, is the real-time turbine cooling valve opening at the jth time point;

[0040] TN seq,j Input the trained first Transformer model to obtain the predicted value of the turbine outer cylinder pressure at the jth time point and the predicted value of the compressor exhaust pressure at the jth time point

[0041] according to Get the time series of the predicted value of the turbine outer cylinder pressure according to Get the time series of compressor exhaust pressure prediction value

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] This paper introduces the Transformer deep learning model, replacing traditional mechanism modeling with a data-driven approach and fully utilizing gas turbine operating data, overcoming the limitations of traditional mechanism modeling in dealing with complex systems. The Transformer algorithm, with its self-attention mechanism, can effectively capture the dynamic interactions between multidimensional features, not only improving prediction accuracy but also significantly enhancing the model's adaptability to complex dynamic characteristics, enabling the model to better cope with nonlinear and dynamic changes in the gas turbine secondary air system. Secondly, the method quantifies the degree of performance degradation of the secondary air system by calculating the residual between the model's predicted and measured values, and dynamically evaluates it based on historical trends, thus achieving intelligent degradation assessment of system performance. The magnitude of the residual directly reflects the degree of performance degradation, providing an intuitive decision-making basis for system maintenance and optimization, and realizing intelligent and refined performance management. This helps to promptly identify potential system problems, rationally arrange maintenance plans, reduce maintenance costs, and improve equipment reliability and operating efficiency. Furthermore, the method fully utilizes steady-state data from the gas turbine's historical operating data, extracts key characteristic parameters as model input, and constructs a Transformer-based performance prediction framework. Through deep learning technology, it is possible to extract implicit long-term trends and interactive characteristics from historical data, achieve accurate modeling and prediction of system performance, and provide strong technical support for the stable operation and performance optimization of gas turbines, which helps to improve energy utilization efficiency and ensure the safe, stable and efficient operation of gas turbines. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 This is a flow chart of the principle of the transformer-based gas turbine secondary air system performance prediction method of the present invention;

[0046] Figure 2 A flow chart of a method for obtaining a key feature sequence in a transformer-based gas turbine secondary air system performance prediction method of the present invention;

[0047] Figure 3 This is a flow chart of a method for quantitatively evaluating the performance degradation degree of the current gas turbine secondary air system and generating a health assessment report in the transformer-based gas turbine secondary air system performance prediction method of the present invention. DETAILED DESCRIPTION

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0049] Example 1

[0050] See also Figure 1 As shown, this embodiment provides a Transformer-based gas turbine secondary air system performance prediction method, including:

[0051] Step S1000: acquiring historical operating data including timestamps of a healthy gas turbine secondary air system under steady-state conditions, and preprocessing the historical operating data; extracting key feature parameters from the preprocessed historical operating data, and time-sequencing the key feature parameters according to the timestamps to obtain a key feature sequence;

[0052] Furthermore, step S1000 includes:

[0053] Step S1100, obtaining historical operating data including timestamps of a healthy gas turbine secondary air system under steady-state conditions, and preprocessing the historical operating data;

[0054] Specifically, we first need to extract historical operating data from the gas turbine's operating records under steady-state conditions (i.e., when the gas turbine's operating parameters are stable and there are no abnormal fluctuations). This data typically includes timestamps, operating parameters (such as temperature, pressure, and flow), and device status information. The purpose of obtaining this data is to provide basic data support for subsequent feature extraction and model training.

[0055] Preprocessing is an important step in data processing, and its purpose is to convert raw data into a format suitable for model training. The specific operations of preprocessing include:

[0056] Data cleaning: Remove outliers and noise from the data. Outliers may be caused by sensor failure or data transmission errors. If these data are not processed, they may affect the training results of the model. For example, if the temperature data at a certain point in time deviates significantly from the normal range, it can be considered an outlier and needs to be corrected or deleted.

[0057] Data alignment: Ensures that all data points have the same timestamp. Since gas turbine operating data is typically collected simultaneously by multiple sensors, inconsistent timestamps may exist. Alignment ensures that each data point has a consistent timestamp, facilitating subsequent time series analysis.

[0058] Data normalization: Converting data to the same scale. Because different operating parameters may have different dimensions and ranges, directly using raw data for model training may cause the model to be overly sensitive to certain parameters. Normalization converts all parameters to the same scale, such as [0, 1] or [-1, 1], thereby improving model training results.

[0059] Data completion: Missing data points can be filled in through interpolation or other methods. Gas turbine operating data may contain some missing data points. If these missing data points are not processed, they may affect model training results. Interpolation can be used to fill in missing data points based on the values ​​of adjacent data points, ensuring data integrity.

[0060] Through the above preprocessing operations, the original historical operation data can be converted into a format suitable for model training, providing high-quality data support for subsequent feature extraction and model training.

[0061] Step S1200: extract key characteristic parameters from the pre-processed historical operation data, wherein the key characteristic parameters include the compressor inlet total temperature T in , compressor inlet wall static pressure P in , compressor outlet pressure P out , compressor outlet temperature T outand turbine cooling valve opening V cool ;

[0062] Specifically, step S1200 extracts key characteristic parameters from the preprocessed historical data. This process is based on the physical mechanism of the gas turbine secondary air system, and selects the five parameters that have the most significant impact on system performance to form a core variable set for model input.

[0063] Compressor inlet total temperature T in Refers to the total temperature of the air at the compressor inlet section, including static and dynamic temperatures, and is a key parameter for calculating air density and compression work. High temperature will reduce air density, resulting in a decrease in compressor mass flow, which in turn affects the cooling capacity of the secondary air system. Compressor inlet wall static pressure P in The static pressure measured at the wall of the compressor inlet duct reflects the intake system resistance (such as the degree of filter blockage) and the ambient air pressure. A decrease in static pressure may indicate an increase in intake resistance, resulting in reduced compressor efficiency. Compressor outlet pressure P out The pressure value of the air after being compressed by the compressor directly determines the driving force of the cooling air to the outer cylinder cavity of the turbine. The outlet pressure fluctuation will affect the flow control accuracy of the cooling valve and is the core indicator for evaluating the pressure balance of the secondary air system. out :The temperature rise caused by power consumption during the compression process affects the thermodynamic state of the cooling air. The increase in outlet temperature will reduce the cooling efficiency and increase the thermal load of the turbine components. Turbine cooling valve opening V cool This refers to the real-time adjustment status (0%-100%) of the turbine cooling valve and is the system's active parameter for controlling cooling air flow. Changes in opening directly affect the turbine outer cylinder pressure, making it a key regulating variable connecting compressor output with turbine cooling requirements.

[0064] Through correlation analysis (e.g., Pearson correlation coefficient > 0.5) and mechanism verification, the above parameters are strongly correlated with the turbine outer cylinder pressure and compressor exhaust pressure. For example, the correlation coefficient between turbine cooling valve opening and cavity pressure reaches 0.85, indicating a direct causal relationship between the two. The extracted parameters are all continuous time series data, strictly aligned with the timestamps pre-processed in step S1100, forming structured data corresponding to "parameter-time", providing standardized input for the time series processing in step S1300.

[0065] For example, Table 1 shows the Pearson correlation coefficient of a certain type of gas turbine. A positive sign indicates that the variables change in the same direction (e.g., P out The increase is accompanied by the increase of the pressure in the outer cylinder of the turbine), and the negative sign indicates the reverse change of the variable (such as V cool The increase in opening causes the pressure in the turbine outer cylinder chamber to drop).

[0066] Table 1: Pearson correlation coefficient results

[0067]

[0068] Through correlation analysis and mechanism verification, the interference of minor parameters (such as bearing temperature and lubricating oil pressure) is eliminated, and the model input dimension is controlled within a reasonable range (5 parameters), avoiding the "curse of dimensionality" that can reduce model training efficiency. For example, if 20 redundant parameters are included, the computational complexity of the self-attention mechanism will increase by 400%. However, the selected key feature set retains 90% of the valid information while reducing computational cost. Parameter selection combines thermodynamic principles (such as energy conversion between compressor inlet and outlet parameters) and control logic (such as the regulation of cooling valve opening), enabling the model to learn feature interactions that conform to actual operating rules. For example, the negative correlation between compressor inlet total temperature and outlet pressure (high temperature → low density → drop in outlet pressure) is determined by aerodynamic equations. By learning this mechanistic relationship, the model improves prediction accuracy. The extracted parameters serve as the core elements of the key feature sequence in step S1300, and their quality directly affects the training results of the first Transformer model. For example, accurate inlet wall static pressure data enables the model to correctly identify the impact of changes in intake system resistance on compressor performance, thereby improving the accuracy of exhaust pressure prediction.

[0069] Step S1300, time-sequencing the key feature parameters according to the timestamp to obtain a key feature sequence;

[0070] Furthermore, if Figure 2 As shown, step S1300 includes:

[0071] Step S1310: The total temperature T at the compressor inlet is calculated based on the timestamp. in Perform time series processing to obtain the compressor inlet total temperature time series T in-seq ;

[0072] Step S1320: Compare the static pressure P of the compressor inlet wall according to the timestamp. in Perform time series processing to obtain the compressor inlet wall static pressure time series P in-seq ;

[0073] Step S1330: Compressor outlet pressure P according to the timestamp. out Perform time series processing to obtain the compressor outlet pressure time series P out-seq ;

[0074] Step S1340: Compressor outlet temperature T out Perform time series processing to obtain the compressor outlet temperature time series T out-seq ;

[0075] Step S1350: The turbine cooling valve opening V is set according to the timestamp. cool Perform time series processing to obtain the turbine cooling valve opening time series V cool-seq ;

[0076] Step S1360: T in-seq 、P in-seq 、P out-seq 、T out-seq and V cool-seq Align by timestamp and build key feature sequence TS seq =(TS seq,1 ,TS seq,2 ,…,TS seq,n ), where TS seq,k For TS seq The kth element in , n is TS seq The total number of elements in , 1≤k≤n.

[0077] TS seq Each element in is a sextuple:

[0078] TS seq,k =(t k ,T in,k ,P in,k ,P out,k ,T out,k ,V cool,k );

[0079] Among them, t k For TS seq The kth element TS in seq,k Timestamp, T in,k is the timestamp t k The total temperature at the compressor inlet under in,k is the timestamp t k The static pressure on the compressor inlet wall under out,k is the timestamp t k The compressor outlet pressure at out,k is the timestamp t k The compressor outlet temperature under cool,k is the timestamp t k Turbine cooling valve opening below.

[0080] Specifically, step S1300 aims to convert the pre-processed historical operating data into a key feature sequence suitable for Transformer model processing. Through time serialization and timestamp alignment, a multivariate time series data structure containing time dimension information is constructed to solve the problem of structured expression of multi-parameter time series data and provide standardized input for subsequent model training. In steps S1310 to S1350, for the total temperature T at the compressor inlet,in , compressor inlet wall static pressure P in , compressor outlet pressure P out , compressor outlet temperature T out and turbine cooling valve opening V cool These five key characteristic parameters are processed in time series. The specific operation is to arrange the measured values ​​of the same parameter at different times into a continuous time series according to the timestamp order of data acquisition. This process solves the discretization problem of the relationship between parameter values ​​and time in the original data, so that the historical change trend of each parameter can be captured by the model. in For example, its time series T in-seq It reflects the long-term impact of ambient temperature on compressor efficiency. During the operation of the gas turbine, the increase in the total inlet temperature will lead to a decrease in air density, which in turn affects the compression efficiency of the compressor and the cooling effect of the secondary air system. Through time series, the model can learn T in The hysteresis effect of temperature changes in different seasons or day and night temperature differences on system performance. Similarly, the static pressure on the compressor inlet wall P in The time series P in-seq Used to characterize the change in intake system resistance, outlet pressure P out The time series P out-seq Directly related to the driving force of cooling air, outlet temperature T out Time series T out-seq Reflects the thermal effect of the compression process, the turbine cooling valve opening V cool Time series V cool-seq Reflects the system's adjustment strategy for cooling needs.

[0081] In step S1360, T in-seq 、P in-seq 、P out-seq 、T out-seq and V cool-seq According to the timestamp alignment, the key feature sequence TS in the form of six tuples is constructed seq Each sextuplet TS seq,k =(t k ,T in,k ,P in,k ,P out,k ,T out,k ,V cool,k ) corresponds to the same time t k The multi-parameter state under t k is the timestamp, T in,k t k The total temperature of the compressor inlet at the time, P in,k t kThe inlet wall static pressure at the moment, and so on. Timestamp alignment solves the clock deviation problem that may exist in multi-sensor data acquisition, ensuring that different parameters are strictly corresponding in the time dimension. For example, if the compressor outlet pressure P out In t k The sudden drop at the same moment and the total inlet temperature T in The aligned sequences allow the model to capture the immediate coupling relationship between the two. By integrating the multivariate parameters into a unified time series structure, TS seq It provides the Transformer model with an input format containing spatiotemporal correlation information, meeting the model's processing requirements for multi-dimensional time series data.

[0082] Steps S1310-S1350 convert the historical data of a single parameter into a time-sequential information stream, solving the problem of expressing the dynamic characteristics of a single parameter; step S1360 establishes a time synchronization relationship between multiple parameters through timestamp alignment, solving the problem of spatiotemporal misalignment of multivariate data. The combination of single parameter time serialization and timestamp alignment makes the key feature sequence TS seq It not only preserves the temporal dependency of each parameter, but also constructs the spatial association of multiple parameters at the same time point, providing structured input for the subsequent self-attention mechanism of the Transformer model.

[0083] The performance of the gas turbine secondary air system is affected by the coupling of multiple parameters, such as the turbine cooling valve opening V cool The adjustment will immediately affect the turbine outer cylinder pressure, and the compressor inlet total temperature T in The change needs to go through the compression process to affect the outlet pressure P out TS seq The construction of TS enables the model to learn such complex associations across parameters and time. seq When inputting the first Transformer model, the self-attention mechanism can calculate T at different time points in With P out The correlation between them, the multi-head attention mechanism analyzes V in parallel cool With P in The local interaction of parameters such as , ultimately improves the model's ability to fit the system's dynamic characteristics.

[0084] Through time serialization and timestamp alignment, key feature sequence TS seq Converting discrete historical data into a continuous sequence with equal time intervals facilitates the model to learn the time trend and periodic characteristics of the parameters, such as capturing the regular decrease in compressor efficiency during the high temperature period in summer. Ensure that the multi-parameter states at the same timestamp correspond one to one to avoid correlation analysis errors caused by time misalignment, such as accurately correlating t kThe change in the cooling valve opening at each moment and the cavity pressure fluctuation at the same time form a multivariate time series format suitable for the Transformer model to process, enabling the model to use the self-attention mechanism to mine cross-parameter and cross-time dependencies, such as finding the compressor outlet temperature T out The law of hysteresis effect on turbine cooling effect.

[0085] Step S2000: Build and train the first Transformer model based on the key feature sequence; obtain the real-time operating data sequence containing timestamps of the current gas turbine secondary air system, and obtain the turbine outer cylinder pressure prediction value time series based on the real-time operating data sequence and the trained first Transformer model. and compressor exhaust pressure prediction time series

[0086] Furthermore, step S2000 includes:

[0087] Step S2100: constructing and training a first Transformer model based on the key feature sequence, wherein the first Transformer model includes an input embedding layer, a position encoding layer, a self-attention mechanism, a multi-head attention mechanism, a feedforward neural network, and an output layer;

[0088] Specifically, step S2100 utilizes the key feature sequence to construct and train the first Transformer model. Through a series of complex yet ordered operations, the model learns the inherent patterns in the gas turbine secondary air system operating data, thereby enabling accurate predictions of turbine outer cylinder pressure and compressor exhaust pressure. This process addresses the difficulty traditional models have in capturing dynamic interactions when processing multi-parameter time series data. By leveraging structured feature input and a deep neural network architecture, the model improves its ability to fit the nonlinear behavior of complex systems.

[0089] Furthermore, step S2100 includes:

[0090] Step S2110, from the key feature sequence TS seq Each element of TS seq,k Extract the model input feature vector TS in,k =(T in,k ,P in,k ,P out,k ,T out,k ,V cool,k );

[0091] Step S2120: Input the model into the feature vector TS in,k Input to the input embedding layer for encoding to obtain the timestamp t k The embedded feature representation Ek ;

[0092] The model input feature vector TS in,k The input to the input embedding layer for encoding consists of:

[0093] The model input feature vector TS is transformed linearly in,k Mapping to high-dimensional space representation:

[0094] E k =W e TS in,k +b e ;

[0095] Among them, W e is the embedding matrix, b e is the bias term.

[0096] Step S2130, obtain the timestamp t through the position coding layer k The position encoding vector P under k ;

[0097] The position encoding vector P k =[P k,0 ,P k,1 ,P k,2 ,P k,3 ,…,P k,2d-1 ], where d represents the dimension of the position encoding vector, Represents the position encoding vector at the jth * The value of the dimension, j * is the dimension index in the position encoding vector, 0≤j * ≤2d-1; the value of even dimensions is calculated by the sine function, that is, The values ​​of odd dimensions are calculated by the cosine function, that is, where i * is the dimension index in the positional encoding vector.

[0098] Step S2140: position encoding vector P k Add to the embedded feature representation E k , get the timestamp t k The model input feature representation Z k =E k +P k ; The model input feature representation Z k Contains five input sub-features, namely timestamp t k The total inlet temperature characteristic T' in,k , timestamp t k The static pressure characteristics of the inlet wall under in,k, timestamp t k Outlet pressure characteristic P' out,k , timestamp t k The outlet temperature characteristic T' out,k and timestamp t k Turbine cooling valve opening characteristic V' under cool,k ;

[0099] Step S2150: Input feature representation Z according to the model k Get the input sequence Z={Z1,Z2,…,Z n}, where 1≤k≤n;

[0100] Step S2160: Calculate the correlation between the model input feature representations corresponding to different timestamps in the input sequence Z through the self-attention mechanism in the first Transformer model to obtain a one-dimensional feature correlation matrix;

[0101] Step S2170: The multi-head attention mechanism in the first Transformer model uses n2 attention heads to parallelly calculate the model input feature representation Z k The relationship between each input sub-feature in is used to obtain the multi-head attention weight matrix;

[0102] The n2 attention heads include at least head 1 head1 = (T' in,k 、P' out,k ), head 2 head2 = (V' cool,k 、P' in,k ) and head 3head3=(V' cool,k , T' out,k );

[0103] In step S2180, the one-dimensional feature correlation matrix and the multi-head attention weight matrix are nonlinearly transformed through a feedforward neural network, and the output layer maps the output of the feedforward neural network to the turbine outer cylinder pressure and the compressor exhaust pressure.

[0104] Specifically, in step S2110, from the key feature sequence TS seq Each element of TS seq,k Extract the model input feature vector TS in,k =(T in,k ,P in,k ,P out,k ,T out,k ,V cool,k ), that is, extract the timestamp t kThe five key parameters of the compressor inlet total temperature, inlet wall static pressure, outlet pressure, outlet temperature and turbine cooling valve opening are obtained. This operation converts the multivariate time series data into a vector form that can be processed by the model, solving the problem that the original data format does not match the model input requirements. Step S2120 embeds the TS data into the vector form that can be processed by the model. in,k Perform linear transformation to obtain the embedded feature representation E k =W e TS in,k +b e , where W e is the embedding matrix, b e is the bias term, W e and b e are pre-set parameters that determine the mapping method and effect. This transformation maps the original features to a high-dimensional space, enhancing the expressive power of the features. For example, if the original feature vector is five-dimensional, the embedding matrix W e The dimensions can be set to 5×d (where d is the embedding dimension). Each parameter's scalar value is converted into a d-dimensional vector through a linear transformation, allowing the characteristics of different parameters to form a distinguishable distribution in the high-dimensional space. This operation resolves the problem of model training bias caused by dimensional differences in the original data. For example, the compressor inlet total temperature (unit: K) and turbine cooling valve opening (unit: %) are scaled uniformly through embedded coding, making it easier for the model to learn the relationships between parameters.

[0105] The purpose of position encoding in step S2130 is to provide the model with sequential information of the time series, solving the problem that the Transformer model itself does not have the ability to perceive time series. For example, for timestamps t1 and t2, the corresponding position encoding vector P k By distinguishing the order by different sine and cosine values, the model can identify the order of "t1" and "t2", thereby capturing the time dependence of parameter changes, such as the hysteresis effect of the increase in the total temperature of the compressor inlet at time t1 on the outlet pressure at time t2. Step S2140 converts the position encoding vector P k and the embedded feature E k Add together to get the model input feature representation Z k =E k +P k This operation deeply integrates the temporal sequence information with the parameter features to form an input representation containing spatiotemporal information. For example, if E k Represents the parameter feature vector at time t1, P k represents the position code of t1, then Z kIt also includes the parameter state and time position at that moment, so that the model can distinguish the dependencies between different time points when processing the features of the subsequent moment t2, such as the influence of the change in the opening of the turbine cooling valve at moment t1 on the pressure of the turbine outer cylinder at moment t2. n}Form a sequence structure suitable for Transformer processing. The sequence retains the temporal order and spatial correlation of key feature parameters, providing an input basis for the self-attention mechanism. For example, the input sequence Z can be regarded as an n×d-dimensional matrix, where each row corresponds to the spatiotemporal fusion features of a time point, enabling the model to calculate the correlation of features at different time points on a global scale. The elements in the single-dimensional feature correlation matrix represent the feature correlation strength of two time points. Step S2170 calculates the relationship between the input sub-features in parallel through n2 attention heads. For example, head 1 focuses on the inlet total temperature and outlet pressure (T' in,k 、P' out,k ), the first 2 focuses on the turbine cooling valve opening and the inlet wall static pressure (V' cool,k 、P' in,k ), the first 3 focuses on the turbine cooling valve opening and outlet temperature (V' cool,k , T' out,k ). This design solves the problem of capturing complex interactions between multiple parameters. For example, the self-attention mechanism can discover the cross-time correlation between the increase in the total inlet temperature at time t1 and the decrease in the outlet pressure at time t3, while the multi-head attention captures the local interactions of different parameter pairs in parallel (such as the immediate impact of the cooling valve opening adjustment on the intake static pressure). Step S2180 inputs the single-dimensional feature association matrix and the multi-head attention weight matrix into the feedforward neural network, extracts high-order features through nonlinear transformation, and finally maps them into turbine outer cylinder cavity pressure and compressor exhaust pressure prediction values ​​by the output layer. The nonlinear transformation capability of the feedforward network solves the problem of modeling the nonlinear behavior of the system, such as the nonlinear relationship between the compressor outlet pressure and the total inlet temperature and the cooling valve opening, and realizes complex mapping through a multi-layer perceptron structure. The output layer uses a regression model to convert high-dimensional features into specific pressure prediction values ​​to ensure dimensional consistency with the actual physical quantities.

[0106] Step S1300 provides structured multi-parameter input through time serialization and timestamp alignment. Step S2100 implements deep feature processing through embedded coding, position coding and attention mechanism. The fusion of position coding and embedded features enables the model to perceive time sequence and capture parameter associations across time points in combination with the self-attention mechanism. For example, it accurately identifies the impact of the three time points after the compressor inlet total temperature rises on the outlet pressure, solving the limitation of the traditional model that relies on fixed time window analysis. The multi-head attention mechanism processes different parameter pairs (such as V') in parallel. cool,k With P' in,kThe model captures local interactions and global dependencies, such as discovering the immediate feedback mechanism of turbine cooling valve opening adjustment on the static pressure of the intake system, improving the model's ability to fit complex coupling relationships. The nonlinear transformation of the feedforward neural network adapts to the nonlinear characteristics of the gas turbine system, such as the nonlinear curve of compressor efficiency changing with temperature. Through multi-layer activation functions, precise mapping is achieved, significantly improving prediction accuracy compared to linear models.

[0107] Step S2200: Obtain the real-time operating data sequence of the secondary air system of the current gas turbine, including the timestamp, and obtain the turbine outer cylinder pressure prediction value time series based on the real-time operating data sequence and the trained first Transformer model. and compressor exhaust pressure prediction time series

[0108] Furthermore, step S2200 includes:

[0109] Step S2210: extract the real-time compressor inlet total temperature from the real-time operation data sequence. Real-time compressor inlet wall static pressure Real-time compressor outlet pressure Real-time compressor outlet temperature and real-time turbine cooling valve opening

[0110] Step S2220, respectively and Perform time series processing to obtain the real-time compressor inlet total temperature sequence Real-time compressor inlet wall static pressure sequence Real-time compressor outlet pressure series Real-time compressor outlet temperature series and real-time turbine cooling valve opening sequence

[0111] Step S2230: and According to the timestamp alignment, construct the real-time key feature vector set TN seq =(TN seq,1 ,TN seq,2 ,…,TN seq,m ), where TN seq,j is the real-time key feature vector at the jth time point, m is the total number of time points, 1≤j≤m; in, is the real-time compressor inlet total temperature at the jth time point, is the real-time static pressure of the compressor inlet wall at the jth time point, is the real-time compressor outlet pressure at the jth time point, is the real-time compressor outlet temperature at the jth time point, is the real-time turbine cooling valve opening at the jth time point;

[0112] Step S2240: TN seq,j Input the trained first Transformer model to obtain the predicted value of the turbine outer cylinder pressure at the jth time point and the predicted value of the compressor exhaust pressure at the jth time point

[0113] Step S2250, according to Get the time series of the predicted value of the turbine outer cylinder pressure according to Get the time series of compressor exhaust pressure prediction value

[0114] Specifically, step S2200 aims to process the real-time operating data of the current gas turbine secondary air system based on the trained first Transformer model, and generate a time series of predicted values ​​of the turbine outer cylinder pressure and the compressor exhaust pressure. This step solves the format matching problem between real-time operating data and model input, as well as the real-time analysis problem of multi-parameter coupling relationship, by extracting real-time key feature parameters, constructing a feature vector set that conforms to the model input format, and using the trained model to achieve dynamic prediction. Step S2210 aims to extract five key parameters that are closely related to the performance of the secondary air system from the current operating data stream of the gas turbine. The real-time operating data sequence is collected through a distributed sensor network and contains a triplet of timestamp, parameter type and measurement value. Real-time compressor inlet total temperature Reflects the total temperature of the air at the compressor inlet, which is a key boundary condition affecting air density and compression efficiency; real-time compressor inlet wall static pressure Represents the static pressure at the compressor inlet wall, used to evaluate the influence of intake system resistance and ambient pressure; real-time compressor outlet pressure It is the pressure value of the air after compression by the compressor, which directly determines the driving force and flow distribution of the cooling air; the real-time compressor outlet temperature The temperature of the compressor outlet air reflects the thermal effect of the compression process and affects the cooling capacity of the secondary air; the real-time turbine cooling valve opening The real-time adjustment status of the turbine cooling valve determines the flow rate of secondary air distributed to the turbine outer cylinder. These parameters are collected in real time by sensors installed at key locations on the gas turbine, such as temperature sensors, pressure sensors, and valve position sensors, ensuring real-time and accurate data.

[0115] Step S2220 performs time serialization processing on the five extracted real-time parameters, that is, arranging the measured values ​​of the same parameter at consecutive time points into a time series according to the timestamp order of data acquisition. This process solves the time sequence expression problem of real-time data, so that the dynamic change trend of each parameter can be captured by the model. Step S2230 aligns the five real-time parameter sequences after serialization according to the timestamps to construct the real-time key feature vector set TN seq This step is consistent with the construction logic of the key feature sequence in step S1300, ensuring that the format of real-time data and historical training data is unified, which is convenient for model processing. By aligning the timestamps, the problem of asynchronous multi-sensor data is solved. For example, it avoids the misalignment of the acquisition time of the compressor outlet pressure and the turbine cooling valve opening, which causes the model to misjudge the coupling relationship between the two. Step S2240 will TN seq,j Input the trained first Transformer model, each feature vector TN seq,j First, a linear transformation is performed on the input embedding layer, mapping it to a high-dimensional feature space to generate an embedded feature representation. Positional encoding vectors are then superimposed to impart temporal order information, consistent with the feature processing used during model training in step S2100. The model uses a self-attention mechanism to calculate the correlation between feature vectors at different time points, capturing dynamic interactions between parameters (such as the immediate impact of changes in turbine cooling valve opening on compressor outlet pressure). A multi-head attention mechanism concurrently analyzes local correlations between different parameter pairs (such as the temporal dependence of inlet and outlet total temperatures), ultimately mapping them to specific pressure predictions through a feedforward neural network and output layer. The prediction results for each time point are aggregated to generate a complete prediction time series.

[0116] Steps S2210-S2230 form a unified data processing flow with the feature extraction, time serialization and alignment method of historical data in step S1300, ensuring the format compatibility of real-time data and training data. For example, the real-time key feature vector set TN seq The five-tuple structure and historical key feature sequence TS seqConsistency enables the first Transformer model to process real-time input without additional adjustment, avoids prediction deviations caused by data format differences, and improves the generalization ability and real-time response speed of the model. By constructing real-time key feature vectors through timestamp alignment, the clock deviation problem that may exist in gas turbine multi-sensor data acquisition (such as inconsistent acquisition times due to hardware differences between different sensors) is solved. For example, if there is a slight difference in the timestamps of the compressor outlet pressure sensor and the turbine cooling valve opening sensor, the alignment operation ensures that the parameter data at the same time point is strictly corresponding, so that the model can accurately capture the instantaneous coupling relationship between parameters (such as the cavity pressure change trend during the same period when the cooling valve opening increases), avoid prediction errors caused by time and space misalignment, and improve the accuracy of real-time prediction. The application of time series processing and position encoding enables the time-dependent information of real-time data to be retained. For example, the real-time compressor inlet total temperature series The dynamic changes of the inlet total temperature (such as gradually increasing with the increase of ambient temperature) are passed to the model through position encoding. Combined with the correlation calculation of historical time points by the self-attention mechanism, the model can learn the lagged effect of the change of the inlet total temperature on the outlet pressure (such as 2 time points after the inlet total temperature increases, the outlet pressure decreases slightly due to the decrease in air density), thereby more accurately predicting future pressure changes and solving the problem that traditional models are difficult to capture long-distance time series dependencies. The generated predicted value time series is strictly aligned with the measured value time series in the time dimension, providing direct comparative data for the residual calculation and performance evaluation in the subsequent step S3000. For example, the predicted value of the turbine outer cylinder cavity pressure and measured values The difference at the jth time point directly reflects the degree of fit of the model to the current operating conditions, while the trend analysis of the long-term prediction value series (such as continuous increase or decrease) can quantify the direction of system performance decline and provide data support for operation and maintenance decisions.

[0117] Step S3000: Obtain the time series of the actual measured value of the turbine outer cylinder pressure of the current gas turbine P cyl and the measured value time series of compressor exhaust pressure P comp , according to P cyl 、P comp 、 and Get the current residual time series of the performance ratio; Based on the current residual time series and the pre-built second Transformer model, predict the residual time series of the next Y time points The residual time series Contains Y forecast residual values; based on the residual time series of the predicted Y time points in the future Quantitatively evaluate the performance degradation of the current gas turbine secondary air system and generate a health assessment report.

[0118] Furthermore, step S3000 includes:

[0119] Step S3100: Obtain the time series of the actual measured value of the turbine outer cylinder pressure of the current gas turbine P cyl =(P cyl,1 ,P cyl,2 ,…,P cyl,m ) and the time series of the measured values ​​of the compressor exhaust pressure P comp =(P comp,1 ,P comp,2 ,…,P comp,m ), according to P cyl 、P comp 、 and Get the current residual time series of performance ratio; where P cyl,m is the measured value of the mth turbine outer cylinder pressure, P comp,m is the measured value of the exhaust pressure of the mth compressor;

[0120] According to P cyl 、P comp 、 and The current residual time series for obtaining the performance ratio includes:

[0121]

[0122] in:

[0123] r j : The residual value at the jth time point, 1≤j≤m;

[0124] P cyl,j : The measured value of the turbine outer cylinder pressure at the jth time point;

[0125] P comp,j : The measured value of the compressor exhaust pressure at the jth time point;

[0126] According to r j The current residual time series r = (r1, r2, ..., r m ).

[0127] Specifically, step S3100 obtains the current measured time series of the turbine outer cylinder pressure and compressor exhaust pressure in the gas turbine secondary air system, combines them with previously obtained predicted values, and calculates the current residual time series of the performance ratio. This residual time series serves as a key basis for evaluating the current system performance. The residual time series provides a visual representation of the degree of deviation between the model's predicted values ​​and the actual measured values. A larger residual indicates a greater degree of deviation from a healthy system performance. For example, if the residual value continues to increase over a period of time, this indicates a widening gap between the model's predicted and actual values, potentially indicating an abnormality in the gas turbine secondary air system, such as component wear or operating parameter drift. By monitoring the residual time series in real time, operators can promptly identify potential system issues and implement appropriate maintenance measures to prevent system failures and ensure safe and stable operation of the gas turbine. Furthermore, the residual time series provides an important data foundation for subsequent assessments of system performance degradation, facilitates further analysis of system performance trends, and provides strong support for the development of maintenance plans and optimized operation.

[0128] Step S3200: Based on the current residual time series and the pre-built second Transformer model, predict the residual time series at the next Y time points. The residual time series Contains Y prediction residual values;

[0129] Specifically, the construction of the second Transformer model is similar to the first, consisting of components such as an input embedding layer, a position encoding layer, a self-attention mechanism, a multi-head attention mechanism, a feedforward neural network, and an output layer. During the construction process, the model parameters, such as the number of network layers, hidden layer dimensions, and number of attention heads, must be appropriately determined based on the characteristics of the residual time series data and the requirements of the prediction task. For example, the appropriate position encoding vector dimension d is determined based on the length and complexity of the historical residual data to ensure that the model accurately captures the position information in the time series.

[0130] When training the second Transformer model, historical residual time series data is used as the training set. The model learns patterns and trends in this historical residual time series data to improve its ability to predict future residuals. During training, an optimization algorithm is used to continuously adjust the model parameters to minimize the error between the model's predicted values ​​and the actual residual values. The optimization objective is typically to minimize mean squared error (MSE) or cross-entropy loss. Backpropagation is used to calculate the gradient and update the model's weights and biases. Through multiple iterations of training, the model gradually learns the pattern of residual changes over time.

[0131] After model training is complete, the current residual time series is used as input, and the position encoding layer in the model incorporates the temporal information of each residual data point into the model input. Position encoding is calculated using specific sine and cosine functions to assign a unique code to the residual data at each time point, enabling the model to understand the temporal order and dependencies between residual data at different time points.

[0132] Next, the self-attention mechanism takes effect, calculating the mutual influence between each time point in the input data, thereby learning the trend patterns between historical residual data. The self-attention mechanism calculates the attention weights of the residual data at different time points to determine the importance of each time point in predicting future residuals. For example, for a sequence containing residual data at multiple time points, the self-attention mechanism can automatically identify which time points' residual data are most critical for future predictions, thereby more accurately capturing the trend of residual changes.

[0133] Based on the results of the self-attention calculation, the feedforward neural network further processes the data and extracts higher-level features through nonlinear transformations. Finally, the model's output layer generates a prediction of the residual for Y time points in the future, forming a predicted residual time series containing Y predicted residual values.

[0134] By predicting future residual time series, it's possible to predict performance trends in gas turbine secondary air systems. For example, if a predicted increasing residual indicates a potential decline in system performance, operators can proactively schedule equipment maintenance and adjust operating parameters to avoid equipment failures and production interruptions caused by deteriorating system performance. This helps improve equipment reliability and availability, reduces maintenance costs, and ensures efficient gas turbine operation. Furthermore, the prediction results provide a scientific basis for formulating maintenance plans, making maintenance more targeted and proactive, and improving the efficiency and effectiveness of equipment management.

[0135] Step S3300: Based on the predicted residual time series of Y future time points Quantitatively evaluate the performance degradation of the current gas turbine secondary air system and generate a health assessment report.

[0136] Furthermore, if Figure 3 As shown, step S3300 includes:

[0137] Step S3310, Perform linear regression and calculate the trend slope k';

[0138] The calculation of the trend slope k' comprises:

[0139]

[0140] in:

[0141] Y: the number of predicted future time points;

[0142] i: time point index from 1 to Y;

[0143] The predicted residual value at time m+i;

[0144] Linear regression reveals the changing trend of data by building a linear relationship model between variables. The slope k' depends on the prediction residual and time index i. With As time i increases or time i advances, k' will change accordingly. The slope k' represents the rate of change of the residual over time. If k' is a positive number, it means that the residual is increasing, which may mean that system performance is declining. If k' is a negative number, it means that the residual is decreasing and the system performance is improving; if k' is close to zero, it indicates that the system performance is relatively stable. By calculating the trend slope, the rate of system performance degradation can be quantified, providing an important indicator for evaluating system performance changes. This helps operators to promptly understand the changing trends of system performance and take measures in advance to address possible problems, such as adjusting equipment operating parameters and arranging equipment maintenance, to ensure the stable operation of the system.

[0145] Step S3320, calculating the cumulative value S of the prediction residual value at the next Y time points;

[0146]

[0147] Specifically, step S3320 calculates the cumulative value S of the predicted residuals at Y future time points. This cumulative value serves as a key indicator for assessing the total amount of performance degradation in the gas turbine secondary air system, providing an important basis for a comprehensive understanding of system performance changes. Cumulative value S comprehensively considers the overall deviation of system performance from a healthy state over the future period. The principle is that each predicted residual value reflects the difference between the predicted system performance value and the theoretical value under a healthy state at the corresponding time point. By summing these differences, a comprehensive indicator is generated to measure the overall degree of system performance degradation over the next Y future time points. A larger cumulative value S indicates a greater deviation from a healthy state during the predicted period, and a more significant overall performance degradation. For example, if S gradually increases from a smaller value, it indicates that system performance is deteriorating over time, potentially due to issues such as increased component wear or operating parameters gradually deviating from optimal values.

[0148] From a system maintenance and management perspective, the cumulative value S provides operators with a visual and quantitative indicator. By continuously monitoring changes in the S value, operators can gain a more comprehensive understanding of the long-term trends in system performance. If the S value approaches or exceeds a pre-set threshold, this alerts operators to the need for more in-depth inspection and maintenance of the system, such as scheduling equipment maintenance and adjusting operating parameters. This can prevent further deterioration in system performance, ensure stable operation of the gas turbine, and reduce the risk of downtime and maintenance costs caused by equipment failure.

[0149] Step S3330, calculating the performance degradation index PDI according to the trend slope k' and the cumulative value S;

[0150]

[0151] Among them, α is the weight coefficient of the trend slope, and β is the weight coefficient of the ratio of the cumulative value S to the predicted number of future time points Y;

[0152] Specifically, the trend slope k' reflects the rate of change of the residual over time and reflects the speed of system performance degradation; the cumulative value S comprehensively considers the total amount of system performance degradation in the future period. By combining the two, PDI can more comprehensively reflect the degradation of system performance. The setting of weight coefficients α and β allows the influence of trend slope and cumulative value to be adjusted according to actual conditions during the evaluation process; α and β are calibrated through expert experience; for example, in some application scenarios that are more sensitive to the speed of performance degradation, the value of α can be appropriately increased to highlight the impact of trend slope on the evaluation results; and when paying more attention to the total amount of performance degradation, the proportion of β can be increased.

[0153] As a comprehensive evaluation metric, PDI's advantage lies in its ability to consolidate multiple factors reflecting system performance degradation into a single, quantitative value, providing operators with an intuitive, unified evaluation standard. By comparing PDI with pre-set thresholds, operators can quickly assess the system's performance status, determine whether maintenance measures are necessary, and to what extent. This helps improve the targeted and effective nature of equipment maintenance, rationally allocate maintenance resources, and avoid over- or under-maintenance, thereby ensuring the safe, stable, and efficient operation of gas turbines.

[0154] Step S3340: classify the health level of the current gas turbine secondary air system performance according to the performance degradation index PDI; and provide maintenance suggestions based on the health level.

[0155] PDI<θ1: healthy state, where θ1 is the health threshold of the secondary air system;

[0156] θ1≤PDI<θ2: Warning state, where θ2 is the warning threshold of the secondary air system;

[0157] PDI≥θ2: dangerous state.

[0158] Specifically, when the system is in a healthy state, performance is stable and no additional intervention is required. When in a warning state, it is recommended to increase monitoring frequency and closely monitor changes in system performance. When in a critical state, immediate shutdown and maintenance are required to prevent serious system failure. The secondary air system healthy threshold θ1 and secondary air system warning threshold θ2 are set by those skilled in the art based on experience. Preferably, θ1 = 0.3 and θ2 = 0.7.

[0159] The PDI-based health classification method can help maintenance personnel perform equipment maintenance according to regular maintenance plans, rationally allocate maintenance resources, and reduce unnecessary maintenance costs. For example, when a gas turbine is in a healthy state, the interval between equipment inspections can be appropriately extended to reduce maintenance workload while ensuring normal equipment operation. For systems in a warning state, increasing the monitoring frequency can promptly detect subtle changes in system performance, allowing measures to be taken before problems escalate. By more frequently monitoring key parameters such as compressor inlet total temperature and outlet pressure, maintenance personnel can more accurately determine whether the system is developing into a critical state. This allows for proactive maintenance plans, such as adjusting operating parameters and scheduling preventive maintenance, to avoid system failures and reduce the economic losses caused by downtime. For systems in a critical state, immediate shutdown and maintenance is crucial to ensure equipment safety and prevent major accidents. In such cases, timely and comprehensive equipment inspection and repair, replacement of damaged components, and adjustment of system parameters can restore the system to normal operation and ensure the reliability and safety of the gas turbine.

[0160] Through clear health level classification and targeted maintenance recommendations, this method can help operation and maintenance personnel respond to system performance changes in a timely and accurate manner, effectively improve the reliability and operating efficiency of the gas turbine secondary air system, reduce equipment failure rate, extend equipment service life, and provide strong guarantees for the stable operation of the gas turbine.

[0161] Step S3350: Generate a health assessment report, which includes the residual time series The trend slope k', the cumulative value S of the predicted residual value at the next Y time points, the performance degradation index PDI and the health level of the current gas turbine secondary air system.

[0162] Specifically, the residual time series This records the time-varying differences between predicted system performance at Y future time points and the theoretical values ​​in a healthy state. By analyzing this series, operations and maintenance personnel can visually observe the degree of deviation in system performance at each time point, as well as the changing trend of this deviation. For example, if the values ​​in the residual time series gradually increase, it indicates that system performance is continuously deteriorating; if the values ​​fluctuate, it may indicate system instability, requiring further analysis.

[0163] The trend slope k' represents the rate of change of the residual over time, reflecting the rate of system performance degradation. As mentioned earlier, a positive k' indicates an increasing residual and a downward trend in system performance; its value reflects the speed of performance degradation. By analyzing the trend slope, operations and maintenance personnel can determine the urgency of system performance changes. For example, a large k' value indicates a rapid system performance degradation, necessitating prompt intervention to prevent system failure.

[0164] The cumulative value S combines the predicted residual values ​​at Y future time points, measuring the total amount of system performance degradation. A larger S value indicates that system performance has deviated significantly from a healthy state during this period, reflecting long-term system performance degradation. For example, when S exceeds a certain threshold, even if the system is still able to operate, the risk of long-term equipment failure is high, and comprehensive inspection and maintenance are required.

[0165] The Performance Degradation Index (PDI) is a quantitative indicator derived from the combined trend slope and cumulative value. It comprehensively reflects the degree of system performance degradation. Using the PDI, operations and maintenance personnel can quickly determine the system's health level and, consequently, decide on the appropriate maintenance strategy. When the PDI is within the healthy range, routine maintenance can be maintained; when in a warning state, monitoring frequency should be increased; and when in a critical state, the system should be immediately shut down for maintenance.

[0166] The health level of a gas turbine's secondary air system is determined by comparing the PDI with a set threshold, providing operators with an intuitive indicator of system performance. The health level clearly informs operators of the system's current condition and indicates whether and what action is needed.

[0167] The health assessment report integrates this information, providing a comprehensive and systematic reference for gas turbine operation and maintenance management. Based on the report content, operation and maintenance personnel can develop targeted maintenance plans, rationally allocate maintenance resources, promptly identify potential problems, and take appropriate measures to ensure the safe and stable operation of the gas turbine, improve equipment reliability and service life, reduce operating costs, and mitigate the risk of production interruptions due to equipment failure. For example, when formulating a monthly maintenance plan, the operation and maintenance team can use the health assessment report to arrange different levels of maintenance work for gas turbines in different health levels. They can conduct routine inspections for equipment in healthy states, add key monitoring items for equipment in warning states, and prioritize comprehensive overhauls for equipment in dangerous states, thereby achieving efficient and scientific equipment management.

[0168] In addition, the parts of the above technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0169] The above-described specific embodiments further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above description is merely a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A Transformer-based gas turbine secondary air system performance prediction method, characterized in that: The method comprises: Obtaining historical operating data containing timestamps of a healthy gas turbine secondary air system under steady-state conditions and preprocessing the historical operating data; extracting key feature parameters from the preprocessed historical operating data, and time-sequencing the key feature parameters according to the timestamps to obtain a key feature sequence; Based on the key feature sequence, the first Transformer model is constructed and trained; the real-time operating data sequence containing timestamps of the current gas turbine secondary air system is obtained, and the turbine outer cylinder cavity pressure prediction value time series is obtained based on the real-time operating data sequence and the trained first Transformer model. and compressor exhaust pressure prediction time series ; Obtain the time series of the actual measured value of the turbine outer cylinder pressure of the current gas turbine and the time series of measured values ​​of compressor exhaust pressure ,according to 、 、 and , get the current residual time series of performance ratio; where, is the measured value of the mth turbine outer cylinder pressure, is the measured value of the exhaust pressure of the mth compressor; based on the current residual time series and the pre-built second Transformer model, the residual time series of the next Y time points is predicted ;according to , quantitatively evaluate the performance degradation of the current gas turbine secondary air system and generate a health assessment report.

2. The Transformer-based gas turbine secondary air system performance prediction method according to claim 1, characterized in that: The key characteristic parameters include the compressor inlet total temperature T in , compressor inlet wall static pressure P in , compressor outlet pressure P out , compressor outlet temperature T out and turbine cooling valve opening V cool ; The key feature sequence is obtained as follows: The total temperature T at the compressor inlet is calculated based on the timestamp. in Perform time series processing to obtain the compressor inlet total temperature time series T in-seq ; Compressor inlet wall static pressure P according to timestamp in Perform time series processing to obtain the compressor inlet wall static pressure time series P in-seq ; Compressor outlet pressure P according to timestamp out Perform time series processing to obtain the compressor outlet pressure time series P out-seq ; Compressor outlet temperature T according to timestamp out Perform time series processing to obtain the compressor outlet temperature time series T out-seq ; According to the timestamp, the turbine cooling valve opening V cool Perform time series processing to obtain the turbine cooling valve opening time series V cool-seq ; T in-seq 、P in-seq 、P out-seq 、T out-seq and V cool-seq Align by timestamp and build key feature sequence ,in For TS seq The kth element in , n is TS seq The total number of elements in , 1≤k≤n.

3. The Transformer-based gas turbine secondary air system performance prediction method according to claim 2, characterized in that: The TS seq Each element in is a sextuple: ； in, For TS seq The timestamp of the kth element in , Timestamp The total temperature at the compressor inlet is Timestamp The static pressure on the compressor inlet wall under Timestamp The compressor outlet pressure under Timestamp The compressor outlet temperature under Timestamp Turbine cooling valve opening below.

4. The Transformer-based gas turbine secondary air system performance prediction method according to claim 3, characterized in that: The first Transformer model includes an input embedding layer, a position encoding layer, a self-attention mechanism, a multi-head attention mechanism, a feedforward neural network, and an output layer; The constructing and training of the first Transformer model includes: From the key feature sequence Each element of Extract the model input feature vector ; Input the model into the feature vector Input to the input embedding layer for encoding to obtain the timestamp Embedded feature representation under ; Get timestamps through the position encoding layer The position encoding vector P under k ; The position encoding vector P k Add to embedding feature representation , get the timestamp The model input feature representation is ; According to the model input feature representation Get the input sequence .

5. The Transformer-based gas turbine secondary air system performance prediction method according to claim 4, characterized in that: The constructing and training the first Transformer model further includes: Through the self-attention mechanism in the first Transformer model, the input sequence is calculated The correlation between the model input feature representations corresponding to different timestamps in is used to obtain a one-dimensional feature correlation matrix; The multi-head attention mechanism in the first Transformer model uses n2 attention heads to parallelly calculate the model input feature representation The relationship between each input sub-feature in is used to obtain the multi-head attention weight matrix; The single-dimensional feature correlation matrix and the multi-head attention weight matrix are nonlinearly transformed through a feedforward neural network, and the output layer maps the output of the feedforward neural network to the turbine outer cylinder pressure and the compressor exhaust pressure.

6. The Transformer-based gas turbine secondary air system performance prediction method according to claim 4, characterized in that: The position encoding vector , where d represents the dimension of the position encoding vector, Indicates the position encoding vector in The values ​​in the dimensions, is the dimension index in the positional encoding vector, 0≤ ≤ ; The values ​​of even dimensions are calculated by the sine function, and the values ​​of odd dimensions are calculated by the cosine function.

7. The Transformer-based gas turbine secondary air system performance prediction method according to claim 5, characterized in that: The model input feature representation Contains five input sub-features, namely timestamp Total inlet temperature characteristics under , timestamp The static pressure characteristics of the inlet wall under , timestamp Outlet pressure characteristics under , timestamp Outlet temperature characteristics under and timestamp Turbine cooling valve opening characteristics under .

8. The Transformer-based gas turbine secondary air system performance prediction method according to claim 1, characterized in that: The time series of the predicted value of the turbine outer cylinder pressure is obtained based on the real-time running data sequence and the trained first Transformer model. and compressor exhaust pressure prediction time series include: Extracting real-time compressor inlet total temperature from real-time operating data series , real-time compressor inlet wall static pressure , Real-time compressor outlet pressure , Real-time compressor outlet temperature and real-time turbine cooling valve opening ; Respectively 、 、 、 and Perform time series processing to obtain the real-time compressor inlet total temperature sequence , Real-time compressor inlet wall static pressure sequence , Real-time compressor outlet pressure series , Real-time compressor outlet temperature series and real-time turbine cooling valve opening sequence ; according to 、 、 、 、 And the first Transformer model, get the time series of the turbine outer cylinder pressure prediction value and compressor exhaust pressure prediction time series .

9. The Transformer-based gas turbine secondary air system performance prediction method according to claim 8, characterized in that: The basis 、 、 、 、 And the first Transformer model, get the time series of the turbine outer cylinder pressure prediction value and compressor exhaust pressure prediction time series include: Will 、 、 、 and Build a real-time key feature vector set based on timestamp alignment ,in is the real-time key feature vector at the jth time point, 1≤j≤m; ,in, is the real-time compressor inlet total temperature at the jth time point, is the real-time static pressure of the compressor inlet wall at the jth time point, is the real-time compressor outlet pressure at the jth time point, is the real-time compressor outlet temperature at the jth time point, is the real-time turbine cooling valve opening at the jth time point; Will Input the trained first Transformer model to obtain the predicted value of the turbine outer cylinder pressure at the jth time point and the predicted value of the compressor exhaust pressure at the jth time point ; according to Get the time series of the predicted value of the turbine outer cylinder pressure ,according to Get the time series of compressor exhaust pressure prediction value .

10. The Transformer-based gas turbine secondary air system performance prediction method according to claim 7, characterized in that: The n2 attention heads include at least head 1 head1 = ( 、 ), head 2head2=( 、 ) and head3head3=( 、 ).

Images