Gas turbine compressor outlet pressure on-line monitoring method and system

Abstract

The present application relates to the technical field of gas turbine performance monitoring, and discloses a gas turbine compressor outlet pressure online monitoring method and system, which comprises the following steps: obtaining an offline health data set to construct an offline benchmark model, dividing a load level to establish an offline submodel, and then generating an online submodel through incremental learning, and combining online and offline prediction sequences to evaluate performance degradation; through the synergistic effect of the offline benchmark model and the online submodel, the present application solves the problems of inaccurate outlet pressure prediction and performance degradation monitoring lag or misjudgment of the existing model under multiple load conditions, and realizes high-precision dynamic monitoring of the gas turbine compressor outlet pressure and timely early warning of performance degradation.

Description

Technical Field

[0001] This invention relates to the field of gas turbine performance monitoring technology, and more specifically, to a method and system for online monitoring of gas turbine compressor outlet pressure. Background Technology

[0002] As complex power equipment, the accurate monitoring of the compressor outlet pressure of gas turbines is crucial for equipment health management. While the gas turbine health status prediction method proposed in Chinese patent application CN109520740A collects data from multiple sensors and constructs a health analysis model, it lacks refined modeling for the differences in aerodynamic characteristics under different load conditions (e.g., low load, partial load, full load). Because the mapping relationship between compressor efficiency, pressure ratio, and outlet pressure differs significantly under different loads (e.g., IGV angle adjustment is more sensitive to pressure changes at low loads, while pressure ratio dominates efficiency changes at full loads), a single model cannot guarantee prediction accuracy across the entire load range. This may lead to overestimation of pressure under low load conditions due to overfitting to high load data, or underestimation of pressure under full load conditions due to underfitting to low load characteristics, ultimately resulting in a large error in outlet pressure prediction. Furthermore, this method lacks a dynamic model update mechanism, making it unable to adapt to the slow performance degradation caused by factors such as fouling and blade wear during gas turbine operation, resulting in monitoring results lagging behind the actual degradation process.

[0003] Chinese patent application CN117588312A proposes a gas turbine fault prediction and diagnosis method based on deep transfer learning. While it improves the fault diagnosis capability of new units through data transfer between the source and target domains, it mainly focuses on fault category identification and does not address high-precision prediction of continuous parameters such as outlet pressure. Its model adjustment process relies on fixed batches of target domain data, lacking an online incremental learning mechanism and failing to dynamically update model parameters using the latest operating data in real time. When the performance of the gas turbine compressor gradually declines due to long-term operation, the traditional transfer learning model cannot capture the residual change trend in a timely manner. This may lead to missed early decline signals due to delayed model parameter updates, or misjudging normal operating fluctuations as performance anomalies, resulting in insufficient reliability of the early warning logic.

[0004] Existing technologies cannot construct refined prediction models for the specific characteristics of gas turbines under multiple load conditions, and lack online dynamic learning mechanisms to track the performance degradation process in real time. This results in a significant increase in the compressor outlet pressure prediction error as the load changes, and there is a risk of lag or misjudgment in performance degradation monitoring. Summary of the Invention

[0005] To overcome the aforementioned shortcomings of existing technologies, this invention provides a method and system for online monitoring of gas turbine compressor outlet pressure. By acquiring offline health datasets to construct a baseline model and dividing the data into load levels to establish offline sub-models, and combining incremental learning to generate online sub-models, high-precision prediction of gas turbine compressor outlet pressure under multiple load conditions is achieved. Through residual analysis of online and offline prediction sequences, performance degradation signals can be captured and warned in a timely manner, significantly improving the timeliness and reliability of gas turbine operation and maintenance, and reducing the risk of unplanned downtime.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Online monitoring methods for gas turbine compressor outlet pressure include:

[0008] Obtain the offline health dataset when the gas turbine starts running, and construct an offline benchmark model for outlet pressure prediction based on the offline health dataset; divide the load of the gas turbine into n1 load levels, and establish an offline sub-model for outlet pressure prediction for each load level based on the offline benchmark model for outlet pressure prediction.

[0009] Based on the offline sub-model for predicting outlet pressure corresponding to each load level, an online sub-model for predicting outlet pressure corresponding to each load level is established through incremental learning.

[0010] Based on the online sub-model for export pressure prediction, the online export pressure prediction sequence for the current operating condition is obtained; based on the offline sub-model for export pressure prediction, the offline export pressure prediction sequence for the current operating condition is obtained; based on the online and offline export pressure prediction sequences, the performance degradation of the gas turbine is assessed and an early warning is issued.

[0011] Furthermore, the offline health dataset includes first health condition data directly measured under different loads and second health condition data calculated; the first health condition data includes at least the output power timing; and the second health condition data includes the compressor efficiency timing.

[0012] Furthermore, the method for calculating the compressor efficiency time series includes: constructing a compressor efficiency mathematical model, obtaining the compressor efficiency based on the first healthy operating condition data and the constructed compressor efficiency mathematical model; and sorting the compressor efficiency by timestamp to obtain the compressor efficiency time series.

[0013] Furthermore, the offline benchmark model for export pressure prediction adopts a two-layer LSTM neural network structure, including an input layer, a hidden layer, and a fully connected layer; the hidden layer contains two LSTM layers.

[0014] Furthermore, the method for dividing the load of the gas turbine into n1 load levels includes:

[0015] Extract the maximum output power P from the output power time series of the offline health dataset within a given monitoring time window. max The maximum output power P max With the rated full-load power P of the gas turbine rated Compare the data to determine the load level.

[0016] Furthermore, the method for establishing offline sub-models for outlet pressure forecasting corresponding to each load level based on the offline benchmark model for outlet pressure forecasting includes:

[0017] Filter offline data subsets for each load level from the offline health dataset;

[0018] Based on the offline data subsets of each load level, initialize the offline sub-model for predicting outlet pressure corresponding to each load level;

[0019] The offline benchmark model for outlet pressure prediction was used to pre-train the offline sub-models for outlet pressure prediction corresponding to each load level, and the initial parameters of the offline sub-models for outlet pressure prediction corresponding to each load level were obtained.

[0020] By using a subset of offline data corresponding to each load level, the offline sub-models for outlet pressure prediction with initial parameters are trained in a specialized manner to obtain the final parameters of the offline sub-models for outlet pressure prediction corresponding to each load level.

[0021] Furthermore, the method for establishing an online sub-model for predicting outlet pressure corresponding to each load level includes:

[0022] Based on the final parameters of the offline sub-model for forecasting outlet pressure corresponding to each load level, an online sub-model for forecasting outlet pressure corresponding to each load level is constructed.

[0023] Real-time operating condition data of the gas turbine at each load level is acquired. Using the real-time operating condition data of each load level, the parameters of the online sub-model for predicting the outlet pressure at each load level are updated in real time through incremental learning, so as to obtain the final n1 online sub-models for predicting the outlet pressure at each load level.

[0024] Furthermore, the method for obtaining the online prediction sequence of the outlet pressure under the current operating condition includes:

[0025] Obtain the current operating data of the gas turbine, and determine the current load level based on the current operating data;

[0026] Based on the current operating data and the compressor efficiency mathematical model, the current compressor efficiency time series is obtained;

[0027] Based on the current load level determined, the corresponding online sub-model for outlet pressure prediction is invoked. According to the current operating data, the current compressor efficiency time series, and the invoked corresponding online sub-model for outlet pressure prediction, the online prediction sequence for outlet pressure is obtained.

[0028] Furthermore, the method for assessing and providing early warning of gas turbine performance degradation based on the online and offline export pressure prediction sequences includes:

[0029] Based on the online and offline export pressure prediction sequences, the export pressure residual sequence is obtained; based on the export pressure residual sequence, it is determined whether a compressor performance degradation warning is triggered.

[0030] A gas turbine compressor outlet pressure online monitoring system, used to implement the above-mentioned gas turbine compressor outlet pressure online monitoring method, the system comprising:

[0031] Offline benchmark model construction module: used to obtain the offline health dataset when the gas turbine is first put into operation, and to build an offline benchmark model for outlet pressure prediction based on the offline health dataset; the load of the gas turbine is divided into n1 load levels, and an offline sub-model for outlet pressure prediction corresponding to each load level is established based on the offline benchmark model for outlet pressure prediction.

[0032] Online Prediction Model Construction Module: Based on the offline sub-model for predicting outlet pressure corresponding to each load level, an online sub-model for predicting outlet pressure corresponding to each load level is established through incremental learning;

[0033] Performance degradation assessment module: Based on the online sub-model for export pressure prediction, the online export pressure prediction sequence for the current operating condition is obtained; based on the offline sub-model for export pressure prediction, the offline export pressure prediction sequence for the current operating condition is obtained; based on the online and offline export pressure prediction sequences, the gas turbine performance degradation is assessed and an early warning is issued.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] This invention achieves accurate prediction and real-time monitoring of gas turbine compressor outlet pressure by constructing an offline baseline model and offline sub-models for different load levels, and by using incremental learning to establish an online sub-model. First, during the offline phase, the data characteristics of the gas turbine during healthy operation are fully explored to construct a baseline model and sub-models that reflect different load conditions, providing a reliable reference for online monitoring. Second, the online sub-model is continuously updated based on the offline sub-model through incremental learning, dynamically adapting to the gradual degradation of gas turbine performance and ensuring the timeliness and accuracy of the prediction results. Finally, by comparing the online and offline prediction sequences, a residual sequence is obtained, and performance degradation assessment and early warning are performed accordingly. This effectively solves the problems of large outlet pressure prediction errors, delayed or misjudged performance degradation monitoring caused by the inability to adapt to multiple load conditions and insufficient dynamic tracking in existing technologies. This significantly improves the efficiency and reliability of gas turbine operation and maintenance, and reduces the risk and maintenance costs of unplanned shutdowns. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart illustrating the principle of the online monitoring method for gas turbine compressor outlet pressure in this invention.

[0038] Figure 2 This is a flowchart illustrating the principle of establishing an offline sub-model for predicting outlet pressure at each load level in the online monitoring method for gas turbine compressor outlet pressure of the present invention.

[0039] Figure 3 This is a flowchart illustrating the principle of the gas turbine performance degradation assessment and early warning method in the online monitoring method for gas turbine compressor outlet pressure of the present invention.

[0040] Figure 4 This is a functional block diagram of the online monitoring system for gas turbine compressor outlet pressure in this invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Example 1:

[0043] Please see Figure 1 As shown, this embodiment provides a method for online monitoring of the outlet pressure of a gas turbine compressor, including:

[0044] Step S1000: Obtain the offline health dataset when the gas turbine starts running; construct an offline benchmark model for outlet pressure prediction based on the offline health dataset; divide the load of the gas turbine into n1 load levels; establish an offline sub-model for outlet pressure prediction for each load level based on the offline benchmark model for outlet pressure prediction; each load level corresponds to an offline sub-model for outlet pressure prediction.

[0045] Please see Figure 2 As shown, step S1000 further includes:

[0046] Step S1100: Obtain first health condition data directly measured under different loads and second health condition data calculated when the gas turbine starts to run.

[0047] Further, step S1100 includes:

[0048] Step S1110: Obtain first health condition data directly measured under different loads when the gas turbine starts running; the first health condition data includes compressor inlet temperature timing, compressor outlet temperature timing, compressor inlet pressure timing, compressor outlet pressure timing, IGV angle timing, VGV angle timing, output power timing, turbine cooling air valve position display timing, compressor running time, and water washing status indicator timing; the turbine cooling air valve position display includes level 2, level 3, and level 4 display, and the water washing status indicator is a binary variable, where 0 indicates no water washing and 1 indicates water washing;

[0049] Specifically, in the early stages of gas turbine manufacturing, i.e., when there is no performance degradation and the component clearances are normal, various sensors and systems are used to collect real-time operating data under multiple load conditions (including low load, partial load, and base load) and multiple environmental conditions (including temperature, humidity, and atmospheric pressure).

[0050] Compressor inlet temperature: This data is collected by a temperature sensor installed at the compressor inlet. It is expressed as a time-series representation of temperature changes, reflecting the impact of the intake conditions on the compression process. It is one of the fundamental parameters for calculating compressor efficiency and analyzing compression power consumption. For example, under different loads, the inlet temperature affects the gas density and initial energy state, thus altering the energy demand and outlet pressure during compression.

[0051] Compressor outlet temperature: Collected by a temperature sensor installed at the compressor outlet, synchronously with the inlet temperature. It is also presented as a time-series temperature value, reflecting actual compression power consumption and heat loss. During actual compression, the degree of temperature rise is directly related to power consumption; combined with the inlet temperature, the energy conversion efficiency of the compression process can be calculated.

[0052] Compressor inlet and outlet pressures: These are measured using high-precision pressure transmitters and recorded as time-series pressure values. These two parameters are used to calculate the pressure ratio, a key indicator of compressor compression capacity that directly affects compressor efficiency. For example, a higher pressure ratio indicates a higher degree of gas compression by the compressor, reflecting its performance to some extent.

[0053] IGV and VGV angle timing: The opening of the inlet guide vanes (IGV) and compressor adjustable stationary vanes (VGV) is monitored in real time by angle encoders and expressed as a time sequence of angle values. These are used to quantify the impact of airflow regulation on pressure. Changes in the opening of the IGV and VGV alter the direction and flow rate of the airflow entering the compressor, thus affecting the compressor's efficiency and outlet pressure. For example, at low loads, appropriately adjusting the IGV angle can increase the intake air volume, maintaining stable compressor operation and stable outlet pressure.

[0054] Output power: Real-time electrical power is obtained from the unit control system and expressed as a power value time series, reflecting the load status. Changes in load directly lead to changes in the compressor's operating conditions. The output power time series can be used to classify load levels, providing a basis for subsequent model construction based on load levels.

[0055] Turbine cooling air valve position display sequence: Normalized values ​​(0-1) are obtained through valve opening sensors (such as potentiometers) to characterize the coupled effect of cooling air flow on compressor load. Stages 2, 3, and 4 of the turbine cooling air valve represent different cooling intensities; only one stage is active at a time, and the valve position display value reflects the opening status of the currently active stage. For example, a fully open stage 2 valve corresponds to 0.3, a fully open stage 3 valve corresponds to 0.6, and a fully open stage 4 valve corresponds to 1.0. An example data structure, such as collecting data once per hour, might show a valve position display sequence of [0.3, 0.3, 0.6, 0.6, 0.6, 1.0, 1.0, 1.0, ...]. This data reflects the real-time opening of the cooling valves at different stages of the turbine. Cooling air flow affects turbine efficiency, which in turn impacts compressor load, making it one of the key parameters affecting compressor outlet pressure.

[0056] Compressor runtime: Based on the cumulative operating hours recorded in the system operation log, it is expressed as a time value and used to analyze the time dependence of performance degradation. As the operating time increases, compressor components may experience wear, fouling, and other problems, leading to performance degradation. Runtime is an important time-dimensional parameter for assessing performance changes.

[0057] Water washing status indicator sequence: Recorded in binary variable form (0 indicates no water washing, 1 indicates water washing), marking the effect of maintenance operations on compressor efficiency recovery. Water washing can remove scale buildup inside the compressor and restore its efficiency. This indicator is used in subsequent LSTM models to distinguish compressor performance data under different maintenance states, enabling the model to more accurately reflect the impact of maintenance operations on outlet pressure.

[0058] Existing technologies suffer from incomplete initial operating data collection for gas turbines, focusing only on some key parameters while neglecting parameters that significantly impact compressor outlet pressure, such as turbine cooling air valve position displays and water washing status indicators. This results in LSTM models that cannot fully and accurately reflect the actual operating state of the compressor. This new step, by comprehensively collecting various direct measurement data, solves the data gap problem, enabling the model to consider more influencing factors and improving its accuracy and reliability.

[0059] Step S1120: Obtain second health condition data including different load conditions when the gas turbine starts to run;

[0060] The second health condition data includes compressor efficiency timing;

[0061] The calculation method for compressor efficiency time series includes: constructing a compressor efficiency mathematical model; obtaining the compressor efficiency based on the compressor inlet temperature, compressor outlet temperature, compressor inlet pressure, compressor outlet pressure, and the constructed compressor efficiency mathematical model in the first health condition data; and sorting the compressor efficiency by timestamp to obtain the compressor efficiency time series.

[0062] The second health condition data is not directly obtainable and needs to be calculated using formulas.

[0063] Specifically, compressor efficiency is a key indicator of compressor performance, representing the degree to which the compressor effectively converts mechanical energy into gas pressure energy and kinetic energy. Its calculation is based on the first law of thermodynamics and the energy conversion principle of the gas compression process. The specific calculation process is as follows:

[0064] Considering the ideal adiabatic compression process of gas in a compressor, the power consumption under ideal conditions is expressed by the formula for ideal adiabatic compression work. This formula involves the outlet temperature after ideal adiabatic compression, the compressor inlet temperature, the compressor outlet pressure, the compressor inlet pressure, and the specific heat ratio of the gas, specifically:

[0065]

[0066] Among them, T 2,is The outlet temperature after ideal adiabatic compression is T1, the compressor inlet temperature is P2, the compressor outlet pressure is P1, and the compressor inlet pressure is P2. γ is the specific heat ratio of the gas, which is dimensionless and can be obtained by looking up a table based on the gas composition and operating temperature range, or by measuring it with a gas composition analyzer. It is approximately 1.4 at normal temperature and pressure.

[0067] Calculate the specific work (ideal work) W of an ideal adiabatic compression process. is Ideal work is related to the isobaric specific heat capacity of the gas, the inlet temperature, and the pressure ratio, specifically:

[0068]

[0069] Among them, C p This is the specific heat capacity of the gas at constant pressure.

[0070] Calculate the power consumption (actual work) W during the actual compression process. a :

[0071] In actual compression, due to factors such as friction, turbulence, and heat loss, the actual power consumption is greater than the ideal power consumption. The actual power consumption is calculated by the difference between the compressor outlet temperature and the inlet temperature after actual compression. Specifically:

[0072] W a =C p ×(T2-T1);

[0073] Where T2 is the actual compressor outlet temperature after compression. T2-T1 reflects the energy conversion and heat loss during the compression process.

[0074] The definition of compressor efficiency: Compressor efficiency η can be expressed as the ratio of ideal work to actual work:

[0075]

[0076] After simplification and rearrangement, the mathematical model for compressor efficiency is obtained:

[0077]

[0078] This determines the sensitivity to efficiency. As the pressure ratio increases, As the pressure ratio increases, η also increases, quantifying the contribution of the pressure ratio to the compressor efficiency. An increase in the compressor outlet temperature T2 leads to an increase in the actual temperature rise (T2-T1), which in turn reduces the compressor efficiency. This reflects the heat dissipation effect caused by irreversible losses such as friction and leakage. An abnormal increase in outlet temperature can provide an early warning of blade corrosion, increased clearance, or fouling failure.

[0079] The aforementioned compressor efficiency mathematical model is directly derived from thermodynamic principles, accurately reflecting the compressor's efficiency in converting mechanical energy into gas pressure and kinetic energy, providing a reliable basis for evaluating compressor performance. By monitoring the trend of compressor efficiency changes, performance degradation can be detected in a timely manner. A decrease in efficiency may indicate problems such as fouling, blade damage, or seal leakage inside the compressor. When the compressor tip clearance increases, gas leakage increases, leading to an increase in actual compression work and a decrease in compressor efficiency. Therefore, compressor efficiency is a direct physical mapping of clearance changes, and the two are negatively correlated. The compressor efficiency mathematical model calculates compressor efficiency using measurable temperatures (T1, T2) and pressures (P1, P2), eliminating the need for direct clearance measurement and solving the problem of "unmeasurable clearance." Through the dynamic calculation of compressor efficiency η, the impact of tip clearance changes on aerodynamic performance is transformed into an observable quantitative indicator. For example, when the clearance increases, leakage increases, leading to an increase in actual temperature rise (T2-T1), a significant decrease in compressor efficiency, a reduction in pressure ratio, and further deterioration of compressor efficiency. By using η as the input feature of the LSTM model, the neural network can learn the nonlinear relationship between gap change and outlet pressure, overcoming the limitation of traditional data-driven models that cannot model structural parameters.

[0080] In existing technologies, compressor efficiency is often difficult to measure directly, requiring complex measurement equipment or indirect estimation, resulting in insufficient accuracy and real-time performance. Step S1120 solves the problem of obtaining compressor efficiency by constructing a mathematical model of compressor efficiency and calculating it using collected direct measurement data, providing a reliable quantitative indicator for evaluating compressor performance. The compressor efficiency time series, as an important input feature, is used for subsequent model learning of the nonlinear relationship between gap changes and outlet pressure. This overcomes the limitation of traditional data-driven models in modeling structural parameters and serves as a crucial bridge connecting structural parameter changes and outlet pressure prediction. Without this data, the model cannot reflect the impact of structural parameters such as gap changes on outlet pressure, leading to a significant decrease in the model's predictive power and accuracy.

[0081] Step S1130: The first health condition data and the second health condition data form an offline health dataset for constructing an offline benchmark model for predicting export pressure.

[0082] Specifically, the first and second health condition data are integrated chronologically to form a dataset containing multiple time-series features. Since the different data have different dimensions and value ranges, a normalization method (such as min-max normalization or Z-score normalization) is used to process the data and scale each data feature to a suitable range (e.g., 0-1) to eliminate the influence of dimensions and improve the convergence speed and model training effect of the subsequent LSTM network. In existing technologies, multi-source data often have different dimensions; directly inputting them into the model may lead to difficulties in model training, slow convergence speed, and low prediction accuracy. This data integration and normalization process solves the problem of inconsistent data dimensions, making data from different dimensions comparable and improving the model's training efficiency and prediction accuracy. The normalized data forms the basis for building an offline benchmark model for export pressure prediction, ensuring that the subsequent model can effectively learn the relationship between data features and export pressure. Without the data normalization step, the model may not converge effectively, and the mapping relationship between features and export pressure cannot be accurately established, leading to the failure of the entire model building process.

[0083] Step S1200: Based on the offline health dataset, construct an offline benchmark model for predicting export pressure;

[0084] Specifically, nine types of features—compressor inlet temperature timing, inlet pressure timing, IGV angle timing, VGV angle timing, output power timing, turbine cooling air valve position display timing, compressor running time, water wash status indicator timing, and compressor efficiency timing—are sampled using a sliding window method within a set time window (e.g., 1 hour). For example, if the time window length is 60 minutes (sampling once per minute, i.e., 60 time steps), each sample consists of 9-dimensional features from 60 time steps, forming a three-dimensional tensor of [number of samples, window length, number of features]. The corresponding output is the compressor outlet pressure at the last time step of each window.

[0085] The offline benchmark model for export pressure prediction employs a two-layer LSTM neural network structure, specifically composed of:

[0086] Input layer: Receives three-dimensional tensors. The window length of the input layer corresponds to the window length of the sliding window sampling. For example, the window length is set to 60, which corresponds to 1 hour of time series data, enabling the model to capture short-term time dependencies (such as the immediate impact of load fluctuations on outlet pressure).

[0087] Hidden Layer: Contains two LSTM layers, each with 64 neurons. The LSTM layers (containing weight matrices and bias vectors) effectively capture long-term time dependencies through forget gates, input gates, and output gates, such as the impact of the slow accumulation of compressor fouling on efficiency over operating time. The weight matrix and bias vector of the LSTM layer are used to calculate the output of each state gate. The weight matrix learns the nonlinear relationships between features, and the bias vector adjusts the neuron activation threshold.

[0088] Fully connected layer: The output of the LSTM layer is mapped to the predicted value of the compressor outlet pressure. A linear activation function is used, which is suitable for continuous value prediction.

[0089] During training, mean squared error (MSE) is used as the loss function to measure the deviation between the predicted value and the actual export pressure. The Adam optimizer is used to automatically adjust the learning rate to improve training stability. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for updating model parameters, the validation set is used to prevent overfitting, and the test set is used to evaluate the model's generalization ability.

[0090] In step S1410, the hidden layer parameters (such as the weight matrix) of the offline baseline model are transferred to the offline sub-models for each load level, realizing cross-domain learning of "general laws in the source domain (full load) → personalized features in the target domain (specific load)". For example, the law learned under full load that "for every 1% decrease in compressor efficiency, the outlet pressure decreases by 0.5 kPa" can be used as the initial knowledge for the low-load offline sub-model. After fine-tuning with low-load data, it can adapt to the characteristic that "efficiency decrease is more sensitive to the impact on outlet pressure under low load".

[0091] In existing technologies, the outlet pressure of a gas turbine compressor is affected by multiple coupled factors (such as load, temperature, and maintenance status), making it difficult for traditional models to capture the complex dependencies in long-term time-series data. Step S1200 uses nine types of features, including directly measured temperature, pressure, valve opening, and calculated compressor efficiency, as inputs, covering the physical processes of compressor operation (such as compression power consumption, airflow regulation, and maintenance intervention). Compared to single-feature modeling, this improves the model's ability to represent complex operating conditions. For example, the timing of the turbine cooling air valve position display reflects the impact of cooling air flow on the compressor load. Combined with the timing of compressor efficiency, the energy coupling relationship between the cooling system and the compression process can be quantified. The LSTM network, through memory cells and gating mechanisms, can distinguish between short-term load fluctuations (such as instantaneous changes in output power) and long-term performance degradation (such as efficiency decline due to fouling), avoiding the gradient vanishing problem of traditional recurrent neural networks (RNNs). For example, using compressor running time as a timing feature input, the LSTM can learn the rule that "the longer the running time, the more significant the efficiency decline and the greater the deviation of the outlet pressure from the baseline value." The offline baseline model is trained based on the initial health data of the gas turbine at the time of manufacture, learning the "health baseline" patterns under conditions without performance degradation, such as the positive correlation between pressure ratio and compressor efficiency, and the negative correlation between actual temperature rise and compressor efficiency. These patterns, as general knowledge across load levels, provide a foundation for parameter transfer in the offline sub-model of subsequent step S1400, reducing the cost of repeated training. Without an offline baseline model as a starting point for transfer learning, the offline sub-model of step S1400 would need to be trained from scratch, requiring a large amount of data support, and would be difficult to capture general physical patterns across loads, resulting in slow model convergence and a high risk of overfitting. It would also cause the residual calculation (online predicted value - offline baseline value) in subsequent step S3000 to lose the "health baseline" reference, making it unable to effectively distinguish between normal operating condition fluctuations and performance degradation, potentially leading to false or missed warnings.

[0092] Step S1300: Divide the load of the gas turbine into n1 load levels and filter the offline data subsets of each load level from the offline health dataset;

[0093] The method for dividing the load of the gas turbine into n1 load levels includes:

[0094] Extract the maximum output power P from the output power time series of the offline health dataset within a given monitoring time window. max The maximum output power P max With the rated full-load power P of the gas turbine rated A comparison is performed to determine the load level.

[0095] Maximum output power P max With the rated full-load power P of the gas turbine rated Methods for determining load levels through comparison include:

[0096] Low load: P max <k4×P rated ;

[0097] Partial load 1: k4×P rated ≤P max <k3×P rated ;

[0098] Partial load 2: k3×P rated ≤P max <k2×P rated ;

[0099] Partial load 3: k2×P rated ≤P max <k1×P rated ;

[0100] Full load: P max ≥k1×P rated .

[0101] Among them, k1 is the full load threshold coefficient, k2 is the high load threshold coefficient, k3 is the medium load threshold coefficient, k4 is the low load threshold coefficient, and 0 < k4 < k3 < k2 < k1 < 1; in industry, the coefficients are usually determined according to the statistical distribution of the actual operation data of gas turbines. For example:

[0102] Collect the output power data within 1 year, calculate P for each hour max , and draw a load distribution histogram;

[0103] Determine the power points with cumulative frequencies of 10%, 30%, 50%, and 70%, corresponding to k4×P rated (upper limit of low load), k3×P rated (upper limit of partial load 1), k2×P rated (upper limit of partial load 2), k1×P rated (lower limit of full load);

[0104] If the statistical results show that 30MW (30%), 50MW (50%), 70MW (70%), and 90MW (90%) are the boundary points of each interval, then k1 = 0.9, k2 = 0.7, k3 = 0.5, and k4 = 0.3.

[0105] This data-driven division method makes the load level boundaries conform to the actual operation law of the unit, is more adaptable than the fixed threshold method, ensures that each level data subset has sufficient sample size and representativeness, and provides a reliable data basis for subsequent sub-model training.

[0106] Step S1300 involves dividing the gas turbine load into n1 levels and selecting offline data subsets for each level. This addresses the problem of insufficient model generalization ability due to load variations in existing technologies. By refining the load classification, subsequent models can model for different operating conditions, improving prediction accuracy. In existing technologies, the aerodynamic characteristics of gas turbines differ significantly under different loads (e.g., compressor surge is prone to occur at low loads, while efficiency approaches design values ​​at full loads), making it difficult for a single model to maintain accuracy across the entire load range. Step S1300 divides the full load range into multiple load levels, allowing each offline sub-model to focus on feature learning within a specific load range. For example, the IGV angle is more sensitive to outlet pressure regulation at low loads, while the turbine cooling air valve position has a more significant coupling effect on the load at full loads. The offline sub-models after load classification can specifically strengthen the weights of relevant features. When selecting data subsets for each level, samples strongly correlated with that load are retained to avoid the model learning incorrect feature associations due to the mixing of data from different loads. For example, compressor efficiency is generally high in full-load data (due to proximity to design conditions). Mixing this with low-load data might mask the pattern of "efficiency reduction leading to lower outlet pressure." Compared to a single full-load model, load-level offline sub-models can be trained in parallel, reducing the number of parameters in a single model and increasing training speed. For instance, the total number of parameters in five load-level offline sub-models may be less than a single model, and the input data for each sub-model has higher operating condition homogeneity (e.g., a narrower range of IGV angle values ​​under low load), resulting in a more concentrated feature distribution and reducing the computational complexity of model training. Since the IGV angle under low-load conditions is primarily used to prevent surge, its adjustment range is limited (e.g., only varying between 20° and 40°). Compared to full-load conditions (30°-50°), this feature exhibits smaller numerical fluctuations, and the sub-model does not need to learn large-scale nonlinear relationships, thus reducing computational load. Furthermore, the load-level offline data subset only contains samples within specific power ranges, avoiding the discretization of feature distribution caused by mixing different load data, further improving computational efficiency.

[0107] If step S1300 is missing, a single model must simultaneously fit the complex characteristics of the entire load range. This may lead to overfitting of high-load data at low loads, resulting in higher predicted values, or underfitting of low-load characteristics at full loads, leading to increased errors. Performance degradation manifests differently under different loads (e.g., fouling has a more significant impact on efficiency at low loads). Ungraded models struggle to distinguish between pressure fluctuations caused by load changes and actual performance degradation, potentially delaying warnings or triggering false alarms.

[0108] Step S1400: Based on the offline data subsets of each load level and the offline baseline model for outlet pressure prediction, obtain n1 offline sub-models for outlet pressure prediction corresponding to n1 load levels.

[0109] Preferably, n1 = 5, with 5 load levels corresponding to low load, partial load 1, partial load 2, partial load 3, and full load; the n1 offline sub-models for outlet pressure prediction are respectively the offline sub-model for outlet pressure prediction corresponding to low load, the offline sub-model for outlet pressure prediction corresponding to partial load 1, the offline sub-model for outlet pressure prediction corresponding to partial load 2, the offline sub-model for outlet pressure prediction corresponding to partial load 3, and the offline sub-model for outlet pressure prediction corresponding to full load.

[0110] Step S1400 uses transfer learning and hierarchical training mechanisms to achieve refined modeling of "inheritance of general rules + fine-tuning of specific loads", solving the problem in existing technologies that "a single model cannot adapt to the characteristics of multiple load conditions".

[0111] Further, step S1400 includes:

[0112] Step S1410: Based on the offline data subsets of each load level, initialize the offline sub-model for predicting outlet pressure corresponding to each load level.

[0113] The initialization of the offline sub-model for predicting outlet pressure corresponding to each load level includes:

[0114] Some parameters from the offline baseline model for export pressure prediction are copied to the offline sub-models for export pressure prediction corresponding to each load level; these parameters include the weight matrix and bias vector located in the LSTM layer.

[0115] Specifically, after completing the load level division and filtering of offline data subsets for each level in step S1300, cross-domain transfer learning is used to initialize the corresponding offline sub-model for each load level (e.g., low load, partial load 1-3, full load). The hidden layer parameters (including the weight matrix and bias vector of the LSTM layer) of the offline benchmark model for outlet pressure prediction (source domain model, trained based on full load health data) constructed in step S1200 are copied to the sub-model of the target load level. For example, the low load sub-model inherits the weight parameters capturing the temperature-pressure-efficiency relationship from the benchmark model as initial parameters. The temporal features of the offline data subsets for each load level (e.g., compressor inlet temperature temporal series, IGV angle temporal series, etc., 9-dimensional features) are sampled using a sliding window according to a set time window (e.g., 1 hour) to generate a three-dimensional tensor [number of samples, window length, number of features], consistent with the input format of the offline benchmark model, ensuring compatibility of parameter transfer. In existing technologies, training models separately for different loads requires a large amount of data and cannot reuse general rules, resulting in high training costs and a high risk of overfitting. Step S1410 transfers the hidden layer parameters of the baseline model, enabling the sub-model to inherit general physical laws learned by the source model, such as the "nonlinear relationship between pressure ratio and efficiency" and the "influence of temperature on compression power consumption," thus reducing training instability caused by random parameter initialization during cold starts. For example, the full-load baseline model has learned the rule that "efficiency increases by 2% for every 0.1 increase in pressure ratio," and the low-load sub-model does not need to relearn this basic relationship and can be directly used for initialization, shortening the training cycle. Compared to training from scratch, the sub-model only needs a small amount of data from the target load to start training, solving the model building problem when low-load data is scarce. The transfer of hidden layer parameters ensures that the sub-model inherits the basic physical laws of gas turbine operation, avoiding "model fragmentation" caused by load grading and improving prediction consistency across load conditions.

[0116] Step S1420: Use the offline benchmark model for outlet pressure prediction to pre-train the offline sub-model for outlet pressure prediction corresponding to each load level, and obtain the initial parameters of the offline sub-model for outlet pressure prediction corresponding to each load level.

[0117] Specifically, based on the initialization, and under the supervision of the offline baseline model for predicting outlet pressure in the source domain (full load), each offline sub-model is pre-trained using a subset of offline data corresponding to the load interval. The objective function for pre-training is a hybrid loss function, including:

[0118] Source Domain Prediction Error (MSE): The mean square error between the sub-model's predicted values ​​on full-load data and the actual outlet pressure, ensuring that the sub-model does not lose the core knowledge of the baseline model.

[0119] Elastic Weight Consolidation Regularization (EWC): Constrains the difference between the parameters of the sub-model and the baseline model by calculating the Fisher information matrix of the parameters, preventing the forgetting of source domain knowledge during pre-training.

[0120] L2 regularization: suppresses parameter overfitting and improves the model's generalization ability.

[0121] If the sub-model is trained directly using the target load data, the model may deviate from the general rules due to differences in data distribution (such as more frequent IGV angle adjustments under low load), resulting in "catastrophic forgetting" (i.e., forgetting the basic physical relationships learned from the source domain). Pre-training, through a hybrid loss function, forces the sub-model to consolidate the core parameters of the source domain before adapting to the target load. For example, in the pre-training of the low-load sub-model, it is necessary to simultaneously fit the pressure ratio-efficiency relationship of the full load data to avoid its parameters being overly biased towards the "small pressure ratio-high IGV angle" characteristics unique to low loads, which would prevent the accurate calculation of the basic compression power consumption.

[0122] Pre-training allows the sub-model to inherit knowledge from the source model while initially adapting to the data distribution of the target load (e.g., a lower efficiency distribution range under low load). For example, through pre-training, the low-load sub-model adjusts parameters to adapt to the characteristic of "greater efficiency fluctuations under low pressure ratios" while maintaining the "positive correlation between efficiency and pressure ratio". EWC regularization ensures that the sub-model will not overturn the learned physical laws due to target data noise (e.g., sensor errors) by limiting the variation range of key parameters. For example, it prevents the model from misjudging "random pressure fluctuations under low loads" as "a negative correlation between pressure ratio and efficiency".

[0123] Step S1430: Use the offline data subset corresponding to the load level to perform specialized training on the offline sub-model for predicting outlet pressure for each load level with initial parameters, and obtain the final parameters of the offline sub-model for predicting outlet pressure for each load level.

[0124] Specifically, after pre-training, the sub-model is fine-tuned using a subset of offline data at the target load level. The objective function focuses on the mean squared error in the target domain, and pre-training knowledge is preserved through EWC regularization. The specific process includes:

[0125] Feature-specific learning: Adjust the model's weights on relevant features based on the characteristic distribution of the target load (e.g., turbine cooling air valve position readings are generally low under low load). For example, the low-load sub-model can strengthen the weight of the IGV angle on the outlet pressure, since the IGV opening is a key adjustment mechanism to prevent surge under low load.

[0126] Iterative optimization: Through multiple rounds of training, the sub-model learns the unique patterns under the target load, such as "when the compressor efficiency decreases by 1% at low load, the outlet pressure decreases by 0.8 kPa" (compared to 0.5 kPa at full load), thus quantifying the load-specific effects.

[0127] Offline baseline models and pre-training phases only provide general patterns and cannot capture the differences in sensitivity to features under different loads. For example, at full load, pressure ratio is the dominant factor in efficiency, while at low load, IGV angle and water washing status have a more significant impact on efficiency. Specialized training reveals these specificities through target data, avoiding errors caused by "one-size-fits-all" modeling. For example, for a certain gas turbine at low load, due to the small intake air volume, every 5° increase in IGV angle can increase the outlet pressure by 3 kPa. This pattern can only be captured by the model through specialized training. For the "small pressure ratio, large IGV angle" characteristics of low-load conditions, the offline sub-model can adjust the memory cell state of the LSTM layer to prioritize storing the real-time correlation between IGV angle and pressure, resulting in lower prediction errors compared to sub-models without specialized training.

[0128] If step S1400 is missing, load-specific modeling cannot be achieved. A single baseline model may overfit full-load data at low loads, resulting in generally higher predicted pressure values ​​and failing to reflect the true pressure decay at light loads (e.g., the actual pressure at low loads may be 10-15 kPa lower than the baseline value), leading to delayed performance degradation monitoring. Data utilization efficiency is low. When training with mixed load data, the model needs to learn complex mappings across loads, resulting in parameter redundancy (e.g., simultaneously fitting the patterns of "low-load IGV regulation" and "full-load pressure ratio dominance"), increasing training time and significantly raising the risk of overfitting. When unit load changes frequently, a single model cannot quickly adapt to the switching of operating conditions. For example, when suddenly dropping from full load to low load, the model may need several hours to reconverge, while the tiered sub-model can call the corresponding model in real time through load determination, shortening the response time to minutes.

[0129] Step S2000: Based on the offline sub-model for predicting outlet pressure corresponding to each load level, an online sub-model for predicting outlet pressure corresponding to each load level is established through incremental learning.

[0130] Further, step S2000 includes:

[0131] Step S2100: Based on the final parameters of the offline sub-model for predicting outlet pressure corresponding to each load level, construct the online sub-model for predicting outlet pressure corresponding to each load level.

[0132] Specifically, step S2100 achieves a smooth transition from the "health baseline model" to the "real-time tracking model" through a parameter transfer mechanism. The online sub-model is architecturally identical to the offline sub-model, both employing a two-layer LSTM neural network structure, including an input layer, hidden layers (2 LSTM layers, 64 neurons per layer), and a fully connected layer. The input feature dimensions are the same as the offline sub-model (9-dimensional temporal features, such as compressor inlet temperature, IGV angle, etc.), and the output is a predicted sequence of compressor outlet pressure. The core difference between the two lies in the parameter initialization method:

[0133] Offline sub-model parameters: obtained through pre-training and specialized training in step S1400, reflecting the performance law of the "ideal health state" under the corresponding load level. For example, the full load sub-model parameters learn a strong positive correlation between pressure ratio and efficiency.

[0134] Online sub-model initialization: The final parameters (including hidden layer weight matrix, bias vector, and output layer parameters) of the offline sub-model are directly inherited as the starting point for incremental learning. For example, the LSTM layer weights of the low-load online sub-model are directly copied from the parameters of the low-load offline sub-model, avoiding training instability caused by random initialization from zero.

[0135] In existing technologies, online models require a large amount of real-time data for training from scratch and cannot utilize prior knowledge from historical health data, leading to large prediction errors in the early stages of unit operation. This step reduces cold-start costs and retains health baseline knowledge through parameter transfer; a usable model can be quickly built without relying on a large amount of real-time data. For example, when low-load operating data for a gas turbine is scarce, the online sub-model, using parameters from the offline sub-model, can achieve the training effect of 1000 samples using only 100 real-time samples, similar to traditional methods. The offline sub-model parameters contain the health status patterns of the gas turbine at the initial stage of manufacturing (such as the pressure-efficiency relationship when there is no fouling and the blade clearance is normal). After inheriting these parameters, the online sub-model can compare real-time predicted values ​​with the health baseline, providing a benchmark reference for performance degradation monitoring.

[0136] The offline sub-model serves as a "knowledge source," providing initial parameters for the online sub-model. The online sub-model, acting as a "dynamic adapter," fine-tunes parameters during real-time operation to adapt to performance changes. For example, the offline sub-model establishes a baseline relationship that "for every 1% decrease in compressor efficiency, the outlet pressure decreases by 0.5 kPa." The online sub-model adjusts this coefficient to 0.6 kPa based on real-time data, reflecting the actual rate of degradation after equipment aging. After a unit is newly commissioned or undergoes a major overhaul, the offline sub-model parameters are closer to the actual healthy state. Using this as a starting point, the online sub-model can more accurately capture initial performance fluctuations and avoid misjudging them as degradation.

[0137] Step S2200: Obtain real-time operating condition data for each load level of the gas turbine. Using this data, and through incremental learning, update the parameters of the online sub-model for outlet pressure prediction corresponding to each load level in real time, resulting in n1 online sub-models for outlet pressure prediction corresponding to n1 load levels. These n1 online sub-models are: the online sub-model for outlet pressure prediction corresponding to low load, the online sub-model for outlet pressure prediction corresponding to partial load 1, the online sub-model for outlet pressure prediction corresponding to partial load 2, the online sub-model for outlet pressure prediction corresponding to partial load 3, and the online sub-model for outlet pressure prediction corresponding to full load.

[0138] Specifically, real-time operating data is collected at fixed time intervals (e.g., every 10 minutes or 1 hour), including: compressor inlet temperature, inlet pressure, IGV angle, and output power. The compressor efficiency is calculated using a compressor efficiency mathematical model based on the inlet temperature, inlet pressure, outlet temperature, and outlet pressure. The newly collected data is then sorted by timestamp to generate a standardized data sample D'. (t) The format is consistent with the offline training data, and normalization is performed to maintain dimensional consistency.

[0139] The objective function for incremental learning is a hybrid loss function, consisting of two parts:

[0140] Mean Squared Error (MSE): Measures the error of an online submodel on new data D′. (t) The prediction error is reduced to ensure that the model adapts to real-time performance changes.

[0141] Elastic Weight Consolidation (EWC) regularization term: By constraining the magnitude of parameter changes, it prevents the model from forgetting the healthy baseline knowledge learned in the past. Specifically, it applies a larger penalty to key parameters (such as the weights affecting the pressure ratio-efficiency relationship) by calculating the Fisher information matrix of the parameters, avoiding drastic parameter oscillations caused by real-time data noise.

[0142] The Adam optimizer is used to perform stochastic gradient descent optimization on the objective function, updating the parameters of the online sub-model. For example, when new data shows a decrease in compressor efficiency without a corresponding decrease in outlet pressure, the model adjusts the correlation weights between the IGV angle and pressure, while maintaining the stability of the pressure ratio-efficiency relationship through EWC regularization.

[0143] In existing technologies, offline models cannot adapt to the performance degradation of gas turbines over time (such as blade wear and fouling), leading to increased prediction errors with operating time. This step dynamically tracks performance degradation through an incremental learning mechanism, balancing learning and memory; continuous input of real-time data enables the model to capture slow changes such as efficiency decline and increased clearance. For example, when the compressor operating time exceeds 5000 hours, the online sub-model automatically adjusts the weight of the "operating time" feature through incremental learning, increasing its impact on pressure prediction from 0.1 to 0.3, reflecting the intensified impact of aging on performance. The EWC regularization term ensures that the model does not disrupt established health baseline associations when learning new degradation features. For example, when learning the new rule that "fouling leads to efficiency decline," the model retains historical knowledge that "washing can restore efficiency," avoiding misjudging post-maintenance performance recovery as abnormal fluctuations. The offline sub-model provides fixed health baseline predictions, while the online sub-model provides real-time predictions; the residual between the two (step S3000) directly reflects the degree of performance degradation. For example, the offline sub-model predicts a full-load outlet pressure of 100 kPa, while the online sub-model, due to fouling, predicts 95 kPa, with a residual of 5 kPa triggering a recession warning. Historical data from incremental learning can be fed back into the retraining of the offline sub-model (e.g., periodically updating the offline model with long-term data), forming a closed loop of "offline benchmark - online tracking - data feedback." For instance, the offline sub-model can be fine-tuned quarterly using recession data accumulated through online learning to improve the adaptability of the benchmark model.

[0144] There are two major bottlenecks in the current gas turbine performance monitoring:

[0145] First, offline models lack timeliness: models trained in the initial stage of operation cannot adapt to performance drift during operation, resulting in monitoring lag. For example, after one year of operation, the prediction error of traditional offline models may rise from 3% to 10%, failing to accurately identify early degradation.

[0146] Second, online model training is costly: training online models independently requires a large amount of real-time data and computing resources, and is prone to overfitting to short-term fluctuations.

[0147] Step S2000 achieves efficient real-time modeling and dynamic degradation monitoring through an "offline benchmark + online incremental" mechanism. It initializes the online model using offline sub-model parameters, reducing the initial training data requirements. Simultaneously, it gradually optimizes the model through incremental learning, enabling it to achieve high predictive accuracy in the early stages of operation and improve over time. The real-time updated online model can respond to performance changes within minutes. For example, when compressor efficiency decreases by 0.5% per month, the online model can detect an increase in the residual change rate in the first month, while traditional offline models require an accumulated efficiency decrease of over 3% to identify this.

[0148] If step S2000 is omitted and only the offline sub-model is used for online prediction, it will be unable to adapt to slow time-varying degradation such as equipment aging and fouling, resulting in a continuous increase in residuals and failure to trigger early warnings; it will also be unable to cope with real-time operating condition fluctuations (such as sudden changes in ambient temperature), mistakenly identifying normal fluctuations as degradation, leading to an increased false alarm rate. The real-time predicted value of the online sub-model and the baseline value of the offline sub-model together constitute the input for residual calculation, and their synergy ensures the reliability of the early warning logic. For example, if the online model falsely reports a pressure drop due to sensor noise, the stable output of the offline baseline can filter out random errors through residuals, avoiding false triggering of early warnings. Incremental learning enables the model to capture early degradation signals of efficiency decline, providing an earlier warning time compared to traditional methods, thus gaining a window of opportunity for preventative maintenance. The combination of parameter transfer and incremental learning makes the computational workload of the online model far less than that of an independently trained model, making it suitable for embedded systems deployed locally on the gas turbine, avoiding the latency and cost of data uploads to the cloud. The EWC regularization term ensures that the model retains core health baseline knowledge even after experiencing extreme operating conditions (such as emergency shutdown). For example, after a surge event, the model will not disrupt the fundamental correlation between pressure ratio and efficiency due to abnormal data, maintaining the stability of the monitoring logic. In summary, step S2000, through an incremental learning mechanism, combines prior knowledge from offline modeling with real-time feedback from online data, systematically solving the challenges of timeliness and accuracy in gas turbine performance monitoring.

[0149] Step S3000: Based on the online sub-model for outlet pressure prediction, obtain the online prediction sequence for outlet pressure under the current operating condition; based on the offline sub-model for outlet pressure prediction, obtain the offline prediction sequence for outlet pressure under the current operating condition; based on the online and offline prediction sequences for outlet pressure, conduct a gas turbine performance degradation assessment and issue an early warning.

[0150] Please see Figure 3 As shown, step S3000 further includes:

[0151] Step S3100: Obtain the current operating data of the gas turbine; based on the current operating data and the obtained online sub-model for outlet pressure prediction, obtain the online prediction sequence of outlet pressure for the current operating condition.

[0152] Further, step S3100 includes:

[0153] Step S3110: Obtain the current operating data of the gas turbine, and determine the current load level based on the current operating data;

[0154] Step S3120: Based on the current operating data and the compressor efficiency mathematical model, obtain the current compressor efficiency time series;

[0155] Step S3130: Based on the determined current load level, the corresponding online sub-model for outlet pressure prediction is invoked. Based on the current operating data, the current compressor efficiency time series, and the invoked corresponding online sub-model for outlet pressure prediction, the online prediction sequence for outlet pressure is obtained.

[0156] Specifically, real-time acquisition of current operating data, including output power, is performed. Output power is read in real-time by the unit control system, reflecting the current load status. The maximum output power within the monitoring time window is extracted; for example, with a 1-hour window, power data is collected every 10 minutes, and the maximum value among 6 samples is taken. The current load level is determined according to the load level determination method in step S1300. Through a load classification matching model, the outlet pressure prediction error is reduced, and the factor model parameters are better matched to the aerodynamic characteristics of the current load (e.g., low pressure ratio at low loads, frequent IGV adjustments).

[0157] Using the compressor inlet temperature, outlet temperature, inlet pressure, and outlet pressure from the current operating data, the compressor efficiency mathematical model is input to calculate the current compressor efficiency. The calculated efficiency values ​​are then sorted by timestamp to form a compressor efficiency time series. Step S3120, by integrating the compressor efficiency mathematical model and using the collected temperature and pressure data to calculate the compressor efficiency in real time, solves the problem of "compressor efficiency not being observable in real time," enabling the prediction model to promptly detect efficiency declines caused by blade wear, fouling, etc. The compressor efficiency time series and outlet pressure prediction form a causal chain. When the blade tip clearance increases by 0.1 mm, the compressor efficiency may decrease by 1%. At this time, the control system maintains stable outlet pressure by increasing the IGV angle, resulting in no significant change in outlet pressure. However, the prediction model in this invention detects the performance degradation trend by observing the decrease in compressor efficiency, while traditional methods that only monitor outlet pressure cannot identify early anomalies. Therefore, compared to traditional methods that rely solely on pressure data, this invention, by forming a causal chain between the compressor efficiency time series and outlet pressure prediction, can advance the warning time. This invention indirectly reflects the influence of unmeasurable parameters such as blade tip clearance by changing compressor efficiency. For example, when an increase in clearance leads to a decrease in efficiency of 0.5% per month, the online sub-model for predicting outlet pressure in this invention can be identified by the slope change of the compressor efficiency time series, while traditional methods require periodic disassembly and detection.

[0158] Based on the load level determined by S3110, a corresponding model is selected from n1 online sub-models. For example, if the load level is determined to be partial load 2, the online sub-model for that load level is called. The current operating data and the compressor efficiency time series calculated by S3120 are used to generate a sliding window sequence according to a set time window (e.g., 1 hour), forming a three-dimensional tensor of [number of samples, window length, number of features]. Through forward propagation via an LSTM network, the predicted outlet pressure sequence for a future time period (e.g., 30 minutes) is output.

[0159] Traditional single-model forecasting suffers from high error rates when predicting across loads because it cannot adapt to the aerodynamic characteristics of different operating conditions (such as low-load surge risk and full-load efficiency saturation). This step improves the prediction error across load ranges by dynamically determining the load and matching it with the model. Existing methods often rely on offline efficiency calculations or single pressure parameters, failing to reflect real-time performance changes in a timely manner. Step S3100 calculates compressor efficiency in real time and inputs it into the prediction model, enabling the outlet pressure prediction to reflect efficiency degradation synchronously. For example, when compressor efficiency decreases at a rate of 0.2% / day due to fouling, the model can detect an increase in the prediction residual within 3 days, while traditional methods require an accumulated efficiency decrease of more than 1% to identify this.

[0160] Without load level determination, the full-load model may overestimate the pressure ratio at low loads, leading to an overestimation of the predicted pressure and misjudging it as a healthy state. In reality, the pressure due to surge risk may be underestimated, resulting in missed pressure degradation warnings. Without real-time efficiency calculation, the model cannot distinguish between pressure changes caused by load fluctuations and pressure drops caused by efficiency degradation. For example, pressure fluctuations caused by rising ambient temperature may be misjudged as fouling degradation, increasing the false alarm rate. In summary, step S3100, through dynamic load determination, real-time efficiency calculation, and precise model invocation, systematically improves the real-time performance and accuracy of gas turbine compressor outlet pressure prediction.

[0161] Step S3200: Based on the current operating data, the current compressor efficiency time series, and the offline sub-model for predicting the outlet pressure corresponding to the current load level, obtain the offline prediction sequence for the outlet pressure under the current operating condition.

[0162] Specifically, step S3200 provides a "health baseline value" for real-time monitoring by calling an offline sub-model with fixed parameters, thus solving the problem of "lack of dynamic health baseline" in the prior art.

[0163] The offline sub-model for outlet pressure prediction is trained using offline health data in step S1400. Its parameters (such as LSTM hidden layer weights and bias vectors) remain fixed, representing the outlet pressure pattern when the gas turbine has no performance degradation at the corresponding load level. Nine-dimensional features from the current operating data, including compressor inlet temperature, inlet pressure, and IGV angle, are formatted according to the time window (e.g., 1 hour) used during offline sub-model training, along with the current compressor efficiency time series calculated in step S3120, to generate a three-dimensional tensor ([number of samples, window length, number of features]) in the same format as the training data. This formatted three-dimensional tensor is input into the offline sub-model, and long-term time-dependent features (such as fouling trends) are extracted through calculations in the hidden layers of the LSTM network. The final output is the offline predicted sequence for outlet pressure, processed through a fully connected layer.

[0164] In existing technologies, monitoring systems often use a single fixed value or historical average as a benchmark, which cannot adapt to load changes and equipment aging. This step provides a load-specific benchmark through offline sub-models. For example, the benchmark pressure output by the low-load offline sub-model takes into account the characteristics of a large IGV angle and low pressure ratio under that load, avoiding misjudging normal low pressure as degradation. The full-load sub-model, based on design condition data, reflects the pressure value under ideal conditions, used to compare the deviation from actual operation. The fixed parameter characteristics of the offline sub-model make its output unaffected by real-time noise, effectively distinguishing between "normal fluctuations in operating conditions" and "pressure anomalies caused by degradation." For example, when the ambient temperature rises, causing the inlet temperature to rise, the offline sub-model still outputs a pressure benchmark value based on a healthy state, while the online sub-model adjusts the predicted value due to real-time data input. The residuals of the two can separate the degradation signal after temperature compensation.

[0165] Step S3300: Obtain the outlet pressure residual sequence based on the online and offline outlet pressure prediction sequences;

[0166] Specifically, the outlet pressure residual is the difference between the online predicted value and the offline predicted value of the outlet pressure at the same time step. The sign of the residual reflects the direction of pressure deviation, and the absolute value reflects the degree of deviation. The residual sequence transforms performance degradation into a measurable numerical indicator, solving the subjectivity problem of traditional methods relying on human experience. For example, when the absolute value of the residual is greater than 5 kPa, it indicates that the pressure deviates from the healthy baseline beyond the allowable range, possibly indicating fouling or blade wear; when the residual change rate is greater than 0.3 kPa / hour, it indicates an accelerated rate of degradation, requiring urgent maintenance. The type of degradation can be identified through the residual change pattern, including reversible and irreversible degradation. For reversible degradation, such as fouling caused by lack of water washing, the residual should return to the baseline value after water washing; otherwise, it indicates irreversible damage (such as blade corrosion). For irreversible degradation, such as increased clearance caused by blade wear, the residual continues to increase and cannot be restored by water washing; it is necessary to determine whether to trigger a component replacement warning through the residual trend.

[0167] Step S3400: Based on the outlet pressure residual sequence, determine whether a compressor performance degradation warning has been triggered.

[0168] Further, step S3400 includes:

[0169] Step S3410: Filter the outlet pressure residual sequence to obtain a smooth residual change curve;

[0170] Step S3420: Perform trend prediction on the smoothed residual change curve and estimate the residual change rate;

[0171] Step S3430: Compare the residual change rate with the preset decay rate threshold to determine whether it is necessary to trigger a gas turbine performance decay warning.

[0172] Specifically, step S3400 addresses the problems of "lagging early warning and high false alarm rate" in existing technologies through noise suppression, trend analysis, and threshold determination. The Kalman filter algorithm is used to suppress noise in the outlet pressure residual sequence. The residual sequence consists of the difference between the online predicted value and the offline reference value, and includes interference such as sensor measurement noise (e.g., ±0.5℃ error from the temperature sensor transmitted to the pressure residual) and control system adjustment fluctuations (e.g., pressure transients caused by dynamic adjustment of the IGV angle). The Kalman filter achieves noise removal through the following recursive process:

[0173] State prediction: Based on the residual state estimate and state transition equation of the previous time step (assuming the residual change is a random walk process), predict the residual prior estimate and its covariance at the current time step;

[0174] Observation update: The current measured residual value is combined with the predicted value, and the posterior estimate is calculated by Kalman gain. The gain is determined by the ratio of the predicted covariance to the observation noise variance, reflecting the degree of confidence in the measured value.

[0175] Example: If the residual sequence has measurements of [-2kPa, -1.8kPa, -2.1kPa, -1.9kPa] over 1 hour, the Kalman filter, through iterative calculation, outputs a smoothed estimated value sequence of [-2kPa, -1.95kPa, -2.02kPa, -1.98kPa], effectively suppressing high-frequency fluctuations.

[0176] In existing technologies, direct residual analysis is susceptible to noise interference, leading to misjudgment of trends. For example, occasional sensor impulse noise may cause sudden fluctuations in the residuals, which traditional methods may misinterpret as a decay signal. Step S3410 reduces the noise variance through Kalman filtering, making the residual curve more accurately reflect the decay trend.

[0177] Step S3420 uses the least squares method to perform first-order polynomial fitting on the filtered residual sequence and extracts the trend slope as the residual change rate. Existing techniques often calculate the change rate through fixed-window differencing, which is easily affected by the window size and cannot identify nonlinear trends. This step, through least squares fitting, can adaptively capture linear trends. For example, when the residual change rate gradually decreases from -0.05 kPa / h to -0.1 kPa / h (the absolute value increases from 0.05 kPa / h to 0.1 kPa / h, indicating a faster decay rate), the fitting slope can reflect the accelerating decay trend in real time. Step S3420 upgrades the degree of decay from a "qualitative judgment" to a "quantitative indicator," facilitating the formulation of maintenance priorities. Compared to traditional differencing methods, the least squares method can detect trend changes earlier because the fitting process utilizes the entire sequence data, not just the differences between earlier and later stages.

[0178] The residual change rate is compared with the preset decay rate threshold: if the absolute value of the residual change rate is less than the absolute value of the decay rate threshold, it means that the compressor performance decay rate has not yet reached the warning line, the compressor health status is within an acceptable range, and there is no need to trigger the warning for the time being; if the absolute value of the residual change rate is greater than or equal to the absolute value of the decay rate threshold, it means that the compressor performance decay rate has exceeded the warning line, the compressor health status has deteriorated significantly, and a warning signal needs to be triggered immediately.

[0179] The decay rate threshold is determined based on the gas turbine type, load level, and historical health data. Based on the decay characteristics, the decay rate threshold is set to a negative value. For example, the threshold for low load conditions is set to -0.08 kPa / h (allowing for slower decay); the threshold for full load conditions is set to -0.05 kPa / h (more sensitive to decay).

[0180] Existing technologies often use a uniform threshold, failing to consider load variations. For example, compressor efficiency is more sensitive to pressure changes at full load, and the same degradation rate is more harmful under full load, requiring a stricter threshold. This step uses load-level thresholds to make the early warning logic more aligned with actual operating conditions. Under full load, efficiency decline due to degradation may trigger surge risk; a smaller absolute threshold (-0.05 kPa / h) achieves a stricter early warning standard, allowing for earlier warnings and a response earlier than a uniform threshold. At low loads, a larger absolute threshold (-0.08 kPa / h) is allowed to avoid false alarms triggered by normal efficiency fluctuations. The negative sign represents the residual change trend (the direction of pressure deviation from the healthy baseline), while the numerical value represents the absolute value of the rate.

[0181] Step S3400 transforms the residual sequence from "noisy raw data" into "decision-based degradation indicators" through a three-stage processing of "filtering-fitting-comparison," addressing the problem of existing technologies relying on human experience and lacking quantitative evidence for early warning. For example, when a unit is operating at partial load 2, the filtered residual sequence shows a slope of -0.06 kPa / h, whose absolute value exceeds the absolute value of the load threshold of -0.05 kPa / h, triggering a level-two early warning, indicating a risk of fouling or increased clearance. Traditional methods might fail to identify this trend due to noise interference. Without filtering and trend analysis, residual noise may mask the true degradation signal. For instance, if a unit experiences a slow decrease in residual speed of -0.03 kPa / h due to blade wear, noise fluctuations of ±0.2 kPa would completely obscure the trend, making it undetectable by traditional methods. An alarm might only be triggered in the later stages of degradation when pressure drops sharply, leading to maintenance delays. Without quantified residual change rates, operators must rely on experience to judge, which may lead to false alarms and missed alarms. Short-term pressure fluctuations may be misinterpreted as degradation, increasing unnecessary downtime for maintenance; ignoring slow degradation trends can result in severe equipment damage (such as blade breakage). Early warning systems can shift maintenance planning from "reactive repairs" to "preventative maintenance," reducing unplanned downtime. For example, a power plant used this step to identify compressor fouling in advance and cleaned it during planned maintenance, avoiding power generation losses due to sudden shutdowns. Early intervention in degradation issues (such as timely water washing or gap adjustment) can extend the lifespan of critical compressor components, preventing severe wear in the later stages of degradation. Multi-stage processing (filtering + fitting + thresholding) makes the system robust to sensor failures and short-term disturbances. For example, a brief failure of a temperature sensor may cause a sudden change in residuals; Kalman filtering can suppress this disturbance through state prediction, avoiding false alarm triggers.

[0182] Example 2:

[0183] This embodiment, based on Embodiment 1, provides an online monitoring system for the outlet pressure of a gas turbine compressor, such as... Figure 4 As shown, it includes:

[0184] Offline benchmark model construction module: used to obtain the offline health dataset when the gas turbine is first put into operation, and to build an offline benchmark model for outlet pressure prediction based on the offline health dataset; the load of the gas turbine is divided into n1 load levels, and an offline sub-model for outlet pressure prediction corresponding to each load level is established based on the offline benchmark model for outlet pressure prediction.

[0185] Online Prediction Model Construction Module: Based on the offline sub-model for predicting outlet pressure corresponding to each load level, an online sub-model for predicting outlet pressure corresponding to each load level is established through incremental learning;

[0186] Performance degradation assessment module: Based on the online sub-model for export pressure prediction, the online export pressure prediction sequence for the current operating condition is obtained; based on the offline sub-model for export pressure prediction, the offline export pressure prediction sequence for the current operating condition is obtained; based on the online and offline export pressure prediction sequences, the gas turbine performance degradation is assessed and an early warning is issued.

[0187] In the offline baseline model construction module, the method for establishing offline sub-models for outlet pressure prediction corresponding to each load level includes:

[0188] Step S1410: Based on the offline data subsets of each load level, initialize the offline sub-model for predicting outlet pressure corresponding to each load level.

[0189] Step S1420: Use the offline benchmark model for outlet pressure prediction to pre-train the offline sub-model for outlet pressure prediction corresponding to each load level, and obtain the initial parameters of the offline sub-model for outlet pressure prediction corresponding to each load level.

[0190] Step S1430: Use the offline data subset corresponding to the load level to perform specialized training on the offline sub-model for predicting outlet pressure for each load level with initial parameters, and obtain the final parameters of the offline sub-model for predicting outlet pressure for each load level.

[0191] In the degradation assessment module, the method for assessing and issuing early warnings about gas turbine performance degradation based on online and offline outlet pressure prediction sequences includes:

[0192] Based on the online and offline export pressure forecast sequences, the export pressure residual sequence is obtained.

[0193] The outlet pressure residual sequence is filtered to obtain a smooth residual variation curve;

[0194] Trend prediction is performed on the smoothed residual change curve to estimate the residual change rate;

[0195] The residual change rate is compared with a preset decay rate threshold to determine whether a gas turbine performance degradation warning needs to be triggered.

[0196] The methods and systems of this application may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the method is for illustrative purposes only, and the steps of the method of this application are not limited to the order specifically described above, unless otherwise specifically stated.

[0197] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0198] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method of on-line monitoring of the pressure at the outlet of a gas turbine compressor, characterized in that, The method comprises: An offline health data set when the gas turbine starts to operate is acquired, and an outlet pressure prediction offline benchmark model is constructed according to the offline health data set; the load of the gas turbine is divided into n1 load grades, and an outlet pressure prediction offline sub-model corresponding to each load grade is established based on the outlet pressure prediction offline benchmark model; the offline health data set comprises first health working condition data directly measured under different loads and second health working condition data calculated; the first health working condition data comprises a compressor inlet temperature time sequence, a compressor outlet temperature time sequence, a compressor inlet pressure time sequence, a compressor outlet pressure time sequence, an IGV angle time sequence, a VGV angle time sequence, an output power time sequence, a turbine cooling air valve position display time sequence, a compressor operation time, and a water washing state identification time sequence; and the second health working condition data comprises a compressor efficiency time sequence; Nine types of characteristics, i.e., the compressor inlet temperature time sequence, the compressor inlet pressure time sequence, the IGV angle time sequence, the VGV angle time sequence, the output power time sequence, the turbine cooling air valve position display time sequence, the compressor operation time, the water washing state identification time sequence, and the compressor efficiency time sequence, are sampled in a sliding window according to a set time window to form a three-dimensional tensor with [sample number, window length, and characteristic number] as an input of the outlet pressure prediction offline benchmark model, and the corresponding output is the compressor outlet pressure at the last time step of each window; the outlet pressure prediction offline benchmark model adopts a two-layer LSTM neural network structure, and comprises an input layer, a hidden layer, and a full connection layer; the hidden layer comprises two LSTM layers; The method for establishing the outlet pressure prediction offline sub-model corresponding to each load grade based on the outlet pressure prediction offline benchmark model comprises: screening an offline data subset of each load grade from the offline health data set; initializing the outlet pressure prediction offline sub-model corresponding to each load grade based on the offline data subset of each load grade; copying part of parameters in the outlet pressure prediction offline benchmark model to the outlet pressure prediction offline sub-model corresponding to each load grade; the part of parameters comprises a weight matrix and a bias vector in the LSTM layer; pre-training the outlet pressure prediction offline sub-model corresponding to each load grade by using the outlet pressure prediction offline benchmark model to obtain initial parameters of the outlet pressure prediction offline sub-model corresponding to each load grade; and specializing the outlet pressure prediction offline sub-model corresponding to each load grade with the initial parameters by using the offline data subset corresponding to the load grade to obtain final parameters of the outlet pressure prediction offline sub-model corresponding to each load grade; Based on the outlet pressure prediction offline sub-model corresponding to each load grade, an outlet pressure prediction online sub-model corresponding to each load grade is established through incremental learning. The method for establishing the outlet pressure prediction online sub-model corresponding to each load level comprises: based on final parameters of the outlet pressure prediction offline sub-model corresponding to each load level, constructing the outlet pressure prediction online sub-model corresponding to each load level; obtaining real-time operation condition data of each load level of the gas turbine, and using the real-time operation condition data of each load level to update parameters of the outlet pressure prediction online sub-model corresponding to each load level in real time through incremental learning to obtain n1 outlet pressure prediction online sub-models corresponding to n1 final load levels; the outlet pressure prediction online sub-model is completely consistent with the outlet pressure prediction offline sub-model in architecture, both adopt a two-layer LSTM neural network structure, contain an input layer, a hidden layer and a full connection layer, have the same input feature dimension as the outlet pressure prediction offline sub-model, and output a compressor outlet pressure prediction sequence; According to the outlet pressure prediction online sub-model, an outlet pressure online prediction sequence of the current condition is obtained; according to the outlet pressure prediction offline sub-model, an outlet pressure offline prediction sequence of the current condition is obtained; and according to the outlet pressure online prediction sequence and the outlet pressure offline prediction sequence, gas turbine performance degradation is evaluated and early warning is given.

2. The gas turbine compressor outlet pressure on-line monitoring method of claim 1, wherein, The method for calculating the compressor efficiency time sequence comprises: constructing a compressor efficiency mathematical model, obtaining compressor efficiency based on the first health condition data and the constructed compressor efficiency mathematical model, and sorting the compressor efficiency according to a time stamp to obtain the compressor efficiency time sequence.

3. The gas turbine compressor outlet pressure on-line monitoring method of claim 2, wherein, The method for dividing the load of the gas turbine into n1 load levels comprises: extracting the maximum output power from the output power time series of the offline health dataset within a given monitoring time window comparing the maximum output power to the rated full load power of the gas turbine to determine the load class.

4. The gas turbine compressor outlet pressure on-line monitoring method of claim 3, wherein, The method for obtaining the outlet pressure online prediction sequence of the current condition comprises: obtaining current operation data of the gas turbine, and determining a current load level based on the current operation data; obtaining a current compressor efficiency time sequence according to the current operation data and the compressor efficiency mathematical model; calling a corresponding outlet pressure prediction online sub-model based on the determined current load level, and obtaining an outlet pressure online prediction sequence according to the current operation data, the current compressor efficiency time sequence and the called corresponding outlet pressure prediction online sub-model.

5. The gas turbine compressor outlet pressure on-line monitoring method of claim 4, wherein, The method for evaluating and giving early warning of gas turbine performance degradation according to the outlet pressure online prediction sequence and the outlet pressure offline prediction sequence comprises: obtaining an outlet pressure residual sequence according to the outlet pressure online prediction sequence and the outlet pressure offline prediction sequence; and determining whether to trigger a compressor performance degradation early warning based on the outlet pressure residual sequence.

6. A gas turbine compressor outlet pressure on-line monitoring system for implementing the gas turbine compressor outlet pressure on-line monitoring method according to any one of claims 1 to 5, characterized by, The system comprises: an offline benchmark model construction module configured to obtain offline health data sets of the gas turbine when the gas turbine is just put into operation, construct an outlet pressure prediction offline benchmark model according to the offline health data sets, divide the load of the gas turbine into n1 load levels, and establish outlet pressure prediction offline sub-models corresponding to each load level based on the outlet pressure prediction offline benchmark model; an online prediction model construction module configured to establish outlet pressure prediction online sub-models corresponding to each load level through incremental learning based on the outlet pressure prediction offline sub-models corresponding to each load level. The degradation evaluation module comprises: an outlet pressure online prediction sub-model, which is configured to obtain an outlet pressure online prediction sequence of the current working condition; an outlet pressure offline prediction sub-model, which is configured to obtain an outlet pressure offline prediction sequence of the current working condition; and a gas turbine performance degradation evaluation and early warning module, which is configured to perform gas turbine performance degradation evaluation and early warning according to the outlet pressure online prediction sequence and the outlet pressure offline prediction sequence.

Images