Online monitoring method and system for outlet pressure of gas compressor of gas turbine

Abstract

The invention relates to the technical field of gas turbine performance monitoring, and discloses a gas turbine compressor outlet pressure online monitoring method and system, and the method comprises the steps: obtaining an offline health data set to construct an offline reference model, dividing load levels, building an offline sub-model, and generating an online sub-model through incremental learning; evaluating performance degradation by combining online and offline prediction sequences; through the synergistic effect of the off-line reference model and the on-line sub-model, the problems of inaccurate outlet pressure prediction and lagging or misjudgment of performance degradation monitoring of an existing model under the multi-load working condition are solved, and high-precision dynamic monitoring of the outlet pressure of the gas turbine compressor and timely early warning of performance degradation are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of gas turbine performance monitoring, and more particularly to a method and system for online monitoring of a gas turbine compressor outlet pressure. Background Art

[0002] As a complex power plant, accurate monitoring of the compressor outlet pressure of a gas turbine is crucial for equipment health management. A gas turbine health status prediction method proposed in Chinese patent application publication number CN109520740A uses multiple sensors to collect data and construct a health analysis model, but it does not perform detailed modeling of the differences in aerodynamic characteristics under different gas turbine load conditions (such as low load, partial load, and full load). Because the mapping relationship between compressor efficiency, pressure ratio, and outlet pressure under different gas turbine loads varies significantly (for example, IGV angle adjustment is more sensitive to pressure at low load, while pressure ratio dominates efficiency changes at full load), a single model cannot provide consistent prediction accuracy across the entire load range. This can lead to overestimation of pressure under low-load conditions due to overfitting high-load data, or underestimation of pressure under full load due to underfitting low-load characteristics, ultimately resulting in large outlet pressure prediction errors. Furthermore, this method does not incorporate a dynamic model update mechanism and cannot adapt to the slow performance degradation caused by factors such as fouling and blade wear during gas turbine operation, causing the monitoring results to lag behind the actual degradation process.

[0003] A Chinese patent application with publication number CN117588312A proposes a gas turbine fault prediction and diagnosis method based on deep transfer learning. Although it improves the fault diagnosis capability of new units by migrating data between the source and target domains, it mainly focuses on fault category identification and does not involve high-precision prediction of continuous parameters such as outlet pressure. Its model adjustment process relies on fixed batches of target domain data and lacks an online incremental learning mechanism, making it impossible to dynamically update model parameters using the latest operating data in real time. When the performance of a gas turbine compressor gradually declines due to long-term operation, traditional transfer learning models are unable to capture the trend of residual changes in a timely manner. They may miss early decline signals due to delayed model parameter updates, or mistakenly judge normal operating fluctuations as performance anomalies, resulting in insufficient reliability of the early warning logic.

[0004] Existing technologies are unable to build refined prediction models based on the specific characteristics of gas turbine multi-load conditions, and lack an online dynamic learning mechanism to track the performance degradation process in real time. As a result, the compressor outlet pressure prediction error increases significantly with load changes, and performance degradation monitoring is subject to the risk of lag or misjudgment. Summary of the Invention

[0005] To overcome the aforementioned shortcomings of the prior art, the present invention provides a method and system for online monitoring of gas turbine compressor outlet pressure. This system constructs a baseline model by acquiring an offline healthy dataset, divides load levels into load-based sub-models, and combines incremental learning to generate online sub-models. This system achieves high-precision prediction of gas turbine compressor outlet pressure under multiple load conditions. Residual analysis of online and offline prediction sequences can promptly capture performance degradation signals and provide early warnings, significantly improving the timeliness and reliability of gas turbine operation and maintenance and reducing the risk of unplanned downtime.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A method for online monitoring of a gas turbine compressor outlet pressure includes:

[0008] Obtain an offline health dataset of the gas turbine at the start of operation and construct an offline benchmark model for outlet pressure prediction based on the offline health dataset. Divide the gas turbine load into n1 load levels and, based on the offline benchmark model for outlet pressure prediction, establish an offline sub-model for outlet pressure prediction corresponding to each load level.

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

[0010] Based on the outlet pressure online prediction sub-model, the outlet pressure online prediction sequence of the current operating condition is obtained; based on the outlet pressure offline prediction sub-model, the outlet pressure offline prediction sequence of the current operating condition is obtained; based on the outlet pressure online prediction sequence and the outlet pressure offline prediction sequence, the gas turbine performance degradation assessment and early warning are carried out.

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

[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 outlet pressure prediction offline benchmark model adopts a two-layer LSTM neural network structure, including an input layer, a hidden layer and a fully connected layer; the hidden layer includes 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 within a given monitoring time window from the output power time series of the offline health dataset max , the maximum output power P max The rated full load power P of the gas turbine rated Make a comparison and determine the load level.

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

[0017] Filter offline data subsets of each load level from the offline health data set;

[0018] Based on the offline data subset of each load level, the outlet pressure prediction offline sub-model corresponding to each load level is initialized;

[0019] The outlet pressure prediction offline benchmark model is used to pre-train the outlet pressure prediction offline sub-model corresponding to each load level to obtain the initial parameters of the outlet pressure prediction offline sub-model corresponding to each load level;

[0020] The offline data subset corresponding to the load level is used to perform specialized training on the outlet pressure prediction offline sub-model corresponding to each load level with initial parameters, and the final parameters of the outlet pressure prediction offline sub-model corresponding to each load level are obtained.

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

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

[0023] Real-time operating condition data of the gas turbine at each load level is obtained. Using the real-time operating condition data of each load level, the parameters of the outlet pressure prediction online sub-model corresponding to each load level are updated in real time through incremental learning, and finally n1 outlet pressure prediction online sub-models corresponding to n1 load levels are obtained.

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

[0025] obtaining current operating data of the gas turbine, and determining a current load level based on the current operating data;

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

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

[0028] Furthermore, the method for evaluating and warning gas turbine performance degradation based on the outlet pressure online prediction sequence and the outlet pressure offline prediction sequence includes:

[0029] According to the outlet pressure online prediction sequence and the outlet pressure offline prediction sequence, the outlet pressure residual sequence is obtained; based on the outlet pressure residual sequence, it is determined whether the compressor performance degradation warning is triggered.

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

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

[0032] Online prediction model construction module: Based on the offline sub-model for outlet pressure prediction corresponding to each load level, an online sub-model for outlet pressure prediction corresponding to each load level is established through incremental learning;

[0033] Degradation assessment module: Based on the outlet pressure prediction online sub-model, the outlet pressure online prediction sequence of the current operating condition is obtained; based on the outlet pressure prediction offline sub-model, the outlet pressure offline prediction sequence of the current operating condition is obtained; based on the outlet pressure online prediction sequence and the outlet pressure offline prediction sequence, the gas turbine performance degradation assessment and early warning are carried out.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] The present invention achieves accurate prediction and real-time monitoring of the outlet pressure of the gas turbine compressor by constructing an offline benchmark model and offline sub-models for different load levels, and using incremental learning to establish an online sub-model. First, in the offline stage, the data characteristics of the gas turbine during healthy operation are fully explored, and a benchmark model and sub-model that can reflect different load conditions are constructed, providing a reliable reference benchmark for online monitoring. Secondly, the online sub-model is continuously updated through incremental learning based on the offline sub-model, and can dynamically adapt to the gradual decline of gas turbine performance, ensuring the timeliness and accuracy of the prediction results. Finally, by comparing the online prediction sequence and the offline prediction sequence, a residual sequence is obtained, and performance degradation assessment and early warning are performed based on this. This effectively solves the problems in the existing technology such as large outlet pressure prediction error, performance degradation monitoring lag or misjudgment caused by the inability to adapt to multiple load conditions and insufficient dynamic tracking, significantly improving the efficiency and reliability of gas turbine operation and maintenance, and reducing the risk of unplanned downtime and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0037] Figure 1 This is a principle flow chart of the method for online monitoring of the outlet pressure of a gas turbine compressor according to the present invention;

[0038] Figure 2 A flow chart showing the principle of establishing an offline sub-model for predicting outlet pressure corresponding to each load level in the method for online monitoring outlet pressure of a gas turbine compressor of the present invention;

[0039] Figure 3 This is a flow chart showing the principle of evaluating and providing early warning for gas turbine performance degradation in the method for online monitoring of gas turbine compressor outlet pressure of the present invention;

[0040] Figure 4 This is a functional module diagram of the gas turbine compressor outlet pressure online monitoring system in the present invention. DETAILED DESCRIPTION

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0042] Example 1:

[0043] See also Figure 1 As shown, this embodiment provides a method for online monitoring of a gas turbine compressor outlet pressure, comprising:

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

[0045] See also Figure 2 As shown, further, step S1000 includes:

[0046] Step S1100, obtaining first health condition data directly measured under different loads and second health condition data calculated when the gas turbine starts to operate;

[0047] Furthermore, step S1100 includes:

[0048] Step S1110, obtaining first healthy operating condition data directly measured under different loads when the gas turbine starts to operate; the first healthy operating condition data includes a compressor inlet temperature time series, a compressor outlet temperature time series, a compressor inlet pressure time series, a compressor outlet pressure time series, an IGV angle time series, a VGV angle time series, an output power time series, a turbine cooling air valve position display time series, a compressor operating time, and a water washing status indicator time series; the turbine cooling air valve position display includes a 2-level, 3-level, and 4-level display; 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 stage of gas turbine delivery, that is, when there is no performance degradation and the component gaps are normal, various sensors and systems are used to collect operating data under multiple load conditions (covering low load, partial load, and base load) and multiple environmental conditions (including temperature, humidity, and atmospheric pressure) in real time.

[0050] Compressor inlet temperature: This data is collected by a temperature sensor installed at the compressor inlet. This data, represented as a time series of temperature changes, reflects the impact of intake conditions on the compression process and is a fundamental parameter for calculating compressor efficiency and analyzing compression power consumption. For example, under varying loads, inlet temperature affects gas density and initial energy, which in turn alters the energy demand and outlet pressure during compression.

[0051] Compressor outlet temperature: This temperature is collected simultaneously with the inlet temperature by a temperature sensor installed at the compressor outlet. This temperature is also presented as a time series, reflecting actual compression power consumption and heat loss. During the actual compression process, the 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 are measured using high-precision pressure transmitters and recorded as a time series. These two parameters are used to calculate the pressure ratio, a key indicator of a compressor's compression capacity and a direct factor in compressor efficiency. For example, a higher pressure ratio indicates a higher degree of gas compression, reflecting, to a certain extent, the compressor's performance.

[0053] IGV and VGV angle timing: Angle encoders monitor the opening of the inlet guide vanes (IGVs) and compressor variable stators (VGVs) in real time, expressed as a time series of angle values. These are used to quantify the impact of airflow regulation on pressure. Changes in IGV and VGV openings alter the direction and flow of air entering the compressor, thereby affecting compressor efficiency and outlet pressure. For example, at low loads, appropriately adjusting the IGV angle can increase intake airflow, maintaining stable compressor operation and outlet pressure.

[0054] Output Power: Real-time electrical power is obtained from the unit control system and expressed as a time series of power values, reflecting the load status. Load changes directly alter compressor operating conditions. The output power time series can be used to classify load levels, providing a basis for subsequent load-level model construction.

[0055] Turbine cooling air valve position display sequence: A valve position sensor (e.g., a potentiometer) generates a normalized value (0-1) representing the coupled effect of cooling air flow on compressor load. The turbine cooling air valve's levels 2, 3, and 4 represent different cooling intensities. Only one level is active at a time, and the valve position display reflects the opening state of the currently active level. For example, a fully open level 2 valve corresponds to 0.3, a fully open level 3 valve corresponds to 0.6, and a fully open level 4 valve corresponds to 1.0. For example, if data is collected once per hour, the valve position display sequence might be [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 air valve at different stages of the turbine. Cooling air flow affects turbine efficiency, which in turn affects compressor load, making it a key parameter affecting compressor outlet pressure.

[0056] Compressor Runtime: This data is based on the cumulative operating hours recorded in the system operation log and expressed as a time value. This data is used to analyze the time dependence of performance degradation. As operating time increases, compressor components may experience wear and fouling, leading to performance degradation. Operating time is an important time dimension parameter for evaluating performance changes.

[0057] Water washing status flag time series: This is recorded as a binary variable (0 indicates not washed, 1 indicates washed), marking the effect of maintenance operations on compressor efficiency. Water washing removes internal compressor fouling and restores efficiency. This flag is used in the subsequent LSTM model 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] The existing technology suffers from incomplete data collection on the initial operation of the gas turbine, focusing only on some key parameters while ignoring parameters that have a significant impact on the compressor outlet pressure, such as the turbine cooling air valve position display and the water wash status indicator. This results in the constructed LSTM model being unable to fully and accurately reflect the actual operating status of the compressor. This step comprehensively collects all types of direct measurement data, solving the data missing problem and enabling the model to take into account more influencing factors, thereby improving the accuracy and reliability of the model.

[0059] Step S1120, obtaining second healthy operating condition data including different load operating conditions when the gas turbine starts to operate;

[0060] The second healthy operating condition data includes compressor efficiency time series;

[0061] The method for calculating the 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 in the first healthy operating condition data and the constructed compressor efficiency mathematical model; sorting the compressor efficiencies by timestamp to obtain the compressor efficiency time series;

[0062] The second health condition data cannot be directly obtained and needs to be calculated through a formula.

[0063] Specifically, compressor efficiency is a key indicator of compressor performance, indicating how effectively the compressor 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 the compressor, the ideal power consumption is reflected by the ideal adiabatic compression work formula. 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 is the outlet temperature after ideal adiabatic compression, T1 is the compressor inlet temperature, P2 is the compressor outlet pressure, and P1 is the compressor inlet pressure; γ is the specific heat ratio of the gas, which is dimensionless and can be obtained by looking up the table based on the gas composition and operating temperature range, or measured by a gas composition analyzer. It is approximately 1.4 at normal temperature and pressure.

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

[0068]

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

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

[0071] During the actual compression process, 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 compressor inlet temperature after actual compression, specifically:

[0072] W a =C p ×(T2-T1);

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

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

[0075]

[0076] After simplification and arrangement, the mathematical model of compressor efficiency is obtained:

[0077]

[0078] Determines the sensitivity of efficiency. As the pressure ratio increases, As η increases, it also increases, quantifying the contribution of the pressure ratio to the compressor efficiency. The 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. The abnormal increase in the outlet temperature can provide early warning of blade corrosion, increased clearance, or fouling failure.

[0079] The above-mentioned mathematical model for compressor efficiency, derived directly from thermodynamic principles, accurately reflects 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 changing trends of compressor efficiency, compressor performance degradation can be promptly detected. Decreased efficiency may indicate problems such as fouling, blade damage, or seal leakage within 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 reflection of the clearance change, and the two are negatively correlated. The mathematical model calculates compressor efficiency using measurable temperature (T1, T2) and pressure (P1, P2), eliminating the need for direct clearance measurement and addressing the issue of unmeasurable clearance. Through the dynamic calculation of the compressor efficiency η, the impact of tip clearance changes on aerodynamic performance is converted into an observable, quantitative metric. For example, when the clearance increases, leakage increases, resulting in an increase in the actual temperature rise (T2-T1), a significant decrease in compressor efficiency, a lower pressure ratio, and further deterioration in compressor efficiency. Taking η as the input feature of the LSTM model enables the neural network to learn the nonlinear relationship between gap change and outlet pressure, breaking through the limitation of traditional data-driven models that cannot model structural parameters.

[0080] In the existing technology, compressor efficiency is usually difficult to measure directly and needs to rely on complex measuring equipment or indirect estimation, which lacks accuracy and real-time performance. Step S1120 solves the problem of compressor efficiency being difficult to obtain by constructing a mathematical model of compressor efficiency and using the collected direct measurement data to calculate the compressor efficiency, providing a reliable quantitative indicator for evaluating compressor performance. The compressor efficiency time series is used as an important input feature for subsequent model learning of the nonlinear relationship between clearance changes and outlet pressure, breaking through the limitation of traditional data-driven models that cannot model structural parameters. It is a key bridge connecting structural parameter changes with outlet pressure prediction. If this data is missing, the model will not be able to reflect the impact of structural parameters such as clearance changes on outlet pressure, resulting in a significant decrease in the model's predictive ability and accuracy.

[0081] Step S1130 : The first healthy operating condition data and the second healthy operating condition data are used to form an offline healthy data set for constructing an offline benchmark model for outlet pressure prediction.

[0082] Specifically, the first health condition data and the second health condition data are integrated in chronological order to form a data set containing multiple time series features. Since the dimensions and value ranges of different data are different, in order to eliminate the dimensional effect and improve the convergence speed and model training effect of the subsequent LSTM network, a normalization method (such as minimum-maximum normalization or Z-score normalization) is used to process the data and scale each data feature to an appropriate range (such as the range of 0-1). In the prior art, multi-source data often have dimensional differences. If directly input into the model, it may lead to model training difficulties, slow convergence speed and low prediction accuracy. Through the data integration and normalization processing in this step, the problem of inconsistent data dimensions is solved, data of different dimensions are made comparable, and the training efficiency and prediction accuracy of the model are improved. The normalized data is the basis for constructing an offline benchmark model for outlet pressure prediction, which provides a guarantee for the subsequent model to effectively learn the relationship between data features and outlet pressure. If the data normalization step is missing, the model may not converge effectively, and the mapping relationship between features and outlet pressure cannot be accurately established, resulting in the failure of the entire model building process.

[0083] Step S1200: constructing an offline benchmark model for outlet pressure prediction based on an offline health data set;

[0084] Specifically, nine feature types—compressor inlet temperature, inlet pressure, IGV angle, VGV angle, output power, turbine cooling air valve position, compressor operating time, water wash status indicator, and compressor efficiency—are sampled over 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 3D 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 outlet pressure prediction uses a two-layer LSTM neural network structure, which is specifically composed of:

[0086] Input layer: Receives a three-dimensional tensor. 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 layer (including the weight matrix and bias vector) effectively captures long-term temporal dependencies through a mechanism involving forget gates, input gates, and output gates, such as the impact of compressor fouling accumulation over time on efficiency. The LSTM layer's weight matrix and bias vector are used to calculate the output of each state gate. The weight matrix learns the nonlinear relationship between features, and the bias vector adjusts the neuron activation threshold.

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

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

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

[0091] In existing technologies, gas turbine compressor outlet pressure 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 time series data. Step S1200 uses nine features, including directly measured temperature, pressure, valve opening, and calculated compressor efficiency, as input. These features cover the physical processes of compressor operation (such as compression power consumption, airflow regulation, and maintenance interventions). Compared to single-feature modeling, this improves the model's ability to characterize complex operating conditions. For example, the turbine cooling air valve position display time series reflects the impact of cooling air flow on compressor load. Combined with the compressor efficiency time series, it can quantify the energy coupling between the cooling system and the compression process. Using memory cells and a gating mechanism, the LSTM network can distinguish short-term load fluctuations (such as transient changes in output power) from long-term performance degradation (such as efficiency reduction due to fouling), avoiding the vanishing gradient problem of traditional recurrent neural networks (RNNs). For example, using compressor operating time as a time series feature input, the LSTM can learn the pattern that "longer operating time leads to more significant efficiency decline and greater deviation of outlet pressure from the baseline." The offline benchmark model is trained based on the healthy data of the gas turbine at the initial stage of leaving the factory, and learns the "healthy benchmark" rules in the state without performance degradation, such as the positive correlation between the pressure ratio and the compressor efficiency, the negative correlation between the actual temperature rise and the compressor efficiency, etc. These rules serve as common knowledge across load levels, providing a basis for the migration of offline sub-model parameters in the subsequent step S1400, reducing the cost of repeated training. If there is no offline benchmark model as the starting point for transfer learning, the offline sub-model of step S1400 needs to be trained from scratch, requiring a large amount of data support, and it is difficult to capture the common physical laws across loads, resulting in slow model convergence and a high risk of overfitting. It will also cause the residual calculation (online prediction value - offline benchmark value) of the subsequent step S3000 to lose the "healthy benchmark" reference, and it will not be able to effectively distinguish between normal operating fluctuations and performance degradation, which may lead to false or missed warnings.

[0092] Step S1300 , dividing the load of the gas turbine into n1 load levels, and filtering the offline data subset of each load level from the offline healthy data set;

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

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

[0095] The maximum output power P max The rated full load power P of the gas turbine rated Methods for comparing and determining load levels 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<00000​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​Step S1300 divides the gas turbine load into n1 levels and selects offline data subsets for each level. This addresses the existing problem of "load variations leading to insufficient model generalization" by refining load grading, enabling subsequent models to model specific operating conditions, improving prediction accuracy. In the prior art, gas turbine aerodynamic characteristics vary significantly under different loads (e.g., compressor surge is more likely at low load, while efficiency approaches design values ​​at full load). This makes 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 learning features within a specific load range. For example, the IGV angle is more sensitive to outlet pressure regulation at low load, while the turbine cooling air valve position has a more significant coupling effect on load at full load. Load-graded offline sub-models 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 mixing data from different loads, which could lead to the model learning incorrect feature associations. For example, full-load data generally shows higher compressor efficiency (due to being close to design operating conditions). If mixed with low-load data, this could obscure the pattern of "decreased efficiency leading to lower outlet pressure." Compared to a single full-load model, load-graded offline submodels can be trained in parallel, reducing the number of parameters in each model and improving training speed. For example, the total number of parameters in the five load-graded offline submodels may be smaller than that of a single model. Furthermore, the input data for each submodel has greater operating condition homogeneity (e.g., the range of IGV angle values ​​at low load is narrower), resulting in a more concentrated feature distribution, thus reducing the computational complexity of model training. Because the IGV angle at low load is primarily used to prevent surge, its adjustment range is limited (e.g., varying only between 20° and 40°). Compared to full-load conditions (30° to 50°), the numerical fluctuation of this feature is smaller, eliminating the need for the submodel to learn nonlinear relationships that vary widely, thus reducing computational complexity. Furthermore, the load-graded offline data subsets contain only samples from specific power ranges, avoiding the discretization of feature distributions caused by mixing data from different loads, further improving computational efficiency.

[0107] Without step S1300, a single model would need to fit the complex characteristics of the entire load range. This could lead to overestimation of predictions at low loads due to overfitting high-load data, or increased errors at full load due to underfitting low-load characteristics. Performance degradation manifests differently under different loads (for example, fouling has a more pronounced impact on efficiency at low loads). Unclassified 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 , obtaining n1 offline outlet pressure prediction sub-models corresponding to n1 load levels based on the offline data subsets of each load level and the offline outlet pressure prediction benchmark model;

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

[0110] Step S1400 realizes the refined modeling of "general rule inheritance + specific load fine-tuning" through transfer learning and hierarchical training mechanism, solving the problem of "a single model cannot adapt to the characteristics of multiple load conditions" in the existing technology.

[0111] Furthermore, step S1400 includes:

[0112] Step S1410 , initializing the outlet pressure prediction offline sub-model corresponding to each load level based on the offline data subset of each load level;

[0113] Initializing the outlet pressure prediction offline sub-model corresponding to each load level includes:

[0114] Some parameters in the outlet pressure prediction offline benchmark model are copied to the outlet pressure prediction offline sub-model corresponding to each load level; the said some parameters include the weight matrix and bias vector located in the LSTM layer.

[0115] Specifically, after completing the load level division and filtering the offline data subsets for each level in step S1300, the corresponding offline sub-model is initialized for each load level (e.g., low load, partial load 1-3, full load) using cross-domain transfer learning technology. The hidden layer parameters (including the weight matrix and bias vector of the LSTM layer) of the outlet pressure prediction offline benchmark model (source domain model, trained based on full-load healthy 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 that capture the temperature-pressure-efficiency relationship in the benchmark model as initial parameters. The time series features of the offline data subsets of each load level (such as the compressor inlet temperature time series, IGV angle time series, and other 9-dimensional features) are sampled in a sliding window according to a set time window (e.g., 1 hour). A three-dimensional tensor of [number of samples, window length, number of features] is generated, which is consistent with the input format of the offline benchmark model to ensure the compatibility of parameter transfer. In the prior art, training models for different loads separately requires a large amount of data and cannot reuse common rules, resulting in high training costs and a high risk of overfitting. Step S1410 migrates the hidden layer parameters of the baseline model, allowing the sub-model to inherit the general physical laws learned by the source model, such as the "nonlinear relationship between pressure ratio and efficiency" and the "effect of temperature on compression power consumption," thereby reducing the training instability caused by random parameter initialization during cold start. For example, the full-load baseline model has learned the law that "every increase of 0.1 in pressure ratio increases efficiency by 2%." 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 requires a small amount of data at the target load to start training, solving the model construction problem when low-load operating condition data is scarce. Hidden layer parameter migration ensures that the sub-model inherits the basic physical laws of gas turbine operation, avoids "model fragmentation" caused by load grading, and improves prediction consistency across load conditions.

[0116] Step S1420: pre-training the outlet pressure prediction offline sub-model corresponding to each load level using the outlet pressure prediction offline benchmark model to obtain initial parameters of the outlet pressure prediction offline sub-model corresponding to each load level;

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

[0118] Source domain prediction error (MSE): The mean square error between the predicted value of the sub-model on the full load data and the actual outlet pressure, ensuring that the sub-model does not lose the core knowledge of the benchmark model.

[0119] Elastic Weight Consolidation Regularization (EWC): constrains the difference between sub-model parameters and baseline model parameters, and prevents the source domain knowledge from being forgotten during pre-training by calculating the parameter Fisher information matrix.

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

[0121] Directly training a sub-model using target load data can cause the model to deviate from general principles due to differences in data distribution (e.g., more frequent IGV angle adjustments under low load), leading to "catastrophic forgetting" (i.e., forgetting the fundamental physical relationships learned in the source domain). Pre-training uses a hybrid loss function to force the sub-model to consolidate the core parameters of the source domain before adapting to the target load. For example, during pre-training, the low-load sub-model must simultaneously fit the compression ratio-efficiency relationship of the full-load data to prevent its parameters from being overly biased towards the "low compression ratio-high IGV angle" characteristic unique to low loads, which could lead to inaccurate calculation of basic compression power consumption.

[0122] Pre-training allows the sub-model to inherit the knowledge of the source model while initially adapting to the target load's data distribution (e.g., the efficiency distribution range is lower at low loads). For example, through pre-training, the low-load sub-model adjusts parameters to accommodate the characteristic of "greater efficiency fluctuations at low pressure ratios" while maintaining the "positive correlation between efficiency and pressure ratio." EWC regularization limits the range of variation of key parameters to ensure that the sub-model does not subvert the learned physical laws due to target data noise (e.g., sensor errors). For example, it prevents the model from misinterpreting "occasional pressure fluctuations at low loads" as "negative correlation between pressure ratio and efficiency."

[0123] Step S1430 , using the offline data subset corresponding to the load level, specialized training is performed on the outlet pressure prediction offline sub-model corresponding to each load level with initial parameters to obtain the final parameters of the outlet pressure prediction offline sub-model corresponding to 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 square error in the target domain, and the pre-training knowledge is retained through EWC regularization. The specific process includes:

[0125] Feature-Specific Learning: Based on the characteristic distribution of the target load (e.g., turbine cooling air valve position is generally lower at low loads), the model weights relevant features. For example, the low-load sub-model can emphasize the impact of the IGV angle on the outlet pressure, as IGV opening is a key adjustment to prevent surge at low loads.

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

[0127] The offline baseline model and pre-training phase only provide general rules and cannot capture the differences in feature sensitivity under different loads. For example, the pressure ratio at full load is the dominant factor in efficiency, while the IGV angle and water wash status have a more significant impact on efficiency at low load. Specialized training reveals these specificities through target data, avoiding errors caused by "one-size-fits-all" modeling. For example, at low load, due to the low intake volume of a gas turbine, the outlet pressure can increase by 3kPa for every 5° increase in the IGV angle. This rule requires specialized training to be captured by the model. In response to 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 the storage of the real-time correlation between the IGV angle and pressure, resulting in lower prediction errors compared to the sub-model that has not been specially trained.

[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 high predicted pressure values, which cannot reflect the actual pressure attenuation at light loads (e.g., the actual pressure at low loads may be 10-15 kPa lower than the baseline value), resulting in delayed performance degradation monitoring. Data utilization efficiency is low. When mixed training is performed with different load data, the model needs to learn complex mappings across loads, resulting in parameter redundancy (such as fitting the rules of "low-load IGV regulation" and "full-load pressure ratio dominance" at the same time), increasing training time and significantly increasing the risk of overfitting. When the unit load changes frequently, a single model cannot quickly adapt to the operating condition switching. For example, when the load drops suddenly from full load to low load, the model may take several hours to reconverge. The hierarchical sub-model can call the corresponding model in real time through load judgment, and the response time is shortened to minutes.

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

[0130] Furthermore, step S2000 includes:

[0131] Step S2100: constructing an online sub-model for predicting outlet pressure corresponding to each load level based on the final parameters of the offline sub-model for predicting outlet pressure corresponding to each load level;

[0132] Specifically, step S2100 achieves a smooth transition from the "healthy baseline model" to the "real-time tracking model" through a parameter migration mechanism. The online sub-model is completely consistent with the offline sub-model in architecture, both using a two-layer LSTM neural network structure, including an input layer, a hidden layer (2 layers of LSTM, 64 neurons per layer), and a fully connected layer. The input feature dimension is the same as the offline sub-model (9-dimensional time series features, such as compressor inlet temperature, IGV angle, etc.), and the output is a compressor outlet pressure prediction sequence. 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 "ideal healthy state" performance law 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: Directly inherit the final parameters of the offline sub-model (including the hidden layer weight matrix, bias vector, and output layer parameters) as the starting point for incremental learning. For example, the LSTM layer weights of the low-load online sub-model directly copy the parameters of the low-load offline sub-model, avoiding training instability caused by random initialization from scratch.

[0135] In the existing technology, if the online model is trained from scratch, it needs a large amount of real-time data support, and it cannot use the prior knowledge in the historical health data, which leads to large prediction errors in the model in the early stage of unit operation. This step reduces the cold start cost and retains the health benchmark knowledge through parameter migration; it can quickly build a usable model without relying on a large amount of real-time data. For example, when the low-load operating data of a gas turbine is scarce, the online sub-model uses the offline sub-model parameters and only needs 100 real-time samples to achieve the training effect of 1,000 samples of the traditional method. The offline sub-model parameters contain the health status laws of the gas turbine in the early stage of leaving the factory (such as the pressure-efficiency relationship when there is no fouling and the blade gap is normal). After the online sub-model inherits these parameters, the real-time prediction value can be compared with the health benchmark, 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: "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 degradation rate of the equipment after aging. After a new unit is commissioned or overhauled, the offline sub-model parameters are closer to the true health 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 obtains real-time operating condition data for each load level of the gas turbine. Using this real-time operating condition data for each load level, incremental learning is used to update the parameters of the outlet pressure prediction online sub-model corresponding to each load level in real time, ultimately obtaining n1 outlet pressure prediction online sub-models corresponding to the n1 load levels. The n1 outlet pressure prediction online sub-models are: an outlet pressure prediction online sub-model for low load, an outlet pressure prediction online sub-model for part load 1, an outlet pressure prediction online sub-model for part load 2, an outlet pressure prediction online sub-model for part load 3, and an outlet pressure prediction online sub-model for full load.

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

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

[0140] Mean Squared Error (MSE): measures the online sub-model's performance on new data D′ (t) This ensures that the model adapts to real-time performance changes.

[0141] Elastic Weight Consolidation (EWC) regularization: This term constrains parameter fluctuations to prevent the model from forgetting historically learned baseline knowledge. Specifically, by calculating the Fisher information matrix of the parameters, it applies a larger penalty to key parameters (such as weights that influence the compression-efficiency relationship) to avoid wild fluctuations in parameters caused by real-time data noise.

[0142] The Adam optimizer is used to perform stochastic gradient descent optimization of the objective function and update the online sub-model parameters. For example, when new data indicates a decrease in compressor efficiency without a simultaneous decrease in outlet pressure, the model adjusts the weight associated with the IGV angle and pressure, while maintaining the stability of the pressure ratio-efficiency relationship through EWC regularization.

[0143] In existing technologies, offline models are unable to adapt to the performance degradation of gas turbines over time (such as blade wear and fouling accumulation), resulting in prediction errors that increase with operating time. This step uses an incremental learning mechanism to dynamically track performance degradation, balancing learning and memory. The continuous input of real-time data enables the model to capture slow changes such as efficiency decline and clearance increase. For example, when the compressor has been in operation for more than 5000 hours, the online sub-model automatically adjusts the weight of the "operating time" feature through incremental learning, increasing its influence on pressure prediction from 0.1 to 0.3, reflecting the increasing impact of aging on performance. The EWC regularization term ensures that the model does not disrupt established healthy baseline associations when learning new degradation features. For example, when learning the new rule that "fouling leads to efficiency degradation," the model retains historical knowledge that "water washing can restore efficiency," avoiding misinterpreting performance increases after maintenance as abnormal fluctuations. The offline sub-model provides a fixed healthy baseline prediction value, while the online sub-model provides a real-time prediction value. 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 predicts 95 kPa due to fouling. A residual error of 5 kPa triggers a degradation warning. Historical data from incremental learning can be fed back into the retraining of the offline sub-model (e.g., regularly updating the offline model with long-term data), forming a closed loop of "offline benchmarking - online tracking - data feedback." For example, quarterly degradation data accumulated through online learning can be used to fine-tune the offline sub-model and improve the adaptability of the baseline model.

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

[0145] First, offline models lack timeliness: Models trained initially at the factory cannot adapt to performance drift during operation, resulting in delayed monitoring. For example, after a unit has been in operation for a year, the prediction error of traditional offline models may increase from 3% to 10%, making it impossible to accurately identify early-stage degradation.

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

[0147] Step S2000 achieves efficient real-time modeling and dynamic degradation monitoring through an "offline benchmark + online incremental" mechanism. The online model is initialized using offline sub-model parameters, reducing the need for initial training data. Simultaneously, incremental learning gradually optimizes the model, enabling it to achieve high prediction accuracy early in its operation and improve over time. The real-time updated online model can respond to performance changes within minutes. For example, if compressor efficiency decreases by 0.5% per month, the online model can detect the increasing rate of change of the residual error within the first month, whereas traditional offline models would only recognize this after an efficiency decrease of more than 3% has accumulated.

[0148] Omitting step S2000 and using only the offline sub-model for online prediction will result in an inability to adapt to slow, time-varying degradation due to equipment aging and fouling, resulting in a continuous increase in residuals and a failure to trigger an early warning. The model will also be unable to cope with real-time operating fluctuations (such as sudden changes in ambient temperature), misinterpreting normal fluctuations as degradation, leading to an increased false alarm rate. The real-time predictions of the online sub-model and the baseline values ​​of the offline sub-model together form the inputs for residual calculation. 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 the residuals, avoiding false alarms. Incremental learning enables the model to capture early signs of efficiency degradation, shortening the warning time compared to traditional methods and creating a window for preventive maintenance. The combination of parameter transfer and incremental learning makes the online model computationally much less complex than that of a standalone training model, making it suitable for local deployment of embedded systems on gas turbines and avoiding the delays and costs of uploading data 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 shutdowns). For example, after a surge event, the model will not disrupt the fundamental relationship 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 the prior knowledge of offline modeling with real-time feedback from online data, systematically solving the timeliness and accuracy challenges in gas turbine performance monitoring.

[0149] Step S3000: Obtain an online prediction sequence of the outlet pressure of the current working condition based on the online sub-model of the outlet pressure prediction; obtain an offline prediction sequence of the outlet pressure of the current working condition based on the offline sub-model of the outlet pressure prediction; and perform a gas turbine performance degradation assessment and early warning based on the online prediction sequence of the outlet pressure and the offline prediction sequence of the outlet pressure.

[0150] See also Figure 3 As shown, further, step S3000 includes:

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

[0152] Furthermore, step S3100 includes:

[0153] Step S3110, obtaining current operating data of the gas turbine, and determining a current load level based on the current operating data;

[0154] Step S3120, obtaining a current compressor efficiency time series based on current operating data and a compressor efficiency mathematical model;

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

[0156] Specifically, the current operating data including the output power is obtained in real time, wherein the output power is read in real time by the unit control system to reflect the current load status. The maximum output power within the monitoring time window is extracted. For example, with a window of 1 hour, power data is collected every 10 minutes, and the maximum value of the 6 samples is taken. The current load level is determined according to the load level determination method in step S1300. The load grading matching model is used to reduce the outlet pressure prediction error, and the factor model parameters are more matched with the aerodynamic characteristics of the current load (such as low pressure ratio at low load and frequent IGV adjustment).

[0157] The compressor inlet temperature, outlet temperature, inlet pressure, and outlet pressure from the current operating data are input into the compressor efficiency mathematical model to calculate the current compressor efficiency. The current efficiency calculation values ​​are sorted by timestamp to form a compressor efficiency time series. Step S3120 integrates the compressor efficiency mathematical model and uses the collected temperature and pressure data to calculate the compressor efficiency in real time, solving the problem of "compressor efficiency cannot be observed in real time" and enabling the prediction model to promptly detect efficiency drops caused by blade wear, fouling, and other factors. The compressor efficiency time series forms a causal chain with the outlet pressure prediction. When the tip clearance increases by 0.1 mm, the compressor efficiency may drop by 1%. At this time, the control system maintains the outlet pressure stable by increasing the IGV angle, resulting in no significant change in the outlet pressure. However, the prediction model in the present invention detects performance degradation trends through the decline in compressor efficiency, while traditional methods that only monitor outlet pressure cannot identify early anomalies. Therefore, compared with traditional methods that rely solely on pressure data, the causal chain formed by the compressor efficiency time series and outlet pressure prediction in the present invention can shorten the warning time. The present invention indirectly reflects the influence of unmeasurable parameters such as tip clearance through changes in compressor efficiency. For example, when the increase in clearance causes the efficiency to drop by 0.5% / month, the outlet pressure prediction online sub-model of the present invention can be identified by the slope change of the compressor efficiency time series, while the traditional method requires periodic disassembly and inspection.

[0158] Based on the load level determined by S3110, the corresponding model is selected from the 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 invoked. The current operating data and the compressor efficiency time series calculated by S3120 are combined to generate a sliding window sequence based on 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 through the LSTM network, a series of outlet pressure predictions for a future period (e.g., 30 minutes) is output.

[0159] The traditional single model has a high error rate when predicting across loads because it cannot adapt to the aerodynamic characteristics of different working conditions (such as low-load surge risk and full-load efficiency saturation). This step improves the prediction error of each load interval by matching the load dynamic judgment with the model. Existing methods mostly rely on offline efficiency calculations or single pressure parameters and cannot reflect real-time performance changes in a timely manner. Step S3100 calculates the compressor efficiency in real time and inputs it into the prediction model so that the outlet pressure prediction can synchronously reflect the efficiency decline. For example, when the compressor efficiency decreases at a rate of 0.2% / day due to fouling, the model can detect the increase in prediction residuals within 3 days, while the traditional method requires an efficiency drop of more than 1% to be recognized.

[0160] If the load level judgment is missing, the full-load model may overestimate the pressure ratio at low load, resulting in a high predicted pressure, which may be mistakenly judged as a healthy state. However, the actual pressure may be too low due to surge risk, resulting in missed reporting of decline. If the efficiency is not calculated in real time, the model cannot distinguish between "pressure changes caused by load fluctuations" and "pressure drops caused by efficiency decline". For example, pressure fluctuations caused by increased ambient temperature may be mistakenly judged as fouling decline, increasing the false alarm rate. In summary, step S3100 systematically improves the real-time and accuracy of the gas turbine compressor outlet pressure prediction through dynamic load judgment, real-time efficiency calculation and precise model calling.

[0161] Step S3200, obtaining an offline prediction sequence of the outlet pressure of the current operating condition based on the current operating data, the current compressor efficiency time series, and the outlet pressure prediction offline sub-model corresponding to the current load level;

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

[0163] The outlet pressure prediction offline sub-model is completed through offline health data training in step S1400, and its parameters (such as LSTM hidden layer weights and bias vectors) are fixed and unchanged, representing the outlet pressure law when the gas turbine has no performance degradation under the corresponding load level. The 9-dimensional features such as compressor inlet temperature, inlet pressure, IGV angle in the current operating data and the current compressor efficiency time series calculated in step S3120 are formatted according to the time window (such as 1 hour) during offline sub-model training to generate a three-dimensional tensor ([number of samples, window length, number of features]) in the same format as the training data. The formatted three-dimensional tensor is input into the offline sub-model, and the long-term time-dependent features (such as fouling trend) are extracted through the hidden layer calculation of the LSTM network, and then the outlet pressure offline prediction sequence is output through the fully connected layer.

[0164] In the prior art, monitoring systems often use a single fixed value or historical average value as a benchmark, which cannot adapt to load changes and equipment aging. This step provides a load-specific benchmark through an offline sub-model. For example, the benchmark pressure output by the low-load offline sub-model takes into account the characteristics of the large IGV angle and low pressure ratio under this load, avoiding misjudging the normal low pressure as a recession; the full-load sub-model is based on the design operating condition data, reflecting the pressure value under the ideal state, which is used to compare the degree of deviation from actual operation. The fixed parameter characteristics of the offline sub-model make its output unaffected by real-time noise, and can effectively distinguish between "normal operating condition fluctuations" and "pressure anomalies caused by recession." For example, when the ambient temperature rises and the inlet temperature rises, the offline sub-model still outputs a pressure benchmark value based on the 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 recession signal after temperature compensation.

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

[0166] Specifically, the outlet pressure residual is the difference between the online predicted value of the outlet pressure and the offline predicted value of the outlet pressure at the same time step. The positive or negative residual reflects the direction of pressure deviation, and the absolute value reflects the degree of deviation. The residual sequence converts performance degradation into a measurable numerical indicator, solving the subjective problem of traditional methods relying on manual experience judgment. For example: when the absolute value of the residual is greater than 5kPa, it indicates that the pressure deviates from the healthy baseline beyond the allowable range, and there may be fouling or blade wear; the residual change rate is greater than 0.3kPa / hour, indicating that the degradation rate is accelerating and urgent maintenance is required. The type of degradation can be identified by the residual change pattern, including reversible degradation and irreversible degradation; reversible degradation, such as fouling caused by not washing with water, the residual should return to the baseline value after washing with water, otherwise it indicates irreversible damage (such as blade corrosion); irreversible degradation, such as increased clearance due to blade wear, the residual continues to increase and cannot be restored by washing with water, and it is necessary to determine whether to trigger a warning for replacing parts based on the residual trend.

[0167] Step S3400: Determine whether a compressor performance degradation warning is triggered based on the outlet pressure residual sequence.

[0168] Furthermore, step S3400 includes:

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

[0170] Step S3420, performing trend prediction on the smoothed residual change curve to estimate the residual change rate;

[0171] Step S3430 : Compare the residual change rate with a preset degradation speed threshold to determine whether a gas turbine performance degradation warning needs to be triggered.

[0172] Specifically, step S3400 solves the problem of "degradation warning lag and high false alarm rate" in the prior art through noise suppression, trend analysis and threshold determination. The Kalman filter algorithm is used to suppress the noise of the outlet pressure residual sequence. The residual sequence is composed of the difference between the online prediction value and the offline reference value, which contains interference such as sensor measurement noise (such as the temperature sensor ±0.5℃ error transmitted to the pressure residual), control system adjustment fluctuations (such as pressure transients caused by dynamic adjustment of the IGV angle), etc. Kalman filtering achieves noise filtering through the following recursive process:

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

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

[0175] Example: If the measured values ​​of the residual sequence within 1 hour are [-2kPa, -1.8kPa, -2.1kPa, -1.9kPa], the Kalman filter iteratively calculates and 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 trend misjudgment. For example, occasional sensor pulse noise can cause sudden fluctuations in residuals, which traditional methods may mistakenly interpret as a decay signal. Step S3410 uses Kalman filtering to reduce the noise variance, allowing the residual curve to more accurately reflect the decay trend.

[0177] Step S3420 uses the least squares method to perform a first-order polynomial fit on the filtered residual sequence and extracts the trend slope as the residual change rate. Existing technologies often calculate the change rate through fixed window differences, which is easily affected by the window size and cannot identify nonlinear trends. This step can adaptively capture linear trends through least squares fitting. For example, when the residual change rate gradually decreases from -0.05kPa / h to -0.1kPa / h (the absolute value increases from 0.05kPa / h to 0.1kPa / h, and the decay speed accelerates), the fitting slope can reflect the accelerated decay trend in real time. Step S3420 upgrades the degree of decay from a "qualitative judgment" to a "quantitative indicator", which facilitates the formulation of maintenance priorities. Compared with the traditional difference method, the least squares method can detect trend changes in advance because the fitting process utilizes the entire sequence data, rather than just the difference between the previous and subsequent periods.

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

[0179] The decay rate threshold is determined according to 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 low-load operating threshold is set to -0.08 kPa / h (allowing slower decay); the full-load operating threshold is set to -0.05 kPa / h (more sensitive to decay).

[0180] Existing technologies often use a unified threshold value without considering load differences. For example, the compressor efficiency is more sensitive to pressure changes at full load, and the same decay rate is more harmful at full load, requiring a stricter threshold value. This step uses load classification thresholds to make the warning logic more in line with actual operating conditions. Under full-load conditions, the efficiency drop caused by decay may cause surge risk. A threshold with a smaller absolute value (-0.05kPa / h) achieves a stricter warning standard, can provide early warning, and respond earlier than the unified threshold; a threshold with a larger absolute value (-0.08kPa / h) is allowed at low load to avoid false alarms caused by normal efficiency fluctuations. The negative sign represents the trend of the residual change (the direction in which the pressure deviates from the healthy baseline), and the numerical value represents the absolute value of the rate.

[0181] Step S3400 converts the residual sequence from "noisy raw data" to "decision-making recession indicators" through the three-stage processing of "filtering-fitting-comparison", solving the problem in the existing technology that early warning relies on manual experience and lacks quantitative basis. For example, when a unit is operating at partial load 2, the residual sequence shows a slope of -0.06kPa / h after filtering, and its absolute value exceeds the absolute value of the load threshold -0.05kPa / h, triggering a secondary warning, indicating the risk of fouling or gap increase, while traditional methods may not identify the trend due to noise interference. Without filtering and trend analysis, residual noise may mask the true recession signal. For example, due to blade wear, the residual of a unit slowly decreases at -0.03kPa / h. The noise fluctuation of ±0.2kPa will completely submerge the trend, and the traditional method cannot identify it until the pressure drops sharply in the late stage of recession, resulting in maintenance delays. Without a quantified residual rate of change, operators must rely on empirical judgment, which can lead to false alarms and missed alarms. Short-term pressure fluctuations can be misinterpreted as recession, resulting in unnecessary maintenance downtime. Ignoring slow recession trends can lead to serious equipment damage (such as blade breakage). Early warning can shift maintenance plans from "post-event repairs" to "preventive maintenance," reducing unplanned downtime. For example, a power plant used this step to identify compressor fouling in advance and complete cleaning during planned maintenance, avoiding power generation losses caused by sudden downtime. Early intervention in recession problems (such as timely water washing or gap adjustment) can extend the life of key compressor components and avoid severe wear in the later stages of recession. Multi-stage processing (filtering + fitting + thresholding) makes the system robust to sensor failures and short-term interference. For example, if a temperature sensor fails briefly and causes a sudden change in the residual, the Kalman filter can suppress this interference through state prediction and avoid false warning triggering.

[0182] Example 2:

[0183] This embodiment provides a gas turbine compressor outlet pressure online monitoring system based on embodiment 1, such as Figure 4 Shown, including:

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

[0185] Online prediction model construction module: Based on the offline sub-model for outlet pressure prediction corresponding to each load level, an online sub-model for outlet pressure prediction corresponding to each load level is established through incremental learning;

[0186] Degradation assessment module: Based on the outlet pressure prediction online sub-model, the outlet pressure online prediction sequence of the current operating condition is obtained; based on the outlet pressure prediction offline sub-model, the outlet pressure offline prediction sequence of the current operating condition is obtained; based on the outlet pressure online prediction sequence and the outlet pressure offline prediction sequence, the gas turbine performance degradation assessment and early warning are carried out.

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

[0188] Step S1410 , initializing the outlet pressure prediction offline sub-model corresponding to each load level based on the offline data subset of each load level;

[0189] Step S1420: pre-training the outlet pressure prediction offline sub-model corresponding to each load level using the outlet pressure prediction offline benchmark model to obtain initial parameters of the outlet pressure prediction offline sub-model corresponding to each load level;

[0190] Step S1430 , using the offline data subset corresponding to the load level, specialized training is performed on the outlet pressure prediction offline sub-model corresponding to each load level with initial parameters to obtain the final parameters of the outlet pressure prediction offline sub-model corresponding to each load level.

[0191] In the degradation assessment module, the method for evaluating and warning gas turbine performance degradation based on the outlet pressure online prediction sequence and the outlet pressure offline prediction sequence includes:

[0192] According to the outlet pressure online prediction sequence and the outlet pressure offline prediction sequence, the outlet pressure residual sequence is obtained;

[0193] Filter the outlet pressure residual sequence to obtain a smooth residual change curve;

[0194] Carry out trend forecast on the smooth residual change curve and estimate the residual change rate;

[0195] The residual change rate is compared with the preset degradation speed threshold to determine whether it is necessary to trigger a gas turbine performance degradation warning.

[0196] The methods and systems of the present application may be implemented in many ways. For example, the methods and systems of the present application may be implemented using software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps used in the method is for illustration only, and the steps of the method of the present application are not limited to the order specifically described above unless otherwise specified.

[0197] In addition, the parts of the above technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0198] The above-described specific embodiments further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above description is merely a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for online monitoring of gas turbine compressor outlet pressure, characterized in that: The method comprises: Obtain an offline health dataset of the gas turbine at the start of operation and construct an offline benchmark model for outlet pressure prediction based on the offline health dataset. Divide the gas turbine load into n1 load levels and, based on the offline benchmark model for outlet pressure prediction, establish an offline sub-model for outlet pressure prediction corresponding to each load level. Based on the offline sub-model for outlet pressure prediction corresponding to each load level, an online sub-model for outlet pressure prediction corresponding to each load level is established through incremental learning; Based on the outlet pressure online prediction sub-model, the outlet pressure online prediction sequence of the current operating condition is obtained; based on the outlet pressure offline prediction sub-model, the outlet pressure offline prediction sequence of the current operating condition is obtained; based on the outlet pressure online prediction sequence and the outlet pressure offline prediction sequence, the gas turbine performance degradation assessment and early warning are carried out.

2. The method for online monitoring of gas turbine compressor outlet pressure according to claim 1, characterized in that: The offline health data set includes first health condition data directly measured under different loads and second health condition data calculated; the first health condition data at least includes output power time series; The second healthy operating condition data includes a compressor efficiency time series.

3. The method for online monitoring of gas turbine compressor outlet pressure according to claim 2, characterized in that: The method for calculating the compressor efficiency time series includes: constructing a compressor efficiency mathematical model, obtaining compressor efficiency based on first healthy operating condition data and the constructed compressor efficiency mathematical model; and sorting the compressor efficiency by timestamp to obtain a compressor efficiency time series.

4. The method for online monitoring of gas turbine compressor outlet pressure according to claim 3, characterized in that: The outlet pressure prediction offline benchmark model adopts a two-layer LSTM neural network structure, including an input layer, a hidden layer and a fully connected layer; the hidden layer includes two LSTM layers.

5. The method for online monitoring of gas turbine compressor outlet pressure according to claim 4, characterized in that: The method for dividing the load of the gas turbine into n1 load levels includes: Extract the maximum output power P within a given monitoring time window from the output power time series of the offline health dataset max , the maximum output power P max The rated full load power P of the gas turbine rated Make a comparison and determine the load level.

6. The method for online monitoring of gas turbine compressor outlet pressure according to claim 5, characterized in that: The method for establishing an outlet pressure prediction offline sub-model corresponding to each load level based on the outlet pressure prediction offline benchmark model includes: Filter offline data subsets of each load level from the offline health data set; Based on the offline data subset of each load level, the outlet pressure prediction offline sub-model corresponding to each load level is initialized; The outlet pressure prediction offline benchmark model is used to pre-train the outlet pressure prediction offline sub-model corresponding to each load level to obtain the initial parameters of the outlet pressure prediction offline sub-model corresponding to each load level; The offline data subset corresponding to the load level is used to perform specialized training on the outlet pressure prediction offline sub-model corresponding to each load level with initial parameters, and the final parameters of the outlet pressure prediction offline sub-model corresponding to each load level are obtained.

7. The method for online monitoring of gas turbine compressor outlet pressure according to claim 6, characterized in that: The method for establishing an online sub-model for outlet pressure prediction corresponding to each load level includes: Based on the final parameters of the offline sub-model for outlet pressure prediction corresponding to each load level, an online sub-model for outlet pressure prediction corresponding to each load level is constructed; Real-time operating condition data of the gas turbine at each load level is obtained. Using the real-time operating condition data of each load level, the parameters of the outlet pressure prediction online sub-model corresponding to each load level are updated in real time through incremental learning, and finally n1 outlet pressure prediction online sub-models corresponding to n1 load levels are obtained.

8. The method for online monitoring of gas turbine compressor outlet pressure according to claim 7, characterized in that: The method for obtaining the online prediction sequence of the outlet pressure of the current working condition includes: obtaining current operating data of the gas turbine, and determining a current load level based on the current operating data; According to the current operating data and the compressor efficiency mathematical model, the current compressor efficiency time series is obtained; Based on the determined current load level, the corresponding outlet pressure prediction online sub-model is called, and the outlet pressure online prediction sequence is obtained according to the current operating data, the current compressor efficiency time series and the called corresponding outlet pressure prediction online sub-model.

9. The method for online monitoring of gas turbine compressor outlet pressure according to claim 8, characterized in that: The method for evaluating and providing early warning for gas turbine performance degradation based on an outlet pressure online prediction sequence and an outlet pressure offline prediction sequence includes: According to the outlet pressure online prediction sequence and the outlet pressure offline prediction sequence, the outlet pressure residual sequence is obtained; based on the outlet pressure residual sequence, it is determined whether the compressor performance degradation warning is triggered.

10. A gas turbine compressor outlet pressure online monitoring system, which is used to implement the gas turbine compressor outlet pressure online monitoring method according to any one of claims 1 to 9, characterized in that: The system comprises: Offline benchmark model construction module: This module is used to obtain an offline health dataset of the gas turbine just after it is put into operation and to construct an offline benchmark model for outlet pressure prediction based on the offline health dataset. The gas turbine load is divided into n1 load levels and, based on the outlet pressure prediction offline benchmark model, an offline sub-model for outlet pressure prediction corresponding to each load level is established. Online prediction model construction module: Based on the offline sub-model for outlet pressure prediction corresponding to each load level, an online sub-model for outlet pressure prediction corresponding to each load level is established through incremental learning; Degradation assessment module: Based on the outlet pressure prediction online sub-model, the outlet pressure online prediction sequence of the current operating condition is obtained; based on the outlet pressure prediction offline sub-model, the outlet pressure offline prediction sequence of the current operating condition is obtained; based on the outlet pressure online prediction sequence and the outlet pressure offline prediction sequence, the gas turbine performance degradation assessment and early warning are carried out.

Images