An artificial intelligence gas turbine dynamic parameter regression prediction method

Abstract

The application relates to the technical field of gas turbine parameter analysis, and discloses an artificial intelligence gas turbine dynamic parameter regression prediction method. The method comprises the following steps: collecting gas turbine operation parameters and performing data cleaning, abnormal value processing, missing value completion and Z-score standardization to obtain a pretreatment data set. Compound features, time window features, frequency domain features and correlation features are constructed based on the data set, and a key parameter set is screened out through feature selection. A hierarchical deep neural network is used for layer-by-layer deduction to obtain a dynamic parameter prediction result. A multi-model integration and a meta-learner are used to dynamically adjust weights, combined with uncertainty quantification analysis, and finally a differential update strategy is used to optimize a prediction model. Through dynamic multi-model integration and optimization, the prediction model is updated and adjusted in real time, so that the prediction accuracy is improved and the gas turbine complex working condition change is adapted, which is a key technical problem to be solved by the application.

Description

Technical Field

[0001] This application relates to the field of gas turbine parameter analysis technology, and in particular to an artificial intelligence-based method for regressing and predicting dynamic parameters of gas turbines. Background Technology

[0002] Currently, gas turbines are widely used in power, aviation, and shipbuilding industries, playing a particularly important role in the power and energy sectors. With the increasing automation and intelligence of industry, monitoring and predicting the operation of gas turbines has become crucial for improving their performance and reliability. To ensure efficient operation of gas turbines, various operating parameters, such as temperature, pressure, speed, and vibration, are typically collected in real time and combined with traditional monitoring methods for condition assessment and fault diagnosis. Existing technologies rely on prediction methods based on physical models and empirical data, usually employing expert systems, simplified thermodynamic models, or statistical methods for prediction and monitoring. While these methods can provide predictions to some extent, they generally suffer from insufficient accuracy, poor real-time performance, and an inability to accurately adapt to complex and variable operating conditions.

[0003] However, existing traditional prediction techniques have significant shortcomings. On the one hand, these methods typically rely on relatively simple linear assumptions or empirical models, failing to fully capture the nonlinear characteristics of gas turbine operation, leading to large prediction errors in complex operating environments. On the other hand, with the increasing volume of equipment data, traditional methods struggle to efficiently process large amounts of real-time data, making it impossible to make timely and accurate predictions under rapidly changing operating conditions. Furthermore, existing technologies also have significant problems in providing early warnings of future faults and adapting to changes in operating conditions, making it difficult to effectively predict performance under different load conditions during gas turbine operation. Therefore, improving prediction accuracy and real-time performance, and effectively addressing the impact of changing operating conditions on prediction accuracy, has become an urgent technical challenge. Summary of the Invention

[0004] This application provides an artificial intelligence-based method for regressing and predicting dynamic parameters of gas turbines. This method uses dynamic multi-model integration and optimization to update and adjust the prediction model in real time, thereby improving prediction accuracy and adapting to complex operating conditions of gas turbines. This is the key technical problem that this invention aims to solve.

[0005] This application provides an artificial intelligence-based method for regressing and predicting dynamic parameters of a gas turbine. The method includes: collecting various operating parameters of the gas turbine; obtaining a preprocessed dataset through data cleaning, outlier handling, missing value completion, and Z-score standardization; constructing composite features, time window features, frequency domain features, and correlation features based on the preprocessed dataset; obtaining a key parameter set through a multi-step feature selection process; inputting the key parameter set into a hierarchical deep neural network for processing; obtaining dynamic parameter prediction results through layer-by-layer derivation of a basic parameter prediction layer, an intermediate state estimation layer, and a target parameter regression layer; performing multi-model integration processing on the dynamic parameter prediction results; dynamically adjusting the weights of each model by constructing a meta-learner to obtain an integrated prediction value; performing uncertainty quantification analysis based on the integrated prediction value; obtaining a prediction confidence interval through Monte Carlo Dropout, a Bayesian neural network framework, and time series scenario simulation; and obtaining a continuously optimized prediction model by triggering a differentiated update strategy through a multi-criteria judgment mechanism based on the comparison between the prediction confidence interval and the actual observed values.

[0006] The technical solution provided in this application employs a sliding window approach to manage real-time gas turbine operating data, ensuring that each prediction is based on the latest 12-hour data. This design effectively captures the dynamic characteristics of the gas turbine, guaranteeing the timeliness of predictions and the consistency of data sequence. By labeling and classifying the processed data according to different operating conditions (such as start-up, full-load, and load-change conditions), this invention enables the prediction model to more accurately adapt to changes under different operating conditions, thereby improving the model's adaptability and prediction accuracy to different gas turbine operating states. This data processing method based on operating condition classification effectively improves the model's robustness in complex environments and reduces prediction bias caused by changes in operating conditions. A multi-criteria judgment mechanism monitors the deviation between the prediction confidence interval and the actual observed value, and strict trigger update conditions are set, such as the actual observed value falling outside the prediction interval for five consecutive times, or the moving average of the prediction deviation showing a monotonically increasing trend for three consecutive days. This ensures that when a prediction deviation occurs, an update signal can be triggered in a timely manner. Through this intelligent deviation monitoring and update mechanism, this invention can detect anomalies in model predictions in a short time, thereby ensuring that the model can always adapt to the real-time operating conditions of the gas turbine. The introduction of this mechanism not only improves prediction accuracy but also enables the system to respond more flexibly to various dynamic changes during actual operation. A differentiated update strategy is adopted, employing different optimization measures such as incremental parameter updates, partial sub-model retraining, and full model retraining for mild, moderate, and severe performance degradation. This flexible update strategy allows the model to quickly make adaptive adjustments when facing different degrees of performance degradation, avoiding the waste of computational resources caused by large-scale retraining, while also ensuring efficient model updates and accuracy. By executing model updates in an independent computing environment using a parallel computing framework, this invention ensures that real-time prediction tasks are unaffected, further improving computational efficiency and the system's real-time response capability during model updates. Indicators such as Mean Absolute Percentage Error (MAPE) and prediction interval coverage are used to comprehensively evaluate the model's accuracy and reliability. If performance degrades, the system can automatically roll back to the previous stable version, ensuring that the system is always in an optimal state. This automatic rollback mechanism effectively avoids inaccurate predictions caused by degraded model performance, further improving the stability and reliability of the system. Attached Figure Description

[0007] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is a schematic diagram of an embodiment of the artificial intelligence-based gas turbine dynamic parameter regression prediction method in this application. Detailed Implementation

[0009] This application provides an artificial intelligence-based method for regressing and predicting dynamic parameters of gas turbines. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0010] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the artificial intelligence-based gas turbine dynamic parameter regression prediction method in this application includes: Step S101: Collect various operating parameters of the gas turbine, and obtain a preprocessed dataset through data cleaning, outlier handling, missing value completion, and Z-score standardization. Specifically, the gas turbine's inlet temperature, inlet pressure, exhaust temperature, exhaust pressure, speed, power output, fuel flow rate, cooling water temperature, vibration value, and bearing temperature are collected at a fixed frequency to obtain raw data. Data points exceeding physical limits in the raw data are automatically marked and deleted, and confirmed outliers are replaced using a local linear regression method to obtain preliminary cleaned data. Missing values ​​in the preliminary cleaned data are processed using a corresponding completion algorithm based on the missing time span to obtain complete data. For the portion of the complete data missing within 3 sampling periods, cubic spline interpolation is used for completion; for the portion missing within 3 but not more than 10 sampling periods, a multivariate regression model based on similar operating conditions is used for completion to obtain completed data. The completed data is processed using the Z-score standardization method to convert each parameter into a standard distribution with a mean of 0 and a standard deviation of 1 to obtain standardized data. The standardized data is then divided into training, validation, and test sets according to a ratio of 70%, 15%, and 15% to obtain the preprocessed dataset.

[0011] The system comprehensively collects various operating parameters of the gas turbine, including key parameters such as inlet air temperature, inlet air pressure, exhaust air temperature, exhaust air pressure, speed, power output, fuel flow rate, cooling water temperature, vibration level, and bearing temperature. To ensure data accuracy and comprehensiveness, the data acquisition frequency is dynamically adjusted according to the specific operating state of the gas turbine. Under stable operating conditions, the acquisition frequency is 10 times per second; during non-steady-state processes such as startup, shutdown, or load changes, the acquisition frequency increases to 50 times per second to capture subtle changes in parameters. The data acquisition system adopts a distributed architecture and installs high-precision sensors in key locations to ensure the reliability of data acquisition, while redundant design increases the system's fault tolerance. The next stage is data preprocessing, which mainly includes data cleaning, outlier handling, missing value completion, and data standardization. During data cleaning, the system automatically marks and deletes outlier data points that clearly exceed physical limits. For suspicious data points, the system performs a rationality check based on historical data and the physical relationship between relevant parameters to confirm whether they are outliers. If confirmed as outliers, they are replaced using a local linear regression method. For handling missing values, the system selects different imputation algorithms based on the time span of the missing data. When the missing range is small, interpolation is used for completion; when there are many missing values, a regression model based on similar operating conditions is used. When missing data exceeds a certain period, the data for that period is marked as invalid.

[0012] After cleaning and imputation, the data undergoes standardization to eliminate the influence of differences in parameter units, ensuring all parameters can be analyzed on the same scale. The standardized dataset is then divided into training, validation, and test sets, with proportions of 70%, 15%, and 15%, respectively. This stage lays the foundation for subsequent feature engineering and the construction of deep learning models.

[0013] It should be noted that for a single-shaft gas turbine with a rated power of 100kW, the following parameters are collected at a frequency of 10 times per second during stable operation: Environmental parameters: ambient temperature (25℃), ambient pressure (101.3kPa); Core operating conditions: compressor inlet / outlet temperature and pressure (inlet 25℃ / outlet 200℃), turbine speed (15000r / min), fuel flow rate (0.05kg / s), output power (95kW); Auxiliary monitoring: bearing temperature (85℃), vibration value (4.5mm / s); During non-steady-state processes (start-up phase), the collection frequency is increased to 50 times per second to capture dynamic changes. For outlier handling, if data significantly exceeds physical limits (bearing temperature 200℃), the system automatically marks and deletes it. For suspicious values ​​(vibration value momentarily spikes to 15mm / s but quickly falls back), local linear regression is used to correct it to a reasonable range (corrected to 5.2mm / s). Missing value completion: For short-term missing data (≤3 seconds): cubic spline interpolation is used. If temperature data is missing within a certain second, interpolation is performed based on the data from the preceding and following seconds (24.8℃, 25.1℃) to arrive at 25.0℃. For long-term missing data (3~10 seconds): a multiple regression model under similar operating conditions is used. Under the same engine speed and ambient temperature, historical data is used to estimate the missing fuel flow rate (completed to 0.048 kg / s). Data segments missing for more than 10 seconds will be completely removed.

[0014] Data standardization and partitioning: Z-score standardization was applied to the complete data to make the mean of each parameter 0 and the standard deviation 1 (the original temperature of 25℃ was converted to -0.3, and the pressure of 101.3kPa was converted to 0.1). The standardized data was then randomly partitioned into 70% training set, 15% validation set, and 15% test set to generate a preprocessed dataset.

[0015] Step S102: Based on the preprocessed dataset, construct composite features, time window features, frequency domain features, and correlation features, and filter them through a multi-step feature selection process to obtain a set of key parameters; Specifically, based on the thermodynamic theory of gas turbines, the original parameters in the preprocessed dataset are combined and calculated to construct composite features of compression ratio, expansion ratio, isentropic efficiency, and thermal efficiency, resulting in a composite feature set. For the time-series data in the preprocessed dataset, the moving average, standard deviation, rate of change, and cumulative value for different time windows are constructed, resulting in a time window feature set. Fast Fourier transform is performed on the key vibration parameters and pressure fluctuation data in the preprocessed dataset to extract the power spectral density, main frequency components, and their amplitudes, resulting in a frequency domain feature set. The Pearson correlation coefficient, mutual information value, and partial correlation coefficient among the key parameters in the preprocessed dataset are calculated, resulting in a correlation feature set. Analysis of variance is used to initially screen the composite feature set, time window feature set, frequency domain feature set, and correlation feature set, eliminating features with variance close to zero, resulting in a preliminary feature set. The preliminary feature set is further screened and dimensionality reduced using LASSO regression, random forest feature importance assessment, recursive feature elimination, and principal component analysis, resulting in the key parameter set.

[0016] Based on the thermodynamic theory of gas turbines, composite features such as compression ratio, expansion ratio, isentropic efficiency, and thermal efficiency were constructed by combining calculations of the original parameters in the preprocessed dataset. These composite features can better reflect the operating performance of the gas turbine, thereby improving the performance of the prediction model. Next, for the time-series data in the preprocessed dataset, multiple time window features were constructed through different time windows, including moving average, standard deviation, rate of change, and cumulative value. These features can capture the dynamic characteristics of gas turbine parameters changing over time, enhancing the model's ability to process time-series information. In addition, for the key vibration parameters and pressure fluctuation data in the preprocessed dataset, the Fast Fourier Transform (FFT) technique was used to extract frequency domain features such as power spectral density, dominant frequency components, and their amplitudes. These frequency domain features can help the model identify periodic change patterns, which plays an important role in equipment fault prediction. Finally, the Pearson correlation coefficient, mutual information value, and partial correlation coefficient among the key parameters in the preprocessed dataset were calculated to construct a correlation feature set. Through these correlation features, the linear and nonlinear dependencies between parameters can be quantified, further enriching the model's feature expression. After feature construction, analysis of variance (ANOVA) was used to initially screen the composite feature set, time window feature set, frequency domain feature set, and correlation feature set, removing features with variance close to zero. These features with near-zero variance typically contribute little to the model's predictions and therefore need to be removed. Then, LASSO regression, random forest feature importance assessment, recursive feature elimination (RFE), and principal component analysis (PCA) were applied to the initially selected feature set for further screening and dimensionality reduction. LASSO regression removes unimportant features through L1 regularization; random forest feature importance assessment sorts features based on their importance scores, selecting those with a significant impact on prediction; recursive feature elimination (RFE) progressively removes features, evaluating model performance after each removal to select the optimal feature; and principal component analysis (PCA) uses linear transformation to reduce the dimensionality of high-dimensional data, preserving the maximum amount of information.

[0017] It should be noted that features are constructed from four dimensions based on the preprocessed standardized data.

[0018] The initial feature set constructed has a high dimensionality and needs to be filtered to improve model efficiency and generalization ability.

[0019] Preliminary screening: Using analysis of variance, irrelevant features with variance close to zero are eliminated, and the cooling water temperature is kept constant under stable operating conditions.

[0020] Core Screening: Dimensionality reduction is achieved through a combination of methods. LASSO regression is applied, and L1 regularization is used to compress a batch of regression coefficients to zero, thereby eliminating features that contribute little to the prediction target. Simultaneously, random forest is used to evaluate the importance score of each feature, and recursive feature elimination (RFE) is employed for reverse elimination. Principal component analysis (PCA) is then applied to the retained features to condense information, transforming the original high-dimensional feature space into a set of key parameters with significantly reduced dimensionality (from the initial 500 features to 35 core features) while retaining over 95% of the original information.

[0021] Step S103: Input the key parameter set into a hierarchical deep neural network for processing. Through the step-by-step derivation of the basic parameter prediction layer, intermediate state estimation layer and target parameter regression layer, the dynamic parameter prediction result is obtained. Specifically, the temporal features of the key parameter set are input into the Long Short-Term Memory (LSTM) network of the basic parameter prediction layer. The LSM network contains three LSTM layers and a Dropout layer to obtain predicted values ​​of the basic physical parameters for future timeframes. These predicted values ​​are then input into the Graph Convolutional Neural Network (GCNN) of the intermediate state estimation layer. This network contains three GCNN layers to represent the physical connections between the gas turbine components as a graph structure, thus obtaining intermediate state parameters. Finally, the predicted values ​​of the basic physical parameters and the intermediate state parameters are input into the deep residual network of the target parameter regression layer. This deep residual network contains six residual... The target prediction parameters are obtained by adding a difference block and a fully connected layer. The basic parameter prediction layer, the intermediate state estimation layer, and the target parameter regression layer are pre-trained respectively, and the parameters of each layer are initialized using the corresponding input and output data to obtain an initial model. The initial model is then jointly optimized end-to-end using a weighted multi-objective loss function combined with mean squared error loss and smoothing L1 loss to obtain an optimized model. The optimized model is then trained using the Adam optimizer and cosine annealing strategy, with an early stopping strategy set. Training is stopped when the performance on the validation set shows no improvement for 20 consecutive rounds to obtain the dynamic parameter prediction results.

[0022] The time-series features extracted from the key parameters are input into a Long Short-Term Memory (LSTM) network in the basic parameter prediction layer. LSTM is a network structure specifically designed for processing time-series data, effectively capturing temporal dependencies. In this layer, the LSTM network consists of three LSTM layers and a Dropout layer, which prevents overfitting. Through these three LSTM layers, the model learns the time-series relationships between various parameters during gas turbine operation and predicts fundamental physical parameters such as intake temperature and exhaust pressure at future moments. The predicted values ​​of the fundamental physical parameters obtained from the basic parameter prediction layer are then input into a Graph Convolutional Neural Network (GCN) in the intermediate state estimation layer. GCN is a network architecture capable of processing graph-structured data. It represents the physical connections between gas turbine components as a graph structure, where each component (node) is connected to other components (nodes) via edges. Through three layers of graph convolution operations, GCN can capture the mutual influences between gas turbine components, thereby estimating some intermediate state parameters such as compressor efficiency and combustion chamber temperature distribution. These parameters are often difficult to measure directly but are crucial for subsequent predictions. The predicted values ​​of basic physical parameters and intermediate state parameters are input together into a deep residual network (ResNet) for the target parameter regression layer. ResNet solves the gradient vanishing problem in deep neural network training through residual blocks and contains multiple fully connected layers, enabling it to effectively regress multiple target prediction parameters, such as power, thermal efficiency, and emission levels. The design of deep residual networks allows the model to learn complex nonlinear relationships at a deeper level and improves prediction accuracy.

[0023] During model training, the basic parameter prediction layer, intermediate state estimation layer, and target parameter regression layer are pre-trained separately. The purpose of pre-training is to provide reasonable initial parameters for each layer to avoid gradient explosion or vanishing gradient problems during training. Pre-training uses a dataset corresponding to the input and output data of each layer to initialize the parameters of each layer, thus obtaining an initial model. This initialized model is then jointly optimized end-to-end using a weighted multi-objective loss function. The multi-objective loss function combines mean squared error (MSE) loss and smoothed L1 loss to handle prediction tasks with different objectives. The weighting strategy assigns different weights to different objectives based on their importance, ensuring comprehensive optimization of each objective. The optimization process uses the Adam optimizer combined with a cosine annealing strategy to dynamically adjust the learning rate. This strategy effectively prevents the learning rate from becoming too high or too low during training, thereby improving training efficiency and convergence speed. Simultaneously, an early stopping strategy is implemented: training stops when the performance on the validation set fails to improve after 20 consecutive training iterations, thus avoiding overfitting and ensuring the model's generalization ability.

[0024] It should be noted that the network structure and data flow of this hierarchical deep neural network consist of three layers with different functions connected in series. Its core information is shown in the table below: Pre-training: The three layers are pre-trained independently. The LSTM layer is trained using historical data to accurately predict the basic parameters; the GCN layer is trained using component relationship data to reasonably estimate the intermediate states.

[0025] Joint optimization: Based on pre-training, the entire network is jointly trained end-to-end. The optimization objective is a weighted multi-objective loss function that combines mean squared error (focusing on overall accuracy) and smoothed L1 loss (enhancing robustness to outliers).

[0026] Optimization strategy: The Adam optimizer is used in conjunction with a cosine annealing strategy to dynamically adjust the learning rate, thereby promoting effective model convergence. Simultaneously, an early stopping strategy is implemented: if the performance on the validation set no longer improves within 20 consecutive training iterations, training is automatically terminated to prevent overfitting.

[0027] Step S104: Perform multi-model integration processing on the dynamic parameter prediction results, and dynamically adjust the weights of each model by constructing a meta-learner to obtain the integrated prediction value; Specifically, based on the dynamic parameter prediction results, a prediction model library is constructed, including a hierarchical deep neural network model, a gradient boosting decision tree model, a random forest model, a support vector regression model, and a physical model based on thermodynamic principles, to obtain a basic model layer. Each model in the basic model layer is trained using different hyperparameter configurations and feature subsets, creating multiple variants of each model to obtain a model variant set. An XGBoost algorithm meta-learner is constructed, taking the output of the model variant set and current operating condition features as input, to predict the prediction error of each model under the current conditions, obtaining a confidence score. The confidence score is converted into normalized weights, and the prediction results of the model variant set are weighted and averaged to obtain a comprehensive prediction result. The comprehensive prediction result is adjusted according to the prediction target and prediction time scale; short-term predictions tend to rely more on models with better recent data performance, while long-term predictions increase the weight ratio of the physical model, resulting in a time-domain optimization result. Every 24 hours, the meta-learner is incrementally trained using the latest accumulated actual operating data to update the model evaluation capability, obtaining the integrated prediction value.

[0028] Based on the dynamic parameter prediction results obtained in the preceding steps, a model library containing various prediction models was constructed. This library includes hierarchical deep neural network models, gradient boosting decision tree models, random forest models, support vector regression models, and physical models based on thermodynamic principles. These different models combine their respective advantages to more comprehensively predict the state of the gas turbine. Each model has its strengths: deep neural networks are suitable for handling complex nonlinear relationships, decision tree models can effectively capture the interactions between features, and physical models can provide reliable predictions even with insufficient data. Next, each model is trained with different hyperparameter configurations and feature subsets, generating multiple variants. These variants optimize the performance of each model according to different operating conditions and data characteristics, increasing the model's applicability in different scenarios. Then, a meta-learner is constructed to integrate the prediction results of these model variants. The meta-learner takes the output of each model variant and the current operating condition features as input, predicts the prediction error of each model, and thus derives a confidence score for each model. This process uses the XGBoost algorithm to train the meta-learner, enabling it to evaluate the performance of each model variant under the current conditions. By converting credibility scores into normalized weights, the meta-learner can adjust the contribution of each model to the final prediction result based on its credibility. After obtaining the credibility scores of each model variant, the meta-learner performs a weighted average of their prediction results to obtain a comprehensive prediction result. This weighted averaging method ensures that the final prediction result fully reflects the advantages of each model, making the integrated prediction result more accurate. To make the integrated prediction result more in line with actual needs, the results are adjusted according to different prediction objectives and time scales. For short-term predictions, the system relies more on the model that performs better in recent data to capture short-term trends; while for long-term predictions, the weight of the physical model is increased to ensure that the prediction result conforms to the physical laws and long-term operating trends of the gas turbine. To maintain the model's high adaptability and prediction accuracy, the system incrementally trains the meta-learner every 24 hours using the latest actual operating data. This incremental training allows the model to continuously learn from new data and optimize its prediction capabilities, thereby continuously improving the accuracy of the prediction results during long-term operation. Through this series of steps, the present invention can effectively improve the accuracy and stability of gas turbine state prediction.

[0029] It should be noted that a predictive model library containing five different types of algorithms was established, with each model using a different training strategy to increase diversity.

[0030] Credibility Assessment: A meta-learner based on the XGBoost algorithm is constructed. The inputs to this meta-learner include: ① the power predictions of the five models for the current moment; ② current operating conditions (speed 15000 r / min, fuel flow 0.05 kg / s, etc.). After training, the meta-learner can predict the expected error of each base model under the current operating conditions based on these inputs and output a credibility score (ranging from 0 to 1). Dynamic Weight Allocation: The credibility score is normalized and converted into weights for each model. In a prediction, if the meta-learner determines that the current operating condition is similar to historical data A, and the SVR model performed best on historical data A, it may assign a higher weight (0.35) to SVR and lower weights to other models. Temporal Optimization Strategy: The strategy is adjusted according to the prediction time scale. Short-term predictions (within the next 10 seconds) rely more on data-driven models with good recent performance (hierarchical DNN, GBDT). Long-term predictions (within the next minute) increase the weight of the physical model to ensure that the prediction conforms to thermodynamic laws.

[0031] Every 24 hours, the system incrementally trains the meta-learner using the latest actual operating data accumulated in the past day, updating its ability to evaluate the credibility of the model, so that it can adapt to the slow degradation of gas turbine performance or environmental changes.

[0032] Step S105: Based on the integrated predicted values, perform uncertainty quantification analysis, and obtain the prediction confidence interval through Monte Carlo Dropout method, Bayesian neural network framework and time series scenario simulation. Specifically, the Monte Carlo Dropout method is applied to the integrated predicted values. During the prediction phase, the Dropout layer remains active, and 100 forward propagation calculations are performed on the same input data to obtain the sampling distribution of the deep neural network model output. A measurement error model is established based on the noise characteristics of each measurement parameter in historical data. The output uncertainty caused by input noise is evaluated using error propagation theory to obtain the quantification result of random uncertainty. Using a Bayesian neural network framework, network weights are treated as random variables, and the posterior distribution is approximated using variational inference. 200 sets of weight samples are extracted from the posterior distribution of the weights to obtain the quantification result of cognitive uncertainty. Scenario simulations of possible future operating conditions are performed using time series analysis to generate multiple possible future input sequences. The differences in prediction distribution under different scenarios are evaluated to obtain the quantification result of scenario uncertainty. The quantification results of random uncertainty, cognitive uncertainty, and scenario uncertainty are fused to form a comprehensive probability distribution, resulting in a comprehensive prediction uncertainty result. Prediction intervals at 90%, 95%, and 99% confidence levels are calculated based on the comprehensive prediction uncertainty result. The interval width is adaptively adjusted based on the frequency of actual observations falling into the prediction interval in historical validation data to obtain the prediction confidence interval.

[0033] The Monte Carlo Dropout method is used to quantify random uncertainty. During the prediction phase of the deep neural network model, the Dropout layer is kept active, meaning that some neurons in the network are randomly "dropped" during each prediction. 100 forward propagation calculations are performed on the same input data to obtain the sampled distribution of the network output. This method captures the uncertainty caused by the randomness within the neural network, reflecting the random fluctuations and instability of the model during the prediction process. Next, a measurement error model is established to analyze the noise characteristics of each measurement parameter in historical data. Using error propagation theory, the system assesses how noise in the input data affects the uncertainty of the output results. This method quantifies the random uncertainty caused by measurement errors, thus obtaining the quantification result of this part of the uncertainty. A Bayesian neural network framework is used to quantify cognitive uncertainty. In the Bayesian framework, the weights of the neural network are treated as random variables, rather than fixed constants. The posterior distribution of the network weights is approximated using variational inference methods, and 200 weight samples are drawn from this distribution. This quantifies the prediction instability caused by the uncertainty of model parameters, obtaining the quantification result of cognitive uncertainty. To account for potential future changes in operating conditions, this method also employs time series analysis for scenario simulation. This method generates multiple possible future input sequences based on historical data and evaluates the differences in predicted distributions under different operating conditions. In this way, the scenario uncertainty caused by the uncertainty of future operating conditions can be quantified, thus obtaining a quantitative result of scenario uncertainty.

[0034] The system integrates the quantitative results from three aspects: random uncertainty, cognitive uncertainty, and situational uncertainty, forming a comprehensive probability distribution. Based on this comprehensive probability distribution, the system can calculate prediction intervals at 90%, 95%, and 99% confidence levels. Furthermore, based on the frequency with which actual observations fall within the prediction interval in historical validation data, the system adaptively adjusts the width of the prediction interval, thereby making the prediction interval more accurate and ultimately obtaining a reliable prediction interval.

[0035] It should be noted that the uncertainty of the integrated prediction values ​​is quantified to generate a prediction confidence interval. The following example, using the power prediction of a 100kW gas turbine in the next 3 seconds (current power is 95kW), illustrates the specific implementation process: Monte Carlo Dropout Method: During the prediction phase, the Dropout layer of the deep learning model remains activated, and 100 forward propagation calculations are performed on the same input (current operating parameters). Each time, a portion of neurons are randomly discarded, resulting in 100 predicted power values ​​(94.8kW, 95.2kW, 95.1kW, etc.), forming the output distribution. The standard deviation (0.15kW) is calculated as a measure of the model's internal randomness. Measurement Error Propagation: Based on historical sensor noise data (power measurement error ±0.5%), the impact of input noise on the output is evaluated using error propagation theory. If the fuel flow error is ±0.01kg / s, an additional uncertainty of ±0.3kW in the power output can be derived.

[0036] Bayesian Neural Network Framework: Treating neural network weights as random variables, an approximate posterior distribution is inferred through variational reasoning. 200 samples are drawn from the posterior distribution of the weights, with each weight generating a predicted power value (range 94.5kW to 95.5kW). The standard deviation of these predicted values ​​(0.25kW) is calculated to quantify the uncertainty of the model parameters. Time Series Scenario Simulation: Multiple possible future input sequences are generated based on historical data (10 scenarios including load abrupt changes and ambient temperature fluctuations). The power prediction distribution under each scenario is evaluated, and the prediction differences between different scenarios are calculated (standard deviation 0.2kW).

[0037] The three types of uncertainty are combined into a comprehensive probability distribution (assuming it follows a Gaussian distribution with a mean of 95kW and a total standard deviation of 0.36kW).

[0038] Calculate the prediction intervals at 90%, 95%, and 99% confidence levels: 90% range: 95kW±0.59kW→[94.41kW,95.59kW]; 95% range: 95kW±0.71kW→[94.29kW,95.71kW]; Adaptive adjustment: Based on the frequency of actual values ​​falling into the interval in historical validation data (the actual coverage rate of the 90% interval is 88%), the interval width is finely adjusted (the coefficient is adjusted from 0.59 to 0.61) to make the coverage rate closer to the nominal confidence level.

[0039] Step S106: Based on the comparison between the predicted confidence interval and the actual observed value, a differentiated update strategy is triggered through a multi-criteria judgment mechanism to obtain a continuously optimized prediction model.

[0040] Specifically, the real-time gas turbine operating data is managed using a sliding window method, with the latest window data processed every hour. The window size is 12 hours, resulting in processed new data. This processed new data is then categorized based on data characteristics, distinguishing between start-up conditions, full-load conditions, and load-changing conditions, resulting in a categorized dataset. A multi-criteria judgment mechanism monitors the deviation between the predicted confidence interval and the actual observed values. An update is obtained when the actual observed value falls outside the predicted interval for five consecutive times, or when the moving average of the prediction deviation shows a monotonically increasing trend for three consecutive days, or when an operating condition not covered in the model training history appears. A signal is sent; based on the update trigger signal and the degree of performance degradation, a differentiated update strategy is adopted: incremental parameter updates are performed when there is a slight performance degradation; some sub-models are retrained and the ensemble weights are updated when there is a moderate performance degradation; and full model retraining is performed when there is a severe performance degradation, resulting in an updated model scheme; the updated model scheme is executed in an independent computing environment, and the model is updated through a parallel computing framework to obtain the updated prediction model; the performance of the updated prediction model is evaluated, and the mean absolute percentage error and prediction interval coverage are calculated. If the performance degrades, it is automatically rolled back to the previous stable version, resulting in the continuously optimized prediction model.

[0041] The system employs a sliding window approach to manage real-time collected operational data. Every hour, the system processes data from the latest window, with the window size set to 12 hours. This ensures that each processed data point covers the past 12 hours of operational data, guaranteeing the model can utilize sufficient historical data for accurate predictions. The processed new data is then categorized based on its characteristics, distinguishing between different operating conditions such as startup, full-load, and load-changing conditions. This classification method ensures the system uses appropriate analysis methods when processing data for different operating conditions, improving the accuracy and relevance of predictions. After data classification, the system monitors the deviation between the predicted confidence interval and the actual observed values ​​using a multi-criteria judgment mechanism. If the actual observed value falls outside the predicted interval for five consecutive times, or the moving average of the prediction deviation shows a monotonically increasing trend for three consecutive days, or a new operating condition point not covered in the model's training history appears, the system will trigger an update signal. These conditions help the system detect deviations between predicted results and actual values ​​in real time, allowing for timely responses to ensure prediction accuracy.

[0042] Upon triggering an update signal, the system employs differentiated update strategies based on the degree of performance degradation. In cases of slight performance degradation, only incremental parameter updates are needed, involving fine-tuning some parameters on the existing model. For moderate performance degradation, the system retrains some sub-models and updates the ensemble weights to better adapt to current operating conditions. In cases of severe performance degradation, a full model retraining is required, involving feature engineering and model architecture optimization to ensure the model recovers to its optimal prediction state. All model update operations are performed in an independent computing environment, accelerating the update process through a parallel computing framework to ensure the real-time prediction system remains unaffected. After the update, the new prediction model undergoes performance evaluation, primarily assessing mean absolute percentage error (MAPE) and prediction interval coverage. If the updated model's performance degrades, the system automatically rolls back to the previous stable version, ensuring the system remains in an optimal state. Through these steps, this invention continuously optimizes the prediction model to address various changes during gas turbine operation, ensuring the system's prediction accuracy and stability.

[0043] It should be noted that the system processes the latest 12-hour window of operating data (intake air temperature, speed, power, etc.) every hour, with a window size of 12 hours and a total of 43,200 data points (collected 10 times per second). Operating condition labeling: New data is automatically categorized according to operating characteristics: Start-up condition (speed < 5000 r / min, rapid temperature rise); Full load condition (power > 95 kW, stable speed); Load variation condition (power fluctuation > 10%). The system continuously compares the predicted confidence interval (95% confidence interval) with the actual observed values, and triggers an update signal when any of the following conditions are met: Based on the degree of performance degradation at the time of triggering, the system automatically selects the update scheme: slight performance degradation (bias increase <10%): only the meta-learner (XGBoost) is updated with incremental parameters, and the weights are fine-tuned using the latest 72-hour data. The update takes about 10 minutes.

[0044] Moderate performance degradation (bias increase of 10%-30%): Retrain the worst-performing sub-model (support vector regression model) and update the ensemble weights. A parallel computing framework is used to train the new model simultaneously, ensuring uninterrupted real-time predictions.

[0045] Severe performance degradation (bias increase > 30% or new operating conditions): Perform a full model retraining, including feature engineering and network architecture optimization. Train for 12 hours in an independent environment and deploy after successful validation.

[0046] The updated model needs to be validated on the test set: Mean Absolute Percentage Error (MAPE) < 1.5% (MAPE = 1.2% after update); 95% prediction interval coverage > 93% (coverage = 94% after update); if performance degrades (MAPE increases from 1.2% to 1.8%), the system will automatically roll back to the previous version and mark the abnormal operating conditions for manual analysis.

[0047] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An artificial intelligence-based method for regressing and predicting dynamic parameters of a gas turbine, characterized in that, The method includes: Various operating parameters of the gas turbine are collected, and a preprocessed dataset is obtained through data cleaning, outlier handling, missing value completion, and Z-score standardization. Based on the preprocessed dataset, composite features, time window features, frequency domain features, and correlation features are constructed. The key parameter set is obtained by filtering through a multi-step feature selection process. The key parameter set is input into a hierarchical deep neural network for processing. Through the step-by-step derivation of the basic parameter prediction layer, intermediate state estimation layer and target parameter regression layer, the dynamic parameter prediction result is obtained. The dynamic parameter prediction results are processed by multi-model integration. By constructing a meta-learner, the weights of each model are dynamically adjusted to obtain the integrated prediction value. Uncertainty quantification analysis is performed based on the integrated predicted values. The prediction confidence interval is obtained by using the Monte Carlo Dropout method, Bayesian neural network framework and time series scenario simulation. Based on the comparison between the predicted confidence interval and the actual observed values, a differentiated update strategy is triggered through a multi-criteria judgment mechanism to obtain a continuously optimized prediction model.

2. The artificial intelligence-based gas turbine dynamic parameter regression prediction method according to claim 1, characterized in that, The process involves collecting various operating parameters of the gas turbine, and then performing data cleaning, outlier handling, missing value completion, and Z-score standardization to obtain a preprocessed dataset, including: The gas turbine's inlet temperature, inlet pressure, exhaust temperature, exhaust pressure, speed, power output, fuel flow rate, cooling water temperature, vibration value, and bearing temperature are collected at a fixed frequency to obtain raw data. Data points in the original data that exceed physical limits are automatically marked and deleted, and confirmed outliers are replaced by a local linear regression method to obtain preliminary cleaned data. The missing values ​​in the preliminary cleaned data are processed by selecting the corresponding completion algorithm based on the missing time span to obtain complete data; For the missing portions of the complete data with a range not exceeding 3 sampling periods, cubic spline interpolation is used to complete them; for the missing portions with a range exceeding 3 sampling periods but not exceeding 10 sampling periods, a multiple regression model based on similar working conditions is used to complete them, thus obtaining the completed data. The completed data is processed using the Z-score standardization method to convert each parameter into a standard distribution with a mean of 0 and a standard deviation of 1, thus obtaining standardized data. The standardized data is divided into training, validation, and test sets according to a ratio of 70%, 15%, and 15% respectively, to obtain the preprocessed dataset.

3. The artificial intelligence-based gas turbine dynamic parameter regression prediction method according to claim 1, characterized in that, The process involves constructing composite features, time window features, frequency domain features, and correlation features based on the preprocessed dataset. A multi-step feature selection process is then used to filter these features, resulting in a set of key parameters, including: Based on the thermodynamic theory of gas turbines, the original parameters in the preprocessed dataset are combined and calculated to construct composite features of compression ratio, expansion ratio, isentropic efficiency, and thermal efficiency, thus obtaining a composite feature set. The moving average, standard deviation, rate of change, and cumulative value of different time windows are constructed for the time series data in the preprocessed dataset to obtain the time window feature set; A fast Fourier transform is performed on the key vibration parameters and pressure fluctuation data in the preprocessed dataset to extract the power spectral density, main frequency components and their amplitudes, thus obtaining a frequency domain feature set. Calculate the Pearson correlation coefficient, mutual information value, and partial correlation coefficient among the key parameters in the preprocessed dataset to obtain the correlation feature set; Analysis of variance was used to initially screen the composite feature set, time window feature set, frequency domain feature set and correlation feature set, eliminating features with variance close to zero, and obtaining the initial selected feature set. The initial feature set was further screened and dimensionality reduced using LASSO regression, random forest feature importance assessment, recursive feature elimination, and principal component analysis to obtain the key parameter set.

4. The artificial intelligence-based gas turbine dynamic parameter regression prediction method according to claim 1, characterized in that, The process involves inputting the key parameter set into a hierarchical deep neural network for processing. Through layer-by-layer derivation of the basic parameter prediction layer, intermediate state estimation layer, and target parameter regression layer, dynamic parameter prediction results are obtained, including: The temporal features of the key parameter set are input into the long short-term memory network of the basic parameter prediction layer. The long short-term memory network contains three LSTM layers and a Dropout layer to obtain the predicted values ​​of the basic physical parameters at future time steps. The predicted values ​​of the basic physical parameters are input into the graph convolutional neural network of the intermediate state estimation layer. This network contains three graph convolutional layers, which represent the physical connection relationship between the components of the gas turbine as a graph structure to obtain the intermediate state parameters. The predicted values ​​of the basic physical parameters and the intermediate state parameters are input together into the deep residual network of the target parameter regression layer. The deep residual network contains 6 residual blocks and a fully connected layer to obtain the target predicted parameters. The basic parameter prediction layer, the intermediate state estimation layer, and the target parameter regression layer are pre-trained respectively, and the parameters of each layer are initialized using the corresponding input and output data to obtain the initialized model; The initialization model is jointly optimized end-to-end using a weighted multi-objective loss function that combines mean squared error loss and smoothing L1 loss to obtain the optimized model. The optimization model was trained using the Adam optimizer and cosine annealing strategy. An early stopping strategy was set, and training was stopped when the performance on the validation set showed no improvement for 20 consecutive rounds, thus obtaining the dynamic parameter prediction results.

5. The artificial intelligence-based gas turbine dynamic parameter regression prediction method according to claim 1, characterized in that, The process of performing multi-model ensemble processing on the dynamic parameter prediction results, and dynamically adjusting the weights of each model by constructing a meta-learner to obtain the ensemble prediction value, includes: Based on the prediction results of the dynamic parameters, a prediction model library is constructed, which includes a hierarchical deep neural network model, a gradient boosting decision tree model, a random forest model, a support vector regression model, and a physical model based on thermodynamic principles, thus obtaining the basic model layer. Each model in the base model layer is trained using different hyperparameter configurations and feature subsets to create multiple variants of each model, resulting in a model variant set; An XGBoost meta-learner is constructed with the output of the model variant set and the current working condition features as inputs to predict the prediction error of each model under the current conditions and obtain a credibility score. The credibility score is converted into normalized weights, and the prediction results of the model variant set are weighted and averaged to obtain the comprehensive prediction result. The comprehensive prediction results are adjusted according to the prediction target and prediction time scale. Short-term predictions tend to rely more on models that have performed better in recent data, while long-term predictions increase the weight of the physical model to obtain time-domain optimized results. The meta-learner is incrementally trained every 24 hours using the latest accumulated actual operating data to update the model evaluation capability and obtain the ensemble prediction value.

6. The artificial intelligence-based gas turbine dynamic parameter regression prediction method according to claim 1, characterized in that, The uncertainty quantification analysis based on the integrated predicted values, using the Monte Carlo Dropout method, Bayesian neural network framework, and time series scenario simulation, yields the predicted confidence interval, including: The Monte Carlo Dropout method is applied to the integrated predicted values. During the prediction phase, the Dropout layer is kept active, and 100 forward propagation calculations are performed on the same input data to obtain the sampling distribution of the deep neural network model output. A measurement error model is established based on the noise characteristics of each measurement parameter in historical data. The output uncertainty caused by input noise is evaluated through error propagation theory, and the quantitative result of random uncertainty is obtained. Using a Bayesian neural network framework, the network weights are treated as random variables. The posterior distribution is approximated by variational inference. 200 weight samples are drawn from the posterior distribution of the weights to obtain the quantification result of cognitive uncertainty. By using time series analysis, possible future changes in working conditions are simulated to generate multiple sets of possible future input sequences. The differences in the predicted distribution under different scenarios are evaluated to obtain the quantitative results of scenario uncertainty. By integrating the quantification results of the random uncertainty, the quantification results of the cognitive uncertainty, and the quantification results of the situational uncertainty, a comprehensive probability distribution is formed to obtain the comprehensive result of the prediction uncertainty. Based on the comprehensive results of the prediction uncertainty, the prediction intervals at 90%, 95%, and 99% confidence levels are calculated, and the interval width is adaptively adjusted based on the frequency of actual observations falling into the prediction interval in historical validation data to obtain the prediction confidence interval.

7. The artificial intelligence-based gas turbine dynamic parameter regression prediction method according to claim 1, characterized in that, The step of triggering a differentiated update strategy through a multi-criteria judgment mechanism based on the comparison results between the predicted confidence interval and the actual observed values ​​to obtain a continuously optimized prediction model includes: The real-time gas turbine operating data is managed using a sliding window method. The latest window data is processed every hour, with a window size of 12 hours, to obtain the processed new data. The processed new data is categorized by labeling based on data characteristics to distinguish between startup conditions, full-load conditions, and load variation conditions, resulting in a categorized dataset. The deviation between the predicted confidence interval and the actual observed value is monitored by a multi-criteria judgment mechanism. When the actual observed value falls outside the predicted interval for 5 consecutive times, or the moving average of the prediction deviation shows a monotonically increasing trend for 3 consecutive days, or when there is a working condition point not covered in the model training history, an update trigger signal is obtained. Based on the update trigger signal and the degree of performance degradation, a differentiated update strategy is adopted: incremental parameter updates are performed when there is a slight performance degradation; some sub-models are retrained and the ensemble weights are updated when there is a moderate performance degradation; and full model retraining is performed when there is a severe performance degradation, thus obtaining the updated model scheme. The updated model scheme is executed in an independent computing environment, and the model is updated through a parallel computing framework to obtain the updated prediction model. The updated prediction model is evaluated for performance by calculating the mean absolute percentage error and prediction interval coverage. If the performance deteriorates, it is automatically rolled back to the previous stable version to obtain the continuously optimized prediction model.

Images