NOx emission prediction method for gas turbine

Abstract

The invention provides a NOx emission prediction method for a gas turbine, and the method comprises the steps: obtaining an operation monitoring data set of a gas turbine unit, and carrying out the feature variable screening of the operation monitoring data according to the generation mechanism, thermal process, combustion reaction and operation condition of the NOx emission of the gas turbine, so as to determine a first feature variable set; performing feature variable selection on the first feature variable set based on a LightGBM recursive feature elimination method to determine a second feature variable set; and the operation monitoring data set is processed according to the second characteristic variable set, the processed operation monitoring data set serves as input, and NOx emission prediction is achieved according to the prediction model, the prediction model comprises a multilayer convolutional neural network used for extracting local features of input data, a channel attention mechanism used for capturing a cross-feature channel dependency relationship, a long and short-term memory network and a trend perception channel and time cross attention mechanism.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of gas turbines, in particular to a method for predicting NOx emissions of a gas turbine, a system for predicting NOx emissions of a gas turbine, and a computer-readable storage medium. BACKGROUND

[0002] Heavy-duty gas turbines produce a large amount of exhaust gas during the combustion process, and its emission control is a core link in the environmental protection measures of power plants, which is directly related to pollution control and emission standards. The main exhaust gas produced by heavy-duty gas turbines is NOx, which needs to be reduced by adjusting the combustion process or treated by SCR denitration device. Both of these two ways depend on the accurate prediction of NOx.

[0003] However, due to the complex operating environment of gas turbines, large load fluctuations and variable control strategies, the NOx emission characteristics show significant non-stationarity, local volatility and multivariate coupling, which brings great challenges to accurate modeling and prediction.

[0004] With the continuous development of machine learning technology, this technology has been widely applied in NOx prediction modeling and has achieved many results. Support vector machine, back propagation neural network, random forest and extreme learning machine are widely used in modeling and prediction of NOx emission concentration. However, since NOx emission has dynamic characteristics in nature, its output not only depends on the current input, but also is affected by the historical state. The above existing models generally fail to fully consider the correlation between time series, so the prediction accuracy is still limited.

[0005] In addition, the NOx emission of gas turbine unit remains in a relatively stable range under most operating conditions, but with frequent adjustment of unit operating conditions and changes in external environmental conditions, the emission level may fluctuate sharply in a short time or show a trend change in a long time scale. Therefore, the prediction task of NOx emission not only needs to deal with high-frequency mutation signals, but also needs to accurately depict long-term evolution trends.

[0006] By using CNN-LSTM (Convolutional Neural Network-Long Short Term Memory) network to predict the NOx of gas turbine, the advantages of local feature extraction (CNN) and long-term dependence modeling (LSTM) can be combined to a certain extent. However, the traditional CNN module cannot dynamically distinguish the importance of each feature channel, and the LSTM module cannot select information of different time periods.

[0007] In order to overcome the above-mentioned defects of the prior art, the technical field urgently needs a NOx emission prediction method for a gas turbine, which can improve the prediction accuracy of NOx emission concentration. SUMMARY

[0008] The following presents a simplified summary of one or more aspects in order to provide a basic understanding of such aspects. This summary is not an extensive overview of all contemplated aspects, and is intended to neither identify key or critical elements of all aspects nor delineate the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form as a prelude to the more detailed description that is presented later.

[0009] In order to overcome the above-mentioned defects of the prior art, the present application provides a NOx emission prediction method for a gas turbine, a NOx emission prediction system for a gas turbine, and a computer readable storage medium, which can extract low-frequency features that are smooth and have global perception ability, suppress high-frequency noise, enhance trend modeling ability, highlight key channels to improve feature expression relevance, improve the response ability to changing trends and sudden features, and thus improve the prediction accuracy of NOx emission concentration.

[0010] Specifically, the above-mentioned NOx emission prediction method for a gas turbine according to the first aspect of the present application comprises the steps of: obtaining a running monitoring data set of a gas turbine unit, and performing feature variable screening on the running monitoring data according to the generation mechanism, thermal process, combustion reaction and operating condition of NOx emission of the gas turbine to determine a first feature variable set; performing feature variable selection on the first feature variable set based on a recursive feature elimination method of LightGBM to determine a second feature variable set; and processing the running monitoring data set according to the second feature variable set, and taking the processed running monitoring data set as input, and realizing NOx emission prediction according to a prediction model, wherein the prediction model comprises a multi-layer convolutional neural network for extracting local features of input data, a channel attention mechanism for capturing cross-feature channel dependency, a long short-term memory network, and a trend perception channel and time cross-attention mechanism.

[0011] Further, in some embodiments of the present application, the step of performing feature variable screening on the running monitoring data further comprises: performing secondary screening on the first feature variable set according to a preset variance threshold.

[0012] Further, in some embodiments of the present application, the step of feature variable selection on the screened operation monitoring data by the LightGBM-based recursive feature elimination method comprises: training a LightGBM model using the first feature variable set to obtain the feature importance of each feature variable in the first feature variable set; recursively removing a preset number of feature variables with the lowest feature importance, and evaluating the performance of the LightGBM model in each iteration using cross-validation; and stopping the iteration according to the set termination condition and selecting the subset with the best performance as the second feature variable set.

[0013] Further, in some embodiments of the present application, the first layer of the multi-layer convolutional neural network is a wide kernel convolutional structure.

[0014] Further, in some embodiments of the present application, the channel attention mechanism comprises a global average pooling layer, an adaptive one-dimensional convolutional layer, an activation function layer, and a channel weighting unit.

[0015] Further, in some embodiments of the present application, the trend-aware channel and time cross-attention mechanism is a channel and time cross-attention module with a trend similarity term.

[0016] Further, in some embodiments of the present application, the trend-aware channel and time cross-attention mechanism comprises the steps of: realizing cross-modeling of the time dimension and the channel dimension through the channel and time cross-attention module; and extracting a local trend vector in a differential manner and correcting an attention scoring function based on the similarity of the trend direction between time steps.

[0017] Further, in some embodiments of the present application, the training phase of the prediction model comprises the steps of: obtaining a historical data set of the gas turbine unit, and performing feature variable screening on the historical data set according to the generation mechanism, thermal process, combustion reaction, and operating condition of the NOx emission of the gas turbine to determine the first feature variable set; performing feature variable selection on the first feature variable set by the LightGBM-based recursive feature elimination method to determine the second feature variable set; and processing the historical data set according to the second feature variable set, and training the prediction model by taking the processed historical data set as input.

[0018] In addition, the above-mentioned NOx emission prediction system for a gas turbine according to the second aspect of the present application comprises a memory and a processor. The memory has computer instructions stored thereon. The processor is connected to the memory and is configured to execute the computer instructions stored on the memory to implement the NOx emission prediction method for a gas turbine provided by any one of the above-mentioned embodiments.

[0019] In addition, the computer readable storage medium according to the third aspect of the present application stores computer instructions. When the computer instructions are executed by a processor, the computer instructions implement the method for predicting NOx emission of a gas turbine according to any one of the above embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0020] The above features and advantages of the present application will be better understood from the following detailed description of embodiments of the present application, read in conjunction with the drawings. In the drawings, the components are not necessarily drawn to scale, and components of similar or identical function or structure can be designated with the same or similar reference signs.

[0021] Figure 1 A schematic diagram of a system for predicting NOx emission of a gas turbine according to some embodiments of the present application is shown;

[0022] Figure 2 A flowchart of an offline training phase of a method for predicting NOx emission of a gas turbine according to some embodiments of the present application is shown;

[0023] Figure 3 A bar chart showing the ascending order of variance of feature variables according to some embodiments of the present application is shown;

[0024] Figure 4 A plot showing the ranking of feature importance of each feature variable determined by a LightGBM regression model according to some embodiments of the present application is shown;

[0025] Figure 5 A schematic diagram showing the processing of a data set based on a sliding window according to some embodiments of the present application is shown;

[0026] Figure 6 A structural diagram of a prediction model according to some embodiments of the present application is shown;

[0027] Figure 7 A plot showing the change of MSE loss during the training process of a prediction model according to some embodiments of the present application is shown;

[0028] Figure 8 A schematic diagram showing the prediction results of a LSTM model is shown;

[0029] Figure 9 A schematic diagram showing the prediction results of a GRU model is shown;

[0030] Figure 10 A schematic diagram showing the prediction results of a CNN model is shown;

[0031] Figure 11 A schematic diagram showing the prediction results of a CNN-LSTM model is shown;

[0032] Figure 12 A schematic diagram showing the prediction results of the CNN-ECA-LSTM-Attention model is shown;

[0033] Figure 13 A schematic diagram showing the prediction results of the WDCNN-ECA-LSTM-SA model is shown;

[0034] Figure 14 A schematic diagram showing the prediction results of the WDCNN-ECA-LSTM-TCTCA model provided by some embodiments of the present application is shown;

[0035] Figure 15 A box plot of prediction errors of the prediction model provided by the present application and other comparative models is shown; and

[0036] Figure 16 A flow chart showing the online analysis stage of the NOx emission prediction method for a gas turbine provided by some embodiments of the present application is shown.

[0037] Reference signs:

[0038] 100: a NOx emission prediction system for a gas turbine;

[0039] 110: a memory;

[0040] 111: a computer readable storage medium; and

[0041] 120: a processor. DETAILED DESCRIPTION

[0042] The present application will now be described by way of specific embodiments thereof, and as the skilled person will readily appreciate from the disclosure contained herein, still other advantages and benefits of the present application will become apparent. Although the description will be made in connection with preferred embodiments, it will be understood that the application is not limited to these embodiments. Rather, the purpose of the embodiments will be to cover all alternatives, modifications, and equivalents as can be included within the spirit and scope of the present application as defined by the appended claims. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. The present application can be practiced without the specific details. In other instances, well-known methods, procedures, components, and circuits have not been described in detail since not to unnecessarily obscure aspects of the present application.

[0043] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connecting" should be understood in a broad sense, for example, it can be fixed connection, or detachable connection, or integral connection; it can be mechanical connection, or electrical connection; it can be direct connection, or indirect connection through intermediate medium, or internal communication of two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0044] In addition, "upper", "lower", "left", "right", "top", "bottom", "horizontal", "vertical" used in the following description should be understood as the orientation shown in the section and the related drawings. Such relative terms are only for the convenience of description, and do not mean that the device described thereby must be manufactured or operated in a particular orientation, and therefore should not be understood as a limitation on the present application.

[0045] It can be understood that although the terms "first", "second", "third" and the like are used herein to describe various components, regions, layers and / or parts, these components, regions, layers and / or parts should not be limited by these terms, and these terms are only used to distinguish different components, regions, layers and / or parts. Therefore, the first component, region, layer and / or part discussed below can be referred to as the second component, region, layer and / or part without departing from some embodiments of the present application.

[0046] As described above, due to the complex operating environment, large load fluctuation and variable control strategy of the gas turbine, the NOx emission characteristics show significant non-stationarity, local fluctuation and multivariate coupling, which brings great challenges to accurate modeling and prediction.

[0047] With the continuous development of machine learning technology, the technology is widely used in NOx prediction modeling, and many achievements have been made. Support vector machine, back propagation neural network, random forest and extreme learning machine and other methods are widely used in modeling and prediction of NOx emission concentration. However, since the NOx emission has dynamic characteristics in nature, its output not only depends on the current input, but also is affected by the historical state, and the above existing models generally fail to fully consider the correlation between time series, so the prediction accuracy is still limited.

[0048] In addition, the NOx emission of gas turbine unit remains in a relatively stable range under most operating conditions, but with frequent adjustment of unit operating conditions and changes in external environmental conditions, the emission level may fluctuate sharply in a short time, or show a trend change in a long time scale. Therefore, the prediction task of NOx emission not only needs to deal with high-frequency mutation signals, but also needs to accurately depict long-term evolution trend.

[0049] The CNN-LSTM (Convolutional Neural Network-Long Short Term Memory) network can predict the NOx of the gas turbine, and can combine the advantages of local feature extraction (CNN) and long-term dependence modeling (LSTM) to a certain extent. However, the traditional CNN module cannot dynamically distinguish the importance of each feature channel, and the LSTM module cannot select information of different time periods.

[0050] In order to overcome the above-mentioned defects existing in the prior art, the present application provides a NOx emission prediction method for a gas turbine, a NOx emission prediction system for a gas turbine, and a computer readable storage medium, which can extract low-frequency features that are smooth and have global perception ability, suppress high-frequency noise, enhance trend modeling capability, highlight key channels to improve feature expression relevance, improve the response capability to changing trends and sudden characteristics, and thus improve the prediction accuracy of NOx emission concentration.

[0051] First, please refer to Figure 1 , Figure 1 A schematic diagram of a NOx emission prediction system for a gas turbine according to some embodiments of the present application is shown.

[0052] In some non-limiting embodiments, the above-mentioned NOx emission prediction method for a gas turbine provided by the first aspect of the present application can be implemented via the above-mentioned NOx emission prediction system for a gas turbine provided by the second aspect of the present application.

[0053] As shown in Figure 1 The NOx emission prediction system for a gas turbine 100 can be configured with a memory 110 and a processor 120. The memory 110 includes but is not limited to the above-mentioned computer readable storage medium 111 provided by the third aspect of the present application, on which computer instructions are stored. The processor 120 is connected to the memory 110 and is configured to execute the computer instructions stored on the memory 110 to implement the NOx emission prediction method for a gas turbine provided by the first aspect of the present application.

[0054] The working principle of the above-mentioned NOx emission prediction system for gas turbines will be described below in combination with some embodiments of NOx emission prediction methods for gas turbines. Those skilled in the art can understand that these embodiments of NOx emission prediction methods for gas turbines are only some non-limiting embodiments provided by the present application, which are intended to clearly demonstrate the main concept of the present application and provide some specific schemes for facilitating the public to implement, but not to limit the overall function or overall working mode of the NOx emission prediction system for gas turbines. Similarly, the NOx emission prediction system for gas turbines is also only a non-limiting embodiment provided by the present application, which does not limit the execution subject and execution order of each step in these NOx emission prediction methods for gas turbines.

[0055] The above-mentioned NOx emission prediction method for gas turbines provided by the first aspect of the present application can be implemented in two parts of offline training phase and online analysis phase.

[0056] Please refer to Figure 2 , Figure 2 The flow chart of the offline training phase of the NOx emission prediction method for gas turbines provided according to some embodiments of the present application is shown.

[0057] As Figure 2 shown, the NOx emission prediction system for gas turbines can first process the data to make the data meet the input requirements of the model.

[0058] Processing the data can first obtain the operation monitoring data of the gas turbine unit as input data. Then the characteristic variable set is determined through mechanism analysis. The required characteristic variables are obtained through variance screening and feature selection based on RFECV-LightGBM, and then the data is arranged and cleaned, and the data set is divided into training set, validation set and test set according to the preset proportion (such as 7:1:2).

[0059] In some embodiments, the emission prediction system can collect the operation monitoring data of the gas turbine unit as the historical data set.

[0060] After obtaining the historical data set, the emission prediction system can arrange and clean the data. The emission prediction system can combine the generation mechanism of gas turbine NOx emission, and preliminarily screen the original variables from the aspects of thermal process, combustion reaction and operation condition.

[0061] Here, the emission prediction system can select the subsystem variables highly related to the NOx emission process, and then make further selection to avoid the loss of feature information as much as possible. The selected variables and subsystems include combustion-related variables, air system-related parameters, exhaust system-related parameters, operating condition-related variables, control and regulation parameters, and / or other variables.

[0062] Among them, the combustion-related variables are variables closely related to flame temperature, which are direct driving factors of NOx generation. The combustion-related variables can include but are not limited to compressor inlet temperature, combustion chamber temperature, and / or combustion pressure.

[0063] The air system-related parameters are variables that can affect the combustion efficiency and excess air coefficient, including but not limited to compressor outlet pressure, intake IGV angle, and / or intake flow rate.

[0064] The exhaust system-related parameters not only affect the heat efficiency of the waste heat boiler, but also may feedback to adjust the combustion chamber temperature field, which indirectly affects the generation of NOx. The exhaust system-related parameters include but are not limited to exhaust temperature and / or flue gas pressure.

[0065] The operating condition-related variables can reflect the influence of overall system load change on NOx emission, including but not limited to unit load, power generation power, and / or rotating speed.

[0066] The control and regulation parameters are variables that may cause instability of the combustion region, including but not limited to gas regulating valve opening and / or burner state indication signal.

[0067] Other variables are variables that seem to have no direct thermodynamic relationship, but may indirectly affect the generation behavior of NOx under abnormal structure or combustion instability, such as vibration-related variables. In the operation of gas turbine, combustion instability (such as backfire, flameout, oscillating combustion, etc.) will cause flame structure disorder and abnormal temperature field distribution. Such instability process is often accompanied by mechanical vibration of a specific frequency. Therefore, the vibration change-related variable can be used as a signal of combustion state disturbance, thereby indirectly reflecting the fluctuation of NOx emission.

[0068] Through the above mechanism analysis, the emission prediction system constructs an initial feature pool with physical interpretation, i.e. the first feature variable set.

[0069] However, after the emission prediction system completes the mechanism-driven construction of variables, the initial first feature variable set contains many dimensions, some of which have physical meaning but have very small variation amplitude in actual operation, making it difficult to provide effective discrimination information for the model. For example, a certain feature has almost constant value in the entire sample, so its discrimination ability for the target variable (NOx emission concentration) is limited, and it may even introduce noise or multicollinearity risk.

[0070] Therefore, in some preferred embodiments, to further remove redundant variables and improve modeling efficiency and stability, the emission prediction system performs a secondary screening of the first feature variable set using the variance threshold method. This secondary screening can serve as a preliminary dimensionality reduction step for subsequent feature selection methods.

[0071] Specifically, the emissions prediction system can set an appropriate lower limit threshold for variance. Remove those whose variability across the entire time series is below this threshold. Characteristic variables:

[0072] ,

[0073] in, For the first One characteristic variable The variance.

[0074] In some embodiments, the emission prediction system selects an empirical threshold of 0.001 as the lower limit of variance based on the variance distribution of the analyzed sample data. This approach aims to balance preserving key dynamic features with removing redundant variables, thereby improving the efficiency of subsequent feature evaluation and model training.

[0075] Please refer to Figure 3 , Figure 3 A bar chart showing the ascending order of variances of characteristic variables provided according to some embodiments of the present invention is shown.

[0076] like Figure 3 As shown, the collected historical dataset contains some almost unchanging variables, which the emissions prediction system can remove. Figure 3 In the illustrated embodiment, the emission prediction system deletes feature variables that are below the lower variance limit based on a set threshold.

[0077] In some embodiments, the variables deleted by the emission prediction system may include: No. 1 gas turbine shaft displacement 1, No. 1 gas turbine shaft displacement 2, load gear vibration 1, load gear vibration 2, No. 1 gas turbine temperature matching setpoint temperature change rate, No. 1 gas turbine generator stator voltage, No. 1 gas turbine generator negative sequence current, No. 1 gas turbine generator frequency, #1 high-pressure superheater outlet steam pressure, #1 low-pressure superheater outlet steam pressure, low-pressure economizer outlet flue gas pressure, generator power factor, #1 gas turbine gas consumption rate 1, and No. 1 gas turbine temperature matching setpoint temperature, etc.

[0078] After secondary screening of feature variables using the variance threshold method, the emission prediction system can eliminate some redundant variables with poor numerical stability. However, the remaining feature variables still suffer from problems such as high dimensionality and strong redundancy among features.

[0079] To further compress the feature space and improve the utilization efficiency of the model on the input information, and reduce the redundant information in the input information, the emission prediction system can use the recursive feature elimination method (Recursive Feature Elimination, RFE) based on LightGBM (Light Gradient Boosting Machine), that is, the RFECV-LightGBM model, to further fine-screen the features to obtain a second feature variable set.

[0080] In the RFECV-LightGBM model, RFE provides a structured and controllable iterative screening mechanism, and LightGBM provides a high-quality feature importance evaluation basis. The former ensures that the screening process has systematicness and search ability, and the latter ensures that the evaluation standard has accuracy and robustness.

[0081] Specifically, the implementation of feature fine-screening first trains a LightGBM model using the first feature variable set to obtain the feature importance of each feature variable in the first feature variable set. Then, a preset number of feature variables with the lowest feature importance are recursively removed, and the model performance is evaluated using cross-validation in each iteration. Finally, the iteration is stopped according to the set termination condition, and the optimal subset is selected as the second feature variable set.

[0082] The RFECV-LightGBM model initializes the LightGBM regression model. The emission prediction system will use the feature set filtered by variance as input. The LightGBM regression model uses a gradient boosting decision tree architecture, and the key parameters include the maximum depth of the tree, the number of trees, and the learning rate.

[0083] Please refer to Figure 4 , Figure 4 The figure shows the ranking of the feature importance of each feature variable determined by the LightGBM regression model according to some embodiments of the present application.

[0084] After the LightGBM regression model is trained, the feature importance score is calculated based on the cumulative information gain (split_gain) of the feature in the decision tree node splitting. The feature importance score is shown in Figure 4 , which quantifies the actual contribution of each feature to reducing the prediction error.

[0085] Next, the recursive feature elimination iteration process is performed. In each iteration, the emission prediction system ranks the current feature importance, removes the last 5% of the features, and evaluates the model performance of the remaining feature subset using K-fold cross-validation (e.g., K=5): splits the training data into 5 subsets, trains on 4 subsets and validates on 1 subset, calculates the average validation set determination coefficient (R²) and root mean square error (RMSE), and records the current feature subset and corresponding performance indicators.

[0086] Then, the iteration is stopped according to the set termination condition and the selection of the optimal subset is realized.

[0087] The termination condition can be a double termination condition, including setting a minimum feature number threshold and a performance decline threshold, and setting a minimum feature number value can prevent the loss of key features caused by excessive removal. In some examples, the set minimum feature number threshold can be 25, and the performance decline threshold can be terminated when the determination coefficient (R²) decreases by more than 2% for three consecutive rounds.

[0088] After the iteration is terminated, the emission prediction system can select the feature subset with the highest cross-validation R² from all iteration records as the final result.

[0089] In one or more embodiments, the emission prediction system selects a total of 28 feature variables based on the recursive feature elimination method of LightGBM and combined with expert knowledge screening, including: No. 1 gas turbine IGV angle, No. 1 gas turbine compressor inlet temperature, air filter inlet pressure difference, compressor inlet parameter 1, No. 1 gas turbine compressor exhaust temperature, No. 1 gas turbine exhaust pressure, No. 1 gas turbine load, No. 1 gas turbine natural gas P2 pressure, No. 1 gas turbine inlet air temperature, waste heat boiler inlet flue gas pressure, #1 gas turbine maximum vibration 1, No. 1 gas turbine oil tank temperature 1, No. 1 gas turbine inlet air filter temperature 2, No. 1 gas turbine natural gas flow, No. 1 waste heat boiler chimney outlet original flue gas pressure, No. 1 gas turbine natural gas P2 pressure, etc.

[0090] In addition, in order to evaluate the effectiveness of the selected features, the emission prediction system can use the LSTM model to perform prediction modeling on different feature sets. First, the data set is divided into training set, validation set and test set, with a ratio of 7:1:2. Then, the data is normalized, and the model is trained based on different feature sets under the same model structure and hyperparameter settings. The model performance is evaluated by the determination coefficient and the root mean square error on the test set to judge the pros and cons of the feature selection method.

[0091] In this case, different feature sets include a feature set using variance filtering only, a feature set screened after feature importance ranking by LightGBM, a feature set screened after feature importance ranking by random forest, and a feature set selected by combining LightGBM feature importance ranking and recursive feature elimination method (RFE).

[0092] The results are shown in Table 1:

[0093]

[0094] Table 1

[0095] If the input feature set is only using simple variance filtering method, directly inputting a large number of features into the model will cause the performance of the model to decrease significantly, and the prediction accuracy is low. In contrast, both random forest and LightGBM can provide feature importance ranking based on tree structure, which can effectively eliminate redundant and invalid features. The performance of the two on the data set provided in this embodiment is relatively close, but LightGBM performs better in terms of computational efficiency and faster training speed. Further, combining LightGBM feature importance ranking and recursive feature elimination method (RFE) for feature selection can effectively improve the quality of the feature set, thereby significantly improving the performance of the LSTM model in the prediction task.

[0096] As shown in Figure 2 After determining the second feature variable set, the emission prediction system can also arrange and clean the original historical data set, and determine the feature variables of the historical data set according to the second feature variable set obtained in the above steps. Then, the processed historical data set is divided into a training set, a validation set and a test set according to a preset ratio (such as 7:1:2). In this case, the division of the data set is strictly in time sequence.

[0097] As shown in Figure 2 Preferably, the emission prediction system can normalize the historical data set to convert the data to the interval of (-1, 1). The emission prediction system can also convert the format of the data to meet the needs of the model.

[0098] In some embodiments, the emission prediction system can use a sliding window method to process the data set. The sliding window can divide a long time series into multiple fixed-length subsequences, each containing the features of a time step. This is very important for processing time series data, because the patterns in time series data are often local, that is, each time step may contain information closely related to its adjacent time steps. The sliding window can maintain the dependency between time steps. When using a sliding window, the data in each window is input in time order, and the model can learn how the features of each time step affect subsequent time steps.

[0099] Please refer to Figure 5 , Figure 5 A schematic diagram of processing a data set based on a sliding window is shown according to some embodiments of the present application.

[0100] As Figure 5 shown, all data in the data set is operated by a sliding window to construct the data set. Taking data at time t as an example, a sliding window with a size of T is used to construct the data set At time t, the input data set is input into the subsequent prediction model to calculate the value . Then, the input data set at time t is input into the subsequent prediction model to calculate the value .

[0101] Please continue to refer to Figure 2 , the emission prediction system processes the historical data set according to the selected second feature variable set, and then trains the prediction model (WDCNN-ECA-LSTM-TCTCA model) for predicting NOx emission by taking the processed historical data set as input.

[0102] Please refer to Figure 6 , Figure 6 A structural diagram of a prediction model is shown according to some embodiments of the present application.

[0103] As Figure 6 shown, the prediction model includes a multi-layer convolutional neural network (CNN) for extracting local features of input data, an efficient channel attention (ECA) mechanism for capturing cross-feature channel dependency, a long short-term memory network (LSTM), and a trend-aware channel-wise temporal cross attention mechanism (TCTCA).

[0104] In the training stage, the emission prediction system first sends the processed historical data set into the multi-layer one-dimensional CNN network of the prediction model, and extracts the local features of the input data layer by layer. Among them, the first layer CNN can realize convolution filtering operation. The first layer CNN effectively suppresses high-frequency noise, transient anomalies and measurement errors through convolution filtering operation, realizes preliminary feature denoising and smoothing. The subsequent deep CNN can further focus on the extraction of fine local features, and enhance the recognition ability of the model to key patterns.

[0105] Preferably, as Figure 6 ​​As shown, the first layer of the multi-layer convolutional neural network can be a wide first-layer kernel convolutional neural network (WDCNN), and the emission prediction system uses the wide first-layer kernel convolutional neural network to smooth local changes on the time axis using a large receptive field.

[0106] Sensor data collected under actual operating conditions of a gas turbine often has some degree of noise, including high-frequency jitter, transient outliers, and measurement errors. These high-frequency noises not only affect the feature extraction effect of the model, but also may interfere with the accurate identification of NOx emission mutations of the gas turbine. The emission prediction system introduces a wide first-layer kernel convolutional structure in the first layer of the CNN network, which uses a large receptive field to smooth local changes on the time axis to alleviate the noise problem in the historical data set. The wide first-layer kernel convolutional structure can weaken high-frequency interference signals and highlight low-frequency trends and key structures in the data, thereby achieving a certain degree of noise reduction and improving feature extraction effect in the early stage of feature extraction. In some embodiments, the emission prediction system can add a small convolution kernel in the subsequent network of the wide first-layer kernel convolutional structure, which can show strong sensitivity to mutation signals in time series data, thereby achieving fine local feature extraction.

[0107] Those skilled in the art can understand that the combination of the wide first-layer kernel convolutional structure and the multi-layer one-dimensional CNN described above is only one non-limiting embodiment provided by the present application, which is intended to clearly demonstrate the main idea of the present application and provide a specific scheme for the public to implement, rather than to limit the protection scope of the present application. Alternatively, in other embodiments, those skilled in the art can also apply other types of convolutional networks to one-dimensional CNN based on the concept of the present application to achieve the same technical effects of weakening high-frequency interference signals, achieving a certain degree of noise reduction, and fine local feature extraction.

[0108] Then, the prediction model uses an efficient channel attention (ECA) mechanism to achieve efficient modeling and dynamic adjustment of channel information. The prediction model introduces the channel attention mechanism in the multi-layer convolutional neural network, which can enhance the modeling ability of the prediction model for the dependency relationship between multi-channel features.

[0109] The efficient channel attention mechanism can efficiently capture cross-channel dependencies without dimension reduction through one-dimensional convolution and Sigmoid function, achieve dynamic adjustment of feature channels, and thus improve the feature expression ability of the model.

[0110] The efficient channel attention mechanism is simplified on the basis of the SE (Squeeze-and-Excitation) attention mechanism. The SE attention mechanism extracts channel statistical features through global average pooling, maps through two fully connected layers, and generates a weight vector through sigmoid activation to realize channel weighting of the feature map. The structure of the SE attention mechanism is simple and effective, but introduces more parameters and computational complexity. The efficient channel attention mechanism removes the fully connected layer on the basis of the SE attention mechanism and replaces it with one-dimensional convolution to realize cross-channel local information interaction. This structure has very low parameter amount while retaining the ability to model channel relationships, thereby maintaining high efficiency without increasing model complexity and significantly improving the accuracy and robustness of the prediction model for NOx emission prediction in complex industrial scenarios.

[0111] In some embodiments, the efficient channel attention (ECA) mechanism includes a global average pooling layer, an adaptive one-dimensional convolution layer, an activation function layer, and a channel weighting unit. The activation function layer can be a sigmoid function. As a channel attention module, the ECA can realize efficient modeling and dynamic adjustment of channel information.

[0112] In some embodiments, the formula calculation process of the efficient channel attention (ECA) mechanism is as follows.

[0113] First, the input feature representation. Here, let the input feature map be where, represents the batch size (the size of the data processed at one time in the training process), represents the number of channels (feature dimension), represents the time step (sequence length).

[0114] After that, the efficient channel attention mechanism performs global average pooling through the global average pooling layer to compress the time dimension for each channel:

[0115] (1)

[0116] where, is the input feature of the th channel at the th time step, is the pooling result of the th channel.

[0117] Then, the efficient channel attention mechanism performs adaptive convolution kernel calculation through the adaptive one-dimensional convolution layer to dynamically determine the 1D convolution kernel size :

[0118] (2)

[0119] where, and are hyperparameters, denotes taking the closest odd number.

[0120] Next, the efficient channel attention mechanism performs cross-channel interaction through an activation function layer and applies a 1D convolution to realize local cross-channel interaction:

[0121] (3)

[0122] where, is a learnable weight, is a one-dimensional convolution operation, is a Sigmoid activation function.

[0123] Then, the efficient channel attention mechanism generates channel weights through a channel weighting unit to determine the final attention weights:

[0124] (4)

[0125] As shown in Figure 2 and Figure 6 , after efficiently modeling and dynamically adjusting the channel information through the efficient channel attention mechanism, the weighted features are input into the subsequent LSTM to model long-term temporal dependencies.

[0126] LSTM is a special recurrent neural network (RNN) structure. LSTM introduces memory cells and three gating mechanisms (forget gate , input gate , and output gate ) to retain key information and forget irrelevant information, thus more effectively modeling long-term dependencies and widely used in time series prediction tasks.

[0127] The forget gate can determine which information is "forgotten" from the memory cell:

[0128] (5)

[0129] where, is a Sigmoid function, is a weight, is a bias term, is the previous time step cell output, is the current input.

[0130] The input gate can integrate the previous time step hidden state and the current input feature, and after calculation, the input gate and candidate memory content :

[0131] (6)

[0132] (7)

[0133] wherein, is a state activation function, , is a weight, , is a bias term.

[0134] Through the joint action of the input gate and the forget gate , the LSTM module can update the memory cell state at the current time :

[0135] (8)

[0136] The output gate filters the key information in the memory cell as the network output, and the output value is multiplied by the activated memory state and the activated memory state to generate the current hidden state :

[0137] (9)

[0138] (10)

[0139] wherein, is a weight, is a bias term.

[0140] Then, as shown in Figure 2 and Figure 6 , the prediction model introduces a trend awareness channel and a time cross-attention mechanism (TCTCA) module.

[0141] As shown in Figure 6 , the trend awareness channel and the time cross-attention mechanism (TCTCA) take the output of the LSTM as the Query, dynamically redistribute the Key / Value in combination with the trend similarity, and enhance the modeling ability of the prediction model for the change law of the gas turbine NOx emission. The prediction model outputs the predicted value of NOx emission through a fully connected layer as the output layer, thereby completing the construction of the prediction model (WDCNN-ECA-LSTM-TCTCA model).

[0142] The trend-aware channel and time cross attention mechanism (TCTCA) module refers to a module formed by introducing a trend similarity term on the basis of a channel and time cross attention mechanism (CTCA, Channel-wise Temporal Cross Attention). The TCTCA module extracts a local trend vector in a differential manner, and corrects an attention scoring function based on the similarity of trend directions between time steps.

[0143] In the actual operation of a gas turbine, the NOx emission is affected by multiple variables (such as an inlet temperature, a gas flow, a load, a turbine outlet temperature, etc.) and also by the operation time of the unit, and the importance of different variables may be different in different time periods, having significant time sequence dependence and complex cross-correlation.

[0144] The channel and time cross attention mechanism (CTCA) can further improve the perception ability of a prediction model for fluctuating signals and the modeling ability of dependence between multi-channel features. The CTCA module introduces a cross-channel attention mechanism based on time features on the basis of local features extracted by a CNN, dynamically adjusts the importance of different channel features, and thus enables the prediction model to focus more on key variables in time series data that are closely related to mutation behaviors.

[0145] Specifically, the CTCA module takes the output features of the aforementioned CNN-ECA module as input, generates attention weights between channels, and performs weighted fusion on the channel features to improve the quality of feature representation. Subsequently, the CTCA module inputs the weighted features into an LSTM to model long-term time sequence dependence. This mechanism effectively enhances the response ability of the prediction model to sudden changes in NOx emission of a gas turbine while retaining local and global feature structures.

[0146] However, in the task of predicting NOx emission of a gas turbine, sequence data often presents obvious time trend and mutation response characteristics, such as a slowly varying process caused by fluctuations in operating load or control strategies, and rapid mutation behaviors in peak-shaving states. Therefore, the prediction model not only needs to identify important variable channels at the current time, but also needs to focus on the consistency of trends in the time dimension.

[0147] The CTCA module has a certain time modeling capability by combining time sequence feature information to guide attention weighting on the channel features extracted by the CNN-ECA. However, the attention distribution of the CTCA module is still mainly based on the feature similarity between Query and Key, and it is difficult to explicitly model the consistency of trend directions between time steps. When processing gas turbine NOx emission data containing significant trend disturbances, this attention mechanism that ignores trend evolution may cause the model to fail to focus on the truly key time periods and channels.

[0148] Therefore, the emission prediction system introduces a trend similarity term on the basis of the CTCA module, and proposes a trend-aware channel-time cross-attention mechanism (TCTCA).

[0149] The TCTCA module extracts local trend vectors in a differential manner, and corrects the attention scoring function based on the similarity of trend directions between time steps, thereby enhancing the modeling capability of the prediction model for the change law of NOx emissions, and significantly improving the response effect to long-term trend changes and local mutation behaviors.

[0150] The specific process of TCTCA is described in detail below.

[0151] First, the CTCA module realizes cross-modeling in the time dimension and the channel dimension to enhance the attention ability of the prediction model to mutation time and key channels.

[0152] The CTCA module receives two inputs: Query input, one of which comes from the output sequence of LSTM , used to model time-dependent relationships and capture global temporal context; the other Key / Value input comes from the channel features of CNN-ECA , used to represent local temporal features and preserve inter-channel semantic information. Among them, denotes the batch size, denotes the number of channels, denotes the time step, denotes the hidden layer dimension of LSTM.

[0153] Map the inputs to the attention space with dimension , then:

[0154] , (11)

[0155] Among them, and indicate the sequence number of the time step, , , , respectively represent the query, key and value in the attention space, W Q , W K , W V are the corresponding weight matrices, denotes the feature dimension of the final output of the attention module.

[0156] Next, calculate the attention weight:

[0157] (12)

[0158] The attention weight calculation operation dynamically weights different channels at each time step, thereby achieving inter-channel feature redistribution guided by the LSTM output. The final output features... The obtained output features By combining the global temporal dependency information of LSTM encoding with the channel local features extracted by CNN-ECA, a channel representation with temporal awareness capability is constructed.

[0159] The prediction model incorporates a trend similarity term based on the CTCA module. To improve the original attention scoring function:

[0160] (13)

[0161] in, The score is given to the attention level after the trend is enhanced.

[0162] To enable the attention mechanism to focus on time steps with similar trends, the TCTCA module also introduces a trend vector. :

[0163] (14)

[0164] (15)

[0165] in, Equation (14) is the second-order difference, and Equation (15) is the trend similarity calculation.

[0166] Attention score results after trend enhancement The input is fed into the Softmax layer to obtain the first... Attention distribution across all channels at each time step :

[0167] (16)

[0168] By performing a weighted summation on the Value features, we obtain the output features that fuse channel context and time trend information:

[0169] (17)

[0170] Thus, the emissions prediction system has completed the construction of its prediction model (WDCNN-ECA-LSTM-TCTCA model). Please continue to refer to... Figure 2 The emissions prediction system can feed the processed historical dataset into the constructed prediction model for training.

[0171] To measure the prediction performance of the prediction model, the emission prediction system uses Root Mean Squared Error (RMSE) and Coefficient of Determination (R²) to evaluate the prediction model.

[0172] RMSE is an indicator that measures the difference between the predicted value and the actual value of the model:

[0173] (18)

[0174] wherein, is the actual value, is the predicted value, is the number of data points.

[0175] The smaller the value of RMSE, the smaller the difference between the predicted value and the actual value of the model, i.e., the higher the prediction accuracy of the model.

[0176] R², the Coefficient of Determination, is a statistical measure of the goodness of fit of the model. Its value is between 0 and 1, and the closer the value is to 1, the stronger the model's ability to explain the data. The calculation formula of R² is as follows:

[0177] (19)

[0178] wherein, is the actual value, is the predicted value, is the average value of the actual value, is the number of data points.

[0179] To suppress model overfitting, the emission prediction system uses Dropout mechanism and regularization method during training. Specifically, Dropout is added after the LSTM layer and the fully connected layer, and L2 regularization is added in the loss function. In some embodiments, the emission prediction system uses mean square error as the loss function and uses Adam optimizer to update the parameters.

[0180] During the training process, the emission prediction system checks the training effect of the prediction model according to the divided validation set, and selects the model with the best performance of the validation set as the final prediction model. In some embodiments, the change process of the MSE loss of the prediction model during the training process is as shown in Figure 7 .

[0181] In addition, the size of the sliding window can affect the prediction results when the emission prediction system processes the dataset using the sliding window. A smaller window is suitable for capturing short-term dependencies and local changes, and has a fast training speed, but can lose long-term dependency information. A larger window is suitable for capturing long-term dependencies, providing more context information, but has a slow training speed, consumes a large amount of computing resources, and can cause overfitting. In some embodiments, the emission prediction system can use a grid search to try different sliding window sizes and select the window size that works best on the validation set. The prediction results of different window sizes are shown in Table 2.

[0182]

[0183] Table 2

[0184] In one or more embodiments, in order to verify the prediction performance of the prediction model provided by the present application and the traditional model, the WDCNN-ECA-LSTM-TCTCA model provided by the present application is compared with the LSTM model, the GRU model, the CNN model, the CNN-LSTM model, and the model in which part of the structure of the prediction model provided by the present application is removed or replaced on the validation set.

[0185] Please refer to Figure 8 and Figure 9 , Figure 8 shows a schematic diagram of the prediction results of the LSTM model, Figure 9 shows a schematic diagram of the prediction results of the GRU model.

[0186] As Figure 8 shown, combined with the prediction value curve, the true value curve and the quantitative evaluation index in Table 3, it can be seen that the LSTM model can better fit the change trend of NOx as a whole and has certain tracking ability. However, LSTM has an advantage in capturing long-term dependencies, but there can still be deficiencies in how to select information of different time periods. As Figure 8 can be seen, the prediction curve of LSTM is relatively smooth as a whole, because NOx emissions usually fluctuate within a relatively stable range, and the model only needs to fit this interval to obtain a smaller loss function and a higher prediction score. Similarly, as Figure 9 shown, the results of the GRU model are similar to those of the LSTM model.

[0187]

[0188] Table 3

[0189] Please refer to Figure 10 , Figure 10 shows a schematic diagram of the prediction results of the CNN model.

[0190] To further explore the response ability of the model to local mutations, the skilled person also predicts a one-dimensional convolutional neural network (1D CNN) model on the validation set, and the results are shown in Figure 10 .

[0191] Although the traditional CNN module can extract local time sequence features, it treats all channels equally when processing multi-channel input and cannot dynamically distinguish the importance of each channel under different working conditions. This may cause the CNN to easily amplify the interference information of redundant channels, suppress the expression ability of key variables, and affect the generalization performance of the model.

[0192] From Figure 10 , it can be observed that the prediction results of the CNN model have more obvious fluctuations, indicating that the CNN model can more sensitively capture local feature changes. However, from the overall trend fitting perspective, its prediction ability is relatively insufficient, with an R² of only 0.4848, lower than the LSTM's 0.5899. This shows that CNN is more inclined to short-term signal extraction and has difficulty grasping the global evolution trend of NOx emissions.

[0193] Please refer to Figure 11 , Figure 11 , which shows a schematic diagram of the prediction results of the CNN-LSTM model.

[0194] By predicting the gas turbine NOx through the CNN-LSTM network, the advantages of local feature extraction (CNN) and long-term dependence modeling (LSTM) can be combined to some extent. However, in the gas turbine NOx emission prediction task, the model usually receives input features from multiple sensor channels, and there is redundancy, coupling, and even mutual interference between these features.

[0195] As shown in Figure 11 , the prediction results of the CNN-LSTM model are smoother than those of the CNN model in Figure 10 , with reduced overall fluctuations, indicating that the LSTM has successfully introduced trend modeling capabilities. At the same time, compared to the LSTM model, more fluctuations appear in the prediction, reflecting the enhanced response of the model to NOx mutations.

[0196] However, for combined changes in trends and mutations, the response is still not ideal. This is because the CNN-LSTM lacks an attention mechanism and treats each time step and feature channel equally, failing to highlight key change areas, resulting in deficiencies in modeling trends and local fluctuations.

[0197] Please refer to Figure 12 , Figure 12 , which shows a schematic diagram of the prediction results of the CNN-ECA-LSTM-Attention model.

[0198] As Figure 12 shown, the performance of the CNN-ECA-LSTM-Attention model is significantly improved after introducing the ECA channel attention and the time attention. The R² is increased to 0.8127, and the RMSE is reduced to 1.3483, which indicates that the channel weighting helps to enhance the focusing ability of the model on key variables. However, the time attention optimizes the distribution of the time weight, but lacks the trend perception ability in the channel attention.

[0199] Please refer to Figure 13 and Figure 14 , Figure 13 a schematic diagram of the prediction results of the WDCNN-ECA-LSTM-SA model is shown, Figure 14 a schematic diagram of the prediction results of the WDCNN-ECA-LSTM-TCTCA model provided by some embodiments of the present application is shown.

[0200] As Figure 13 shown, on the basis of the aforementioned model, the R² is further increased to 0.8412, and the effect of the model can be further enhanced. However, since the self-attention mechanism only models the global dependence based on feature similarity, it lacks explicit perception of trend changes and key region guidance, so in the prediction task of NOx emission sequences with significant trend and mutation, the self-attention effect of the WDCNN-ECA-LSTM-SA model is slightly insufficient compared with the prediction model of the TCTCA module of the present application which introduces the trend perception and cross-channel interaction mechanism. As Figure 14 shown, the WDCNN-ECA-LSTM-TCTCA model proposed by the present application performs best in all indicators, with R² reaching 0.8974, RMSE reducing to 0.8617, and MAE being only 0.5561, which is better than other comparative models.

[0201] In addition, please refer to Figure 15 , Figure 15 a box plot of the prediction errors of the prediction model provided by the present application and other comparative models is shown.

[0202] As Figure 15 shown, the horizontal axis in the figure represents different prediction models, and the vertical axis is the prediction error value (predicted value minus true value). The red dotted line represents the zero error baseline. From Figure 15It can be observed that the error distribution of most models is generally symmetrical, but the median is generally negative, indicating a certain degree of underestimation trend. Among them, the traditional LSTM model (Model 1), GRU model (Model 2) and CNN model (Model 3) are generally in terms of error range, upper and lower quartile range and number of outliers. The error fluctuation of the CNN model is the largest, and its R² value is 0.4848 from Table 3. The CNN-LSTM model (Model 4) with time modeling capability introduced has improved the overall prediction performance compared with the traditional model, and the error distribution tends to converge. Further, Model 5 (i.e. CNN-ECA-LSTM-Attention model) and Model 6 (i.e. WDCNN-ECA-LSTM-SA model) with ECA and time attention module (Attention or SA mechanism) can obviously converge the error distribution, and the median is closer to zero, and the prediction accuracy and stability are improved. Compared with the above models, the box range of Model 7 (WDCNN-ECA-LSTM-TCTCA) provided by the application is the smallest, and the error median is almost aligned with zero. This shows that the proposed WDCNN-ECA-LSTM-TCTCA has small error fluctuation, and the prediction result is more concentrated and stable.

[0203] Reference is made to Figure 16 , Figure 16 A flowchart of an online analysis phase of a method for predicting NOx emissions of a gas turbine according to some embodiments of the application is shown.

[0204] As shown in Figure 16 After completing the training of the prediction model in the foregoing offline training phase, the emission prediction system can obtain a set of operation monitoring data of the gas turbine unit under actual operation.

[0205] Then, according to the generation mechanism, thermal process, combustion reaction and operating condition of NOx emissions of the gas turbine, the feature variable screening is performed on the operation monitoring data to determine a first feature variable set. Preferably, the emission prediction system can further perform secondary screening on the first feature variable set by a preset variance threshold.

[0206] After that, the emission prediction system performs feature variable selection on the first feature variable set by the foregoing recursive feature elimination method based on LightGBM to determine a second feature variable set.

[0207] Then, the emission prediction system processes the set of operation monitoring data according to the second feature variable set, and takes the processed set of operation monitoring data as input to realize the prediction of NOx emissions according to the trained prediction model.

[0208] In conclusion, the NOx emission prediction method for the gas turbine provided by the present application eliminates redundant features to reduce the complexity of the model, improves the feature expression pertinence of the prediction model, the attention ability to key information, and the modeling ability to the complex evolution process of NOx emission, effectively improves the prediction accuracy of the prediction model to NOx emission, has the advantages of high precision, stability, and wide applicability, and the like, provides strong technical support for optimizing the operation of the heavy-duty gas turbine and reducing NOx emission, and has important significance for promoting the environmental protection reconstruction and sustainable development of the heavy-duty gas turbine.

[0209] Although the above-described methods are illustrated and described as a series of acts, it will be appreciated that the methods are not limited by the order of acts as some acts can, in accordance with one or more embodiments, occur simultaneously or in different orders than shown and described herein. Moreover, not all acts can be required for the practice or implementation of the methods.

[0210] Those skilled in the art will understand that information, signals, and data can be represented using any of a variety of different technologies and techniques. For example, data, instructions, commands, information, signals, bits, symbols, and chips that can be referenced throughout the above description can be represented by voltages, currents, electromagnetic waves, magnetic fields or particles, optical fields or particles, or any combination thereof.

[0211] Those skilled in the art will further appreciate that the various illustrative logical blocks, modules, circuits, and algorithm steps described in connection with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or combinations of both. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans can implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application.

[0212] The various illustrative logical blocks, circuits, and circuits described in connection with the embodiments disclosed herein can be implemented or performed with a general purpose processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. The general purpose processor can be a microprocessor, but in the alternative, the processor can be any conventional processor, controller, microcontroller, or state machine. A processor can also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration.

[0213] The steps of a method or algorithm described in connection with the embodiments disclosed herein can be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module can reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor such that the processor can read information from, and write information to, the storage medium. In the alternative, the storage medium can be integral to the processor. The processor and the storage medium can reside in an ASIC. The ASIC can reside in a user terminal. In the alternative, the processor and the storage medium can reside as discrete components in a user terminal.

[0214] In one or more exemplary embodiments, the functions described can be implemented in hardware, software, firmware, or any combination thereof. If implemented in software as a computer program product, the functions can be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Computer-readable media includes both computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. A storage media can be any available media that can be accessed by a computer. By way of example, and not limitation, such computer-readable media can comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer. Also, any connection is properly termed a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using a coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of medium. Disk and disc, as used herein, includes compact disc (CD), laser disc, optical disc, digital versatile disc (DVD), floppy disk and blu-ray disc where disks usually reproduce data magnetically, while discs reproduce data optically with lasers. Combinations of the above should also be included within the scope of computer-readable media.

[0215] The previous description of the disclosure is provided to enable any person skilled in the art to make or use the disclosure. Various modifications to the disclosure will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other variations without departing from the spirit or scope of the disclosure. Thus, the disclosure is not intended to be limited to the examples described herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting NOx emissions from gas turbines, characterized in that, Including the following steps: Obtain the operation monitoring dataset of the gas turbine unit, and filter the operation monitoring data for feature variables based on the NOx emission generation mechanism, thermodynamic process, combustion reaction and operating conditions of the gas turbine to determine the first set of feature variables; The recursive feature elimination method based on LightGBM is used to select feature variables from the first feature variable set to determine the second feature variable set; as well as The operation monitoring dataset is processed according to the second set of feature variables, and the processed operation monitoring dataset is used as input to predict NOx emissions according to the prediction model. The prediction model includes a multi-layer convolutional neural network for extracting local features of the input data, a channel attention mechanism for capturing cross-feature channel dependencies, a long short-term memory network, and a trend-aware channel and time cross-attention mechanism.

2. The method for predicting nitrogen oxide emissions as described in claim 1, characterized in that, The step of filtering feature variables in the operation monitoring data also includes: The first set of feature variables is further filtered based on a preset variance threshold.

3. The emission prediction method as described in claim 1, characterized in that, The LightGBM-based recursive feature elimination method includes the following steps for selecting feature variables from the filtered operational monitoring data: The LightGBM model is trained using the first set of feature variables to obtain the feature importance of each feature variable in the first set of feature variables. The LightGBM model is recursively removed from a predetermined number of feature variables with the lowest feature importance, and cross-validation is used to evaluate the performance of the model in each iteration; and The iteration is stopped according to the set termination condition, and the subset with the best performance is selected as the second feature variable set.

4. The emission prediction method as described in claim 1, characterized in that, The first layer of the multilayer convolutional neural network is a wide-kernel convolutional structure.

5. The emission prediction method as described in claim 1, characterized in that, The channel attention mechanism includes a global average pooling layer, an adaptive one-dimensional convolutional layer, an activation function layer, and a channel weighting unit.

6. The emission prediction method as described in claim 1, characterized in that, The trend-aware channel and time cross-attention mechanism is a channel and time cross-attention module that introduces a trend similarity term.

7. The emission prediction method as described in claim 6, characterized in that, The trend-aware channel and time-crossing attention mechanism includes the following steps: Cross-modeling of the time dimension and the channel dimension is achieved through the aforementioned channel and time cross-attention module; as well as Local trend vectors are extracted using a differential method, and the attention scoring function is modified based on the similarity of trend directions between time steps.

8. The emission prediction method as described in claim 1, characterized in that, The training phase of the prediction model includes the following steps: Historical datasets of gas turbine units are obtained, and feature variables are filtered from the historical datasets based on the NOx emission generation mechanism, thermodynamic process, combustion reaction and operating conditions of the gas turbine to determine the first set of feature variables; The recursive feature elimination method based on LightGBM is used to select feature variables from the first feature variable set to determine the second feature variable set; as well as The historical dataset is processed based on the second set of feature variables, and the processed historical dataset is used as input to train the prediction model.

9. A NOx emission prediction system for gas turbines, characterized in that, include: Memory, on which computer instructions are stored; as well as A processor, connected to the memory, and configured to execute computer instructions stored in the memory to implement the NOx emission prediction method for a gas turbine as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When the computer instructions are executed by the processor, the NOx emission prediction method for gas turbines as described in any one of claims 1 to 8 is implemented.

Images