SCR inlet NOx generation concentration prediction method and system based on LSTM and attention mechanism

Abstract

The invention discloses an SCR inlet NOx generation concentration prediction method and system based on LSTM and an attention mechanism, and the method comprises the steps: obtaining observation data of a desulfurization system, carrying out the preprocessing of the observation data, and obtaining the preprocessed observation data; the pre-processed observation data are input into the trained ADS-Forecaster prediction model, the NOx generation concentration of the SCR inlet is obtained through prediction, the NOx generation concentration of the SCR inlet can be accurately predicted through the method, and the NOx generation concentration of the SCR inlet can be accurately predicted through the method and system.

Description

Technical Field

[0001] This invention belongs to the field of energy conservation and environmental protection, and relates to a method and system for predicting the NOx generation concentration at the inlet of an SCR based on LSTM and attention mechanism. Background Technology

[0002] Despite the rapid development of new energy technologies, coal-fired power generation remains the cornerstone of China's electricity supply. To address the harmful nitrogen oxides (NOxides) produced by coal combustion... x Regarding the environmental and health risks associated with NOx emissions, Selective Catalytic Reduction (SCR) systems, with their superior denitrification efficiency and reliability, have become the preferred method for coal-fired power units to achieve emissions standards for pollutants such as NOx. x Concentration <50mg / Nm 3 The core technology of SCR reactor (superior to ultra-low emission standards). SCR reactor inlet NO x Accurate concentration measurement is crucial for optimizing operation (such as controlling ammonia injection to reduce escape).

[0003] NO at the inlet of the SCR reactor x NO concentration is a key parameter for evaluating combustion efficiency and front-end pollution control, but its dynamic response is significantly affected by the inherent physical time delays in boiler combustion and heat transfer processes. This time delay from fuel input and combustion to flue gas transport makes it difficult to accurately reflect transient NO at the SCR inlet under transient conditions. x Concentration hinders the realization of real-time source optimization control. Therefore, there is an urgent need to develop high-precision real-time prediction technology for SCR inlet NO. x An advanced model for concentration.

[0004] Non-temporal machine learning methods can quickly establish input-output mappings, and their powerful nonlinear approximation capabilities are particularly useful in processing large-scale industrial data for predicting NO. x While this method has advantages in generating concentrations, it somewhat ignores the time dependence of process data, relying solely on instantaneous variable relationships for modeling. Furthermore, it often requires training under static process conditions, severely limiting its applicability in the actual dynamic operating environments of power plants (such as rapid load changes). Deep learning methods, through hierarchical feature extraction, utilize hierarchical feature representations to model multivariate nonlinear relationships, demonstrating potential in addressing the challenges of complex industrial systems. Nevertheless, these methods still have shortcomings in effectively identifying time-series information, especially in accurately modeling complex interactive dynamics within multivariate systems.

[0005] Recurrent Neural Network (RNN) architectures and their variants demonstrate stronger temporal modeling capabilities through explicit memory mechanisms, making them particularly effective for dynamic systems exhibiting long-term temporal dependencies. While these enhanced RNN implementations show better time processing capabilities compared to traditional methods, their robustness in sparse, noisy industrial environments still needs improvement.

[0006] Attention-based computational architectures, leveraging adaptive feature saliency weighting mechanisms, have become a prominent research paradigm in deep learning. These frameworks demonstrate a unique ability to solve multivariate interaction problems in industrial systems by dynamically allocating computational resources to key temporal patterns. This architectural innovation effectively focuses on NO... x The concentration curves exhibit strong nonlinear correlations with process variables, while simultaneously suppressing noise interference.

[0007] Although existing neural network methods have NO x While progress has been made in generative predictive modeling and feature interaction, the core limitation lies in the insufficient collaborative modeling of cross-feature interactions and multi-scale temporal dynamics. This lack of "feature-temporal collaborative modeling capability" makes it difficult to effectively address the triple challenges of multivariate interactions, reaction lags, and dynamic operation in SCR systems. Existing attention models and architectures such as LSTM are all "single-focus" designs, making it difficult to achieve parallel analysis of the current state and historical dependencies, as well as adaptive feature-temporal joint optimization under dynamic operating conditions without losing key information, ultimately leading to a decrease in prediction accuracy. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for predicting NOx generation concentration at the SCR inlet based on LSTM and attention mechanism. This method and system can accurately predict the NOx generation concentration at the SCR inlet.

[0009] To achieve the above objectives, this invention discloses a method for predicting NOx generation concentration at the SCR inlet based on LSTM and attention mechanisms, comprising: Acquire observation data from the desulfurization system, and preprocess the observation data to obtain preprocessed observation data; The preprocessed observation data is input into the trained ADS-Forecaster prediction model to predict the NOx generation concentration at the SCR inlet.

[0010] The further improvement of the SCR inlet NOx generation concentration prediction method based on LSTM and attention mechanism described in this invention lies in: Furthermore, the observation data of the desulfurization system include total coal feed, total air volume, unit load, inlet O2 concentration, inlet CO concentration, inlet flue gas temperature, inlet flue gas pressure, and main steam temperature.

[0011] Furthermore, the ADS-Forecaster prediction model includes a dual-path feature encoder and a decoder based on a self-attention fusion layer. The dual-path feature encoder comprises a univariate temporal encoder and a cross-variable dependency encoder. The univariate temporal encoder utilizes an LSTM network to extract the temporal dynamics within a single process variable; the cross-variable dependency encoder utilizes an attention mechanism to explore the nonlinear interactions between different feature variables; and the decoder based on the self-attention fusion layer adaptively fuses the dual-path encoded features through temporal dynamic attention and cross-variable coupling attention layers, driving the LSTM module to generate NO. x Generate concentration prediction values.

[0012] Furthermore, the loss function of the ADS-Forecaster prediction model during training is: ) in, Indicates sample time step The actual measured value; This indicates its corresponding predicted value; Indicates the number of samples; Indicates the range of predictions.

[0013] Furthermore, the ADS-Forecaster prediction model is trained using the Adam stochastic gradient descent algorithm.

[0014] Furthermore, the preprocessing of the observation data includes: outlier handling, time delay analysis, and key feature variable screening.

[0015] This invention discloses a NOx generation concentration prediction system for SCR inlet based on LSTM and attention mechanism, comprising: The acquisition module is used to acquire observation data of the desulfurization system, preprocess the observation data, and obtain preprocessed observation data. The prediction module is used to input the preprocessed observation data into the trained ADS-Forecaster prediction model to predict the NOx generation concentration at the SCR inlet.

[0016] The further improvement of the SCR inlet NOx generation concentration prediction system based on LSTM and attention mechanism described in this invention lies in: Furthermore, the ADS-Forecaster prediction model includes a dual-path feature encoder and a decoder based on a self-attention fusion layer. The dual-path feature encoder comprises a univariate temporal encoder and a cross-variable dependency encoder. The univariate temporal encoder utilizes an LSTM network to extract the temporal dynamics within a single process variable; the cross-variable dependency encoder utilizes an attention mechanism to explore the nonlinear interactions between different feature variables; and the decoder based on the self-attention fusion layer adaptively fuses the dual-path encoded features through temporal dynamic attention and cross-variable coupling attention layers, driving the LSTM module to generate NO. x Generate concentration prediction values.

[0017] Furthermore, the loss function of the ADS-Forecaster prediction model during training is: ) in, Indicates sample time step The actual measured value; This indicates its corresponding predicted value; Indicates the number of samples; Indicates the range of predictions.

[0018] Furthermore, the ADS-Forecaster prediction model is trained using the Adam stochastic gradient descent algorithm.

[0019] The present invention has the following beneficial effects: The SCR inlet NOx generation concentration prediction method and system based on LSTM and attention mechanism described in this invention combines attention-enhanced LSTM with advanced data processing techniques for modeling SCR reactor denitrification systems. This method predicts NOx generation concentration at the SCR reactor inlet. x The evaluation of the maximum information coefficient of time offset between concentration and auxiliary feature variables solves the problem of traditional NO x This addresses the issue of time information loss in prediction, enabling the simultaneous identification of optimal input variables and their feature delay intervals. A grey relational analysis algorithm is employed to evaluate the relationship between each pre-selected auxiliary feature variable and the SCR entry point NO. x The correlation between concentrations is analyzed, and variables with low correlation to the predicted target are removed. The core prediction architecture employs dual-channel time modeling: one branch utilizes an LSTM network to capture the temporal dynamics within a single process variable through sequential pattern recognition; the other branch applies an attention mechanism to analyze the characteristic relationships between various operating parameters through adaptive cross-variable weighting. The proposed method can comprehensively utilize direct operating parameters and derived time-frequency characteristics, thereby improving the NO inlet temperature of the SCR reactor. x Accuracy of concentration prediction. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the boiler SCR system structure; Figure 2 A schematic diagram of the ADS-Forecaster structure; Figure 3 NO x Flowchart of the concentration prediction process; Figure 4(a) is a schematic diagram of the unprocessed data; Figure 4(b) is a schematic diagram of the results after 3σ detection preprocessing; Figure 5 NO for test set 1 x Concentration prediction curve; Figure 6 NO for test set 2 x Concentration prediction curve; Figure 7 A schematic diagram showing the relative prediction errors of each model on test set 1; Figure 8 This is a schematic diagram of the relative prediction errors of each model on test set 2. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] In the description of this invention, it should be understood that the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0024] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0025] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Additionally, the character " / " in this invention generally indicates that the preceding and following objects have an "or" relationship.

[0026] It should be understood that although terms such as first, second, third, etc., may be used in the embodiments of the present invention to describe the preset range, these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from one another. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.

[0027] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0029] The accompanying drawings illustrate various structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0030] Example 1 The SCR inlet NOx generation concentration prediction method based on LSTM and attention mechanism described in this invention includes the following steps: 1) Acquire observation data, wherein the observation data includes at least the total coal feed, total air volume, unit load, inlet O2 concentration, inlet CO concentration, inlet flue gas temperature, inlet flue gas pressure, and main steam temperature; Specifically, the structure of the denitrification system for subcritical coal-fired power units is as follows: Figure 1 As shown, its core equipment is an SCR reactor, which adopts a vertical downward design and operates within a specific temperature window (285-420℃). The catalyst uses a honeycomb V2O5-WO3 / TiO2 system. The system monitors key parameters in real time, including catalyst bed pressure difference, ammonia slip concentration, and NO. x Conversion efficiency.

[0031] SCR reactor inlet NO x NO concentration is a direct product of the boiler combustion process, and its changes are dynamically influenced by the coupling of multiple combustion parameters. Thermal NO x It is mainly related to the high-temperature zone of the furnace (above 1200℃), and can be indirectly reflected by parameters such as secondary air volume and burnout air opening; fuel-type NO x This is related to nitrogen release from coal and requires consideration of parameters such as total coal quantity, mill output, and primary air pressure. Although the power plant's distributed control system (DCS) lacks direct measurement of furnace temperature and real-time coal quality, economizer outlet flue gas temperature and SCR inlet / outlet NO₂ are important parameters. x Concentration gradient, main steam pressure and other parameters can indirectly characterize critical states.

[0032] In the process of selecting characteristic variables, it is necessary to thoroughly analyze the factors affecting NO at the inlet of the SCR reactor. x Key combustion process parameters and their coupling mechanisms for NO concentration. SCR inlet NO x NO concentration is essentially a direct product of the boiler combustion process. Fluctuations in boiler load significantly alter the combustion intensity, temperature distribution, and flue gas flow rate within the furnace, thereby affecting NO concentration. xThe generation path and rate of NO. Furthermore, the dynamic ratio of total coal feed to total air volume directly affects the O2 concentration and temperature in the combustion zone, which is crucial for controlling the generation of fuel-type and thermal-type NO. x The key factors in combustion formation are: the SCR inlet O2 concentration reflects the overall excess air coefficient of combustion, while the SCR inlet CO concentration indicates localized incomplete combustion; both are related to NO. x The formation of NO is closely related; the SCR inlet flue gas temperature and pressure are important indicators of the thermodynamic state of the combustion system; the main steam temperature indirectly reflects the heat absorption balance of the boiler's heating surfaces and is related to the SCR inlet flue gas temperature, which is a thermal NO generation factor. x Key influencing factors for generation. Based on historical data characteristics of the unit and the experience of power plant experts, the preliminary selected characteristic variables are shown in Table 1, specifically including total coal feed, total air volume, unit load, inlet O2 concentration, inlet CO concentration, inlet flue gas temperature, inlet flue gas pressure, and main steam temperature. This feature set aims to comprehensively capture the NO driving SCR inlet. x Multi-physics state information of combustion sources with dynamic concentration changes.

[0033] Table 1

[0034] 2) Construct the ADS-Forecaster prediction model; The inherent complex nonlinear interdependencies in multivariate time series present significant challenges to predictive modeling. To address the inherent limitations of traditional sequence processing architectures, this invention proposes an improved NO... x The concentration generation prediction framework ADS-Forecaster systematically captures the interaction and temporal evolution patterns of cross variables through an advanced dependency learning mechanism.

[0035] The ADS-Forecaster model architecture is as follows: Figure 2 As shown, an encoder-decoder structure with two main components is adopted: a) a dual-path feature encoder based on a univariate temporal encoder and a cross-variable dependency encoder, and 2) a decoder based on a self-attention fusion layer. The dual-path feature encoder consists of parallel processing streams: the univariate temporal encoder uses an enhanced LSTM network to extract the temporal dynamics within a single process variable, while the cross-variable dependency encoder utilizes an attention mechanism to explore the nonlinear interactions between different feature variables; the decoder based on the self-attention fusion layer adaptively fuses the dual-path encoded features through temporal dynamic attention and a cross-variable coupling attention layer, driving the LSTM module to generate NO. x Generate concentration prediction values.

[0036] To achieve high-precision NO xThe prediction process in this invention is based on the ADS-Forecaster model. Specifically, it includes data acquisition, outlier handling, latency analysis, and key feature variable selection. The data processed through these steps is used as model input to complete concentration prediction. The overall prediction process is as follows: Figure 3 As shown.

[0037] The ADS-Forecaster model optimizes its parameters by minimizing the mean square error (MSE) between the predicted and observed time series. Given a set of... Samples and prediction range The training dataset has the following loss function: ) in, Indicates sample time step The actual measured value, This represents the corresponding predicted value, and parameter optimization is performed using the Adam stochastic gradient descent algorithm.

[0038] Example 2 The experimental dataset was collected from the operating records of a 600MW subcritical coal-fired power unit in China. Historical operating parameters were obtained through the power plant's DCS and supervisory information system (SIS), with a sampling interval of 5 seconds. This dataset contains approximately 29 hours of operating data from December 2024, covering load conditions from 260.80 MW to 557.23 MW, and includes 21,000 sample points. Excluding the output of the prediction model, the SCR reactor inlet NO... x In addition to the generated concentrations, the dataset includes boiler system-related parameters: total coal feed, total air volume, unit load, inlet O2 concentration, inlet CO concentration, inlet flue gas temperature, inlet flue gas pressure, and main steam temperature. The dataset is divided into training, validation, and test sets in a 7:1:2 ratio. The test set, designed to ensure the model's general applicability, consists of two parts with similar timeframes: the first part depicts a scenario where the unit load linearly climbs from 311MW to approximately 476MW, remains stable with minor fluctuations, then drops sharply to approximately 327MW, before climbing again to approximately 478MW and gradually declining to approximately 331MW; the second part depicts a scenario where the unit load is concentrated between 301MW and 358MW, exhibiting frequent short-cycle peak-shaving fluctuations. Specifically, the training set consists of 14,700 sample points, while the validation set, test set 1, and test set 2 each consist of 2,100 sample points.

[0039] Outliers in industrial denitrification datasets typically deviate significantly from normal operating trends. In this embodiment, two types of outliers were identified and processed: the first type is boiler shutdown for maintenance, where the coal feed rate is less than 1 ton / hour, indicating inactive combustion and lacking predictive value; therefore, the corresponding data segments were deleted. The second type involves inaccurate values ​​due to measurement point issues; therefore, the 3σ standard was used to detect the transient error of the sensor. For a given variable... Calculate its average value ( ) and standard deviation ( After that, satisfy The samples were marked as outliers and replaced with the moving average of the previous 5 valid measurements. Preprocessed NO x The concentration curve is shown in Figure 4.

[0040] The maximal information coefficient (MIC) is used to quantify the nonlinear dependencies between variables in high-dimensional datasets. It is defined as the normalized mutual information (MI) maximized over a grid partition. For continuous variables X and Y, the MI is calculated as follows:

[0041] in, Denotes the joint probability density. , The formula for calculating the marginal density (MIC) is:

[0042] in, The MIC value represents the sample size, and the closer it is to 1, the stronger the nonlinear correlation.

[0043] SCR reactor inlet NO x As the target for prediction, the dynamic response of the generated concentration is constrained by the physical time lag of the upstream combustion and heat transfer processes. This leads to a direct correlation between the original operating parameters (recorded in the DCS database) and the combustion state. It is difficult to accurately characterize the actual NO at the SCR inlet in real time. x concentration The instantaneous changes. However, under steady-state operation of the unit, the main process path of the boiler system is relatively fixed, making the overall transmission delay from the combustion core area to the SCR inlet flue relatively constant. To eliminate the negative impact of this delay on the modeling accuracy, this prediction process uses MIC quantitative analysis to compare the input feature variables with the target variable NO. xThe nonlinear correlation between concentrations is generated, and the optimal delay time of each parameter relative to the prediction target is identified and calculated. This is then used to reconstruct the time-delay compensation dataset for deep learning model training and validation. The specific process is as follows: Modeling datasets with For the input vector, For the output variable, where, For input dimensions, The sample size and modeling process include: (1) Determining the delay range: Determining each input variable (2) Optimal delay time identification: for each input Generate candidate delay set By maximizing and Choose the optimal delay between the MICs. (2) To ensure maximum nonlinear correlation; (3) Dataset reconstruction: Delayed variables Replace the original input Maintain dimensionality Simultaneously, system lag is compensated. Globally optimal time delay parameters are obtained through iterative calculations, and a time delay compensation modeling dataset is constructed. The final input feature variables and SCR entry point NO are then obtained. x The time delay between concentrations is shown in Table 2.

[0044] Table 2

[0045] Grey relational analysis (GRA) quantifies the correlation between a target sequence and a comparison sequence by evaluating their geometric similarity in time-series evolution patterns. The closer the geometric shapes of the curves and the more synchronized their trends, the stronger the interdependence of the sequences and the higher the grey relational degree. Compared to methods such as mutual information or Pearson correlation coefficient, GRA avoids strict distribution assumptions and sample size limitations through dimensionless transformation. This characteristic ensures that NO... x Physical consistency of variable selection in concentration prediction models.

[0046] The specific analysis steps for screening key feature variables are as follows: 1a) Data normalization; Variables in the time delay compensation modeling dataset Normalize the mean by feature dimension:

[0047] in, This indicates the sample size; after normalization, the data are concentrated around 1.

[0048] 2a) Sequence definition; Normalized SCR entry NO x Generate concentration sequence As the target sequence, the time-delay compensated input variables Constructing a comparison sequence ( ).

[0049] 3a) Calculation of grey relational coefficient;

[0050] in, Represents the resolution coefficient. and The term takes a global extremum in terms of both the variable dimension and the time step.

[0051] 4a) Relationship quantification; The mean correlation coefficient is used to measure the correlation between variables. The calculation formula is as follows:

[0052] The larger the value, the stronger the correlation between the variables.

[0053] To evaluate the resolution coefficient correlation The impact of different resolution coefficients on the eight input feature variables and the SCR entry NO was calculated based on the sample dataset. x The correlation between concentrations (Table 3). The results show that the changing patterns of the correlation between the input variables are significantly consistent. For example, the variable sorting results are shown in Table 4.

[0054] Table 3

[0055] Table 4

[0056] Based on the sorting results in Table 4, the unit load and SCR inlet NO... x Concentration showed the highest correlation, while total air volume showed the lowest. To improve the model's computational efficiency, the six variables with the highest correlation were selected as valid inputs: unit load, inlet O2 concentration, inlet flue gas temperature, inlet flue gas pressure, inlet CO concentration, and main steam temperature. Total coal feed and total air volume were considered redundant variables due to their low correlation and were removed to reduce model complexity.

[0057] To better evaluate the model's predictive performance, we used MSE, mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R²). 2 The model is analyzed using this as an evaluation index, and the calculation formula is as follows:

[0058]

[0059]

[0060]

[0061] in, and They represent the first Predicted and measured values ​​for each sample; Represents the sample mean. Indicates the sample size. MSE can amplify the impact of large errors in forecasting, assessing the severity of the deviation between predicted and actual values; MAE directly reflects the mean absolute deviation between predicted and actual values, providing a robust and easily understood error metric; MAPE measures the relative error of the forecast as a percentage, intuitively reflecting the proportion of the average deviation between the predicted and actual values; R 2 It measures how well the model's predictions fit the actual trend of data changes.

[0062] The auxiliary feature variables and NO after the filtered and delayed operation will be used. x Historical concentration values ​​are generated as model inputs, and the current SCR reactor inlet NO is... x The current concentration value is used as the model output. The ADS-Forecaster model constructed earlier is used for training to generate time-step NO values. x The concentration predictions, model hyperparameters, and training configurations are shown in Table 5.

[0063] Table 5

[0064] To verify the predictive superiority of the ADS-Forecaster model, it was benchmarked against RNN, LSTM, and xLSTM models. All models used the same dataset (containing six GRA-selected variables: unit load, inlet O2 concentration, inlet flue gas temperature / pressure, inlet CO concentration, and main steam temperature) and data partitioning scheme. The model configurations were standardized as follows: LSTM and RNN each contained three hidden layers with 64 neurons per layer; the xLSTM model used an embedding dimension of 512, four attention heads, and consisted of two stacked sLSTM blocks and two mLSTM blocks; all three models used an initial learning rate of 0.001 and were trained for 200 epochs. The experimental platform consisted of an AMD Ryzen 9 9955HX3D CPU, an NVIDIA GeForce RTX 5090 Laptop GPU, 64GB of RAM, CUDA version 12.9, Python version 3.12.9, and PyTorch version 2.7.0.

[0065] Figure 5 and Figure 6 The SCR entry point NO was compared between test set 1 and test set 2. x The predicted and measured values ​​of the generated concentrations were compared. ADS-Forecaster showed a high degree of agreement with the measured values ​​on both test sets, with minimal hysteresis at the inflection point, indicating robust feature extraction capabilities. (The text then abruptly shifts to a description of magnified 100-point curve details.) Figure 5 and Figure 6 (Partial illustration) and relative error distribution ( Figure 7 and Figure 8 The performance of each model was further quantified, and the evaluation indicators were summarized in Tables 6 and 7. The bolded results in the tables represent the best results, and the underlined results represent the second-best results.

[0066] Table 6

[0067] Table 7

[0068] RNN models suffer from severe gradient vanishing and exploding problems when processing long sequences, resulting in the worst performance across all metrics (Test set 1: MSE=131.45, R...). 2 =0.851; Test set 2: MSE=58.234, R 2 =0.835), the prediction curve fluctuates wildly overall, and exhibits significant divergence at extreme points. The standard deviation of the error increases linearly with time, the cumulative error gradually increases, and abnormal oscillations occur during certain load ramp-up phases. The LSTM model, by introducing a gating mechanism and cell state, outperforms RNN (test set 1: MSE=101.68, R2 =0.885; Test set 2: MSE=44.981, R 2 =0.877), but the scalar gating mechanism still has limitations in handling long sequences and complex nonlinear relationships. The predicted curve shows obvious systematic positive shifts in certain data segments under steady-state conditions, and periodic oscillation errors appear in the low concentration range. The xLSTM model replaces the sigmoid gating with exponential gating and introduces matrix memory and normalization, and its performance significantly surpasses that of LSTM (Test set 1: MSE=68.792, R 2 =0.921; Test set 2: MSE=32.098, R 2 =0.911), the overall prediction bias of the curve is small, and high-frequency noise is well suppressed with no obvious lag effect. However, as a single-path time series model, its design mainly focuses on modeling the time-dimensional dependency relationship, lacking explicit modeling of complex nonlinear interactions between multiple variables. The overall fluctuation amplitude of the curve is large and the peak prediction is low, and the maximum negative bias is still large in the high-concentration region. The ADS-Forecaster model achieves the best performance on all indicators (test set 1: MSE=40.435, R 2 =0.946; Test set 2: MSE=18.923, R 2 =0.938), compared with the suboptimal xLSTM, the MSE was significantly reduced by 41.2% and 41.0% on the two test sets, the MAE was reduced by 20.2% and 14.9%, the MAPE was reduced by 36.9% and 30.8%, and the R² was improved by 2.5 and 2.7 percentage points. Time-series trajectory analysis shows that the ADS-Forecaster prediction results have a small fluctuation range, closely track the fluctuations of the true value, and effectively capture NO. x Although a few outliers appeared in a very small number of high-gradient transition segments (presumably due to the instantaneous imbalance of attention weights under transient conditions), the abrupt changes in concentration characteristics resulted in ADS-Forecaster achieving optimal performance across all metrics in both test conditions. The error distribution also confirmed the improvement in overall prediction accuracy and robustness. This fully demonstrates that the innovative dual-path architecture and attention fusion mechanism of the ADS-Forecaster model not only inherits the advantages of xLSTM in long-range time dependency modeling through a univariate temporal encoder, but also systematically captures the complex nonlinear interactions between different process variables using a transvariate dependency encoder. Furthermore, attention fusion decoding overcomes, to some extent, the problems of temporal information distortion and static context representation in traditional encoder-decoder architectures.

[0069] A comparison of the prediction characteristics of each model verifies the effectiveness of incorporating a multi-head attention mechanism in capturing NO. xThe ADS-Forecaster demonstrates superior performance in analyzing transient concentration characteristics while simultaneously highlighting the inherent limitations of traditional recurrent neural networks under complex operating conditions in thermal power units. The results prove that ADS-Forecaster can accurately analyze auxiliary variables and SCR input NO. x It generates a monotonic relationship between concentrations, exhibiting better generalization ability and prediction accuracy.

[0070] This invention has the following characteristics: This invention proposes a dual-path feature encoder that includes a parallel univariate temporal encoder and a cross-variable dependency encoder, and works in conjunction with an attention-based fusion decoder to effectively capture the dynamic characteristics within variables and the coupling relationships between variables, thus overcoming the limitations of traditional sequence modeling methods to a certain extent.

[0071] This invention addresses the inherent process time delays in thermodynamic systems by achieving time-series alignment of feature variables through maximum information coefficient analysis. Grey relational analysis is performed on the reconstructed dataset to filter physically meaningful features, eliminating redundant inputs while preserving key process correlations.

[0072] Comparative verification between this invention and benchmark models (RNN, LSTM, xLSTM) shows that ADS-Forecaster better resolves complex input-output mappings and achieves the best prediction accuracy, with R... 2 The results are as high as 0.946 and 0.938, and they demonstrate robust generalization ability under various operating conditions.

[0073] This invention relates to NO in thermal systems. x The generation of concentration prediction provides a new deep learning approach. Future work could explore other attention mechanism variants and memory enhancement architectures under this framework, or conduct quantization perception training for embedded controllers, and build a cross-unit transfer learning framework to further improve prediction robustness and accuracy.

[0074] Example 3 The SCR inlet NOx generation concentration prediction system based on LSTM and attention mechanism described in this invention includes: The acquisition module is used to acquire observation data of the desulfurization system, preprocess the observation data, and obtain preprocessed observation data. The prediction module is used to input the preprocessed observation data into the trained ADS-Forecaster prediction model to predict the NOx generation concentration at the SCR inlet.

[0075] In this embodiment, the ADS-Forecaster prediction model includes a dual-path feature encoder and a decoder based on a self-attention fusion layer. The dual-path feature encoder comprises a univariate temporal encoder and a cross-variable dependency encoder. The univariate temporal encoder uses an LSTM network to extract the temporal dynamics within a single process variable; the cross-variable dependency encoder utilizes an attention mechanism to explore the nonlinear interactions between different feature variables; and the decoder based on the self-attention fusion layer adaptively fuses the dual-path encoded features through temporal dynamic attention and cross-variable coupling attention layers, driving the LSTM module to generate NO. x Generate concentration prediction values.

[0076] In this embodiment, the loss function of the ADS-Forecaster prediction model during training is: ) in, Indicates sample time step The actual measured value; This indicates its corresponding predicted value; Indicates the number of samples; Indicates the range of predictions.

[0077] In this embodiment, the ADS-Forecaster prediction model is trained using the Adam stochastic gradient descent algorithm.

[0078] The module division in this embodiment is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in each embodiment of this application can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0079] Example 4 A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the LSTM-based NOx generation concentration prediction method for SCR inlet. For example, the method includes: acquiring observation data of a desulfurization system; preprocessing the observation data to obtain preprocessed observation data; and inputting the preprocessed observation data into a trained ADS-Forecaster prediction model to predict the NOx generation concentration at the SCR inlet. The memory may include main memory, such as high-speed random access memory, or it may also include non-volatile memory, such as at least one disk storage device. The processor, network interface, and memory are interconnected via an internal bus, which can be an industry-standard architecture bus, a peripheral component interconnection standard bus, an extended industry-standard architecture bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. The memory stores the program; specifically, the program may include program code, which includes computer operation instructions. The memory may include main memory and non-volatile memory, and provides instructions and data to the processor.

[0080] Example 5 A computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the LSTM-based NOx generation concentration prediction method for SCR inlet. For example, the method includes: acquiring observation data of a desulfurization system; preprocessing the observation data to obtain preprocessed observation data; and inputting the preprocessed observation data into a trained ADS-Forecaster prediction model to predict the NOx generation concentration at the SCR inlet. Specifically, the computer-readable storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. The volatile memory may include random access memory (RAM) and / or cache memory, etc. The non-volatile memory may include read-only memory (ROM), hard disk, flash memory, optical disk, magnetic disk, etc.

[0081] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0082] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0083] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0084] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0085] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and disclosure of the invention. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the following claims.

[0086] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

[0087] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for predicting NOx generation concentration at the SCR inlet based on LSTM and attention mechanism, characterized in that, include: Acquire observation data from the desulfurization system, and preprocess the observation data to obtain preprocessed observation data; The preprocessed observation data is input into the trained ADS-Forecaster prediction model to predict the NOx generation concentration at the SCR inlet.

2. The method for predicting NOx generation concentration at the SCR inlet based on LSTM and attention mechanism according to claim 1, characterized in that, The observation data of the desulfurization system include total coal feed, total air volume, unit load, inlet O2 concentration, inlet CO concentration, inlet flue gas temperature, inlet flue gas pressure, and main steam temperature.

3. The method for predicting NOx generation concentration at the SCR inlet based on LSTM and attention mechanism according to claim 1, characterized in that, The ADS-Forecaster prediction model includes a dual-path feature encoder and a decoder based on a self-attention fusion layer. The dual-path feature encoder comprises a univariate temporal encoder and a cross-variable dependency encoder. The univariate temporal encoder utilizes an LSTM network to extract the temporal dynamics within a single process variable; the cross-variable dependency encoder utilizes an attention mechanism to explore the nonlinear interactions between different feature variables; and the decoder based on the self-attention fusion layer adaptively fuses the dual-path encoded features through temporal dynamic attention and a cross-variable coupled attention layer, driving the LSTM module to generate NO. x Generate concentration prediction values.

4. The method for predicting NOx generation concentration at the SCR inlet based on LSTM and attention mechanism according to claim 1, characterized in that, The loss function of the ADS-Forecaster prediction model during training is: ） in, Indicates sample time step The actual measured value; This indicates its corresponding predicted value; Indicates the number of samples; Indicates the range of predictions.

5. The method for predicting NOx generation concentration at the SCR inlet based on LSTM and attention mechanism according to claim 1, characterized in that, The ADS-Forecaster prediction model is trained using the Adam stochastic gradient descent algorithm.

6. The method for predicting NOx generation concentration at the SCR inlet based on LSTM and attention mechanism according to claim 1, characterized in that, The preprocessing of the observation data includes: outlier handling, time delay analysis, and key feature variable screening.

7. A NOx generation concentration prediction system for SCR inlet based on LSTM and attention mechanism, characterized in that, include: The acquisition module is used to acquire observation data of the desulfurization system, preprocess the observation data, and obtain preprocessed observation data. The prediction module is used to input the preprocessed observation data into the trained ADS-Forecaster prediction model to predict the NOx generation concentration at the SCR inlet.

8. The method for predicting NOx generation concentration at the SCR inlet based on LSTM and attention mechanism according to claim 7, characterized in that, The ADS-Forecaster prediction model includes a dual-path feature encoder and a decoder based on a self-attention fusion layer. The dual-path feature encoder comprises a univariate temporal encoder and a cross-variable dependency encoder. The univariate temporal encoder utilizes an LSTM network to extract the temporal dynamics within a single process variable; the cross-variable dependency encoder utilizes an attention mechanism to explore the nonlinear interactions between different feature variables; and the decoder based on the self-attention fusion layer adaptively fuses the dual-path encoded features through temporal dynamic attention and a cross-variable coupled attention layer, driving the LSTM module to generate NO. x Generate concentration prediction values.

9. The method for predicting NOx generation concentration at the SCR inlet based on LSTM and attention mechanism according to claim 7, characterized in that, The loss function of the ADS-Forecaster prediction model during training is: ） in, Indicates sample time step The actual measured value; This indicates its corresponding predicted value; Indicates the number of samples; Indicates the range of predictions.

10. The method for predicting NOx generation concentration at the SCR inlet based on LSTM and attention mechanism according to claim 7, characterized in that, The ADS-Forecaster prediction model is trained using the Adam stochastic gradient descent algorithm.

Images