NOX concentration prediction method and device based on LSTM

Abstract

The invention provides an NOX concentration prediction method and device based on LSTM. According to the method provided by the embodiment of the invention, the prediction precision and the response speed of the NOx concentration at the denitration outlet can be effectively improved, the risk of excessive emission and the equipment operation and maintenance cost caused by excessive ammonia injection are reduced, and intelligent and stable control of the denitration system is realized.

Description

Technical Field

[0001] This invention relates to the field of NOx concentration prediction technology in denitrification systems of thermal power units, and particularly to a NOx concentration prediction method and device based on LSTM. Background Technology

[0002] With the continuous improvement of environmental protection standards in the power industry, the control of flue gas emissions from thermal power units has become a key link in achieving ultra-low emission retrofitting.

[0003] Accurate prediction and control of NOx concentration at the outlet of selective catalytic reduction (SCR) denitrification systems are particularly crucial and are widely used in pollutant treatment at coal-fired power plants. Related technologies utilize a collaborative approach combining PID control, manual experience-based adjustment, and a fundamental neural network model to construct a basic framework for denitrification control. Specifically, this technology system covers the entire process from flue gas parameter acquisition and control strategy formulation to ammonia injection rate adjustment, including key aspects such as inlet NOx concentration monitoring, catalyst operating status assessment, and control system response. However, existing control methods, which directly employ PID or manual control strategies, do not fully consider the dynamic coupling relationships of multivariate time-series characteristics in the denitrification system. This can lead to problems such as control lag, high prediction error rates, and excessive ammonia injection, or significant overshoot during load changes, thereby affecting denitrification efficiency, increasing equipment maintenance costs, and posing environmental compliance risks. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] Therefore, the first objective of this invention is to propose a NOx concentration prediction method based on LSTM.

[0006] The second objective of this invention is to provide a NOx concentration prediction device based on LSTM.

[0007] To achieve the above objectives, a first aspect of the present invention proposes a NOx concentration prediction method based on LSTM, comprising: S1, acquiring multi-dimensional time-series input parameters of a thermal power unit denitrification system, wherein the input parameters include inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, and catalyst pressure difference; S2, normalizing the multi-dimensional time-series input parameters, mapping the data to the [0,1] interval, and dividing the training samples and test samples using a sliding window method; S3, constructing a prediction model containing a three-layer LSTM network structure, and integrating an attention mechanism into the prediction model, using a QKV structure to dynamically assign weights to each input parameter to enhance the model's ability to identify key features; S4, training the prediction model using the Adam optimizer, automatically adjusting the learning rate to enable the model to converge quickly within a set learning rate and number of hidden nodes and improve prediction stability.

[0008] In one embodiment of the present invention, the acquisition of multi-dimensional time-series input parameters of the denitrification system of the thermal power unit further includes: S11, collecting historical operating data of inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content and catalyst pressure difference, wherein the time interval of the historical operating data is 1-5 seconds; S12, performing data cleaning on the collected multi-dimensional time-series input parameters, removing outliers and missing values, and completing the data sequence by interpolation method.

[0009] In one embodiment of the present invention, the normalization process for the input parameters further includes: S21, using the minimum-maximum normalization method to map the numerical range of each input parameter to the [0,1] interval; S22, performing standardization processing on the normalized data to eliminate the dimensional differences between different parameters.

[0010] In one embodiment of the present invention, the construction of the prediction model including a three-layer LSTM network structure further includes: S31, setting the number of hidden nodes in each LSTM layer to 50-1000, and dynamically adjusting the number of nodes according to the training error; S32, weighting the importance of each input parameter at different time steps through a dot product-Softmax-weighted attention calculation process.

[0011] In one embodiment of the present invention, the method further includes: S5, evaluating the model performance based on the deviation between the predicted result and the actual NOx concentration using root mean square error and mean absolute error, and iteratively optimizing the model parameters based on the evaluation results.

[0012] To achieve the above objectives, a second aspect of the present invention proposes an LSTM-based NOx concentration prediction device, comprising: an input parameter acquisition module for acquiring multi-dimensional time-series input parameters of a thermal power unit denitrification system, wherein the multi-dimensional time-series input parameters include inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, and catalyst pressure difference; a data preprocessing module for normalizing the multi-dimensional time-series input parameters, mapping the data to the [0,1] interval, and using a sliding window method to divide training samples and test samples; a model building module for constructing a prediction model containing a three-layer LSTM network structure, and integrating an attention mechanism into the model, using a QKV structure to dynamically assign weights to each input parameter to enhance the model's ability to identify key features; and a model training module for training the prediction model using the Adam optimizer, automatically adjusting the learning rate to enable the model to converge quickly within a set learning rate and hidden node number range and improve prediction stability.

[0013] In one embodiment of the present invention, the input parameter acquisition module is further configured to: acquire historical operating data of inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, and catalyst pressure difference, wherein the time interval of the historical operating data is 1-5 seconds; perform data cleaning on the acquired input parameters, remove outliers and missing values, and complete the data sequence by interpolation.

[0014] In one embodiment of the present invention, the data preprocessing module is further configured to: map the numerical range of each input parameter to the [0,1] interval using the minimum-maximum normalization method; and perform standardization processing on the normalized data to eliminate the dimensional differences between different parameters.

[0015] In one embodiment of the present invention, the model building module is further configured to: set the number of hidden nodes in each LSTM layer to 50-1000, and dynamically adjust the number of nodes according to the training error; and weight the importance of each input parameter at different time steps through a dot product-Softmax-weighted attention calculation process.

[0016] In one embodiment of the present invention, the system further includes: a model evaluation and optimization module, used to evaluate the model performance based on the deviation between the predicted results and the actual NOx concentration using root mean square error and mean absolute error, and to iteratively optimize the model parameters based on the evaluation results.

[0017] The method and apparatus of this invention can effectively improve the prediction accuracy and response speed of NOx concentration at the denitrification outlet, significantly reduce the risk of exceeding emission standards and excessive ammonia injection, and enhance the stability and economy of system control.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a NOx concentration prediction method based on LSTM according to an embodiment of the present invention; Figure 2 This is another flowchart of the LSTM-based NOx concentration prediction method according to an embodiment of the present invention; Figure 3 This is a diagram of the LSTM neuron structure according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the attention mechanism structure according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a NOx concentration prediction device based on LSTM according to an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] The following description, with reference to the accompanying drawings, illustrates an LSTM-based method and apparatus for predicting NOx concentration according to an embodiment of the present invention.

[0023] Example 1 Figure 1 This is a flowchart of an LSTM-based NOx concentration prediction method according to an embodiment of the present invention.

[0024] like Figure 1 As shown, a NOx concentration prediction method based on LSTM includes the following steps: S1, acquire the multi-dimensional time-series input parameters of the denitrification system of the thermal power unit. The multi-dimensional time-series input parameters include inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, and catalyst pressure difference.

[0025] Specifically, in this invention, the step of "obtaining multi-dimensional time-series input parameters of the denitrification system of thermal power units" is a crucial preliminary step for achieving accurate prediction of NOx concentration based on the LSTM (Long Short-Term Memory) network NOx concentration prediction model. This step aims to collect and integrate multiple input variables with time-series characteristics in real time during the operation of the denitrification system, including inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, and catalyst pressure difference, providing high-quality multi-dimensional time-series data input for subsequent model training and prediction.

[0026] Specifically, this step involves real-time acquisition of historical operational data for the aforementioned parameters through a sensor network deployed at key nodes of the denitrification system. Inlet NOx concentrations are typically provided by a CEMS (Continuous Emission Monitoring System), with sampling frequencies generally ranging from 1 Hz to 10 Hz, and data formatted as follows: The unit load is obtained from the DCS (Distributed Control System), in MW, with a sampling period typically of 1 minute; the flue gas flow rate is measured by a differential pressure flow meter or an ultrasonic flow meter, in units of... Flue gas temperature was measured by thermocouples, in °C; inlet flue gas oxygen content was obtained using a zirconia oxygen analyzer, in vol%; catalyst differential pressure was measured by a differential pressure transmitter, in Pa. The collected data needed to be aligned by timestamps to form a structured time-series dataset.

[0027] Furthermore, to ensure the consistency and comparability of the model input data, the Min-Max Normalization method is used to map each parameter to the [0,1] interval. In addition, to construct the training samples required for the LSTM model, the data is sliced ​​using the sliding window method, with the window length set to one time step, meaning that each sample contains all the input parameters at the current time step, used to predict the NOx concentration at the denitrification outlet at the next time step.

[0028] Specifically, this step is widely used in the denitrification control system of thermal power units, especially in the optimization of SCR (Selective Catalytic Reduction) system operation and ultra-low emission retrofitting. By collecting and processing the above parameters in real time, the system can provide a dynamic and continuous input data stream for the LSTM model, thereby enabling short-term prediction of NOx concentration at the denitrification outlet.

[0029] Furthermore, the precise acquisition and standardized processing of multi-dimensional time-series parameters lay a solid data foundation for the subsequent training and prediction of the LSTM model. Combined with an attention mechanism, the model can dynamically identify the influence weights of each input parameter under different operating conditions, significantly improving prediction accuracy and response speed, and providing reliable support for achieving intelligent control and environmental compliance of the denitrification system.

[0030] Furthermore, S1 includes: S11, collect historical operating data of inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content and catalyst pressure difference, with the time interval of the historical operating data being 1-5 seconds.

[0031] Specifically, in some implementations, collecting historical operating data such as inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, and catalyst pressure differential is a key step in the data preprocessing and modeling input of LSTM-based NOx concentration prediction methods and systems. This step acquires multi-dimensional time-series data of the denitrification system operation process through real-time or historical acquisition, providing high-quality input features for subsequent model training and prediction.

[0032] Specifically, this step typically relies on industrial control system platforms such as DCS (Distributed Control System) or CEMS (Continuous Flue Gas Monitoring System) to periodically collect relevant parameters via PLC or sensor interfaces. The collection frequency is set to 1-5 seconds to ensure that the model can capture the dynamic characteristics of flue gas parameters changing over time. High-frequency data is particularly helpful in improving the model's response to transient processes, especially during unit load fluctuations or denitrification system adjustments. The collected data includes: inlet NOx concentration (unit: Unit load (unit: MW), flue gas flow rate (unit: Flue gas temperature (unit: °C), inlet flue gas oxygen content (unit: ... The parameters, including catalyst pressure difference (unit: Pa), together constitute a multidimensional feature vector of the operating status of the denitrification system.

[0033] Specifically, inlet NOx concentration is typically measured using chemiluminescence or electrochemical sensors, with measurement errors controlled within ±5% to comply with GB 13223-2011 "Emission Standard of Air Pollutants for Thermal Power Plants". Unit load data is obtained from the turbine control system, with sampling accuracy within ±1% of the error range. Flue gas flow rate and temperature are measured using differential pressure flow meters and thermocouples, oxygen content is obtained using a zirconia sensor, and catalyst differential pressure is monitored in real-time by a differential pressure transmitter. All data undergo validity verification, and outliers and missing values ​​are removed to ensure data quality.

[0034] Specifically, this step is widely used in the SCR (Selective Catalytic Reduction) denitrification system of thermal power units. Especially during the operation of the unit under varying loads, catalyst aging, or ammonia injection system adjustment, the high-frequency collected operating data can provide timely and accurate input information for the prediction model, thereby realizing the dynamic prediction and control optimization of NOx concentration at the denitrification outlet.

[0035] Furthermore, by collecting data in a high-frequency, multi-dimensional manner, rich temporal features were provided for the LSTM model, enhancing the model's ability to perceive the operating status of the denitrification system. This laid a solid data foundation for subsequent attention mechanisms and model training, thereby significantly improving the accuracy and real-time performance of NOx concentration prediction.

[0036] S12 performs data cleaning on the collected multidimensional time-series input parameters, removes outliers and missing values, and completes the data sequence through interpolation.

[0037] Specifically, in some implementations, data cleaning of the collected input parameters to remove outliers and missing values, and completion of the data sequence through interpolation methods, are crucial preprocessing steps in the LSTM-based NOx concentration prediction method of this invention. This step aims to improve the quality of the input data, thereby providing a reliable and continuous time-series data foundation for subsequent model training and prediction.

[0038] Specifically, data cleaning begins with identifying outliers through statistical analysis and threshold judgment. Specifically, the Z-score method or the IQR (interquartile range) method is used to detect anomalies in various parameters (such as inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, catalyst pressure difference, etc.). In the Z-score method, a threshold of ±3σ is set; data points exceeding this range are considered outliers and removed. The IQR method calculates the first and third quartiles, considering data exceeding 1.5 times the IQR range as outliers. For missing values, the system uses time-series-based interpolation methods for completion, such as linear interpolation, spline interpolation, or more advanced LSTM-based sequence interpolation algorithms, to maintain data continuity and temporal characteristics.

[0039] Specifically, the normalization process employs Min-Max Normalization, mapping the original data to the [0,1] interval. This means that each time step uses only the data from the current moment as input, facilitating the model's capture of instantaneous changes. High interpolation accuracy is required, typically employing cubic spline interpolation with an interpolation error controlled within ±0.5%, to meet the stringent data stability requirements of thermal power unit denitrification systems.

[0040] Specifically, this step is applied to the real-time monitoring and prediction platform of the denitrification system in thermal power plants, processing multi-source heterogeneous data from DCS (Distributed Control System) or CEMS (Continuous Emission Monitoring System). Due to sensor failures or communication delays, the raw data often contains missing or abrupt values. This step can effectively improve data availability and ensure the integrity and consistency of model input.

[0041] Specifically, this not only improves data quality but also enhances the LSTM model's ability to identify key input features, thereby improving the accuracy and stability of NOx concentration prediction. By removing noise and filling in missing information, the model can more effectively learn the temporal dependencies in the denitrification process, providing solid data support for achieving ultra-low emission control.

[0042] S2, normalize the multidimensional time-series input parameters, map the data to the [0,1] interval, and use a sliding window method to divide the training samples and test samples.

[0043] Specifically, in some implementations, normalizing the input parameters and using a sliding window to divide the training and testing samples are key data preprocessing steps in the LSTM-based NOx concentration prediction method of this invention. This step aims to improve the stability and prediction accuracy of the model training, while ensuring that the model has good generalization ability.

[0044] Specifically, the normalization process employs the Min-Max Normalization method, linearly mapping the original input data (such as inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, catalyst pressure difference, etc.) to the [0,1] interval. This method eliminates dimensional differences between different parameters, preventing problems such as gradient explosion or convergence difficulties caused by inconsistent numerical ranges during model training.

[0045] Furthermore, to construct a time series prediction model, a sliding window method is used to partition the normalized data. The sliding window length is set to 1, meaning each sample contains only the input features of the current time step, used to predict the NOx outlet concentration at the next time step. The window step size is typically set to 1 to ensure the continuity of the time series and the model's sensitivity to short-term dynamic changes. In practice, the dataset is divided into multiple input-output pairs, with the input being the multidimensional feature vector of the current time step and the output being the NOx concentration value at the next time step. This method meets the processing requirements of LSTM models for time series data and helps the model capture the nonlinear relationship between input parameters and time.

[0046] Specifically, normalization must ensure that the numerical range of all input features is consistent. Normalized data should typically be evenly distributed to avoid extreme values ​​or data shifts. When dividing the sliding window, the window step size should be set according to the temporal resolution of the historical data (e.g., per second, per minute), and the ratio of training to test sets should be appropriately divided, generally using a 7:3 or 8:2 ratio to ensure sufficient model training and good validation results.

[0047] Specifically, this step is widely used in real-time data processing of denitrification systems in thermal power units. Input parameters typically come from DCS systems or online monitoring equipment, exhibiting high sampling frequency and strong temporal sequence characteristics. Through normalization and sliding window processing, the raw data can be transformed into standardized time-series samples suitable for LSTM model training, providing a high-quality data foundation for subsequent model training and prediction.

[0048] Specifically, normalization improves the numerical stability of model training, while sliding window partitioning enhances the model's ability to model time-series features. Combined with LSTM and attention mechanisms, this preprocessing method provides reliable data support for achieving high-precision NOx concentration prediction, thereby effectively improving the control accuracy and response speed of the denitrification system.

[0049] Furthermore, S2 includes: S21 uses the minimum-maximum normalization method to map the numerical range of each input parameter to the [0,1] interval.

[0050] Specifically, in this invention, the use of the Min-Max Normalization method to preprocess the input parameters is one of the key steps in constructing an LSTM-based NOx concentration prediction model. This method linearly maps the original numerical range of each input variable to the [0,1] interval, thereby eliminating the problems of dimensional differences and numerical scale inconsistencies between different parameters, and improving the stability and convergence speed of model training.

[0051] Furthermore, in this system, the normalization operation is performed on several key input parameters of the denitrification system, including the inlet NOx concentration (unit: Unit load (unit: MW), flue gas flow rate (unit: Flue gas temperature (unit: °C), inlet flue gas oxygen content (unit: ... The parameters include catalyst pressure differential (unit: kPa). Each parameter needs to have its minimum and maximum values ​​calculated independently and then normalized. During normalization, it is essential to ensure that the same normalization parameters are used for both the training and test sets to avoid data leakage and decreased model generalization ability.

[0052] Furthermore, to enhance the model's adaptability to dynamic changes, normalization can be combined with a sliding window mechanism, with a window length of 1, meaning that only the input data at the current time step is processed at a time. This setting is suitable for real-time prediction scenarios, ensuring that the model can quickly respond to changes in input parameters.

[0053] Specifically, min-max normalization effectively alleviates the gradient update imbalance problem caused by differences in the numerical ranges of different input parameters, enabling the LSTM model to learn temporal features more efficiently during training. Simultaneously, the normalized data better matches the input range of activation functions (such as Sigmoid or Tanh), helping to improve the model's prediction accuracy and stability. This provides a solid data foundation for subsequent feature weighting in the attention mechanism, thereby enhancing the overall performance and practicality of the prediction system.

[0054] S22, standardize the normalized data to eliminate dimensional differences between different parameters.

[0055] Specifically, in some implementations, standardizing the normalized data is one of the key data preprocessing steps in the LSTM-based NOx concentration prediction method of this invention. Its technical purpose is to further eliminate dimensional differences and distribution biases between different input parameters, thereby improving the model's training efficiency and prediction accuracy. This step is usually performed after Min-Max Normalization, using the Z-score standardization method to convert the data into a standard normal distribution with a mean of 0 and a standard deviation of 1, thus enhancing the model's adaptability to multi-source heterogeneous data.

[0056] Specifically, the sample mean and sample standard deviation are calculated for each normalized input parameter (such as inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, catalyst pressure difference, etc.), and then each data point is standardized. In practice, a rolling window can be optionally used to calculate the moving mean and moving standard deviation to adapt to the dynamic changes in the operating parameters of thermal power units. The window length is usually set to 24 time steps, corresponding to a typical cycle of unit operation.

[0057] Furthermore, the parameters used in the standardization process must meet the accuracy requirements of industrial data processing. Typically, the standardized data should be within ±3σ to avoid outliers interfering with model training. Simultaneously, to ensure the model's generalization ability, the standardization process should be based on the statistical characteristics of the training set. The standardization of the test and prediction sets should use the parameters from the training set, rather than the statistical values ​​from their respective datasets.

[0058] Specifically, in practical applications, it is widely used in the real-time data processing of denitrification systems in thermal power units, especially in multivariate time-series prediction models. Standardization helps accelerate the gradient descent process and improve the model convergence speed. Combined with the subsequent training process of the LSTM model, the standardized data can be more effectively learned by the neural network, thereby improving the accuracy and stability of NOx concentration prediction.

[0059] Specifically, standardization, as an important part of data preprocessing, not only improves the consistency and comparability of model input data, but also provides a more stable numerical basis for the weight allocation of the LSTM model in capturing temporal features and attention mechanisms, thereby enhancing the robustness and predictive performance of the entire prediction system.

[0060] S3. Construct a prediction model containing a three-layer LSTM network structure, and integrate an attention mechanism into the prediction model. Use a QKV structure to dynamically assign weights to each input parameter to enhance the model's ability to identify key features.

[0061] Specifically, this step involves constructing a prediction model with a three-layer LSTM (Long Short-Term Memory) network structure and integrating an attention mechanism into the model. A QKV (Query-Key-Value) structure is used to dynamically assign weights to the input parameters to enhance the model's ability to identify key features. In some implementations, this model is used to predict the NOx concentration at the outlet of the SCR denitrification system in thermal power units. Its core objective is to improve the model's expressive power and feature focusing ability in time-series data modeling.

[0062] Specifically, the three-layer LSTM structure consists of multiple stacked LSTM units. Each unit contains an input gate, a forget gate, and an output gate. The gating mechanism controls the flow and storage of information, effectively solving the gradient vanishing problem in traditional RNNs for long sequence modeling. Input parameters include inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, and catalyst pressure difference. These parameters need to be preprocessed before entering the LSTM layer. Min-Max Normalization is typically used to map the data to the [0,1] interval to improve the stability and convergence speed of model training. In the model, the attention mechanism is implemented through a QKV structure, where the Query, Key, and Value vectors are obtained from different linear transformations of the input sequence. Attention scores are calculated through dot products, then normalized using the Softmax function. Finally, the Value vectors are weighted and summed to output the weighted feature representation.

[0063] Specifically, the number of hidden nodes in the model is set to 50-1000 to accommodate different scales of input data and model complexity requirements; the learning rate is set to 0.01, and the Adam optimizer is used for parameter updates, balancing training speed and stability. The introduction of the attention mechanism allows the model to dynamically adjust the weights of each input parameter. For example, during high-load operation, unit load and inlet NOx concentration may be given higher weights, thereby improving prediction accuracy.

[0064] Furthermore, the model is deployed in the denitrification control system of thermal power units to receive real-time time-series data from sensors, predict the outlet NOx concentration, and assist operators in optimizing the control of ammonia injection to achieve ultra-low emission targets.

[0065] Specifically, by capturing the temporal dependencies of input parameters using LSTM and combining an attention mechanism to dynamically focus on key features, the prediction accuracy and generalization ability of the model are significantly improved, the risk of NOx over-emission and ammonia escape rate are reduced, and it has significant engineering practical value.

[0066] Furthermore, S3 includes: S31, set the number of hidden nodes in each LSTM layer to 50-1000, and dynamically adjust the number of nodes according to the training error.

[0067] Specifically, in some implementations, setting the number of hidden nodes in each LSTM layer to 50-1000 and dynamically adjusting the number of nodes based on training error is one of the key steps in constructing an efficient NOx concentration prediction model in this invention. LSTM (Long Short-Term Memory) is an improved structure of Recurrent Neural Networks (RNNs). Its core lies in effectively handling long-term dependencies in time-series data through gating mechanisms of input gates, forget gates, and output gates. In this invention, the number of hidden nodes in each LSTM layer is set to 50 to 1000. This range is based on a comprehensive consideration of the complexity and temporal characteristics of the input parameters (such as inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, catalyst pressure difference, etc.) of the denitrification system of thermal power units. Too few nodes may lead to insufficient model expressive power and an inability to fully fit the nonlinear relationship between input variables; while too many nodes may cause overfitting and increase computational resource consumption.

[0068] Furthermore, this invention employs a dynamic adjustment mechanism to automatically optimize the number of hidden nodes in each LSTM layer based on the model's error performance during training (such as root mean square error (RMSE) and mean absolute error (MAE)). This mechanism typically combines early stopping with grid search or Bayesian optimization strategies to evaluate model performance under different node configurations on the validation set. In the initial training phase, the initial number of nodes can be set to 256 or 512, and then the number of nodes can be gradually increased or decreased based on the changing trend of the validation error to achieve a balance between model accuracy and computational efficiency.

[0069] Specifically, this step is applicable to real-time prediction scenarios for denitrification systems in thermal power units, especially under conditions of large fluctuations in flue gas emission concentrations and complex and variable operating conditions. By dynamically adjusting the number of hidden nodes, the model's adaptability to different operating conditions can be enhanced. This technique significantly improves the model's generalization ability and prediction accuracy, providing reliable data support for the stable control of NOx concentration at the denitrification outlet and the optimization of ammonia injection.

[0070] S32 uses a dot product-Softmax-weighting process to weight the importance of each input parameter at different time steps.

[0071] Specifically, in this invention, through dot product - The weighted attention calculation process, which weights the importance of each input parameter at different time steps, is one of the key technical steps for achieving high-precision NOx concentration prediction. This step is based on the core idea of ​​the attention mechanism, which dynamically assigns different attention weights to the input features at different time steps during the LSTM model's processing of time-series data, thereby enhancing the model's ability to capture key information.

[0072] Specifically, this attention mechanism employs a Query-Key-Value (QKV) structure, where each time-step feature vector in the input sequence is first mapped to three vectors—Query, Key, and Value—using a learnable weight matrix. Subsequently, the similarity between Query and Key is calculated through a dot product operation to obtain an attention score matrix. Each element of this matrix represents the degree of attention the current time step pays to other time steps. To further enhance the probability distribution properties of these scores, an attention mechanism is used... The function normalizes the score matrix to obtain the attention weights at each time step.

[0073] Optionally, the attention weights are summed with the corresponding value vectors to generate a weighted context vector, which serves as the augmentation input for the LSTM model at the current time step.

[0074] Specifically, the input parameters include inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, and catalyst pressure differential. The degree of influence of these parameters on the NOx outlet concentration varies at different time steps. Through an attention mechanism, the model can automatically identify and strengthen features that are highly relevant to the current prediction task. For example, under conditions of sudden load changes or large fluctuations in flue gas temperature, the attention weights will be adjusted accordingly to highlight the role of key variables.

[0075] Specifically, the attention mechanism plays a role in feature selection and information focusing in the entire prediction model, effectively mitigating the gradient vanishing problem that may occur when LSTM processes long sequence data, and improving the model's ability to model time-series features. Combined with the three-layer stacked structure of LSTM and the Adam optimizer (learning rate 0.01), the attention mechanism further enhances the model's generalization ability and prediction accuracy, providing solid data support for the intelligent control of denitrification systems in thermal power units.

[0076] S4. The prediction model is trained using the Adam optimizer, and the learning rate is automatically adjusted to enable the model to converge quickly within the set learning rate and number of hidden nodes, thereby improving prediction stability.

[0077] Specifically, in some implementations, this invention employs the Adam optimizer to train the LSTM-based NOx concentration prediction model, enabling adaptive adjustment of the learning rate, thereby accelerating model convergence and improving prediction stability. Adam (Adaptive Moment Estimation) is a method that combines momentum estimation and... The first-order optimization algorithm with its advantages is based on the principle of dynamically adjusting the learning rate of each parameter by calculating the first-order moment estimate (mean) and second-order moment estimate (uncentered variance) of the gradient, thereby achieving more efficient parameter updates in different dimensions.

[0078] Specifically, the Adam optimizer is applied to a three-layer stacked LSTM network structure with the number of hidden nodes adjustable between 50 and 1000 to accommodate different scales of input data and model complexity. Through an adaptive learning rate mechanism, the model can converge quickly in the early stages of training and then stably adjust parameters in the later stages, avoiding oscillations or overfitting problems caused by excessively high learning rates. This step plays a crucial role in predicting NOx concentration at the denitrification outlet of thermal power units, significantly improving the model's fitting ability and generalization performance on time-series data, and providing reliable technical support for achieving ultra-low emission control.

[0079] S5. Based on the deviation between the predicted results and the actual NOx concentration, the root mean square error and the mean absolute error are used to evaluate the model performance, and the model parameters are iteratively optimized according to the evaluation results.

[0080] Specifically, in some implementations, this step evaluates model performance using root mean square error (RMSE) and mean absolute error (MAE) based on the deviation between the predicted results and the actual NOx concentration. It then iteratively optimizes the model parameters based on the evaluation results, making it a crucial step in achieving adaptive model improvement within the entire LSTM-based NOx concentration prediction system. This step quantifies the difference between the model output and the true value, providing data support for subsequent model parameter adjustments, thereby improving prediction accuracy and system stability.

[0081] Specifically, the model output is the predicted NOx concentration at the denitrification outlet, and the input is time-series data processed by min-max normalization, including parameters such as inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, and catalyst pressure difference. Training samples are constructed using a sliding window method, with a window length of 1, meaning each sample only contains the input features at the current moment, to meet real-time prediction requirements.

[0082] Specifically, RMSE is more sensitive to larger errors and is suitable for scenarios emphasizing prediction accuracy; while MAE is more robust to outliers and is suitable for scenarios emphasizing error distribution stability. In this system, RMSE and MAE are used together to comprehensively evaluate the model's performance at different error scales. During model optimization, the Adam optimizer is used with an initial learning rate of 0.01 to balance convergence speed and stability. The number of hidden layer nodes is set to 50-1000, and three layers of LSTM are stacked to enhance the model's ability to model complex temporal features.

[0083] Specifically, this step is widely used in the real-time control of denitrification systems in thermal power units, especially in the context of ultra-low emission retrofitting, where the accuracy of NOx concentration prediction is extremely important. By continuously iterating and optimizing model parameters, the system can adapt to changes in operating conditions, such as load fluctuations and catalyst aging, thereby improving the generalization ability and robustness of the prediction model.

[0084] Specifically, this step effectively improves the model's predictive performance, reduces the risk of excessive NOx emissions at the denitrification outlet, and also reduces air preheater blockage and operating costs caused by excessive ammonia injection, providing a reliable guarantee for the intelligent, efficient, and environmentally friendly operation of thermal power units.

[0085] This invention discloses a method for predicting NOx concentration at the denitrification outlet based on LSTM and attention mechanism. By introducing root mean square error (RMSE) and mean absolute error (MAE) to quantitatively evaluate the prediction results, and iteratively optimizing the model parameters based on the evaluation results, the stability and adaptability of the NOx concentration prediction model are further improved, thereby achieving more accurate dynamic control and reducing control errors and resource waste caused by model bias.

[0086] Example 2 To achieve the above embodiments, such as Figure 2 As shown, the embodiments of the invention also provide another workflow for the LSTM-based NOx concentration prediction method, including: S101, collect input parameters.

[0087] Specifically, according to the formation mechanism of NOx concentration at the denitrification outlet of thermal power units, fuel-type and thermal-type NOx formation accounts for about 95% of the total NOx concentration. The main factors affecting NOx concentration include inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, catalyst pressure difference, etc. Therefore, the above parameters are selected as the main input parameters of this prediction model.

[0088] S102, train the LSTM model.

[0089] Specifically, LSTM is a special type of recurrent neural network (such as...) Figure 3 As shown, this model addresses the long-term dependency problem of traditional RNNs through a gating mechanism. Its core structure includes memory cells and a gating mechanism, which consists of an input gate, a forget gate, and an output gate. When training this model, the learning rate is set to 0.01; a rate that is too high will lead to system instability, while a rate that is too low will result in slow convergence. The number of hidden nodes is set to 50-1000, and three layers of LSTM are stacked. The Adam optimizer is used to automatically adjust the learning rate.

[0090] S103, the attention mechanism assigns different weights to input factors.

[0091] Specifically, attention mechanisms (such as Figure 4 As shown, different weights are assigned to each input factor. The long dependency problem of LSTM is solved by using the core structure (QKV) and the calculation process (dot product-Softmax-weighted). This enables the model to effectively capture key information and focus on the important parts of the sequence information of input parameters such as unit load, inlet NOx concentration, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, and catalyst pressure difference, while ignoring irrelevant parts. This effectively improves the accuracy and efficiency of the prediction model.

[0092] S104 processes the input data.

[0093] Specifically, a normalization method is used to classify the original NOx concentration and other input parameters into a specified range. The commonly used normalization method is min-max normalization, which classifies the data into a fixed interval of [0, 1]. A sliding window method is used to divide the dataset into multiple training and test samples, with the sliding window length set to 1. In order to measure the accuracy and effectiveness of the prediction results, the root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation indicators.

[0094] This invention discloses a NOx concentration prediction model for denitrification outlets based on LSTM and attention mechanisms. The model uses LSTM and attention mechanisms to predict NOx concentrations at the denitrification outlets of thermal power units. By capturing temporal patterns with LSTM and dynamically focusing on key features with attention mechanisms, the model significantly improves prediction accuracy, response speed, and generalization ability. Simultaneously, it reduces the operation and maintenance costs of air preheaters, effectively achieving stable NOx concentration control and ammonia slip rate optimization, providing an efficient and intelligent solution for ultra-low emission retrofitting.

[0095] Example 3 To achieve the above embodiments, such as Figure 5 As shown, this embodiment also provides an LSTM-based NOx concentration prediction device 10, including: The input parameter acquisition module 100 is used to acquire multi-dimensional time-series input parameters of the denitrification system of the thermal power unit. The input parameters include inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, and catalyst pressure difference.

[0096] Specifically, historical operating data such as inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, and catalyst pressure difference are collected, with the time interval of the historical operating data being 1-5 seconds; the collected input parameters are cleaned to remove outliers and missing values, and the data sequence is completed by interpolation.

[0097] The data preprocessing module 200 is used to normalize the input parameters, map the data to the [0,1] interval, and divide the training samples and test samples using a sliding window method.

[0098] Specifically, the minimum-maximum normalization method is used to map the numerical range of each input parameter to the [0,1] interval; the normalized data is then standardized to eliminate the dimensional differences between different parameters.

[0099] The model building module 300 is used to build a prediction model containing a three-layer LSTM network structure and integrate an attention mechanism into the model. The QKV structure is used to dynamically assign weights to each input parameter to enhance the model's ability to identify key features.

[0100] Specifically, the number of hidden nodes in each LSTM layer is set to 50-1000, and the number of nodes is dynamically adjusted according to the training error; the importance of each input parameter at different time steps is weighted through the dot product-Softmax-weighted attention calculation process.

[0101] The model training module 400 is used to train the prediction model using the Adam optimizer, automatically adjust the learning rate, and enable the model to converge quickly within the set learning rate and number of hidden nodes, thereby improving prediction stability.

[0102] This invention discloses an LSTM-based NOx concentration prediction device that achieves high-precision prediction and dynamic optimization of NOx concentration at the outlet of a thermal power unit's denitrification system through modular collaboration. The system first uses an input parameter acquisition module to acquire multi-dimensional time-series input parameters in real time, including inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, and catalyst differential pressure. Outliers and missing values ​​are removed through data cleaning, and the data sequence is completed using interpolation methods to ensure data quality and continuity. The data preprocessing module uses min-max normalization to map parameters to the [0,1] interval and performs standardization to eliminate dimensional differences. Then, a sliding window method is used to divide training and testing samples, providing structured input for model construction. The model construction module integrates a three-layer LSTM network structure. The number of hidden nodes is dynamically adjusted within the range of 50-1000 based on the training error. An attention mechanism is incorporated, and a QKV structure dot product-Softmax-weighted process is used to dynamically assign weights to the importance of each input parameter at different time steps, significantly enhancing the model's ability to identify and capture key features. The model training module uses the Adam optimizer to automatically adjust the learning rate, enabling the model to converge quickly within the set parameter range and improving prediction stability and generalization performance. The overall solution organically integrates data preprocessing, deep learning, and attention mechanisms, effectively reducing the risk of overfitting while improving prediction accuracy and enhancing the model's adaptability to complex operating conditions. This provides reliable technical support for the intelligent control and environmental emission management of denitrification systems in thermal power units.

[0103] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0104] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A NOx concentration prediction method based on LSTM, characterized in that, include: S1, acquire the multi-dimensional time-series input parameters of the denitrification system of the thermal power unit, including the inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content and catalyst pressure difference; S2, normalize the multidimensional time-series input parameters, map the data to the [0,1] interval, and use a sliding window method to divide the training samples and test samples; S3. Construct a prediction model containing a three-layer LSTM network structure, and integrate an attention mechanism into the model. Use a QKV structure to dynamically assign weights to each input parameter to enhance the model's ability to identify key features. S4. The prediction model is trained using the Adam optimizer, and the learning rate is automatically adjusted to enable the model to converge quickly within the set learning rate and number of hidden nodes, thereby improving prediction stability.

2. The method as described in claim 1, characterized in that, S1 further includes: S11, collect historical operating data of inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content and catalyst pressure difference, with the time interval of the historical operating data being 1-5 seconds; S12 performs data cleaning on the collected input parameters, removing outliers and missing values, and completes the data sequence using interpolation methods.

3. The method as described in claim 1, characterized in that, S2 further includes: S21, using the minimum-maximum normalization method, maps the numerical range of each input parameter to the [0,1] interval; S22, standardize the normalized data to eliminate dimensional differences between different parameters.

4. The method as described in claim 1, characterized in that, S3 further includes: S31, set the number of hidden nodes in each LSTM layer to 50-1000, and dynamically adjust the number of nodes according to the training error; S32 uses a dot product-Softmax-weighted attention calculation process to weight the importance of each input parameter at different time steps.

5. The method as described in claim 1, characterized in that, Also includes: S5. Based on the deviation between the predicted results and the actual NOx concentration, the root mean square error and the mean absolute error are used to evaluate the model performance, and the model parameters are iteratively optimized according to the evaluation results.

6. A NOx concentration prediction device based on LSTM, characterized in that, include: The input parameter acquisition module is used to acquire multi-dimensional time-series input parameters of the denitrification system of thermal power units. The input parameters include inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, and catalyst pressure difference. The data preprocessing module is used to normalize the input parameters, map the data to the [0,1] interval, and divide the training samples and test samples using a sliding window method; The model building module is used to build a prediction model containing a three-layer LSTM network structure and integrate an attention mechanism into the model. The QKV structure is used to dynamically assign weights to each input parameter to enhance the model's ability to identify key features. The model training module is used to train the prediction model using the Adam optimizer, automatically adjusting the learning rate to enable the model to converge quickly within a set learning rate and number of hidden nodes, thereby improving prediction stability.

7. The apparatus as claimed in claim 6, characterized in that, The input parameter acquisition module is also used for: Historical operating data of inlet NOx concentration, unit load, flue gas flow rate, flue gas temperature, inlet flue gas oxygen content, and catalyst pressure difference are collected, with the time interval of the historical operating data being 1-5 seconds; The collected input parameters are cleaned to remove outliers and missing values, and the data sequence is completed using interpolation methods.

8. The apparatus as claimed in claim 6, characterized in that, The data preprocessing module is also used for: The minimum-maximum normalization method is used to map the numerical range of each input parameter to the interval [0,1]. The normalized data is standardized to eliminate the dimensional differences between different parameters.

9. The apparatus as claimed in claim 6, characterized in that, The model building module is also used for: Set the number of hidden nodes in each LSTM layer to 50-1000, and dynamically adjust the number of nodes according to the training error; The importance of each input parameter at different time steps is weighted through a dot product-Softmax-weighted attention calculation process.

10. The apparatus as claimed in claim 6, characterized in that, Also includes: The model evaluation and optimization module is used to evaluate the model performance based on the deviation between the predicted results and the actual NOx concentration, using root mean square error and mean absolute error, and to iteratively optimize the model parameters based on the evaluation results.

Images