Denitration predictive control method based on LSTM and PINN

Abstract

According to the denitration predictive control method based on the LSTM and the PINN, production data are collected in real time, a complex time sequence dynamic state is captured through an LSTM time sequence prediction model, the NOx concentration and the ammonia escape amount in 2-3 minutes in the future are predicted, and the lag problem in a traditional control method is solved. Meanwhile, in combination with the physical constraint of the PINN, it is ensured that the prediction result conforms to the physical law, and the prediction precision and stability are improved. In addition, an ammonia spraying amount adjusting strategy is optimized by adopting reinforcement learning, the frequency of an ammonia water pump is dynamically adjusted, the maximum use efficiency of ammonia water is ensured, and waste is reduced. Data acquisition adopts MQTT and OPC protocols, data real-time performance and accuracy are ensured, and a Kalman filtering algorithm is used for removing abnormal values. The method can effectively improve the precision, the stability and the response speed of a denitration system, optimizes denitration control, meets the ultra-low emission requirement of the iron and steel industry, and has remarkable environmental protection and economic benefits.

Description

Technical Field

[0001] The present invention relates to the field of environmental pollution, and in particular to a denitrification predictive control method based on LSTM and PINN. Background Art

[0002] The steel industry is a typical high-energy-consuming and high-pollution industry. During the sintering process, sintering flue gas generates large amounts of nitrogen oxides (NOx). With the Ministry of Ecology and Environment's latest "Technical Specifications for Ultra-Low Emission Control of Sintering Waste Gas in the Steel Industry," which mandates full implementation of ultra-low emissions of pollutants from sintering waste gas from the steel industry starting April 1, 2025, including a requirement for NOx concentrations in sintering waste gas to be below 50 mg / Nm³, current denitrification control technologies face the following challenges: Adjusting ammonia injection based on the changing trends of inlet and outlet NOx levels suffers from hysteresis, which can easily lead to large fluctuations in outlet NOx levels and excessive NOx levels. Furthermore, factors such as catalyst activity and reactor temperature (200-300°C) significantly influence denitrification efficiency, resulting in low reactor inlet temperature and ammonia waste. Furthermore, the ammonia control valve opening is not linearly linked to the ammonia flow rate, leading to erratic fluctuations in ammonia flow, resulting in excessive ammonia injection and a high risk of ammonia slip at the reactor outlet. Summary of the Invention

[0003] The purpose of the present invention is to solve the shortcomings of the prior art and propose a denitrification predictive control method based on LSTM and PINN.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a denitrification predictive control method based on LSTM and PINN, the specific steps are as follows: S1: Data acquisition and preprocessing; Edge devices collect production data in real time through protocols such as MQTT and OPC, store it in a time series database, and use algorithms such as Kalman filtering to remove missing and outliers in the data; S2: Build the model; The inlet NOx concentration, outlet NOx concentration, inlet flue gas flow rate, oxygen content, ammonia injection amount, catalyst activity, reactor temperature, dilution air flow fluctuation, and hot blast furnace heating efficiency collected in real time in step S1 are input as characteristic factors into the MPC prediction model to obtain the outlet NOx concentration and ammonia slip amount in the next 2-3 minutes; S3: training model; Utilize LSTM to capture complex time series dynamics in real time, addressing the lag issues of manual control and PID control. Furthermore, PINN (physical hard constraints) prevents unreasonable predictions, making the prediction data more reliable. S4: Confirm the results; Through steps S1-S3, the final predicted value of NOx is confirmed to be: .

[0005] Preferably, in step S1, the edge device collects the ammonia injection amount, catalyst activity and reactor temperature in the SCR denitrification system through the OPC protocol, and collects the inlet NOx concentration, outlet NOx concentration and oxygen content through the MQTT protocol; the observation noise covariance matrix of the Kalman filter algorithm adopts an adaptive adjustment strategy to dynamically update the covariance parameters according to the volatility of the time series data.

[0006] Preferably, in step S2, the number of nodes in the input layer of the MPC prediction model is consistent with the number of characteristic factors, the hidden layer adopts a 3-layer LSTM network structure, and the number of neurons in each layer is 64, 32, and 16, respectively; the output layer of the MPC prediction model is set with two nodes, corresponding to the predicted values of the outlet NOx concentration and ammonia escape in the next 2 minutes and 3 minutes, respectively.

[0007] Preferably, in step S2, the MPC prediction model solves the optimal ammonia injection valve or ammonia pump frequency sequence for the future period every 5 minutes based on the latest prediction results, wherein the constraints are the valve opening limit and the ammonia slip threshold <3 ppm.

[0008] Preferably, the MPC prediction model can also be designed using reinforcement learning (RL).

[0009] Preferably, the reinforcement learning RL design is based on the reward function: R=-(NOx excess penalty + λ*ammonia escape + μ*ammonia consumption); the state changes of inlet NOx concentration, outlet NOx concentration, inlet flue gas flow, oxygen content, ammonia injection amount, catalyst activity, reactor temperature, valve opening, and ammonia delivery pump frequency are adjusted by the opening of the ammonia injection amount.

[0010] Preferably, in step S3, the LSTM captures the dynamic coupling relationship between the inlet NOx concentration and the ammonia injection amount through a sliding window with a time step of 5 seconds; the physical constraint condition of the PINN includes adding the residual term of the SCR reaction equation to the PINN loss function, including .

[0011] Compared with existing technologies, this invention offers the following benefits: It proposes a predictive denitrification control method based on a long short-term memory network (LSTM) and a physical information neural network (PINN). This method aims to address issues such as control lag, low denitrification efficiency, and excessive ammonia slip in the denitrification process of sintering waste gas in the steel industry. This method effectively improves the accuracy and stability of the denitrification system and optimizes the control of ammonia spraying, thereby achieving ultra-low emissions.

[0012] First, this invention employs an LSTM-based time series prediction model to capture complex time series dynamics in real time, resolving the hysteresis issues common in traditional manual and PID control. By learning the temporal dependencies in historical data, the LSTM model can predict NOx concentrations and ammonia slip for the next 2-3 minutes based on multiple input characteristic factors (such as inlet NOx concentration, outlet NOx concentration, and catalyst activity). This allows the system to proactively implement optimization measures to avoid excessive fluctuations in outlet NOx concentration and prevent NOx overshoot.

[0013] Secondly, the integration of PINN's physical hard constraints further enhances the reliability of prediction results. By introducing physical constraints into the SCR reaction equation, PINN ensures that the prediction data is optimized based on the laws of physics. This effectively avoids unreasonable fluctuations in prediction results and improves the stability and accuracy of the model.

[0014] Furthermore, this invention further optimizes the control strategy through reinforcement learning (RL). RL iteratively adjusts the control strategy based on a reward function, enabling the regulation of ammonia injection to fully account for multiple factors, including NOx emissions, ammonia slip, and ammonia consumption. In practical applications, RL can dynamically adjust the ammonia injection rate and rationally schedule the ammonia pump frequency, thereby ensuring optimal ammonia usage efficiency, avoiding waste, and reducing environmental pollution.

[0015] In terms of data collection, this invention uses edge devices to collect key data in real time through MQTT and OPC protocols, ensuring real-time and accurate data. During data preprocessing, the Kalman filter algorithm is used to remove missing values and outliers, further improving data quality and providing a reliable data foundation for model training.

[0016] In summary, this invention integrates advanced technologies such as LSTM, PINN, and reinforcement learning to achieve intelligent prediction and optimized control of a steel sintering waste gas denitrification system. Compared to existing technologies, this invention offers higher prediction accuracy, real-time response capabilities, and operational stability. It can effectively improve denitrification efficiency, reduce ammonia consumption, and minimize ammonia escape, meeting the steel industry's strict waste gas emission control requirements and providing significant environmental and economic benefits. DETAILED DESCRIPTION

[0017] In order to provide a further understanding of the purpose, structure, features, and functions of the present invention, the present invention is described in detail below with reference to the embodiments.

[0018] This invention takes "data-driven + physical mechanism fusion" as its core and constructs a three-layer control architecture: prediction layer: predicts future NOx concentration changes based on LSTM (long short-term memory network); optimization layer: embeds the SCR reaction kinetic equation through PINN (physical information neural network) to generate physically compliant ammonia injection amount recommendations; control layer: uses reinforcement learning (RL) and model predictive control (MPC) to dynamically adjust the valve opening, and collects feedback data to update the model.

[0019] The present invention provides a denitrification predictive control method based on LSTM and PINN, and the specific steps are as follows: S1: Data acquisition and preprocessing; Edge devices collect production data in real time through protocols such as MQTT and OPC, store it in a time series database, and use algorithms such as Kalman filtering to remove missing and outliers in the data; S2: Build the model; The inlet NOx concentration, outlet NOx concentration, inlet flue gas flow rate, oxygen content, ammonia injection amount, catalyst activity, reactor temperature, dilution air flow fluctuation, and hot blast furnace heating efficiency collected in real time in step S1 are input as characteristic factors into the MPC prediction model to obtain the outlet NOx concentration and ammonia slip amount in the next 2-3 minutes; S3: training model; Utilize LSTM to capture complex time series dynamics in real time, addressing the lag issues of manual control and PID control. Furthermore, PINN (physical hard constraints) prevents unreasonable predictions, making the prediction data more reliable. S4: Confirm the results; Through steps S1-S3, the final predicted value of NOx is confirmed to be: .

[0020] Combining LSTM (Long Short-Term Memory Network) and PINN (Physical Information Neural Network) can effectively capture complex dynamic changes when processing time series data. Especially in control systems with delays and lags, LSTM can better learn and process the regularities in time series data; while PINN introduces physical constraints to ensure that the model output conforms to actual physical laws, avoiding unreasonable fluctuations in the predicted values.

[0021] By collecting real-time data through protocols like MQTT and OPC, and combining it with MPC (Model Predictive Control) strategies, the system can promptly update predictions and meet the demand for rapid response. This is particularly important for denitrification control systems, allowing for faster and more precise adjustment of operating parameters to ensure the effectiveness of the denitrification process.

[0022] Traditional PID control methods often suffer from lag and inability to respond promptly, especially when dealing with systems with large dynamic changes. LSTM can effectively alleviate this problem, making control more precise. PINN helps the system better adapt to the constraints of the physical system, further improving the robustness of the model.

[0023] By integrating data collected by multiple sensors (such as NOX concentration, flue gas flow, oxygen content, etc.), the method can comprehensively reflect the different changing factors in the denitrification process, enhance the comprehensive prediction ability of the model, and help the system make optimal decisions from multiple dimensions.

[0024] The physical hard constraints in the PINN model (such as thermodynamics and reaction kinetics) help ensure the model's reliability in real-world environments, ensuring that its output is not only data-driven but also consistent with physical laws. This enhances the credibility of the predictions and facilitates interpretation and optimization when necessary.

[0025] Preferably, in step S1, the edge device collects the ammonia injection amount, catalyst activity and reactor temperature in the SCR denitrification system through the OPC protocol, and collects the inlet NOx concentration, outlet NOx concentration and oxygen content through the MQTT protocol; the observation noise covariance matrix of the Kalman filter algorithm adopts an adaptive adjustment strategy to dynamically update the covariance parameters according to the volatility of the time series data.

[0026] Edge devices use OPC and MQTT protocols to collect different types of data, including important variables such as ammonia injection rate, catalyst activity, and reactor temperature, as well as key indicators such as NOx concentration and oxygen content. This ensures that the system can monitor and collect multi-dimensional data related to SCR denitrification system operation in real time, providing accurate input for subsequent analysis, prediction, and optimization.

[0027] The Kalman filter algorithm plays a key role in data processing, effectively filtering out noise and improving system accuracy. By adaptively adjusting the observed noise covariance matrix, the Kalman filter can dynamically adjust filter parameters based on fluctuations in time series data, allowing for more flexible adaptation to noise variations in diverse operating environments. This significantly improves data accuracy and reliability, avoiding the inadaptability issues associated with traditional fixed covariance matrices.

[0028] The adaptive adjustment strategy enables the system to dynamically adjust the noise covariance based on data changes, ensuring accurate estimation and prediction under various operating conditions. This is particularly important in highly variable industrial environments, such as the complex dynamics of SCR denitrification systems, where the system's noise characteristics may change under different operating conditions. The adaptive adjustment strategy can optimize the filtering effect in real time.

[0029] Preferably, in step S2, the number of nodes in the input layer of the MPC prediction model is consistent with the number of characteristic factors, the hidden layer adopts a 3-layer LSTM network structure, and the number of neurons in each layer is 64, 32, and 16, respectively; the output layer of the MPC prediction model is set with two nodes, corresponding to the predicted values of the outlet NOx concentration and ammonia escape in the next 2 minutes and 3 minutes, respectively.

[0030] The number of input layer nodes is consistent with the number of characteristic factors, ensuring that the model can accept all relevant feature inputs and avoid missing key information, thereby providing sufficient training data to accurately predict future NOx concentrations and ammonia slip.

[0031] The number of feature factors usually corresponds to the different characteristics of the system. In this way, the model can integrate various input features to make comprehensive predictions.

[0032] The hidden layer adopts a 3-layer LSTM network structure: LSTM (Long Short-Term Memory): This network effectively processes time series data, particularly in tasks with temporal dependencies, such as NOx concentration and ammonia slip. LSTM can capture both long-term and short-term dependencies when processing time series data.

[0033] Multi-layer LSTM architecture: By adding hidden layers, higher-order features in the data can be gradually extracted, enabling the model to learn more complex patterns and nonlinear relationships. A network structure with more layers can effectively improve prediction accuracy, especially for complex time series data, such as industrial emissions data.

[0034] The number of neurons in each layer is 64, 32, and 16 respectively: This design allows the model to reduce the number of neurons layer by layer, helping to extract important high-dimensional features while avoiding overfitting. This gradually reducing the number of neurons also helps simplify the model and improve generalization ability.

[0035] By setting up two nodes, outlet NOx concentration and ammonia slip can be simultaneously predicted for the next two and three minutes. This multi-target output approach enables more precise adjustments in the control system, especially in dynamically changing environments, providing short- and medium-term forecasts.

[0036] Preferably, in step S2, the MPC prediction model solves the optimal ammonia injection valve or ammonia pump frequency sequence for the future period every 5 minutes based on the latest prediction results, wherein the constraints are the valve opening limit and the ammonia slip threshold <3 ppm.

[0037] Preferably, the MPC prediction model can also be designed using reinforcement learning (RL).

[0038] Preferably, the reinforcement learning RL design is based on the reward function: R=-(NOx excess penalty + λ*ammonia escape + μ*ammonia consumption); the state changes of inlet NOx concentration, outlet NOx concentration, inlet flue gas flow, oxygen content, ammonia injection amount, catalyst activity, reactor temperature, valve opening, and ammonia delivery pump frequency are adjusted by the opening of the ammonia injection amount.

[0039] Preferably, in step S3, the LSTM captures the dynamic coupling relationship between the inlet NOx concentration and the ammonia injection amount through a sliding window with a time step of 5 seconds; the physical constraint condition of the PINN includes adding the residual term of the SCR reaction equation to the PINN loss function, including .

[0040] The present invention has been described with reference to the above embodiments. However, the above embodiments are merely exemplary embodiments of the present invention. It should be noted that the disclosed embodiments do not limit the scope of the present invention. On the contrary, modifications and improvements that do not depart from the spirit and scope of the present invention are intended to be protected by the present invention.

Claims

1. A denitrification predictive control method based on LSTM and PINN, characterized by: S1: Data acquisition and preprocessing; Edge devices collect production data in real time through protocols such as MQTT and OPC, store it in a time series database, and use algorithms such as Kalman filtering to remove missing and outliers in the data; S2: Build the model; The entry NO collected in real time in step S1 X Concentration, outlet NO X Concentration, inlet flue gas flow, oxygen content, ammonia injection amount, catalyst activity, reactor temperature, dilution air flow fluctuation, and hot blast furnace heating efficiency are input into the MPC prediction model as characteristic factors to obtain the outlet NOx concentration and ammonia slip in the next 2-3 minutes. S3: training model; Utilize LSTM to capture complex time series dynamics in real time, addressing the lag issues of manual control and PID control. Furthermore, PINN (physical hard constraints) prevents unreasonable predictions, making the prediction data more reliable. S4: Confirm the results; Through steps S1-S3, confirm the final NO X The predicted value is: 。 2. The denitrification predictive control method based on LSTM and PINN according to claim 1, characterized in that: In step S1, the edge device collects the ammonia injection amount, catalyst activity and reactor temperature in the SCR denitrification system through the OPC protocol, and collects the inlet NOx concentration, outlet NOx concentration and oxygen content through the MQTT protocol; the observation noise covariance matrix of the Kalman filter algorithm adopts an adaptive adjustment strategy, and dynamically updates the covariance parameters according to the volatility of the time series data.

3. The denitrification predictive control method based on LSTM and PINN according to claim 1, characterized in that: In step S2, the number of nodes in the input layer of the MPC prediction model is consistent with the number of characteristic factors, and the hidden layer adopts a 3-layer LSTM network structure, with the number of neurons in each layer being 64, 32, and 16, respectively; the output layer of the MPC prediction model is set with two nodes, corresponding to the predicted values of the outlet NOx concentration and ammonia escape in the next 2 minutes and 3 minutes, respectively.

4. The denitrification predictive control method based on LSTM and PINN according to claim 1, characterized in that: In step S2, the MPC prediction model solves the optimal ammonia injection valve or ammonia pump frequency sequence for the future period every 5 minutes based on the latest prediction results, where the constraints are the valve opening limit and the ammonia slip threshold <3 ppm.

5. The denitrification predictive control method based on LSTM and PINN according to claim 4, characterized in that: The MPC prediction model can also be designed using reinforcement learning (RL).

6. The denitrification predictive control method based on LSTM and PINN according to claim 5, characterized in that: The reinforcement learning RL design is based on the reward function: R=-(NOx excess penalty + λ*ammonia escape + μ*ammonia consumption); the state changes of inlet NOx concentration, outlet NOx concentration, inlet flue gas flow, oxygen content, ammonia injection amount, catalyst activity, reactor temperature, valve opening, and ammonia delivery pump frequency are adjusted by the opening of the ammonia injection amount.

7. The denitrification predictive control method based on LSTM and PINN according to claim 1, characterized in that: In step S3, the LSTM captures the dynamic coupling relationship between the inlet NOx concentration and the ammonia injection amount through a sliding window with a time step of 5 seconds; the physical constraint condition of the PINN includes adding the residual term of the SCR reaction equation to the PINN loss function, including 。

Images