Ammonia injection predictive control method and system based on Transform and SVR hybrid algorithm

Abstract

The invention provides an ammonia injection predictive control method and an ammonia injection predictive control system based on a Transform and SVR (Support Vector Regression) hybrid algorithm. According to the method, data driving and mechanism fusion are taken as a core, and a four-layer control framework is constructed: a data preprocessing layer acquires data through edge equipment, carries out Kalman filtering processing on the data, extracts time sequence characteristics by utilizing Transform, and supplements static parameter regression by SVR (Support Vector Regression); the prediction layer integrates a time sequence prediction result of the Transform and a static parameter prediction result of the SVR through a fusion formula, and outputs an outlet NOx concentration prediction value 2-3 minutes in the future; an SCR reaction kinetic equation residual error is introduced into a loss function by the optimization layer, and a prediction result is restrained to conform to a physical mechanism; and the control layer generates a dynamic ammonia spraying instruction based on reinforcement learning to realize closed-loop optimization. The ammonia spraying amount can be accurately regulated and controlled, it is ensured that the NOx concentration of an outlet is smaller than 50 mg / Nm, meanwhile, the ammonia escape rate and the ammonia water consumption are reduced, and the method is suitable for complex and changeable sintering production working conditions.

Description

Technical Field

[0001] This invention relates to the field of industrial environmental protection control technology, specifically to a predictive control method and system for ammonia injection based on a hybrid Transformer and SVR algorithm. Background Technology

[0002] The steel industry is a typical high-energy-consuming and high-polluting industry. The sintering flue gas produced during the sintering process generates a large amount of nitrogen oxides (NOx). With the Ministry of Ecology and Environment's latest "Technical Specification for Ultra-Low Emission Treatment Engineering of Sintering Waste Gas in the Steel Industry," which mandates ultra-low emissions of pollutants from sintering waste gas from the steel industry, to be fully implemented from April 1, 2025, including a requirement that the NOx emission concentration in sintering waste gas be <50 mg / Nm³, current denitrification control technologies face the following challenges.

[0003] Adjusting the ammonia injection rate based on the changing trends of inlet and outlet nitrogen oxides has a lag effect, easily leading to large fluctuations in outlet NOx and potential over-injection. It is also significantly affected by factors such as catalyst activity and reactor temperature (200-300℃). Low reactor inlet temperature can easily reduce denitrification efficiency and waste ammonia. Furthermore, the ammonia regulating valve opening is not linearly related to the ammonia flow rate; fluctuating ammonia flow rate can lead to over-injection, easily causing excessive ammonia escape at the reactor outlet. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a predictive control method and system for ammonia injection based on a hybrid Transformer and SVR algorithm.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a predictive control method for ammonia injection based on a hybrid Transformer and SVR algorithm, comprising the following steps: S1: Data Acquisition and Preprocessing Production data of the sintering flue gas denitrification system is collected in real time using edge devices via MQTT or OPC protocols, and the production data is stored in a time-series database. Edge computing devices (such as industrial gateways) are used to collect production data of 11 types of sintering flue gas denitrification systems in real time via MQTT (Message Queuing Telemetry Transmission) or OPC (Object Linking and Embedding for Process Control) protocols. The collection frequency is set to 1Hz to ensure data timeliness. The production data includes inlet NOx concentration (mg / Nm³), outlet NOx concentration (mg / Nm³), inlet flue gas flow rate (Nm³ / h), inlet oxygen content (%), ammonia injection rate (L / h), catalyst activity K value (L / h), reactor temperature (°C), dilution air flow rate (Nm³ / h), hot blast furnace heating efficiency (°C / min), ammonia water regulating valve opening (%), and ammonia water delivery pump frequency (Hz). The Kalman filter algorithm is used to remove missing and outlier values ​​from the production data to obtain preprocessed data; The collected production data is stored in a time series database (such as InfluxDB), and the Kalman filter algorithm is used to clean the data: for missing values, the state estimation function of the Kalman filter is used to fill in the missing values; for outliers (such as jump data caused by sensor failure), a 3σ threshold (σ is the standard deviation of the data) is set to remove outliers that exceed the threshold, so as to ensure the data quality of the input model.

[0006] S2: Construction and Training of Hybrid Prediction Models S21: Construct the Transformer model: A four-layer encoder-decoder architecture is stacked, with 128 hidden units per layer; a multi-head attention mechanism (8 heads) is used to capture long-range dependencies in the temporal data. The time-series data in the preprocessed data is input into the Transformer model to train the Transformer model to output the predicted NOx concentration at the outlet for the next 2-3 minutes and the time-series high-dimensional features. Seven types of time-series data were selected, including inlet NOx concentration, outlet NOx concentration, inlet flue gas flow rate, ammonia injection rate, reactor temperature, dilution air flow rate, and hot air furnace heating efficiency. An input sequence was constructed according to the time step (1 minute / step), and the sequence length was set to 30 (i.e., predicting the future based on the past 30 minutes of data). The Adam optimizer was used with a learning rate of 0.005 and 1500 iterations. The training objective was to output the predicted NOx concentration at the outlet for the next 2-3 minutes and the time series high-dimensional features (such as the slope of concentration change and periodic fluctuation features).

[0007] S22: Constructing the SVR model: A Gaussian kernel is selected as the kernel function, and the hyperparameters are optimized through grid search. The time-series high-dimensional features are input into the SVR model to train the SVR model and output steady-state predictions of key static parameters such as catalyst activity K value and inlet oxygen content threshold. The Gaussian kernel (RBF) is used because it has a strong ability to fit nonlinear data in high-dimensional space, and can effectively handle nonlinear regression problems of static parameters such as catalyst activity K value and oxygen content threshold. The penalty coefficient C (value of 50) and kernel bandwidth γ (value of 0.01) are optimized using a grid search method to avoid model overfitting. Using the temporal high-dimensional features extracted by Transformer as input, the steady-state predicted values ​​of catalyst activity K and inlet oxygen content threshold are output, supplementing the blind spot of Transformer in static parameter prediction.

[0008] S23: Hybrid Prediction Output: The final prediction result is calculated using a fusion formula, which is: Ypred = γ* yTransformer + (1−γ) * ySVR; Where: Ypred is the final predicted value; yTransformer is the predicted output of the Transformer model; ySVR is the predicted output of the SVR model; γ is the weight coefficient, with a value range of 0.6-0.8.

[0009] S3: Prediction result optimization: A loss function with physical mechanism constraints is constructed to optimize the final prediction result. The loss function is as follows: Loss=α* MSE(ypred,ytrue)+β * Residual(physical equation)+λ*SVR; Where α, β, and λ are weighting coefficients; MSE(ypred,ytrue) is the mean square error between the predicted and true values; Residual (physical equation) is the residual of the SCR reaction kinetic equation, including the residual of the NH3 / NOx molar ratio constraint and the residual of the catalyst activity equation; Where α = (mean squared error weight, to ensure the good fit between the predicted value and the true value), β = 0.3 (physical residual weight, to constrain the prediction to conform to the process mechanism), and λ = 0.1 (SVR regularization term weight, to prevent overfitting).

[0010] The physical equation consists of two parts: one is the NH3 / NOx molar ratio constraint residual (the molar ratio needs to be controlled between 1.0 and 1.2, and the residual is calculated as (|actual molar ratio - target molar ratio|). Second, the residual of the catalyst activity equation; the catalyst activity K value must satisfy: K = K0 * e- Ea / (R -T)The residual is calculated as (|predicted K value - theoretical K value|).

[0011] The loss function is minimized by gradient descent, and the parameters of the Transformer and SVR models are iteratively adjusted to ensure that the prediction results are both accurate and in line with actual process constraints (such as avoiding unreasonable predictions such as "NH3 / NOx molar ratio = 0.8").

[0012] S4: Dynamic ammonia injection control: Reinforcement learning is introduced to construct a closed-loop control strategy, and an ammonia injection rate adjustment command is generated based on the optimized final prediction result. The state space of the reinforcement learning includes inlet NOx concentration, outlet NOx concentration, inlet flue gas flow rate, inlet oxygen content, ammonia injection rate, catalyst activity K value, reactor temperature, ammonia water regulating valve opening, ammonia water delivery pump frequency, and static parameters output by the SVR model; the action space is set as the ammonia injection rate ±5% opening adjustment command. The reward function is set as: R = -(NOx excess penalty + λ * ammonia escape + μ * ammonia water consumption + η * control fluctuation penalty), where λ, μ, and η are penalty coefficients.

[0013] (When the NOx output is ≥50mg / Nm³, the penalty value is 100; when the standard is met, the penalty value is 0), λ = 50 (ammonia escape penalty coefficient), μ = 10 (ammonia water consumption penalty coefficient), η = 20 (control fluctuation penalty coefficient, calculated based on the adjustment range of the ammonia injection amount between two adjacent times).

[0014] The reinforcement learning agent reads the state space data every 5 seconds, combines it with the outlet NOx prediction value of the hybrid prediction layer, and selects the optimal action (ammonia injection rate adjustment command). After the command is issued, the actual outlet NOx concentration and ammonia escape amount are collected by sensors, the reward value is calculated and fed back to the agent, and the learning strategy is updated to achieve dynamic optimization.

[0015] Preferably, in step S21, the time-series data includes inlet NOx concentration, outlet NOx concentration, inlet flue gas flow rate, ammonia injection rate, reactor temperature, dilution air flow rate, and hot air furnace heating efficiency. The Transformer model is trained and iterated at no less than 1000 times, and the learning rate is set to 0.001-0.01.

[0016] Preferably, in step S22, the parameter range of the grid search includes a Gaussian kernel bandwidth γ of 0.001-0.1; and the training sample size of the SVR model is not less than 1000 sets.

[0017] Preferably, the control method further includes step S5: lightweight deployment of the model. The trained Transformer model is converted to TensorFlowLite format, and the SVR model is deployed in ONNX format and integrated into the energy-efficient all-in-one machine. The GPU acceleration engine and CPU multi-threaded parallel computing are used to ensure that the time for a single prediction is ≤5 seconds.

[0018] Preferably, the control method further includes step S6: model verification. A / B testing was conducted on some sintering production lines to compare the outlet NOx concentration fluctuation, ammonia slip rate, ammonia water consumption, and denitrification efficiency between traditional PID control and this method. The control target for the ammonia slip rate was ≤3ppm.

[0019] Preferably, a predictive control system for ammonia injection into sintering flue gas based on a hybrid Transformer and SVR algorithm is provided, the system comprising: Data acquisition module: used to acquire production data of the sintering flue gas denitrification system in real time via MQTT or OPC protocol. The production data includes inlet NOx concentration, outlet NOx concentration, inlet flue gas flow rate, inlet oxygen content, ammonia injection rate, catalyst activity K value, reactor temperature, dilution air flow rate, hot blast furnace heating efficiency, ammonia water regulating valve opening degree and ammonia water delivery pump frequency. Data preprocessing module: used to store the production data into a time series database, and use the Kalman filter algorithm to remove missing and outlier values, and output the preprocessed data; The hybrid prediction module includes a Transformer submodule, an SVR submodule, and a fusion submodule. The Transformer submodule processes time-series data and outputs predicted NOx concentrations at the outlet for the next 2-3 minutes, along with time-series high-dimensional features. The SVR submodule processes the time-series high-dimensional features and outputs steady-state predicted values ​​of static parameters. The fusion submodule outputs the final prediction result using a fusion formula. Optimization module: used to optimize the final prediction result using a loss function with physical mechanism constraints; Control module: used to generate ammonia injection adjustment commands based on reinforcement learning to achieve closed-loop control; Deployment module: used to deploy the Transformer model and SVR model in a lightweight manner to ensure real-time prediction.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention predicts the NOx concentration at the outlet 2-6 minutes in the future using the Transformer model, achieving "advance adjustment", resulting in a decrease in the NOx concentration at the outlet and an improvement in the compliance rate; secondly, based on the NH3 / NOx molar ratio constraint and reinforcement learning reward function, the ammonia escape rate is reduced, avoiding environmental penalties and reducing secondary pollution; Furthermore, precise ammonia injection control reduces ammonia consumption compared to traditional PID control; the hybrid model can capture dynamic fluctuations in parameters such as inlet flue gas flow and reactor temperature in real time, maintaining stable control even under conditions where the sintering machine material layer thickness changes by ±10% and the reactor temperature fluctuates by ±20℃; in addition, the reinforcement learning feedback mechanism can automatically adjust parameters as the equipment deteriorates (e.g., catalyst activity decreases by 5% per year), eliminating the need for manual readjustment and extending the effective lifespan of the model. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0022] To provide a further understanding of the purpose, structure, features, and functions of the present invention, detailed descriptions are provided below with reference to specific embodiments.

[0023] This invention provides a predictive control method and system for ammonia injection based on a hybrid algorithm of Transformer and SVR.

[0024] A CEMS analyzer (measuring inlet / outlet NOx concentration, accuracy ±2%), a temperature sensor (measuring reactor temperature, accuracy ±1℃), and a flow sensor (measuring inlet flue gas flow / dilution air flow, accuracy ±3%) were installed. Data was collected via an industrial gateway (Huawei AR550) using the MQTT protocol and stored in the InfluxDB time-series database. A Kalman filter algorithm was used to process the data, with a missing value completion threshold (completed when consecutive missing values ​​are ≤3 seconds) and an outlier removal threshold (3σ, σ=8mg / Nm³). The Transformer model was trained using the TensorFlow framework, with a 4-layer encoder-decoder structure, 128 hidden units, an input sequence length of 30 (30 minutes of data), and an output of the predicted outlet NOx value for the next 2 minutes. The SVR model was trained using the Scikit-learn framework, with a Gaussian kernel (C=50, γ=0.01), inputting the Transformer's time-series high-dimensional features (64 dimensions in total), and outputting the catalyst activity K. The values ​​and oxygen content thresholds were used; the fusion weight γ=0.7, and the fusion calculation was performed using a Python script.

[0025] The loss function was set with α=0.6, β=0.3, and λ=0.1, the target molar ratio of NH3 / NOx was 1.1, and the catalyst activity equation parameters were Ea=80kJ / mol and R=8.314J / (mol·K). Gradient descent optimization was implemented using the PyTorch framework. The reinforcement learning algorithm adopted the DQN (Deep Q-Network) algorithm, with a 10-dimensional state space, 3 discrete actions in the action space, and reward function coefficients λ=50, μ=10, and η=20. It was deployed to an energy efficiency integrated machine (model: Advantech UNO-2484G) via TensorFlow Lite. The Transformer model inference was accelerated using an NVIDIA Jetson Nano GPU, and the CPU (Intel Core i5) was used for multi-threaded processing of SVR calculation and reinforcement learning decision-making, with a single prediction time of 3.2 seconds.

[0026] Running result: Prediction accuracy: The average value of traditional PID control is 8.2 mg / Nm³, while that of Transformer+SVR+RL intelligent control is 5.6 mg / Nm³. Ammonia slip rate: Traditional PID control has an average value of 5-8 ppm, while Transformer+SVR+RL intelligent control has a value of ≤2.5 ppm; Ammonia consumption: Transformer+SVR+RL intelligent control reduces ammonia consumption by 10%-20% compared to traditional PID control; Operating condition adaptability: Traditional PID control relies on manual experience for adjustment, while Transformer+SVR+RL intelligent control can automatically adapt to fluctuations in flow rate and reactor temperature.

[0027] By combining "deep features + regression algorithms", Transformer focuses on long-sequence time series modeling. It can leverage Transformer's powerful time series feature extraction capabilities and SVR's good regression performance on high-dimensional data, as well as solve static parameter nonlinear regression. The two complement each other, thus achieving better prediction results.

[0028] The present invention has been described in the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.

Claims

1. A method of ammonia injection predictive control based on a hybrid algorithm of Transformer and SVR, characterized in that: The method comprises the following steps: S1: data acquisition and preprocessing: Real-time acquisition of production data of the sintering flue gas denitration system through edge devices using MQTT or OPC protocol, and storage of the production data into a time series database; Using Kalman filtering algorithm to eliminate missing values and abnormal values in the production data to obtain preprocessed data; The production data includes inlet NOx concentration, outlet NOx concentration, inlet flue gas flow, inlet oxygen content, ammonia injection amount, catalyst activity K value, reactor temperature, dilution air flow, hot blast stove temperature rising efficiency, ammonia water regulating valve opening degree and ammonia water delivery pump frequency; S2: mixed prediction model construction and training: S21: constructing a Transformer model: Stacking 4 layers of encoder-decoder, and setting the number of hidden units in each layer to 128; Inputting time series data in the preprocessed data into the Transformer model, training the Transformer model to output 2-3 minutes of future outlet NOx concentration prediction value and time series high-dimensional features; S22: constructing an SVR model: Selecting a Gaussian kernel as a kernel function, optimizing hyperparameters through grid search, inputting the time series high-dimensional features into the SVR model, and training the SVR model to output steady-state prediction values of catalyst activity K value and inlet oxygen content threshold key static parameters; S23: mixed prediction output: Using a fusion formula to calculate the final prediction result, and the fusion formula is: Y pred = γ* y Transformer + (1−γ) * y SVR ; wherein: Y pred Final final prediction value; y Transformer is the prediction output of the Transformer model; y SVR is the prediction output of the SVR model; γ is a weight coefficient, and the value range is 0.6-0.8; S3: prediction result optimization: Constructing a loss function containing physical mechanism constraints to optimize the final prediction result, and the loss function is: Loss=α* MSE(ypred,ytrue)+β * Residual(physical equation)+λ*SVR; Wherein α, β, λ are weight coefficients; MSE(ypred,ytrue) is the mean square error of the prediction value and the true value; Residual(physical equation) is the SCR reaction kinetics equation residual, including NH3 / NOx molar ratio constraint residual and catalyst activity equation residual; S4: dynamic ammonia injection control: Introducing reinforcement learning to construct a closed-loop control strategy, and generating ammonia injection amount adjustment instructions based on the optimized final prediction result; The state space of the reinforcement learning includes inlet NOx concentration, outlet NOx concentration, inlet flue gas flow, inlet oxygen content, ammonia injection amount, catalyst activity K value, reactor temperature, ammonia water regulating valve opening degree, ammonia water delivery pump frequency and static parameters output by the SVR model; the action space is set to ammonia injection amount ±5% opening degree adjustment instruction; The reward function is set to: R = -(NOx over-standard penalty +λ* ammonia escape amount + μ* ammonia water consumption amount + η* control fluctuation penalty), wherein λ, μ, η are penalty coefficients.

2. The ammonia injection predictive control method based on a hybrid algorithm of Transformer and SVR according to claim 1, wherein: In step S21, the time series data includes inlet NOx concentration, outlet NOx concentration, inlet flue gas flow, ammonia injection amount, reactor temperature, dilution air flow, and hot blast stove temperature rising efficiency, the number of training iterations of the Transformer model is not less than 1000 times, and the learning rate is set to 0.001-0.

01.

3. The ammonia injection predictive control method based on a hybrid algorithm of Transformer and SVR according to claim 1, wherein: In step S22, the parameter range of the grid search includes a Gaussian kernel bandwidth γ of 0.001-0.1, and the number of training samples of the SVR model is not less than 1000 groups.

4. The ammonia injection predictive control method based on a hybrid algorithm of Transformer and SVR according to claim 1, wherein: The control method further comprises step S5: model lightweight deployment: The trained Transformer model is converted into TensorFlowLite format, the SVR model is deployed through ONNX format, and is integrated into an energy efficiency all-in-one machine; GPU acceleration engine and CPU multi-thread parallel computing are adopted to ensure that the time consumption of single prediction is ≤5 seconds.

5. The ammonia injection predictive control method based on a hybrid algorithm of Transformer and SVR according to claim 4, wherein: The control method further comprises step S6: model verification: A / B test is carried out on a part of sintering production line, the outlet NOx concentration fluctuation amplitude, ammonia escape rate, ammonia water consumption and denitration efficiency of the traditional PID control and the method are compared, and the control target of the ammonia escape rate is ≤3 ppm.

6. A sintering flue gas ammonia injection predictive control system based on a hybrid algorithm of Transformer and SVR, characterized by: The system comprises A data acquisition module is configured to acquire production data of a sintering flue gas denitration system in real time through MQTT or OPC protocol, the production data includes inlet NOx concentration, outlet NOx concentration, inlet flue gas flow, inlet oxygen content, ammonia injection amount, catalyst activity K value, reactor temperature, dilution air flow, hot blast stove temperature rising efficiency, ammonia water regulating valve opening degree, and ammonia water delivery pump frequency; A data preprocessing module is configured to store the production data into a time series database, and remove missing values and abnormal values by using Kalman filtering algorithm, and output preprocessed data; A hybrid prediction module comprises a Transformer submodule, an SVR submodule, and a fusion submodule; the Transformer submodule is configured to process time series data and output future 2-3 minute outlet NOx concentration prediction value and time series high-dimensional features; The SVR submodule is configured to process the time series high-dimensional features and output steady-state prediction value of static parameters; the fusion submodule is configured to output final prediction result through a fusion formula; An optimization module is configured to optimize the final prediction result by using a loss function containing physical mechanism constraints; A control module is configured to generate ammonia injection amount adjustment instructions based on reinforcement learning to realize closed-loop control; and a deployment module is configured to lightweight deploy the Transformer model and the SVR model to ensure prediction real-time performance.

Images