Closed-loop control method for pollutant emission of thermal power plant

Abstract

The invention belongs to the technical field of automatic control of thermal power plants, particularly discloses a closed-loop control method for pollutant emission of a thermal power plant, and aims to solve the problems of low pollutant emission control precision, high agent consumption and poor subsystem collaboration in the prior art. The method comprises the following steps: constructing a multi-source data acquisition and fusion platform to form a unified time sequence data stream; establishing a multi-step advanced prediction model of nitrogen oxide and sulfur dioxide concentration based on a long-short-term memory neural network; designing a multi-target collaborative optimization control strategy which is constrained by up-to-standard emission, minimum cost and equipment safety; online correction of model parameters and dynamic adjustment of a control strategy are realized through Kalman filtering and an expert rule base; a denitration subsystem, a desulfurization subsystem and a dust removal subsystem are integrated to construct an intelligent environment-friendly island collaborative management and control center, and whole-process closed-loop optimization is achieved. According to the method, the emission control precision and the system robustness are remarkably improved, the consumption of ammonia water and limestone is reduced, an information island is broken through, and efficient cooperative operation of environmental protection facilities is achieved.

Description

Technical Field

[0001] This invention relates to the field of automatic control technology for thermal power plants, specifically to a closed-loop control method for pollutant emissions from thermal power plants. Background Technology

[0002] With increasingly stringent national requirements for air pollution control, thermal power plants, as key pollutant emitters, are facing higher standards in their pollutant emission control technologies. Nitrogen oxides (NOx) produced during coal-fired power generation... x Pollutants such as sulfur dioxide (SO2) and particulate matter not only cause serious harm to the ecological environment, but may also trigger complex environmental problems such as acid rain, photochemical smog, and fine particulate matter (PM2.5) pollution. To address this challenge, the state has successively issued the "Emission Standard of Air Pollutants for Thermal Power Plants" (GB 13223) and the "ultra-low emission" retrofit policy, clearly requiring thermal power plants to reduce NO2 emissions to a minimum under a baseline oxygen content of 6%. x The concentrations of SO2 and particulate matter emissions shall not exceed 50 mg / m³. 3 35mg / m 3 and 10mg / m 3 Against this backdrop, environmental protection facilities such as denitrification, desulfurization, and dust removal have become an indispensable core component of thermal power plants, and must be operated efficiently and in a coordinated manner to achieve stable and compliant emissions of pollutants.

[0003] However, the actual operating conditions of thermal power plants are extremely complex. They are affected by multiple factors, including fluctuations in coal composition (such as coal type switching, changes in calorific value, and differences in sulfur and nitrogen content), frequent adjustments to unit load (such as peak-shaving operation and frequent start-ups and shutdowns), equipment aging (such as catalyst deactivation, nozzle blockage, and scale buildup in absorption towers), and environmental temperature and humidity. This results in the pollutant generation and removal processes exhibiting significant nonlinearity, large hysteresis, strong coupling, and time-varying characteristics. For example, during rapid load increases and decreases, the boiler combustion state changes drastically, leading to increased NO... x The generation rate fluctuates rapidly, and the SCR denitrification system experiences a long transmission delay (typically tens of seconds to several minutes) as the flue gas flows through the catalyst bed, making it difficult for traditional control strategies to respond in a timely manner, which easily leads to excessive NO at the outlet. x Excessive concentration or excessive ammonia injection. Similarly, there is a complex dynamic coupling relationship between the reaction rate and pH control of limestone slurry in wet desulfurization systems and SO2 absorption efficiency. If not properly controlled, it will not only affect the desulfurization efficiency, but may also cause problems such as a decline in gypsum quality or system blockage.

[0004] Currently, denitrification systems in thermal power plants generally employ selective catalytic reduction (SCR) technology. This involves injecting ammonia water or liquid ammonia into the flue gas as a reducing agent, which, under the action of a catalyst, reduces NO.x It is reduced to harmless nitrogen and water. Traditional control strategies mostly rely on conventional PID (proportional-integral-derivative) controllers, based on the SCR outlet NO... x The ammonia injection valve opening is adjusted based on the real-time feedback signal of the concentration. However, this method has significant limitations: firstly, due to the large inertia and long time delay characteristics of the SCR reactor itself, the PID controller struggles to effectively compensate for the system's dynamic lag, resulting in control actions lagging behind actual requirements; secondly, this strategy lacks consideration of boiler combustion status, flue gas flow rate, and inlet NO₂ levels. x The lack of utilization of feedforward information such as concentration change trends makes it impossible to predict the impact of operating condition disturbances on pollutant formation, resulting in the control process exhibiting typical problems of "lagging adjustment and overshoot oscillation." (Export NO...) x The concentration fluctuates drastically and has poor stability.

[0005] Furthermore, existing emission control systems generally suffer from "information silos." The three main subsystems—denitrification, desulfurization, and dust removal—are typically provided by different manufacturers, employing independent control logic and optimization objectives, lacking a unified collaborative scheduling mechanism. For example, the denitrification system aims for NO... x While achieving compliance, excessive ammonia injection may lead to increased ammonia slip (NH3 slip). This slipping ammonia, entering subsequent air preheaters or desulfurization towers, reacts with SO3 to form ammonium bisulfate, causing equipment blockage or corrosion, and increasing the burden on the desulfurization system. Similarly, to ensure SO2 compliance, the desulfurization system may over-add limestone, wasting materials and potentially affecting the quality of gypsum byproducts. This fragmented control model makes it difficult to achieve coordinated emission reduction of multiple pollutants and optimal resource allocation, increasing operating costs and creating environmental compliance risks.

[0006] A deeper problem lies in the lack of in-depth mining and intelligent analysis capabilities of existing control methods for massive historical operating data. Thermal power plant DCS (Distributed Control Systems) and environmental monitoring systems have accumulated a large amount of high-dimensional, multi-source, and time-series operating data, including boiler parameters, flue gas composition, equipment status, and environmental variables. However, traditional control models are mostly based on simplified mechanisms or empirical rules, making it difficult to accurately characterize nonlinear dynamic relationships under complex operating conditions. Although some power plants have attempted to introduce advanced algorithms such as fuzzy control, neural networks, or model predictive control (MPC) in recent years, they still face bottlenecks in practical applications, such as weak model generalization ability, poor online adaptability, and complex engineering deployment, making it difficult to maintain long-term stable and efficient control performance under varying operating conditions.

[0007] Therefore, there is an urgent need for a method for the coordinated control of pollutant emissions from thermal power plants that integrates artificial intelligence prediction capabilities with advanced closed-loop control strategies. This method should be able to construct a high-precision, adaptive pollutant concentration prediction model based on multi-source heterogeneous data, and predict NO concentrations in advance. xThe study analyzes the changing trends of key indicators such as SO2. Simultaneously, by establishing a coupling relationship model among the denitrification, desulfurization, and dust removal subsystems, a multi-objective collaborative optimization control strategy is designed. This strategy aims to minimize the consumption of reducing agents and absorbents, suppress ammonia escape, and improve the overall energy efficiency and environmental performance of the system, while ensuring stable and compliant emissions of all pollutants. This not only helps thermal power plants cope with increasingly stringent environmental regulations but also provides key technological support for achieving clean and efficient coal-fired power generation under the "dual carbon" target. Summary of the Invention

[0008] The present invention aims to at least solve one of the technical problems existing in the prior art, and provides a closed-loop control method for pollutant emissions from thermal power plants.

[0009] This invention provides a closed-loop control method for pollutant emissions from thermal power plants, comprising the following steps: Step 1: Construct a multi-source data acquisition and fusion platform. Through a distributed control system, collect multi-source heterogeneous data in real time, including boiler combustion parameters, flue gas composition data, and environmental protection facility operation status. Then, preprocess and standardize the data to form a time-series data stream in a unified format. Step 2: Establish an intelligent prediction model for pollutant concentrations. Based on the time-series data stream, use a long short-term memory neural network to perform multi-step advance prediction of the concentration change trends of nitrogen oxides and sulfur dioxide to generate prediction output. Step 3: Design a multi-objective collaborative optimization control strategy. Use the predicted output as feedforward information and combine it with the model predictive control framework to dynamically find the optimal control sequence of the denitrification, desulfurization and dust removal system under the constraints of emission compliance, minimum operating cost and equipment safety. Step 4: Implement closed-loop feedback and adaptive correction. By comparing the actual emission concentration monitoring values ​​with the model prediction values ​​in real time, the Kalman filter algorithm is used to correct the prediction model parameters online, and the rolling optimization mechanism is used to dynamically adjust the control strategy to cope with operating condition fluctuations. Step 5: Construct a smart environmental protection island collaborative management and control center, integrating the control loops of denitrification, desulfurization, and dust removal subsystems. Through a unified communication protocol, it realizes the synchronous issuance of control commands and centralized monitoring of operating status, forming a closed-loop management and control system for the entire process.

[0010] Furthermore, the multi-source data collected in step 1 specifically includes boiler main steam pressure, main steam temperature, coal feed rate, primary air volume, secondary air volume, flue gas oxygen content, nitrogen oxide concentration at the inlet and outlet of the selective catalytic reduction system, sulfur dioxide concentration at the inlet and outlet of the wet desulfurization system, flue gas particulate matter concentration, absorber slurry pH value, slurry density, and flue gas temperature parameters, and the sampling frequency of each parameter is 1 time per second.

[0011] Furthermore, in step 1, the preprocessing uses a sliding window method to handle missing data, with a window length of 10 sampling points. Principal component analysis is used to reduce the dimensionality of high-dimensional features, and the principal component features with a cumulative contribution rate greater than 95% after screening are retained as input to the intelligent prediction model.

[0012] Furthermore, the long short-term memory neural network in step 2 contains three hidden layers, with 128, 64 and 32 neurons in each layer, respectively. The neural network has an input feature dimension of 50, and its output is the predicted concentration of nitrogen oxides and sulfur dioxide for the next 10 sampling points. The training process of the neural network is optimized using the mean squared error loss function and the adaptive moment estimation algorithm is used as the optimizer.

[0013] Furthermore, the intelligent pollutant concentration prediction model in step 2 is updated online every 24 hours using the latest operating data to dynamically control the absolute value of the model prediction error within ±3 mg / m³, ensuring that the prediction model can continuously adapt to long-term operating condition evolution such as equipment aging, coal quality changes, or seasonal environmental fluctuations, and maintain the long-term stability of prediction accuracy.

[0014] Furthermore, in step 3, the model predictive control framework adopts a quadratic programming solver. Its optimization objective function simultaneously considers the square error of the nitrogen oxide emission concentration deviating from the set value, the square of the ammonia injection flow rate adjustment, the square error of the sulfur dioxide emission concentration deviating from the set value, and the square of the limestone slurry addition rate. Hard constraints are applied, including the ammonia injection valve opening range, the upper and lower limits of the slurry circulation pump frequency, and the ammonia escape concentration not exceeding 2.5 mg / m³.

[0015] Furthermore, the control strategy optimization cycle in step 3 is 15 seconds. Each optimization solves for the optimal control sequence within the next 60 seconds, but only the first control variable of the sequence is executed. In the next cycle, the next optimization is performed based on the updated system state, thus forming a rolling optimization mechanism to balance the forward-looking nature of the control with the ability to respond quickly to disturbances.

[0016] Furthermore, the closed-loop feedback correction in step 4 employs the extended Kalman filter algorithm, treating the long short-term memory neural network as a nonlinear dynamic system. The hidden state variables and output layer weight matrix of the network are estimated and updated in real time online to compensate for systematic biases caused by model simplification or insufficient training data. The correction period is 5 minutes.

[0017] Furthermore, the adaptive correction in step 4 also includes an expert rule base. When the prediction deviation exceeds 5 consecutive sampling periods and its absolute value is greater than 5 milligrams per cubic meter, the expert rule base automatically triggers a model reconstruction mechanism. The model reconstruction mechanism includes local fine-tuning, full retraining, or switching to a backup model.

[0018] Furthermore, in step 5, the smart environmental protection island collaborative management and control center communicates with the controllers of each subsystem through the industrial Ethernet protocol, with a data refresh cycle of 100 milliseconds, and is equipped with a redundant hot backup server to ensure the reliability of continuous system operation. The method also includes establishing an operating cost and energy efficiency assessment module to calculate in real time the pollutant emission reduction cost per unit of power generation, reagent consumption, and system power consumption, and to generate a daily operation optimization report to support collaborative optimization with power generation load commands.

[0019] Compared with existing technologies, the closed-loop control method for pollutant emissions from thermal power plants in this invention significantly improves the accuracy and stability of pollutant emission control. By constructing a multi-step predictive model based on LSTM, the system can predict trends within 5 to 30 sampling periods (i.e., 5 to 30 seconds) before pollutant concentrations actually exceed standards, achieving proactive intervention rather than passive response. Actual measurement data shows that NO... x Emission fluctuation range reduced from ±15 mg / m³ under conventional control 3 Compressed to ±5 mg / m 3 SO2 control also achieves high precision and stability, effectively avoiding the risk of instantaneous exceedances due to control lag, and ensuring that emissions throughout the year are consistently better than the national ultra-low emission standards.

[0020] Furthermore, the closed-loop control method for pollutant emissions from thermal power plants in this invention significantly reduces operating costs and resource consumption. Through a multi-objective collaborative optimization strategy, the system automatically balances reagent dosage and equipment movement amplitude while meeting emission constraints. In practical applications, annual ammonia consumption is reduced by 180 tons, limestone consumption by 2500 tons, and slurry pump power consumption is reduced by approximately 8%, resulting in comprehensive annual economic benefits exceeding 4 million yuan. Simultaneously, smooth control commands significantly reduce mechanical wear on actuators (such as regulating valves and variable frequency pumps), lowering equipment failure rates by 35%, extending the service life of key components, and reducing maintenance costs.

[0021] Furthermore, the closed-loop control method for pollutant emissions from thermal power plants in this embodiment of the invention effectively suppresses the risk of secondary pollution. This is achieved by embedding ammonia escape hard constraints (≤2.5 mg / m³) into the optimization objective function. 3 (), and combined with real-time feedback correction, the system reduced the ammonia escape concentration from 3.5 mg / m³. 3 Stable control at 1.8 mg / m³ 3The following levels are significantly lower than industry-standard levels (typically 2.5–5 mg / m²). 3 This fundamentally avoids problems such as air preheater blockage, catalyst poisoning, and ammonium bisulfate deposition caused by excessive ammonia escape, ensuring the safe and long-term operation of boilers and environmental protection equipment.

[0022] Furthermore, the closed-loop control method for pollutant emissions from thermal power plants in this embodiment of the invention significantly enhances the robustness and adaptability of the system under varying operating conditions. This embodiment of the invention effectively addresses complex scenarios such as rapid load adjustments, coal quality fluctuations, and equipment aging through a triple mechanism of rolling optimization, adaptive correction, and expert rule base. Under a 30% deep peak-shaving condition, NO... x Emissions can still be stably controlled at 35±3 mg / m³ 3 It exhibits excellent low-load control performance; during fuel switching or start-stop processes, the system response delay is reduced from 90 seconds to 30 seconds, avoiding the "over-adjustment" or "under-adjustment" phenomena commonly found in traditional control.

[0023] Furthermore, the closed-loop control method for pollutant emissions from thermal power plants in this embodiment of the invention achieves efficient coordination among the three subsystems of the environmental protection island. Through the intelligent environmental protection island collaborative management center, information barriers between the denitrification, desulfurization, and dust removal systems are broken down, achieving millisecond-level command synchronization and status sharing. For example, when the boiler load suddenly drops, the system can simultaneously coordinate to reduce ammonia injection, lower the slurry circulation pump frequency, and adjust the electrostatic precipitator voltage, avoiding chain fluctuations caused by excessive adjustment of a single subsystem, thus improving overall control efficiency by more than 20%.

[0024] Furthermore, the closed-loop control method for pollutant emissions from thermal power plants in this embodiment of the invention promotes the transformation of power plant operation modes from "compliance and standards" to "smart, economical, and environmentally friendly." By integrating operating cost and energy efficiency assessment modules, the system not only focuses on emission compliance but also quantitatively analyzes the environmental cost per kilowatt-hour, supports linkage with grid dispatch instructions, and achieves coordinated optimization of the plant's economic and environmental aspects. Operators can perform refined operations based on daily optimization reports, truly achieving the sustainable development goal of "both blue skies and economic efficiency."

[0025] In summary, the closed-loop control method for pollutant emissions from thermal power plants in this invention, through the deep integration of data-driven approaches, intelligent prediction, multi-objective optimization, and closed-loop adaptive control, constructs a high-precision, low-cost, robust, and easily coordinated closed-loop control system for pollutant emissions from thermal power plants. This not only solves many bottleneck problems in existing technologies but also provides key technical support for the green and low-carbon transformation of coal-fired power plants under the "dual carbon" objectives, demonstrating significant engineering application value and broad prospects for promotion. Attached Figure Description

[0026] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0027] Figure 1 This is a schematic diagram of the overall technical solution architecture of a closed-loop control method for pollutant emissions from a thermal power plant according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the multi-step advance prediction and adaptive correction mechanism for pollutant concentration based on long short-term memory neural network in this embodiment of the invention. Figure 3 This is a logical flow diagram of the multi-objective collaborative optimization control strategy in this embodiment of the invention; Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between the smart environmental protection island collaborative management and control center and various subsystems in an embodiment of the present invention. Detailed Implementation

[0028] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0029] Unless otherwise specifically stated, the technical or scientific terms used in the embodiments of this invention should be understood in their ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. The terms "comprising" or "including," as used in the embodiments of this invention, do not limit the shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof mentioned, nor do they exclude the appearance or addition of one or more other different shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof, or the inclusion of these.

[0030] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale, and techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail; however, where appropriate, the illustrated techniques, methods, and apparatus should be considered part of the specification. In all the examples shown and discussed herein, any other specific example may have different values. It should be noted that similar symbols and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0031] In the description of the embodiments of the present invention, the terms "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 the embodiments of the present invention, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in the embodiments of the present invention, as well as the features of different embodiments or examples.

[0032] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0033] Currently, under increasingly stringent regulations on air pollutant emissions from thermal power plants, their denitrification, desulfurization, and dust removal systems suffer from complex and variable operating conditions, lack of coordination between subsystems, lagging control strategies, and insufficient model adaptability. This results in poor stability of nitrogen oxide, sulfur dioxide, and particulate matter emissions, high reagent consumption, and a high risk of ammonia escape, making it difficult to achieve long-term stable compliance with ultra-low emissions under disturbances such as variable loads, coal quality fluctuations, and equipment aging. To address these technical problems, this invention proposes a closed-loop control method for pollutant emissions from thermal power plants that integrates multi-source data-driven approaches, intelligent prediction, multi-objective collaborative optimization, and closed-loop adaptive correction, and applies it to a closed-loop control method for pollutant emissions from thermal power plants.

[0034] refer to Figure 1The overall technical architecture of this invention includes a multi-source data acquisition and fusion platform, an intelligent pollutant concentration prediction model, a multi-objective collaborative optimization control strategy, a closed-loop feedback and adaptive correction module, and a smart environmental protection island collaborative management and control center. The following will provide a detailed engineering explanation of each technical aspect based on the five core steps of the invention, in conjunction with the accompanying drawings.

[0035] In the aforementioned closed-loop control method for pollutant emissions from thermal power plants, step 1 involves constructing a multi-source data acquisition and fusion platform. Specifically, step 1 uses a distributed control system to collect boiler combustion parameters, flue gas composition data, environmental protection facility operating status, and environmental variables in real time. Data cleaning and feature engineering modules are then used to preprocess and standardize the multi-source heterogeneous data, forming a time-series data stream in a unified format. This platform is deployed in the central control room of the thermal power plant and is connected to the sensor networks of key equipment such as the boiler body, selective catalytic reduction system, wet desulfurization system, and electrostatic precipitator via an industrial Ethernet network. See [link to relevant documentation]. Figure 1 .

[0036] The multi-source data acquisition covers boiler main steam pressure, main steam temperature, coal feed rate, primary air volume, secondary air volume, flue gas oxygen content, nitrogen oxide concentration at the inlet and outlet of the selective catalytic reduction system, sulfur dioxide concentration at the inlet and outlet of the wet desulfurization system, flue gas particulate matter concentration, absorber slurry pH value, slurry density, and flue gas temperature parameters, with a sampling frequency of 1 time per second. This high-frequency sampling ensures high-fidelity capture of boiler combustion dynamics and environmental protection facility response characteristics. For example, the main steam pressure signal is input to the DCS analog input card as a 4-20 mA current signal, and stored with 16-bit precision after A / D conversion; flue gas composition data is measured in real time by a laser spectrometer or ultraviolet differential absorption spectrometer, and the raw spectral data is calculated by an internal algorithm to output concentration values, which are transmitted to the data acquisition server via the Modbus TCP protocol. All data streams are timestamped with precision, and the IEEE 1588 precision time protocol is used for plant-wide clock synchronization to ensure strict alignment of data from different sources on the time axis, with time synchronization errors controlled within 1 millisecond.

[0037] The data acquisition module supports seamless integration with plant-level monitoring information systems, historical data databases, and DCS systems. The data transmission link employs a dual-ring network redundancy design, with automatic switching time between primary and backup links less than 50 milliseconds. It also utilizes the AES-256 industrial-grade encryption protocol to prevent data tampering or theft during transmission. Upon arrival at the fusion platform, data first enters a buffer queue with a depth of 300 seconds to handle momentary network congestion.

[0038] Subsequently, the data cleaning and feature engineering module initiates the processing flow. For short-term data gaps caused by sensor malfunctions or communication interruptions, a sliding window method is used for imputation, with a window length of 10 sampling points. Specifically, for a parameter x(t) missing at a certain time t, the system calculates the weighted average of the five valid sampling points before and after it, with the weights decaying exponentially over time, with a decay coefficient of 0.9, thus generating a smooth interpolated value that conforms to physical trends. For long-term missing values ​​exceeding 10 seconds, the system marks the parameter as "unavailable" and masks this feature dimension in subsequent model inputs, while simultaneously triggering a sensor health status alarm.

[0039] The feature engineering module further performs outlier detection. Using a statistical method based on the 3σ principle, the mean μ and standard deviation σ are calculated for each parameter's historical sliding window (3600 seconds). If the current sampled value exceeds the interval [μ-3σ, μ+3σ], it is identified as an outlier. For parameters with non-Gaussian distributions (such as ammonia escape concentration), the Isolation Forest algorithm is used to construct 100 isolation trees. Samples with a path length less than a threshold of 0.6 are identified as outliers. All identified outliers are removed and replaced with the median within the window.

[0040] To reduce the dimensionality of the model input and improve computational efficiency, the module performs principal component analysis (PCA) for dimensionality reduction. The original feature dimension is 60. After standardization, a covariance matrix is ​​constructed, and its eigenvalues ​​and eigenvectors are solved. The system sorts the eigenvalues ​​from largest to smallest, accumulates the contribution rate of each principal component until the cumulative contribution rate exceeds 95%, and finally retains the top 50 principal components as the model input features. These 50-dimensional eigenvectors not only retain more than 95% of the information from the original data but also effectively remove collinearity between sensors and measurement noise.

[0041] Furthermore, segmented feature mapping rules are established for different operating conditions. The system determines the current unit load status in real time: high load (>80% of rated load), medium load (50%-80%), low load (<50%), and start-up / shutdown phases. Each operating condition corresponds to an independent set of feature scaling parameters and principal component basis vectors. For example, under low load conditions, boiler combustion stability is poor, and NO... x When power generation fluctuates drastically, the feature mapping amplifies the weights of key combustion parameters such as the air-coal ratio and oxygen content, making the input features more closely match the dynamic characteristics under low load. Operating condition identification is based on the moving average of the load command and the actual power generation, with a switching lag of 5% to prevent frequent feature mapping switching due to misjudgment of operating conditions.

[0042] After the above processing, the original multi-source heterogeneous data is transformed into a unified format, high-fidelity, low-noise, and condition-adaptive 50-dimensional time-series data stream, which is output to the next stage at a frequency of 1 Hz, providing a solid data foundation for pollutant concentration prediction.

[0043] In the aforementioned closed-loop control method for pollutant emissions from thermal power plants, step 2 involves establishing an intelligent pollutant concentration prediction model. Specifically, step 2, based on the time-series data stream generated in step 1, employs a long short-term memory neural network structure to perform multi-step forward prediction of the concentration change trends of key pollutants such as nitrogen oxides and sulfur dioxide, covering a time domain of 5 to 30 sampling periods. This model is deployed on a high-performance computing server; see [link to relevant documentation]. Figure 2 .

[0044] The Long Short-Term Memory (LSTM) neural network comprises three hidden layers, with 128, 64, and 32 neurons per layer, respectively. The input layer receives a 50-dimensional feature vector sequence, the length of which represents 60 historical sampling points (i.e., data from the past 60 seconds). The output layer generates predicted concentrations of nitrogen oxides (NOx) and sulfur dioxide (SO2) for the next 10 sampling points (i.e., the next 10 seconds), resulting in a total of 20 output nodes (10 NOx predictions and 10 SO2 predictions). x + 10 SO2). The network internally adopts a standard LSTM cell structure, including input gate, forget gate, output gate and cell state. The weight matrices of each gate are learned through the backpropagation algorithm.

[0045] The model training employs the mean squared error loss function, defined as the average of the squared differences between predicted and true values. The optimizer uses the adaptive moment estimation algorithm, with an initial learning rate of 0.001, β1=0.9, and β2=0.999. To prevent overfitting, L2 regularization is introduced during training with a regularization coefficient of 0.0001, and an early stopping mechanism is used: training terminates when the validation set loss no longer decreases for 10 consecutive training epochs. Furthermore, to improve model robustness, adversarial examples are injected into the training data; these are small perturbations conforming to a Gaussian distribution (with a standard deviation of 5% of the original data's standard deviation) superimposed on the original input, forcing the model to learn feature representations insensitive to noise.

[0046] The model employs a sliding time window approach for incremental training. The historical training data window is 30 days long and is updated daily, retaining only the most recent 30 days of high-quality data. An online update mechanism executes every 24 hours, using the newly added 24-hour data to fine-tune the existing model, rather than training from scratch. During fine-tuning, the weights of the first two layers are frozen, and only the third and output layers are updated, significantly reducing computational resource consumption; a single update takes less than 30 minutes.

[0047] The absolute value of the model prediction error is controlled within ±3 milligrams per cubic meter. To achieve this accuracy, a model performance monitoring module is set up in the system to calculate the mean absolute error, root mean square error, and coefficient of determination R in real time within a rolling window (1 hour in length). 2 If the mean absolute error exceeds 3 milligrams per cubic meter for one consecutive hour, or R 2If the concentration falls below 0.95, the system automatically triggers an early warning and prepares to execute the model update process. The predicted output includes not only the concentration value but also a 95% confidence interval for risk assessment of subsequent control strategies.

[0048] refer to Figure 2 The core of this predictive model lies in its multi-step advance capability. For example, when the system detects that the boiler load command is about to drop from 600 MW to 400 MW, the model, based on current combustion parameters and historical load change patterns, predicts the SCR outlet NO for 10 seconds in advance. x The concentration will rise from 40 mg / m³ to 65 mg / m³. This forecast provides the control layer with a sufficient decision window, enabling it to proactively increase the ammonia injection rate before the concentration actually exceeds the limit, thus avoiding a reactive response.

[0049] In the aforementioned closed-loop control method for pollutant emissions from thermal power plants, step 3 involves designing a multi-objective collaborative optimization control strategy. Specifically, step 3 uses the predicted output from step 2 as feedforward information and, in conjunction with a model predictive control framework, dynamically calculates the optimal control sequence for denitrification ammonia injection, desulfurization slurry dosing, and dust removal system adjustment, simultaneously satisfying emission compliance, minimum operating costs, and equipment safety constraints. The logical flow of this strategy is described below. Figure 3 .

[0050] The model predictive control framework employs a quadratic programming solver. Its objective function simultaneously considers the squared error of nitrogen oxide emission concentration deviating from the setpoint, the square of the ammonia injection flow rate adjustment, the squared error of sulfur dioxide emission concentration deviating from the setpoint, and the square of the limestone slurry dosing rate. Hard constraints are imposed, including the ammonia injection valve opening range, the upper and lower limits of the slurry circulation pump frequency, and ammonia slip concentration not exceeding 2.5 mg / m³. This objective function can be expressed as: J=w1 (C NOx C set ) 2 +w2 (Δu NH3 ) 2 +w3 (C SO2 C set ) 2 +w4 (Δu lime ) 2 Where J is the objective function; w1 is the weighting coefficient of the nitrogen oxide emission concentration deviation; C NOx C represents the nitrogen oxide emission concentration predicted by the model. set The setpoint for nitrogen oxide emission concentration; w2 is the weighting coefficient for the change in ammonia injection flow rate; ΔuNH3 w3 is the difference between the current ammonia injection flow rate and the ammonia injection flow rate of the previous cycle; w3 is the weighting coefficient for the sulfur dioxide emission concentration deviation; C SO2 w4 represents the sulfur dioxide emission concentration predicted by the model; w4 is the weighting coefficient for the change in limestone slurry addition rate; Δu lime This is the difference between the current limestone slurry addition acceleration rate and the previous cycle.

[0051] The weighting coefficients w1 to w4 can be dynamically adjusted according to the power plant's operation strategy. The system has three preset operating modes: environmental priority mode (w1=10, w3=10, w2=1, w4=1), economic operation mode (w1=1, w3=1, w2=10, w4=10), and balancing mode (w1=5, w3=5, w2=5, w4=5). Operators can manually select these modes through a human-machine interface, or the system can automatically switch modes based on grid dispatch instructions, electricity price signals, and environmental assessment cycles.

[0052] The system operating boundaries are strictly limited by the following constraints: the ammonia injection valve opening range is 0% to 100%; the slurry circulation pump frequency is limited to 30 Hz to 50 Hz; and the ammonia slip concentration must not exceed 2.5 mg / m³. The ammonia slip concentration is estimated in real time using a soft-sensor model based on the SCR inlet NO₂. x The concentration, ammonia injection rate, flue gas temperature, and catalyst activity factor are constructed, and their output is used as a hard-constrained embedding optimization problem.

[0053] The solver employs a sequential quadratic programming algorithm, completing optimization calculations within 15 seconds. The control strategy optimization cycle is 15 seconds, with each optimization solving the control sequence for the next 60 seconds but only executing the first control variable. This rolling optimization mechanism ensures that the system can both plan ahead and respond promptly to sudden disturbances. For example, when a step-down load is predicted for the next 60 seconds, the optimizer generates a smoothly decreasing ammonia injection sequence, avoiding mechanical shocks caused by large valve openings and closings.

[0054] Under drastic operating conditions (such as a load change rate greater than 5% per minute), the system dynamically shortens the prediction time domain to 30 seconds to enhance response speed; under stable operating conditions (a load change rate less than 1% per minute), it extends to 90 seconds to improve economy. Before the control sequence is executed, it must be verified by a safety check module: checking whether the command change rate exceeds the maximum allowable slope of the actuator (such as the change rate of the ammonia injection valve opening not exceeding 5% per second). If it exceeds the limit, smoothing is performed to ensure equipment safety.

[0055] In the aforementioned closed-loop control method for pollutant emissions from thermal power plants, step 4 involves implementing closed-loop feedback and adaptive correction. Specifically, step 4 compares the actual emission concentration monitoring values ​​with the model prediction values ​​in real time, performs online correction of the prediction model parameters using a Kalman filter algorithm, and dynamically adjusts the control strategy using a rolling optimization mechanism to cope with fluctuations in operating conditions.

[0056] The closed-loop feedback correction employs an extended Kalman filter algorithm to estimate and update the hidden state variables and output layer weight matrix of the Long Short-Term Memory (LSTM) neural network in real time, with a correction cycle of 5 minutes. The system treats the LSTM network as a nonlinear dynamic system, whose state vector contains the cell states and hidden states of all hidden layers, as well as the weight matrix of the output layer. The observation equation is the actual monitored NO... x The process noise covariance matrix Q and the observation noise covariance matrix R are set according to the statistical characteristics of historical data, along with SO2 concentration.

[0057] In each calibration cycle, the extended Kalman filter first linearizes the nonlinear state transition and observation equations of the LSTM and calculates the Jacobian matrix; then, it performs a prediction step, predicting the current state based on the state estimate from the previous cycle and the current input; finally, it performs an update step, correcting the predicted state using the actual observations to obtain the optimal state estimate. During the calibration process, the system synchronously updates the internal state of the LSTM network and fine-tunes the output layer weights to ensure that the predicted output remains highly consistent with the measured values.

[0058] The adaptive correction module also includes an expert rule base. When the prediction deviation exceeds 5 consecutive sampling periods and its absolute value is greater than 5 milligrams per cubic meter, a model reconstruction mechanism is automatically triggered. The expert rule base covers more than 20 typical scenarios, each associated with specific diagnostic logic. For example, when NO... x When predicted values ​​are consistently higher than actual values, accompanied by increased ammonia slip, the system determines that catalyst activity has decreased, automatically suggests reducing the ammonia injection gain coefficient, and triggers the catalyst performance evaluation process. Model reconstruction can be performed through local fine-tuning (updating only the output layer), full retraining (using data from the last 7 days), or switching to a backup model (a pre-trained model specifically for low-activity catalysts) to ensure control continuity.

[0059] In the aforementioned closed-loop control method for pollutant emissions from thermal power plants, step 5 involves constructing a smart environmental protection island collaborative management and control center. Specifically, step 5 integrates the control loops of the three subsystems: denitrification, desulfurization, and dust removal. A unified communication protocol is used to achieve synchronous issuance of control commands and centralized monitoring of operational status, forming a complete closed-loop management and control system. The multi-level interaction relationships and data flow of this center are detailed below. Figure 4 .

[0060] The Smart Environmental Protection Island Collaborative Management and Control Center communicates with the controllers of each subsystem via the Industrial Ethernet protocol, with a data refresh cycle of 100 milliseconds. The center is equipped with dual-machine hot backup servers and employs virtualization technology to achieve seamless fault switching, ensuring a continuous operating reliability exceeding 99.99%. Communication protocols support Modbus TCP and OPC UA, ensuring compatibility with equipment from different manufacturers.

[0061] The center adopts a microservice architecture, with each functional module deployed independently: the data access service is responsible for interfacing with systems such as DCS and CEMS; the model service encapsulates LSTM prediction models and correction algorithms; the optimization engine performs MPC calculations; and the human-machine interface provides visual monitoring and operation. Services communicate asynchronously through message queues (such as Kafka) to ensure system stability under high concurrency.

[0062] The core of collaborative control lies in synchronized commands. When the optimization engine generates new control commands, the center simultaneously sends the commands to the denitrification ammonia injection controller, the desulfurization slurry pump frequency converter, and the electrostatic precipitator high-voltage power controller via multicast, ensuring that the three systems operate in tandem within the same control cycle. For example, when the boiler load suddenly drops by 20%, the system simultaneously issues commands: reduce the ammonia injection valve opening by 15%, reduce the slurry circulation pump frequency by 5 Hz, and lower the electrostatic precipitator secondary voltage by 10 kV, avoiding a chain reaction caused by excessive adjustment of a single system.

[0063] In addition, the system integrates an operating cost and energy efficiency assessment module. This module calculates in real time the pollutant reduction cost per unit of electricity generated, ammonia consumption, limestone consumption, and system power consumption. The cost calculation model comprehensively considers the unit price of reagents (3000 yuan per ton for ammonia and 80 yuan per ton for limestone), electricity price (peak-valley time-of-use pricing), equipment depreciation, and maintenance costs, and uses a dynamic weighted algorithm to assess the overall operating cost. A daily optimization report is automatically generated, including trend charts of key indicators, deviation analysis, energy-saving potential assessment, and operational suggestions, such as "It is recommended to appropriately increase the desulfurization pH setting to 5.8 during the off-peak electricity price period from 22:00 to 06:00, which can reduce limestone consumption by approximately 8%."

[0064] This module also supports collaborative optimization with power generation load commands. Upon receiving the load curve for the next 24 hours from the power grid dispatch, the system performs rolling optimization, adjusting the operating parameters of the environmental protection system in advance. For example, if the load is predicted to decrease by 20% in the next two hours, the system will gradually reduce the ammonia injection rate and slurry circulation pump frequency 15 minutes in advance to avoid reagent waste due to delayed adjustments. Simultaneously, it supports multi-objective optimization scenario simulation, allowing operators to compare cost and emission differences under different modes to aid decision-making.

[0065] In an implementation case involving two 600 MW supercritical coal-fired power generating units, the method of this invention reduced the fluctuation range of nitrogen oxide emission concentration from ±15 mg / m³ to ±5 mg / m³, ammonia slip concentration from 3.5 mg / m³ to 1.8 mg / m³, limestone slurry consumption by 12%, and system response delay from 90 seconds to 30 seconds. Under a 30% deep peak-shaving condition, NO x Emissions remained stable at 35±3 mg / m³, far exceeding national standards. Annual ammonia consumption was reduced by approximately 180 tons and limestone consumption by approximately 2,500 tons, resulting in economic benefits exceeding 4 million yuan; the failure rate of actuators decreased by 35%.

[0066] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A closed-loop control method for pollutant emissions from a thermal power plant, characterized in that, Includes the following steps: Step 1: Construct a multi-source data acquisition and fusion platform. Through a distributed control system, collect multi-source heterogeneous data in real time, including boiler combustion parameters, flue gas composition data, and environmental protection facility operation status. Then, preprocess and standardize the data to form a time-series data stream in a unified format. Step 2: Establish an intelligent prediction model for pollutant concentrations. Based on the time-series data stream, use a long short-term memory neural network to perform multi-step advance prediction of the concentration change trends of nitrogen oxides and sulfur dioxide to generate prediction output. Step 3: Design a multi-objective collaborative optimization control strategy. Use the predicted output as feedforward information and combine it with the model predictive control framework to dynamically find the optimal control sequence of the denitrification, desulfurization and dust removal system under the constraints of emission compliance, minimum operating cost and equipment safety. Step 4: Implement closed-loop feedback and adaptive correction. By comparing the actual emission concentration monitoring values ​​with the model prediction values ​​in real time, the Kalman filter algorithm is used to correct the prediction model parameters online, and the rolling optimization mechanism is used to dynamically adjust the control strategy to cope with operating condition fluctuations. Step 5: Construct a smart environmental protection island collaborative management and control center, integrating the control loops of denitrification, desulfurization, and dust removal subsystems. Through a unified communication protocol, it realizes the synchronous issuance of control commands and centralized monitoring of operating status, forming a closed-loop management and control system for the entire process.

2. The closed-loop control method for pollutant emissions from a thermal power plant according to claim 1, characterized in that, The multi-source data collected in step 1 specifically includes boiler main steam pressure, main steam temperature, coal feed rate, primary air volume, secondary air volume, flue gas oxygen content, nitrogen oxide concentration at the inlet and outlet of the selective catalytic reduction system, sulfur dioxide concentration at the inlet and outlet of the wet desulfurization system, flue gas particulate matter concentration, absorber slurry pH value, slurry density, and flue gas temperature parameters, and the sampling frequency of each parameter is 1 time per second.

3. The closed-loop control method for pollutant emissions from a thermal power plant according to claim 2, characterized in that, The preprocessing in step 1 uses the sliding window method to handle missing data, with a window length of 10 sampling points. Principal component analysis is used to reduce the dimensionality of high-dimensional features, and the principal component features with a cumulative contribution rate greater than 95% after screening are retained as the input of the intelligent prediction model.

4. A closed-loop control method for pollutant emissions from a thermal power plant according to any one of claims 1 to 3, characterized in that, The long short-term memory neural network in step 2 contains three hidden layers, with 128, 64 and 32 neurons in each layer, respectively. The neural network has an input feature dimension of 50, and its output is the predicted concentration of nitrogen oxides and sulfur dioxide for the next 10 sampling points. The training process of the neural network is optimized using the mean squared error loss function and the adaptive moment estimation algorithm is used as the optimizer.

5. The closed-loop control method for pollutant emissions from a thermal power plant according to claim 4, characterized in that, In step 2, the intelligent pollutant concentration prediction model is updated online every 24 hours using the latest operating data to dynamically control the absolute value of the model prediction error within ±3 mg / m³, ensuring that the prediction model can continuously adapt to long-term operating conditions such as equipment aging, coal quality changes, or seasonal environmental fluctuations, and maintain the long-term stability of prediction accuracy.

6. A closed-loop control method for pollutant emissions from a thermal power plant according to any one of claims 1 to 3, characterized in that, In step 3, the model predictive control framework uses a quadratic programming solver. Its optimization objective function simultaneously considers the square error of the nitrogen oxide emission concentration deviating from the set value, the square of the ammonia injection flow rate adjustment, the square error of the sulfur dioxide emission concentration deviating from the set value, and the square of the limestone slurry addition rate. Hard constraints are applied, including the ammonia injection valve opening range, the upper and lower limits of the slurry circulation pump frequency, and the ammonia escape concentration not exceeding 2.5 mg / m³.

7. The closed-loop control method for pollutant emissions from a thermal power plant according to claim 6, characterized in that, The control strategy optimization cycle in step 3 is 15 seconds. Each optimization solves for the optimal control sequence within the next 60 seconds, but only the first control variable of the sequence is executed. In the next cycle, the next optimization is performed based on the updated system state, thus forming a rolling optimization mechanism to balance the forward-looking nature of the control with the ability to respond quickly to disturbances.

8. A closed-loop control method for pollutant emissions from a thermal power plant according to any one of claims 1 to 3, characterized in that, The closed-loop feedback correction in step 4 uses the extended Kalman filter algorithm, treating the long short-term memory neural network as a nonlinear dynamic system. The hidden state variables and output layer weight matrix of the network are estimated and updated in real time to compensate for systematic biases caused by model simplification or insufficient training data. The correction period is 5 minutes.

9. A closed-loop control method for pollutant emissions from a thermal power plant according to claim 8, characterized in that, The adaptive correction in step 4 also includes an expert rule base. When the prediction deviation exceeds 5 consecutive sampling periods and its absolute value is greater than 5 milligrams per cubic meter, the expert rule base automatically triggers the model reconstruction mechanism. The model reconstruction mechanism includes local fine-tuning, full retraining, or switching to a backup model.

10. A closed-loop control method for pollutant emissions from a thermal power plant according to any one of claims 1 to 3, characterized in that, In step 5, the smart environmental protection island collaborative management and control center communicates with the controllers of each subsystem through the industrial Ethernet protocol. The data refresh cycle is 100 milliseconds, and it is equipped with a redundant hot backup server to ensure the reliability of continuous system operation. The method also includes establishing an operating cost and energy efficiency assessment module to calculate in real time the pollutant emission reduction cost per unit of power generation, reagent consumption, and system power consumption, and to generate a daily operation optimization report to support collaborative optimization with power generation load commands.

Images