Flue gas waste heat-water recovery collaborative pollutant emission reduction system and regulation method

Abstract

The application discloses a flue gas waste heat-water recovery and pollutant emission reduction system and a regulation and control method, and belongs to the technical field of industrial waste heat recovery and automatic control. The system comprises a desulfurization tower, a slurry circulating subsystem, a flash tank, a heat pump unit and an intelligent regulation and control system. The method acquires historical operation data and pre-processes the same, constructs a double-layer coupled network prediction model of a Transformer-GRU, and outputs a water supply temperature prediction value. With temperature tracking precision and control amount change amplitude as targets, a control variable and a temperature safety constraint are introduced, and a spiral envelope element heuristic algorithm is used to solve the optimal driving steam regulating valve opening degree and slurry circulating pump rotating speed combination. The system can remove more than 40% of condensable particulate matters, and the SO3 removal efficiency is more than 92%. In the heating season, the waste heat recovery power is 36 MW, and 100,000 tons of condensate water is recovered. The method has high prediction precision and strong stability, and effectively overcomes the influence of the volume inertia of the flash tank and the multi-stage heat exchange lag of the heat pump.

Description

Technical Field

[0001] This invention belongs to the field of industrial waste heat recovery and automatic control technology, specifically relating to a flue gas waste heat-water recovery synergistic pollutant emission reduction system and control method. Background Technology

[0002] Currently, the flue gas from coal-fired combustion plants, after wet desulfurization, contains a large amount of low-grade waste heat. After wet desulfurization, the flue gas is in a saturated wet state at 50-55℃, and neither the latent heat of water vapor nor the sensible heat of the flue gas is effectively recovered. Heat loss from exhaust accounts for more than 10% of total primary energy consumption, resulting in serious energy waste. Simultaneously, the condensable particulate matter and dust carried by the flue gas, after emission, condense into fine particulate matter in the atmosphere, becoming a key factor in smog formation. Therefore, promoting deep recovery of waste heat from flue gas and synergistic treatment of pollutants is of great significance for improving energy efficiency and controlling compound pollution.

[0003] However, achieving these goals faces three major technical challenges. First, deep system integration and operational stability. Recovering latent heat and water resources requires embedding new devices into the existing desulfurization system, coordinating multiple balances of water, heat, and mass, and avoiding fluctuations in downstream processes such as desulfurization efficiency and dust removal due to changes in makeup water or heat recovery. Second, adaptability to wide load conditions and peak-shaving flexibility. When power units participate in deep peak shaving, flue gas flow, temperature, and composition fluctuate drastically. The device must still operate efficiently and stably under low loads without increasing unit losses or affecting environmental compliance. Third, multi-objective synergy and comprehensive energy efficiency improvement. It is necessary to balance reducing energy consumption and drag, while simultaneously removing multiple pollutants such as condensable particulate matter, SO2, and dust to achieve synergistic carbon emission reduction.

[0004] Existing single-function retrofits, such as simply adding low-temperature economizers or condensation dehumidification devices, are insufficient to achieve a unified approach to stabilizing the desulfurization system, adapting to varying loads, and synergistically removing multiple pollutants. Therefore, there is an urgent need to develop a system and control method that deeply couples waste heat recovery, water conservation, and multi-pollutant removal to achieve highly coordinated and energy-efficient stable operation of multiple objectives under complex and variable operating conditions.

[0005] Reference Figure 1 As shown, the desulfurization slurry flash evaporation waste heat recovery system refers to a complete process system in which desulfurization slurry drawn from the primary desulfurization tower is evaporated under negative pressure in a flash tank, and the resulting waste steam is recovered by a heat pump unit and used for urban heating. This system achieves the cascade recovery of the latent heat of condensation of desulfurization flue gas and the sensible heat of the slurry, resulting in significant energy-saving and carbon-reduction benefits.

[0006] The outlet water supply temperature of a heat pump unit is a key parameter for evaluating the system's waste heat recovery rate and operational economy. This temperature exhibits characteristics such as large delay, nonlinearity, and high coupling. Its dynamic response is simultaneously affected by the coupled changes of multiple variables, including flash steam volume, driving steam parameters, heating return water temperature, and desulfurization load. The evaporation of slurry in the flash tank has significant volumetric inertia and heat transfer hysteresis effects. The influence of driving steam regulation can only be reflected after three stages of heat exchange: generator, condenser, and absorber. The superposition of these two types of hysteresis factors can easily lead to overshooting and oscillation of the water supply temperature.

[0007] Currently, most flash waste heat recovery systems use the PID method to regulate the supply water temperature. However, under complex operating conditions such as fluctuating desulfurization flue gas load and changing heating network demand, PID controllers struggle to achieve precise and stable temperature control. In actual operation, the supply water temperature setpoint is often conservatively maintained at a low level, resulting in insufficient utilization of the system's waste heat recovery capacity, high steam consumption, and hindering further improvements in system energy efficiency.

[0008] Therefore, it is necessary to construct a predictive control method that can comprehensively utilize the multi-source operating status and the dynamic characteristics of water supply temperature to achieve stable regulation and precise control of water supply temperature under complex operating conditions.

[0009] Currently, most flash evaporation waste heat recovery systems use the PID method to regulate the water supply temperature. For example, Chinese invention patent CN202010359001.0 discloses a waste heat and moisture recovery system for flue gas from a coal-fired unit. This system uses a flash tank to maintain the balance of gas outlet parameters, but it does not construct a predictive model for the large delay characteristic of the water supply temperature, and the response lag problem remains unresolved. Chinese invention patent CN202111435985.7 discloses a multi-functional, multi-stage flash evaporation combined heat pump system for the comprehensive utilization of desulfurization tower solutions. This system achieves multi-functional coupling of flue gas waste heat recovery, desulfurization wastewater concentration, and district heating; however, its control method lacks system design for decoupling multiple variables and dynamic feedforward, and the temperature control accuracy under multi-objective operating conditions needs improvement.

[0010] However, under complex operating conditions such as fluctuating flue gas load and changing heating network demand, PID controllers struggle to achieve precise and stable temperature control. In actual operation, to avoid overshoot risks, the water supply temperature setpoint is often conservatively maintained at a low level, resulting in underutilization of the system's waste heat recovery capacity, high steam consumption, and hindering further improvements in system energy efficiency. Therefore, there is an urgent need to develop a predictive control method that can comprehensively utilize multi-source operating status information and the dynamic characteristics of water supply temperature to achieve stable regulation and precise control of water supply temperature under complex operating conditions, thereby fully releasing the system's waste heat recovery potential. Summary of the Invention

[0011] The purpose of this invention is to provide a flue gas waste heat-water recovery synergistic pollutant emission reduction system and control method to solve the problems of low accuracy and stability, weak waste heat recovery capacity, and high driving steam consumption when adjusting water supply temperature using the existing PID method.

[0012] To achieve the above objectives, the technical solution of the present invention is as follows: This invention relates to a flue gas waste heat-water recovery synergistic pollutant emission reduction system, comprising a desulfurization tower, a slurry circulation subsystem, a flash tank, a vacuum pump unit, a heat pump unit, a condensate recovery subsystem, and an intelligent control system. The slurry circulation subsystem includes a slurry circulation pump, a slurry circulation pipeline, and a return pump. A slurry circulation pump is installed in the slurry circulation pipeline on the third layer of the desulfurization tower, using this pump to guide the slurry to the flash tank for negative pressure evaporation. An outlet regulating valve is installed on the circulation pipeline between the slurry circulation pump and the flash tank to control the amount of slurry drawn to the flash tank. The resulting low-temperature desulfurization slurry... The liquid is pumped back to the third-layer slurry circulation pipeline of the desulfurization tower by the return pump and sprayed into the tower to continue the desulfurization reaction; the exhaust steam generated in the flash tank enters the heat pump unit, and a steam regulating valve is installed on the pipeline between the flash tank and the heat pump unit to control the exhaust steam output; driven by externally supplied superheated steam, the heat pump unit completely absorbs the waste heat of the exhaust steam generated by flash evaporation and converts it into a heat source to achieve waste heat recovery; the exhaust steam after heat extraction condenses into water in the heat pump evaporator, and is then collected by the condensate recovery subsystem and reused as makeup water for the desulfurization tower or as boiler makeup water. The intelligent control system includes: The data acquisition module is used to acquire historical operating data of the flue gas waste heat-water recovery and pollutant emission reduction system; The data processing module is used to preprocess historical operational data to obtain the feature set of the prediction model; The water supply temperature prediction module is a prediction model embedded with a Transformer-GRU dual-layer coupled network. It includes a feature extraction layer and a temporal capture layer. The feature extraction layer uses a Transformer encoder, and the temporal capture layer uses a dual-layer GRU network. It is used to input the feature set into the prediction model and output the predicted water supply temperature value. The collaborative optimization and regulation module is used to construct an optimization problem based on the predicted water supply temperature, with the objective functions being temperature tracking accuracy and the magnitude of control quantity changes. It introduces control variable constraints and temperature safety constraints. The helical envelope meta-heuristic algorithm is used to solve the optimal combination of the driving steam regulating valve opening and the slurry circulation pump speed regulation. The driving steam regulating valve is the main control variable, and the slurry circulation pump speed or the outlet regulating valve opening is the secondary control variable, thus performing collaborative regulation.

[0013] This invention also relates to a control method for the above-mentioned flue gas waste heat-water recovery synergistic pollutant emission reduction system, which includes the following steps: S1. Obtain historical operational data of the flue gas waste heat-water recovery synergistic pollutant emission reduction system; S2. Preprocess historical operating data to obtain the feature set of the prediction model; S3. Construct a prediction model using a Transformer-GRU two-layer coupled network, including a feature extraction layer and a temporal capture layer. The feature extraction layer uses a Transformer encoder, and the temporal capture layer uses a two-layer GRU network. Input the feature set into the prediction model and output the predicted water supply temperature. S4. Based on the predicted water supply temperature, an optimization problem is constructed with temperature tracking accuracy and control quantity change range as objective functions, and control variable constraints and temperature safety constraints are introduced. The optimal combination of the opening degree of the driving steam regulating valve and the speed regulation of the slurry circulation pump is solved by using the driving steam regulating valve as the main control variable and the speed of the slurry circulation pump as the secondary control variable to perform coordinated regulation.

[0014] Preferably, the historical operating data in S1 includes the following key variables: slurry side parameters, steam side parameters, flue gas side parameters, and heating side parameters.

[0015] Preferably, the preprocessing of historical operating data in S2 includes data cleaning, time delay correction, construction of derived features, and feature screening of historical operating data. The specific steps are as follows: S2.1. Use the absolute deviation of the middle value in the sliding window method to identify and correct outliers in historical running data, and replace outliers with the weighted average of adjacent data; S2.2. The cross-correlation method is used to estimate the pure time lag between the historical operating data after cleaning and the water supply temperature; S2.3. Calculate the difference between the current opening degree of the driving steam regulating valve and the minimum allowable opening degree of the system, the real-time deviation between the measured value and the set value of the water supply temperature, and the difference between the slurry temperature at the outlet of the first-stage flash evaporator and the set value of the water supply temperature, and then form the feature set of the water supply temperature prediction model.

[0016] Preferably, in step S2.3, after calculating the difference between the current opening degree of the driving steam regulating valve and the minimum allowable opening degree of the system, the real-time deviation between the measured value and the set value of the water supply temperature, and the difference between the slurry temperature at the first-stage flash outlet and the set value of the water supply temperature, the sliding window method is used to statistically analyze the statistical characteristics of key variables, candidate variables are selected from the key variables, and the maximum information coefficient method is used to calculate the correlation strength between each candidate variable and the water supply temperature, eliminating low-correlation and redundant variables, and simplifying the input feature dimensions of the prediction model.

[0017] Preferably, in step S3, the feature extraction layer captures the nonlinear coupling relationship and long-term dependency characteristics among variables in the feature set, and the time-series capture layer outputs the predicted water supply temperature based on the nonlinear coupling relationship and long-term dependency characteristics. The specific steps are as follows: S3.1. Calculate the query, key, and value of the Transformer encoder using the following formulas: , , , in, Indicates the number of the feature set. The feature set represents the historical dataset of the candidate variables. Indicates a query. The key representing the feature set, Represents the value of the feature set. , and These are the weight matrices for the query, key, and value, respectively. S3.2. Calculate the attention score between each element and all other elements. Different elements correspond to different types of historical operating data. The attention score quantifies the relative importance of each element at different historical moments when predicting water supply temperature. The formula for calculating the attention score is as follows: , in, This represents the attention score. (·) represents the normalized exponential function. Indicates the dimension of the key vector; Alternatively, the attention mechanism can employ a multi-head attention mechanism, with the attention score calculated using the following formula: , , in, This represents the overall operation of the multi-head attention mechanism. This means concatenating the outputs of all attention heads. This represents the output weight matrix. Indicates the input number The output of the attention head for each feature set. This represents the total number of heads of attention. These represent the results of the query, key, and value after a linear transformation, respectively. S3.3. Calculate the gating value of the GRU network reset gate. The calculation formula is as follows: , in, This indicates that the door's gate control value is being reset. It is the Sigmoid activation function. It is the weight matrix of the reset gate. It is the hidden state of the previous time step. It is the first The input vector at each time point corresponds to the input feature set; S3.4. Based on the gate control value of the reset gate, calculate the candidate hidden layer state using the following formula: , in, These are candidate hidden layer states. surface Activation function It is the weight matrix of the candidate hidden layer states; S3.5. Calculate the update gate. The calculation formula is as follows: , in, This indicates updating the value of the gate. It is the Sigmoid activation function. It is the weight matrix of the updated gate; S3.6. Calculate the final hidden layer state using the following formula: , in, This represents the final hidden layer state, i.e., the hidden layer state at the current time step.

[0018] Preferably, the objective function in S4 is: , in, Describe the objective function. Indicates the length of the prediction time domain. Indicates the length of the control time domain. Indicates the current t The future after a moment t + i Predicted water supply temperature at all times Indicates the preceding t The future after a moment t + i The set value of the water supply temperature at all times. and They represent the coefficients of the penalty term, and They represent the current t The future after a moment t + j The adjustment increments of the primary and secondary control variables at any given time.

[0019] Preferably, the constraints in S4 include: , in, and These represent the lower and upper limits of the manipulated variable, respectively. Indicates the current t The future after a moment t + j The control variables at any given time, including primary control variables and secondary control variables. Indicates the current t The future after a moment t + j The change in the manipulative variable at any given time. This represents the maximum permissible change between adjacent time points. express t Predicted water supply temperature at all times and These represent the lower and upper limits of the safe water supply temperature, respectively.

[0020] Preferably, the specific steps in S4 for solving the optimal combination of the opening degree of the driving steam regulating valve and the speed regulation amount of the slurry circulation pump using the spiral envelope meta-heuristic algorithm are as follows: S4.1. Initialize the position and calculate the objective function to determine the optimal position; S4.2. Update the coefficient vector A sum coefficient vector C The updated formula is: , , in, A random number within the interval [0,1]. The convergence factor decreases linearly from 2 to 0 as the number of optimization iterations increases, and the decreasing formula is: , in, The initial maximum number of times, This represents the current iteration number; S4.3. Set the random selection probability. When the random selection probability... p When the value is greater than 0.5, proceed to step S4.4.1 and update the position using a spiral envelope tracking strategy; when the probability is selected... p ≤0.5 and When the probability is selected, proceed to S4.4.2 and update the position using a contraction envelope cooperative approximation strategy; p ≤0.5 and When the time comes, proceed to S4.4.3 and update the position using a stochastic reference solution exploration strategy; S4.4.1. The position is updated using a spiral envelope tracking strategy. The position update formula is as follows: , S4.4.2. The position is updated using a contraction envelope cooperative approximation strategy. The position update formula is: , S4.4.3. The position is updated using a stochastic reference solution exploration strategy. The position update formula is as follows: , in, For the current moment t The individual position, For the current moment t The optimal position, For the current moment t The next moment's position of the individual, where L is a random number in the range [-1, 1]. The logarithmic spiral shape constant is... This is a random reference solution; S4.5. Calculate the objective function, update the optimal position and the corresponding objective value. If the optimal objective value is not greater than the set threshold, output the optimal position. If the optimal objective value is greater than the set threshold and the number of iterations is less than the set number of iterations, return to S4.2. If the optimal objective value is greater than the set threshold and the number of iterations reaches the set number of iterations, output a non-convergence signal.

[0021] Compared with the prior art, the technical solution provided by this invention has the following advantages: (1) This invention addresses the problem of low accuracy in predicting water supply temperature caused by large delays, strong coupling, and multiple disturbances in flash-absorption heat pump coupled systems by constructing a Transformer-GRU dual-layer coupled network prediction model. This model uses the maximum information coefficient method to select key features from a large number of variables on the slurry, steam, and flue gas sides, and utilizes cross-correlation delay correction to eliminate the pure lag effects of flash tank volume inertia and multi-stage heat exchange in the heat pump. It employs a Transformer encoder to capture long-term nonlinear dependencies among multiple variables and a GRU network to extract short-term dynamic evolution patterns of the water supply temperature, fully leveraging the complementary advantages of the two networks. Actual operating data shows that the prediction model achieves a mean absolute error as low as 0.1139℃ and a coefficient of determination as high as 0.9854 on the independent test set, significantly outperforming single-structure models such as LSTM, GRU, and Transformer. It maintains high-precision and stable prediction even under complex operating conditions such as flash pressure fluctuations and frequent changes in heat network load, providing a reliable basis for subsequent optimization control.

[0022] (2) This invention addresses the problem of overshooting and oscillation in water supply temperature caused by the fast response of the driving steam regulating valve and the large lag in the regulation of the slurry circulation pump. It proposes a control strategy that combines primary and secondary control variables with a real-time solution using a helical envelope element heuristic algorithm. The driving steam regulating valve is set as the primary control variable, and the slurry circulation pump outlet regulating valve is set as the secondary control variable. A constrained multi-objective optimization problem is solved in real-time using a helical envelope element heuristic algorithm. A collaborative regulation logic is designed: when the primary valve approaches its upper saturation limit, the secondary valve opening is increased to increase flash steam production, and the optimal primary valve opening is recalculated; when the primary valve approaches its lower limit, the secondary valve opening is decreased, and the solution is recalculated. This proactively coordinates the flash steam production with the heat pump's driving heat, fundamentally resolving the temperature overshoot and oscillation caused by the evaporation inertia of the flash tank and the lag in multi-stage heat exchange of the heat pump. Actual operating data shows that under rapid load changes and disturbance conditions, the standard deviation of the water supply temperature is reduced by more than 50% compared to traditional PID control, while effectively avoiding the thermal shock to the generator caused by frequent large-scale movements of the driving steam regulating valve.

[0023] (3) To address the problems of large fluctuations in water supply temperature, significant steady-state residual error, and high steam consumption under complex operating conditions, this invention constructs a collaborative optimization control method for driving steam and slurry circulation based on Transformer-GRU prediction. Under steady-state load conditions, the standard deviation of water supply temperature decreased from 1.169℃ in the actual PID control to 0.871℃, the maximum deviation was compressed from 8.014℃ to 4.236℃, and the mean was accurately restored from 55.63℃ to 59.65℃, eliminating a steady-state residual error of 4.37℃. Under rapid load change conditions, the standard deviation of predictive control was only 0.849℃, and steam consumption was reduced by 5.1%. Under disturbance conditions, the standard deviation of predictive control was only 0.829℃, and steam consumption was reduced by 5.2%. Through two-sample t-test and F-test, the differences in mean temperature and variance between predictive control and actual PID control reached extremely significant levels (p<0.001). Meanwhile, precise temperature control effectively avoids the risks of heat pump solution crystallization and pipeline overheating caused by excessively high water supply temperature, as well as the problem of reduced heating quality caused by excessively low temperature. It realizes safe, efficient, and long-term coordinated operation between the waste heat recovery system and the main desulfurization process, and has good industrial application value.

[0024] (4) The flash evaporation system of this invention not only cools the slurry, but also extracts relatively clean water vapor from the circulating slurry and condenses and discharges it, effectively avoiding the water quality deterioration problem caused by independent condensation outside the tower, while reducing the cold source grade requirements. This technology does not require an additional tower body or increase the flue resistance, and achieves a condensable particulate matter removal efficiency of over 40%, an SO3 removal efficiency of over 92%, an SO2 removal efficiency of over 93%, and a comprehensive filterable particulate matter removal efficiency of over 67%, realizing deep synergistic purification of multiple pollutants in flue gas.

[0025] (5) The control method of the flue gas waste heat-water recovery synergistic pollutant emission reduction system combined with the intelligent control system of the present invention is applied to a 350MW coal-fired cogeneration unit. During the heating season, the rated waste heat recovery power is 36MW, the waste heat recovered is about 300,000 GJ, saving more than 10,000 tons of standard coal, reducing CO2 emissions by about 30,000 tons, and recovering 100,000 tons of high-quality condensate. Attached Figure Description

[0026] Figure 1 A schematic diagram of the hardware equipment structure of a flue gas waste heat-water recovery and pollutant emission reduction system; Figure 2 Flowchart of the regulation method for a flue gas waste heat-water recovery and synergistic pollutant emission reduction system; Figure 3 A flowchart illustrating the helical envelope meta-heuristic algorithm for solving the problem; Figure 4 A comparison chart of predicted and measured values ​​of outlet water supply temperature for heat pump units using five models: RNN, LSTM, GRU, Transformer, and Transformer-GRU. Figure 5 A graph showing the correspondence between predicted and measured values; Figure 6 This is a diagram illustrating the load operation stability of the intelligent control system of the present invention. Figure 7 This is a diagram illustrating the prediction accuracy of the intelligent control system of the present invention. Detailed Implementation

[0027] To further understand the content of this invention, the invention will be described in detail with reference to the embodiments. The following embodiments are used to illustrate the invention, but are not intended to limit the scope of the invention.

[0028] Example: Refer to Appendix Figure 1 As shown, the flue gas waste heat-water recovery synergistic pollutant emission reduction system involved in this embodiment includes a desulfurization tower, a slurry circulation subsystem, a flash tank, a vacuum pump unit, a heat pump unit, a condensate recovery subsystem, and an intelligent control system. The slurry circulation subsystem includes a slurry circulation pump, a slurry circulation pipeline, and a return pump. A slurry circulation pump is installed in the slurry circulation pipeline on the third layer of the desulfurization tower. Using the existing slurry circulation pump in the third layer of the desulfurization tower, the slurry is led to the flash tank through an opening in the third layer slurry circulation pipeline. The slurry then enters the flash tank for negative pressure evaporation. An outlet regulating valve is installed on the circulation pipeline between the slurry circulation pump and the flash tank to control the amount of slurry led to the flash tank. The low-temperature desulfurization slurry after flash evaporation is pumped back to the slurry circulation pipeline in the third layer of the desulfurization tower by a newly added slurry return pump, and then sprayed into the tower to continue the desulfurization reaction. The exhaust steam generated in the flash tank will become a supplementary heat source for urban heating. It enters the heat pump unit from the flash tank. A steam regulating valve is installed on the pipeline between the flash tank and the heat pump unit to control the output of exhaust steam. Driven by externally supplied superheated steam, the heat pump unit will completely absorb the waste heat of the exhaust steam generated by flash evaporation and convert it into a heat source for heating the return water of urban heating, raising the temperature of the return water to 60°C. This saves some of the superheated steam used in the original heat exchange station system and realizes waste heat recovery. The exhaust steam after heat extraction condenses into water in the heat pump evaporator. After being collected by the condensate recovery subsystem, it can be reused as process water for the desulfurization system or treated and used as boiler feedwater, realizing high-quality recovery and cascade utilization of water resources.

[0029] The intelligent control system of this system includes: The data acquisition module is used to acquire historical operating data of the sulfur slurry flash evaporation waste heat recovery system; The data processing module is used to preprocess historical operational data to obtain the feature set of the prediction model; The water supply temperature prediction module is a prediction model embedded with a Transformer-GRU dual-layer coupled network. The prediction model includes a feature extraction layer and a temporal capture layer. The feature extraction layer adopts a Transformer encoder, and the temporal capture layer adopts a dual-layer GRU network. It is used to input the feature set into the prediction model and output the predicted water supply temperature value. The collaborative optimization and regulation module is used to construct an optimization problem based on the predicted water supply temperature, with the objective functions being temperature tracking accuracy and the magnitude of control quantity changes. It introduces control variable constraints and temperature safety constraints. The helical envelope meta-heuristic algorithm is used to solve the optimal combination of the driving steam regulating valve opening and the slurry circulation pump speed regulation. The driving steam regulating valve is the main control variable, and the slurry circulation pump speed or the outlet regulating valve opening is the secondary control variable, thus performing collaborative regulation.

[0030] See attached document Figure 2 As shown, the present invention relates to a method for predicting and controlling the water supply temperature of a desulfurization slurry flash evaporation waste heat recovery system, comprising the following steps: S1. Obtain historical operating data of the sulfur slurry flash evaporation waste heat recovery system. This historical operating data includes slurry-side parameters, steam-side parameters, flue gas-side parameters, and heating-side parameters. The heating-side parameters are used to compare the predicted model with actual values. The slurry-side parameters include flash slurry volume, slurry inlet temperature, and slurry outlet temperature. The steam-side parameters include first-stage flash evaporation pressure, first-stage flash evaporation steam volume, second-stage flash evaporation pressure, second-stage flash evaporation steam volume, and total flash steam production. The flue gas-side parameters include the desulfurization tower inlet flue gas temperature and flue gas flow rate. The heating-side parameters include the driving steam flow rate, heat network return water temperature, and heat network circulating water flow rate. The slurry-side parameters, steam-side parameters, and flue gas-side parameters are the main operating variables affecting the outlet water supply temperature of the heat pump unit, as shown in Table 1. Table 1: Historical Operational Data Parameter Table

[0031] It should be noted that the MI values ​​of operating variables such as desulfurization tower liquid level, slurry pH value, and circulating pump current are very small compared to the outlet water supply temperature of the heat pump unit, and therefore are not used as input variables for the prediction model.

[0032] S2. To ensure precise control of the outlet water supply temperature of the heat pump unit, a comprehensive and systematic analysis of the multi-dimensional disturbance factors on the slurry side, steam side, and flue gas side is required. Based on this analysis, appropriate input feature optimization measures should be implemented. Therefore, this step involves preprocessing historical operating data, calculating the main variables that may affect the water supply temperature in the table, and removing low-correlation and redundant variables. This simplifies the input features of the prediction model, improves the model's training efficiency and prediction accuracy, and obtains the feature set of the prediction model. The specific steps are as follows: S2.1. The sliding window middle absolute deviation method is used to identify and correct outliers in the slurry side parameters, steam side parameters, and flue gas side parameters in the historical operation data, and the outliers are replaced with the weighted average of adjacent data. The sliding window middle absolute deviation method is a conventional method in the art. S2.2. The cross-correlation method is used to estimate the pure time delay between the historical operating data after cleaning and the water supply temperature, wherein the cross-correlation method for estimating the pure time delay is a conventional method in this field; S2.3. Calculate the difference between the current opening degree of the driving steam regulating valve and the minimum allowable opening degree of the system to reflect the available operating margin of the steam-side regulation; calculate the real-time deviation between the measured value and the set value of the supply water temperature to quantify the instantaneous control accuracy of the system; calculate the difference between the slurry temperature at the outlet of the first-stage flash evaporator and the set value of the supply water temperature to construct a feedforward index reflecting the evolution trend of flash evaporation steam generation capacity and heat pump heat source quality; use the sliding window method to statistically analyze the statistical characteristics of key variables, then select candidate variables from the key variables, and use the maximum information coefficient method to calculate the correlation strength between each candidate variable and the supply water temperature, eliminating low-correlation and redundant variables to form the feature set of the supply water temperature prediction model, thus simplifying the input feature dimensions of the prediction model. The sliding window method and the maximum information coefficient method are commonly used methods in this field.

[0033] The maximum information coefficient method can stably capture the overall dependence between features in complex nonlinear and multivariate coupling backgrounds. It is applicable to water supply temperature, a key operating parameter that also has nonlinear, strong coupling and large delay characteristics, and can provide a reliable feature selection basis for subsequent predictive control models.

[0034] S3. Construct a prediction model using a Transformer-GRU dual-layer coupled network. The prediction model includes a feature extraction layer and a temporal capture layer. The feature extraction layer uses a Transformer encoder to capture the nonlinear coupling relationship and long-term dependency characteristics between variables in the feature set. The temporal capture layer uses a dual-layer GRU network to output the predicted water supply temperature based on the nonlinear coupling relationship and long-term dependency characteristics.

[0035] The Transformer encoder achieves efficient modeling of dependencies between elements at any position in sequence data through a self-attention mechanism. It maps the multi-source feature sequences from the slurry, steam, and flue gas sides into hidden representations rich in contextual information. Through a multi-layer self-attention mechanism and a feedforward neural network, it extracts deep coupling features from flash steam generation, heat pump heat exchange, and heating network load. This effectively captures multi-timescale features formed by the volumetric inertia of the flash tank, the lag in driving steam regulation, and the periodic fluctuations in heating network demand during changes in supply water temperature. Simultaneously, it analyzes the complex nonlinear relationships between desulfurization slurry circulation, exhaust steam recovery, and heat output, and then calculates the correlation strength of multiple variables on the slurry, steam, and flue gas sides over any time span in parallel. This not only overcomes the inherent limitations of traditional sequence models but also demonstrates powerful modeling capabilities in various complex tasks. Its self-attention mechanism adaptively focuses on information such as slurry temperature, flash pressure, and driving steam state at different historical moments, accurately identifying key factors affecting changes in supply water temperature and their cross-time-series dependencies.

[0036] The encoder in the Transformer model can fully analyze and integrate the nonlinear coupling information from multiple input sources on the slurry, steam, and flue gas sides. It is suitable for handling prediction tasks involving multiple stages and variables, such as flash tank negative pressure evaporation, multi-stage heat exchange in heat pump units, and fluctuations in heating network load, as seen in this system. The decoder is responsible for decoding all encoded information and gradually generating the output sequence in an autoregressive manner. However, its design is geared towards sequence-to-sequence generation tasks and is not suitable for directly handling single-step or multi-step prediction tasks like water supply temperature, which require high real-time performance and accuracy. It struggles to fully reflect the short-term states and trends of modeling features such as instantaneous fluctuations in flash steam production and the response of driving steam regulation. Furthermore, the decoder's generic structure introduces additional computational modules such as cross-attention, leading to unnecessary complexity and computational burden, affecting the model's prediction response speed when deployed online in the control system. Therefore, the Transformer model needs improvement, abandoning the standard decoder structure to better meet the practical requirements of high accuracy and low latency in water supply temperature prediction. This invention uses a GRU to replace the Transformer decoder, which offers significant advantages. First, the GRU structure is relatively simple, and its computational complexity is lower compared to complex self-attention mechanisms, which can significantly reduce the resource consumption of the model in the embedded control system and improve online computing efficiency. Second, the GRU can effectively capture the recent state and short-term trends in the water supply temperature sequence, which are composed of rapid fluctuations in flash pressure, delays in driving steam regulation, and periodic changes in heating network load, thus better reflecting the dynamic evolution characteristics of system operating data.

[0037] The specific steps for inputting the feature set into the prediction model and outputting the predicted water supply temperature include: S3.1. Calculate the query, key, and value of the Transformer encoder using the following formulas: , , , in, Indicates the number of the feature set. The feature set represents the historical dataset of candidate variables. Indicates a query. The key representing the feature set, Represents the value of the feature set. , and These are the weight matrices for the query, key, and value, respectively. S3.2. Calculate the attention score between each element and all other elements. Different elements correspond to different types of historical operating data. The attention score quantifies the relative importance of each element at different historical moments when predicting water supply temperature. The formula for calculating the attention score is as follows: , in, This represents the attention score. (·) represents the normalized exponential function. Indicates the dimension of the key vector; Alternatively, the attention mechanism can employ a multi-head attention mechanism, with the attention score calculated using the following formula: , , in, This represents the overall operation of the multi-head attention mechanism. This means concatenating the outputs of all attention heads. This represents the output weight matrix. Indicates the input number The output of the attention head for each feature set. This indicates the total number of heads of attention. These represent the results of the query, key, and value after a linear transformation, respectively. S3.3. Calculate the gating value of the GRU network reset gate. The calculation formula is as follows: , in, This indicates that the door's gate control value is being reset. It is the Sigmoid activation function. It is the weight matrix of the reset gate. It is the hidden state of the previous time step. It is the first The input vector at each time point corresponds to the input feature set; S3.4. Based on the gate control value of the reset gate, calculate the candidate hidden layer state using the following formula: , in, These are candidate hidden layer states. surface Activation function It is the weight matrix of the candidate hidden layer states; S3.5. Calculate the update gate. The calculation formula is as follows: , in, This indicates updating the value of the gate. It is the Sigmoid activation function. It is the weight matrix of the updated gate; S3.6. Calculate the final hidden layer state using the following formula: , in, This represents the final hidden layer state, i.e., the hidden layer state at the current time step.

[0038] S4. Based on the predicted water supply temperature, an optimization problem is constructed with temperature tracking accuracy and the magnitude of control variable change as the objective functions, and constraints on manipulated variables and temperature safety constraints are introduced; the objective function is: , in, Describe the objective function. Indicates the length of the prediction time domain. Indicates the length of the control time domain. Indicates the current t The future after a moment t + i Predicted water supply temperature at all times Indicates the preceding t The future after a moment t + i The set value of the water supply temperature at all times. and They represent the coefficients of the penalty term, and They represent the current t The future after a moment t + j The adjustment increments of the primary and secondary control variables at any given time.

[0039] The constraints include: , in, and These represent the lower and upper limits of the controllable variables, respectively, covering the opening degree of the steam regulating valve and the speed of the slurry circulation pump; Indicates the current t The future after a moment t + j The control variables at any given time, including primary control variables and secondary control variables, are determined by equipment physical limitations and process safety requirements; Indicates the current t The future after a moment t + j Changes over time This indicates the maximum permissible change between adjacent moments, preventing frequent and large-scale valve and pump operations from impacting the vacuum stability of the flash tank and the thermal stress of the heat pump generator. express t Predicted water supply temperature at all times and These represent the lower and upper limits of the safe water supply temperature, respectively. The lower limit is determined by the minimum requirements of heating network users for heating quality, while the upper limit is jointly constrained by the heat pump unit's solution crystallization protection and pipeline thermal stress limits.

[0040] A spiral envelope meta-heuristic algorithm is employed to find the optimal combination of the steam control valve opening and the slurry circulation pump speed regulation. With the steam control valve as the primary control variable and the slurry circulation pump speed as the secondary control variable, coordinated regulation is implemented. The spiral envelope meta-heuristic algorithm first randomly initializes multiple sets of step responses within the feasible solution space, including lag time, amplification factor, and time constant. Then, it continuously updates the solution position through three core search mechanisms. The solution process is as follows: Figure 3 As shown, it specifically includes: S4.1. Initialize the position and calculate the objective function to determine the optimal position; S4.2. Update the coefficient vector A sum coefficient vector C The updated formula is: , , in, A random number within the interval [0,1]. The convergence factor decreases linearly from 2 to 0 as the number of optimization iterations increases, and the decreasing formula is: , in, The initial maximum number of attempts is set to 1000. This represents the current iteration number; S4.3. Set the random selection probability. When the random selection probability... p When the value is greater than 0.5, proceed to step S4.4.1 and update the position using a spiral envelope tracking strategy; when the probability is selected... p ≤0.5 and When the position is updated, proceed to S4.4.2 and use a contraction envelope cooperative approximation strategy; when the probability is selected... p ≤0.5 and When that happens, proceed to S4.4.3 and update the position using a stochastic reference solution exploration strategy.

[0041] S4.4.1. The position is updated using a spiral envelope tracking strategy. The position update formula is as follows: , The spiral envelope tracking strategy searches agents that update their positions along a logarithmic spiral with the current optimal solution as the pole. Multiple search agents move simultaneously from different initial positions in the solution space along this parameterized spiral, and their trajectories naturally intertwine to form a cluster of dynamic spiral envelope surfaces centered on the optimal solution. This allows for a fine characterization and in-depth mining of the local attraction domain of the optimal solution while maintaining population diversity.

[0042] S4.4.2. The position is updated using a contraction envelope cooperative approximation strategy. The position update formula is: , The shrinking envelope cooperative approximation strategy, given the current global optimum and a linearly decaying convergence factor, constrains the next position of all search agents to a hypersphere centered on the global optimum. As iteration progresses, the radius of this neighborhood monotonically shrinks, ensuring that the solution vector sequence of all agents inevitably converges to any small neighborhood of the global optimum.

[0043] S4.4.3. The position is updated using a stochastic reference solution exploration strategy. The position update formula is: , The random reference solution exploration strategy randomly selects another different point from the current solution set as a random reference solution, and uses this random reference solution as an anchor point to perform a large-amplitude random displacement on the search point. The displacement direction is guided by the difference vector, and the step size is scaled by a random factor uniformly sampled in the interval (0,1).

[0044] In the above update formula, For the current moment t The individual position, For the current moment t The optimal position, For the current moment t The next moment's position of the individual, where L is a random number in the range [-1, 1]. The logarithmic spiral shape constant is... This is a random reference solution.

[0045] S4.5. Calculate the objective function, update the optimal position and the corresponding objective value. If the optimal objective value is not greater than the set threshold, output the optimal position. If the optimal objective value is greater than the set threshold and the number of iterations is less than the set number of iterations, return to S4.2. If the optimal objective value is greater than the set threshold and the number of iterations reaches the set number of iterations, output a non-convergence signal.

[0046] Example of effectiveness: In a 300 MW unit's wet desulfurization flash evaporation waste heat recovery system, the dual-layer coupled predictive control model for water supply temperature proposed in this chapter was deployed on an industrial control computer. As described in the examples, the accuracy of the predictive model was evaluated using MAE and R² indices, and the performance of water supply temperature control under different control methods was evaluated using SD.

[0047] Multiple model experiments were conducted to predict the outlet water supply temperature of the heat pump unit. The training and performance of models such as RNN, LSTM, GRU, Transformer, and Transformer-GRU on the same dataset were compared. There is a significant time lag of about 3 minutes between the opening degree of the driving steam regulating valve and the water supply temperature. Therefore, the prediction data of each model at the 3rd minute (step 36) were selected for comparative analysis.

[0048] Figure 4 The curves showing the predicted and measured values ​​of the outlet water supply temperature of the heat pump unit for five prediction models, as well as the absolute error of the prediction, are presented. Figure 5 The correspondence between the predicted values ​​and the measured values ​​of the above prediction model is shown in Table 2. The specific prediction performance evaluation indicators are shown in Table 2.

[0049] Table 2 Results of the Prediction Model

[0050] These models can all capture the basic trend of water supply temperature fluctuations with flash evaporation and changes in heating network load. However, the predicted curves of the four models, RNN, GRU, LSTM and Transformer, show different degrees of fit with the measured curves. The deviation is particularly obvious when the water supply temperature changes rapidly, reflecting the limitations of a single-structure network in handling cross-timescale dynamics such as the volumetric inertia of the flash tank and the multi-stage heat exchange of the heat pump.

[0051] The Transformer-GRU hybrid model significantly outperformed the other four models in prediction bias on the test set, with a MAE of only 0.1139℃, surpassing the Transformer model's 0.1435℃, RNN model's 0.1541℃, GRU model's 0.1869℃, and LSTM model's 0.2091℃. In terms of prediction accuracy, the Transformer-GRU model achieved an R² of 0.9854, higher than the Transformer model's 0.9754, RNN model's 0.9701, GRU model's 0.9558, and LSTM model's 0.9386. On the validation set, the performance ranking of the models was largely consistent, with the Transformer-GRU model's R² of 0.9901, comparable to the RNN's 0.9904, but exhibiting the best generalization ability on the independent test set. It is worth noting that the LSTM model performed better on the validation set (R²=0.9871), but the MAE on the test set increased to 0.2091℃ and the R² decreased to 0.9386, indicating obvious overfitting. In contrast, the Transformer-GRU hybrid model maintained good performance consistency between the validation and test sets, indicating that the hybrid architecture has a significant advantage in avoiding overfitting.

[0052] The results show that the long-term coupling features extracted by the Transformer encoder and the short-term temporal dynamics captured by the GRU network complement each other, enabling a more accurate characterization of the complex evolution of water supply temperature under multi-source disturbances. This hybrid model exhibits higher prediction accuracy and robustness under the complex operating conditions of flash waste heat recovery systems with coexisting multi-source disturbances, making it more suitable for practical industrial applications of predicting the outlet water supply temperature of heat pump units.

[0053] (2) Water supply temperature control effect: Due to the operational safety regulations and limitations of the desulfurization slurry flash evaporation waste heat recovery system, it is difficult to operate and validate all comparative prediction models in a real production environment. Therefore, the Transformer-GRU model with the best prediction performance was selected for online operation and used to verify the effectiveness of the control strategy.

[0054] To evaluate the practical effect of the proposed prediction-based control method, approximately 46 hours of continuous actual operating data were extracted from the historical DCS database of the 300 MW wet desulfurization flash waste heat recovery system. Disturbance sequences for three typical operating conditions were extracted and verified through experiments: (1) Steady-state load condition: evaluating the regulation performance under internal disturbances such as flash slurry volume fluctuations and measurement errors; (2) Rapid load change condition: evaluating the dynamic response capability to track changes in the heating network load; (3) Disturbance condition: evaluating the anti-interference performance under the coupling effect of multiple internal and external disturbances on the slurry side, steam side, and flue gas side. The proposed intelligent control method was compared and analyzed with the original PID control method under each operating condition.

[0055] (2.1) Steady-state load operation: Under the original PID control method, the outlet water temperature of the heat pump unit fluctuates widely around the set value of 60℃, with an actual average operating temperature of only 55.63℃. There is a significant steady-state residual error of 4.37℃ from the set value, a standard deviation (SD) of 1.169℃, and a maximum absolute deviation of 8.014℃. The temperature fluctuates drastically within the range of 51.99℃ to 68.01℃, exceeding the safe allowable range of ±5℃. The root cause of this phenomenon is that the flash tank negative pressure evaporation process has significant volumetric inertia and heat transfer lag. In addition, the random variation of non-condensable gas content and small disturbances in slurry circulation flow rate cause the controlled object to exhibit large delay and strong coupling characteristics. Due to the lack of predictive ability, the PID controller always lags behind the disturbance in adjustment action and is forced to adopt conservative parameters to avoid drastic oscillations, thus sacrificing control accuracy.

[0056] See attached document Figure 6 and attached Figure 7 As shown, in contrast, the predictive control method of this invention uses the Transformer-GRU model to predict future water supply temperatures in advance. Combined with the algorithm's synergistic optimization of the opening of the driving steam regulating valve and the slurry circulation pump outlet regulating valve, it reduces the SD to 0.871℃ under steady-state load conditions, compresses the maximum deviation to 4.236℃, and accurately maintains the mean at 59.65℃. A two-sample t-test showed a highly significant difference in the mean temperatures between the two methods (t = -439.07, p < 0.001); the F-test further indicated that the temperature variance of the predictive control was significantly smaller (F = 1.40, p < 0.001), demonstrating that the improvement in accuracy and stability of predictive control has a solid statistical basis.

[0057] (2.2) Rapid Load Change Operation Condition: In actual operation, the continuous changes in the return water temperature and flow rate of the heating network cause continuous disturbances to the supply water temperature. Since the PID controller only relies on the current temperature feedback, it cannot predict the trend of heating network load changes, and the adjustment action has obvious phase lag, which easily leads to response delay and large overshoot. In the simulation verification based on the real historical disturbance sequence, the SD of the simulated PID under this condition is as high as 3.866℃, the maximum deviation is 5.782℃, the valve opening frequently impacts the full amplitude between 0% and 100%, and the steam consumption is as high as 75.10 GJ.

[0058] The proposed predictive control method utilizes the time-series predictive capability of the Transformer-GRU model to detect the trend of heating network load changes in advance. Before the disturbance effect manifests, it coordinates the adjustment of the opening of the drive steam regulating valve and the opening of the slurry circulation pump outlet regulating valve, reducing the SD to 0.849℃, controlling the maximum deviation to 4.341℃, and reducing steam consumption to 71.30 GJ, which is 5.1% lower than PID control. This significantly improves the control accuracy, response stability, and operating economy of water supply temperature under variable load conditions.

[0059] (2.3) Disturbance Conditions: Faced with complex conditions involving the coupling of multiple internal and external disturbances from the slurry, steam, and flue gas sides, the original PID control method exhibits significant temperature oscillations and overshoot due to the large valve adjustment range and lag response. The simulated PID under these conditions has a SD of 3.869℃, a maximum deviation of 5.809℃, and steam consumption as high as 75.18 GJ.

[0060] Table 3: Comparison of Predictive Control and PID Control Results

[0061] As shown in Table 3, the predictive control method proposed in this paper identifies the coupled effects of multi-source disturbances on water supply temperature in advance based on the Transformer-GRU model. The optimal control sequence is solved in real time by the algorithm, which effectively suppresses temperature oscillation and overshoot, reduces SD to 0.829℃, controls the maximum deviation at 4.479℃, and consumes only 71.27 GJ. It can still keep the temperature within a safe range under complex disturbances, showing excellent anti-interference performance and energy-saving potential.

[0062] The present invention has been described in detail above with reference to the embodiments, but the content described is only a preferred embodiment of the present invention and should not be considered as limiting the scope of the present invention. All equivalent changes and improvements made in accordance with the scope of the present invention should still fall within the patent coverage of the present invention.

Claims

1. A flue gas waste heat-water recovery synergistic pollutant emission reduction system, comprising a desulfurization tower, a slurry circulation subsystem, a flash tank, a vacuum pump unit, a heat pump unit, a condensate recovery subsystem, and an intelligent control system; the slurry circulation subsystem includes a slurry circulation pump, a slurry circulation pipeline, and a return pump. A slurry circulation pump is installed in the slurry circulation pipeline on the third layer of the desulfurization tower, and the slurry is drawn to the flash tank for negative pressure evaporation. An outlet regulating valve is installed on the circulation pipeline between the slurry circulation pump and the flash tank to control the amount of slurry drawn to the flash tank; the low-temperature desulfurization after flash evaporation... The sulfur slurry is pumped back to the third-layer slurry circulation pipeline of the desulfurization tower by a return pump and then sprayed into the tower to continue the desulfurization reaction. The exhaust steam generated in the flash tank enters the heat pump unit. A steam regulating valve is installed on the pipeline between the flash tank and the heat pump unit to control the exhaust steam output. Driven by externally supplied superheated steam, the heat pump unit completely absorbs the waste heat generated by flash steam and converts it into a heat source, achieving waste heat recovery. After the heat is extracted, the exhaust steam condenses into water in the heat pump evaporator, which is then collected by the condensate recovery subsystem and reused as makeup water for the desulfurization tower or as boiler makeup water. Its key feature is... The intelligent control system includes: The data acquisition module is used to acquire historical operating data of the flue gas waste heat-water recovery and pollutant emission reduction system; The data processing module is used to preprocess historical operational data to obtain the feature set of the prediction model; The water supply temperature prediction module is a prediction model embedded with a Transformer-GRU dual-layer coupled network. It includes a feature extraction layer and a temporal capture layer. The feature extraction layer uses a Transformer encoder, and the temporal capture layer uses a dual-layer GRU network. It is used to input the feature set into the prediction model and output the predicted water supply temperature value. The collaborative optimization and regulation module is used to construct an optimization problem based on the predicted water supply temperature, with the objective functions being temperature tracking accuracy and the magnitude of control quantity change. It also introduces control variable constraints and temperature safety constraints. The helical envelope meta-heuristic algorithm is used to solve the optimal combination of the opening degree of the driving steam regulating valve and the speed regulation of the slurry circulation pump. The driving steam regulating valve is the main control variable, and the speed of the slurry circulation pump is the secondary control variable, thus performing collaborative regulation. The specific steps for solving the optimal combination of the driving steam regulating valve opening and the slurry circulation pump speed regulation amount using the spiral envelope meta-heuristic algorithm are as follows: S4.

1. Initialize the position and calculate the objective function to determine the optimal position; S4.

2. Update the coefficient vector A sum coefficient vector C The updated formula is: ， ， in, A random number within the interval [0,1]. The convergence factor decreases linearly from 2 to 0 as the number of optimization iterations increases, and the decreasing formula is: ， in, The initial maximum number of times, This represents the current iteration number; S4.

3. Set the random selection probability. When the random selection probability... p When the value is greater than 0.5, proceed to step S4.4.1 and update the position using a spiral envelope tracking strategy; when the probability is selected... p ≤0.5 and When the position is updated, proceed to S4.4.2 and use a contraction envelope cooperative approximation strategy; when the probability is selected... p ≤0.5 and When the time comes, proceed to S4.4.3 and update the position using a stochastic reference solution exploration strategy; S4.4.

1. The position is updated using a spiral envelope tracking strategy. The position update formula is as follows: ， S4.4.

2. The position is updated using a contraction envelope cooperative approximation strategy. The position update formula is: ， S4.4.

3. The position is updated using a stochastic reference solution exploration strategy. The position update formula is as follows: ， in, For the current moment t The individual position, For the current moment t The optimal position, For the current moment t The next moment's position of the individual, where L is a random number in the range [-1, 1]. The logarithmic spiral shape constant is... This is a random reference solution; S4.

5. Calculate the objective function, update the optimal position and the corresponding objective value. If the optimal objective value is not greater than the set threshold, output the optimal position. If the optimal objective value is greater than the set threshold and the number of iterations is less than the set number of iterations, return to S4.

2. If the optimal objective value is greater than the set threshold and the number of iterations reaches the set number of iterations, output a non-convergence signal.

2. A control method for the flue gas waste heat-water recovery synergistic pollutant emission reduction system as described in claim 1, characterized in that, It includes the following steps: S1. Obtain historical operational data of the flue gas waste heat-water recovery synergistic pollutant emission reduction system; S2. Preprocess historical operating data to obtain the feature set of the prediction model; S3. Construct a prediction model using a Transformer-GRU two-layer coupled network, including a feature extraction layer and a temporal capture layer. The feature extraction layer uses a Transformer encoder, and the temporal capture layer uses a two-layer GRU network. Input the feature set into the prediction model, and output the predicted water supply temperature; S4. Based on the predicted water supply temperature, an optimization problem is constructed with temperature tracking accuracy and control quantity change range as objective functions, and control variable constraints and temperature safety constraints are introduced; the optimal combination of the opening of the driving steam regulating valve and the speed regulation of the slurry circulation pump is solved by using the driving steam regulating valve as the main control variable and the speed of the slurry circulation pump as the secondary control variable to perform coordinated regulation. The specific steps for solving the optimal combination of the driving steam regulating valve opening and the slurry circulation pump speed regulation amount using the spiral envelope meta-heuristic algorithm are as follows: S4.

1. Initialize the position and calculate the objective function to determine the optimal position; S4.

2. Update the coefficient vector A sum coefficient vector C The updated formula is: ， ， in, A random number within the interval [0,1]. The convergence factor decreases linearly from 2 to 0 as the number of optimization iterations increases, and the decreasing formula is: ， in, The initial maximum number of times, This represents the current iteration number; S4.

3. Set the random selection probability. When the random selection probability... p When the value is greater than 0.5, proceed to step S4.4.1 and update the position using a spiral envelope tracking strategy; when the probability is selected... p ≤0.5 and When the position is updated, proceed to S4.4.2 and use a contraction envelope cooperative approximation strategy; when the probability is selected... p ≤0.5 and When the time comes, proceed to S4.4.3 and update the position using a stochastic reference solution exploration strategy; S4.4.

1. The position is updated using a spiral envelope tracking strategy. The position update formula is as follows: ， S4.4.

2. The position is updated using a contraction envelope cooperative approximation strategy. The position update formula is: ， S4.4.

3. The position is updated using a stochastic reference solution exploration strategy. The position update formula is as follows: ， in, For the current moment t The individual position, For the current moment t The optimal position, For the current moment t The next moment's position of the individual, where L is a random number in the range [-1, 1]. The logarithmic spiral shape constant is... This is a random reference solution; S4.

5. Calculate the objective function, update the optimal position and the corresponding objective value. If the optimal objective value is not greater than the set threshold, output the optimal position. If the optimal objective value is greater than the set threshold and the number of iterations is less than the set number of iterations, return to S4.

2. If the optimal objective value is greater than the set threshold and the number of iterations reaches the set number of iterations, output a non-convergence signal.

3. The control method for the flue gas waste heat-water recovery synergistic pollutant emission reduction system according to claim 2, characterized in that: The historical operating data in S1 includes the following key variables: slurry side parameters, steam side parameters, flue gas side parameters, and heating side parameters.

4. The control method for the flue gas waste heat-water recovery synergistic pollutant emission reduction system according to claim 3, characterized in that: The preprocessing of historical operating data in S2 includes data cleaning, time delay correction, construction of derived features, and feature selection. The specific steps are as follows: S2.

1. Use the absolute deviation of the middle value in the sliding window method to identify and correct outliers in historical running data, and replace outliers with the weighted average of adjacent data; S2.

2. The cross-correlation method is used to estimate the pure time lag between the historical operating data after cleaning and the water supply temperature; S2.

3. Calculate the difference between the current opening degree of the driving steam regulating valve and the minimum allowable opening degree of the system, the real-time deviation between the measured value and the set value of the water supply temperature, and the difference between the slurry temperature at the outlet of the first-stage flash evaporator and the set value of the water supply temperature, and then form the feature set of the water supply temperature prediction model.

5. The control method for the flue gas waste heat-water recovery synergistic pollutant emission reduction system according to claim 4, characterized in that: In step S2.3, after calculating the difference between the current opening degree of the driving steam regulating valve and the minimum allowable opening degree of the system, the real-time deviation between the measured value and the set value of the water supply temperature, and the difference between the slurry temperature at the first-stage flash outlet and the set value of the water supply temperature, the sliding window method is used to statistically analyze the statistical characteristics of key variables, candidate variables are selected from the key variables, and the maximum information coefficient method is used to calculate the correlation strength between each candidate variable and the water supply temperature, eliminating low-correlation and redundant variables, and simplifying the input feature dimensions of the prediction model.

6. The control method for the flue gas waste heat-water recovery synergistic pollutant emission reduction system according to claim 5, characterized in that: In step S3, the feature extraction layer captures the nonlinear coupling relationships and long-term dependency characteristics among variables in the feature set, and the time-series capture layer outputs the predicted water supply temperature based on the nonlinear coupling relationships and long-term dependency characteristics. The specific steps are as follows: S3.

1. Calculate the query, key, and value of the Transformer encoder using the following formulas: ， ， ， in, Indicates the number of the feature set. The feature set represents the historical dataset of the candidate variables. Indicates a query. The key representing the feature set, Represents the value of the feature set. , and These are the weight matrices for the query, key, and value, respectively. S3.

2. Calculate the attention score between each element and all other elements. Different elements correspond to different types of historical operating data. The attention score quantifies the relative importance of each element at different historical moments when predicting water supply temperature. The formula for calculating the attention score is as follows: ， in, This represents the attention score. (·) represents the normalized exponential function. Indicates the dimension of the key vector; Alternatively, the attention mechanism can employ a multi-head attention mechanism, with the attention score calculated using the following formula: ， ， in, This represents the overall operation of the multi-head attention mechanism. This means concatenating the outputs of all attention heads. This represents the output weight matrix. Indicates the input number The output of the attention head for each feature set. This indicates the total number of heads of attention. These represent the results of the query, key, and value after a linear transformation, respectively. S3.

3. Calculate the gating value of the GRU network reset gate. The calculation formula is as follows: ， in, This indicates that the door's gate control value is being reset. It is the Sigmoid activation function. It is the weight matrix of the reset gate. It is the hidden state of the previous time step. It is the first The input vector at each time point corresponds to the input feature set; S3.

4. Based on the gate control value of the reset gate, calculate the candidate hidden layer state using the following formula: ， in, These are candidate hidden layer states. surface Activation function It is the weight matrix of the candidate hidden layer states; S3.

5. Calculate the update gate. The calculation formula is as follows: ， in, This indicates updating the value of the gate. It is the Sigmoid activation function. It is the weight matrix of the updated gate; S3.

6. Calculate the final hidden layer state using the following formula: ， in, This represents the final hidden layer state, i.e., the hidden layer state at the current time step.

7. The control method for the flue gas waste heat-water recovery synergistic pollutant emission reduction system according to claim 2, characterized in that: The objective function in S4 is: ， in, Describe the objective function. Indicates the length of the prediction time domain. Indicates the length of the control time domain. Indicates the current t The future after a moment t + i Predicted water supply temperature at all times Indicates the preceding t The future after a moment t + i The set value of the water supply temperature at all times. and They represent the coefficients of the penalty term, and They represent the current t The future after a moment t + j The adjustment increments of the primary and secondary control variables at any given time.

8. The control method for the flue gas waste heat-water recovery synergistic pollutant emission reduction system according to claim 7, characterized in that: The constraints in S4 include: ， in, and These represent the lower and upper limits of the manipulated variable, respectively. Indicates the current t The future after a moment t + j The control variables at any given time, including primary control variables and secondary control variables. Indicates the current t The future after a moment t + j The change in the manipulative variable at any given time. This represents the maximum permissible change between adjacent time points. express t Predicted water supply temperature at all times and These represent the lower and upper limits of the safe water supply temperature, respectively.

Images