Ammonia water injection quantity accurate regulation and control method for flue gas denitration system

Abstract

The invention relates to the technical field of flue gas denitration, in particular to an ammonia water injection quantity accurate regulation and control method for a flue gas denitration system. Comprising the following steps: S1, constructing a function relational expression between the ammonia water injection amount and the denitration efficiency; s2, constructing a comprehensive objective function based on the ammonia water injection amount and the denitration efficiency; s3, collecting historical operation data of the flue gas denitration system, constructing a reinforcement learning model, and iteratively training the reinforcement learning model by using the historical operation data to minimize the comprehensive objective function; s4, acquiring the real-time operation state data of the current flue gas denitration system, and predicting the ammonia water injection amount by using the reinforcement learning model output model; and S5, constructing a correction relational expression, and obtaining the optimal ammonia water injection amount based on the correction relational expression. By constructing the reinforcement learning model and the correction relational expression and combining the real-time working condition data and the historical operation data, the system can dynamically adjust the ammonia water injection amount, the denitration efficiency is ensured to be maximized, and meanwhile ammonia escape is effectively controlled.

Description

Technical Field

[0001] The invention relates to the technical field of flue gas denitration, and in particular to a method for accurately controlling the injection amount of ammonia water for a flue gas denitration system. Background Art

[0002] With the continuous deepening of industrialization and the rapid expansion of industrial scale, the problem of air pollution is becoming increasingly serious, among which nitrogen oxides (NO x ) emissions have become one of the main pollutants affecting ambient air quality. As a by-product that is difficult to completely avoid in industrial production, NO x The control of flue gas NOx has become a key research topic in the field of environmental protection, among which flue gas denitrification technologies such as selective catalytic reduction (SCR) are the key technical paths to solve this problem.

[0003] However, the traditional flue gas denitrification system has significant technical limitations in the regulation of reducing agent (ammonia water) injection, which seriously restricts the system's denitrification efficiency and operating performance. Specifically, it manifests itself in the following technical aspects: First, the traditional control system relies too much on static empirical models and fixed parameter settings for adjustment. This method based on deterministic control logic is difficult to adapt to the complex and changeable operating conditions of industrial sites, and cannot achieve real-time optimization and adjustment for the dynamic changes of process parameters; second, in the quantitative regulation of ammonia water injection, the closed-loop control accuracy of the traditional system is insufficient, and deviations such as excessive injection or insufficient supply of reducing agents frequently occur, which not only affects the denitrification reaction efficiency, but also causes excessive consumption of chemical agents; third, the system's dynamic response characteristics are poor, and it is difficult to adapt to the flue gas temperature, flow rate, NO x When process parameters such as concentration fluctuate rapidly, timely closed-loop adjustment cannot be achieved, resulting in a significant decrease in denitrification efficiency. Summary of the invention

[0004] In view of this, the present invention proposes a method for accurately controlling the injection amount of ammonia water for a flue gas denitrification system, so as to solve the problem in the prior art that the injection amount of ammonia water is difficult to accurately adjust, resulting in reduced denitrification efficiency.

[0005] The technical solution of the present invention is implemented as follows: The present invention provides a method for accurately controlling the amount of ammonia water injection for a flue gas denitrification system, comprising the following steps:

[0006] S1. Constructing a flue gas denitration reaction model, and constructing a functional relationship between the ammonia injection amount and the denitration efficiency based on the flue gas denitration reaction model;

[0007] S2. Based on the functional relationship between the ammonia injection amount and the denitrification efficiency and the constraints of the flue gas denitrification system, a comprehensive objective function is constructed;

[0008] S3. Collect historical operation data of the flue gas denitrification system and build a reinforcement learning model, in which the environmental state includes flue gas flow, nitrogen oxide concentration and ammonia escape, the action space is the adjustment amount of ammonia injection, and a reward function is built based on denitrification efficiency and ammonia escape; use historical operation data to iteratively train the reinforcement learning model, and maximize the reward function to make the model learn the optimal strategy, so that the comprehensive objective function value on the historical operation data is minimized;

[0009] S4, obtaining the real-time operating status data of the current flue gas denitrification system, inputting the real-time operating status data into the trained reinforcement learning model, and outputting the model-predicted ammonia injection amount;

[0010] S5. Construct a correction relationship, correct the ammonia injection amount predicted by the model based on the correction relationship, and calculate the optimal ammonia injection amount.

[0011] On the basis of the above technical solution, preferably, step S1 includes:

[0012] Constructing flue gas denitrification reaction model:

[0013] NO x +NH3→N2+H2O

[0014] According to the analysis of the flue gas denitrification reaction model, the ammonia injection amount, nitrogen oxide concentration, temperature and flue gas flow rate all affect the denitrification efficiency; based on the analysis results of the flue gas denitrification reaction model, the functional relationship between the ammonia injection amount and the denitrification efficiency is constructed:

[0015]

[0016] Among them, η denox It represents the denitrification efficiency. Indicates the effect of ammonia injection amount, represents the influence of nitrogen oxide concentration, f T (T) represents the effect of temperature, f v (V) represents the influence of flue gas flow rate, k1, k2, k3 and k4 are weight coefficients, and k5 is a constant term, which represents the basic denitrification efficiency.

[0017] On the basis of the above technical solution, preferably, in step S2, the calculation formula of the comprehensive objective function is:

[0018]

[0019] Among them, η denox is the denitrification efficiency, Q max is the maximum ammonia injection amount, [NO x ] in is the concentration of nitrogen oxides in flue gas before denitrification, [NO x ]out The concentration of nitrogen oxides in the flue gas after denitrification, w1, w2, w3 are weight coefficients

[0020] On the basis of the above technical solution, preferably, in step S3, a reward function is constructed based on the denitration efficiency and the ammonia escape amount, and the calculation expression of the reward function is:

[0021]

[0022] Among them [NO x ] in is the concentration of nitrogen oxides before denitrification, [NO x ] out is the concentration of nitrogen oxides after denitrification, NH3 is the amount of ammonia escape, is the maximum allowed ammonia escape, ω1 and ω2 are weight coefficients.

[0023] Based on the above technical solution, preferably, in step S3, building a reinforcement learning model specifically includes:

[0024] S31, defining the environment state and action space based on historical operation data;

[0025] S32. Constructing a reward function based on denitration efficiency and ammonia escape;

[0026] S33. Construct an evaluation network Q(S,A;θ) and a target network Q(S,A;θ'), and establish an experience replay pool for storing state transition samples (S,A,r,S'), where S represents the environment state, A represents the action space, θ represents the parameter set of the evaluation network, θ' represents the parameter set of the target network, r represents the immediate reward, and S' represents the next environment state after the action is executed;

[0027] S34, execute model training: use the ε-greedy strategy to select actions, execute the selected actions and obtain immediate rewards r, store state transition samples in the experience replay pool, randomly sample batch data from the experience replay pool, and update the Q value through the Q value update formula;

[0028] S35. Based on the Q value, determine whether the convergence condition is met. If not, return to S34 to continue training. If so, the final reinforcement learning model is obtained.

[0029] On the basis of the above technical solution, preferably, the calculation expression of the Q value update formula is:

[0030]

[0031] Among them, Q(S t , A t ) indicates that in state S t Next, execute action At The expected cumulative reward; α represents the learning rate; r t Indicates execution of action A t The immediate reward after γ represents the discount factor; A′ represents the possible action for the next environment state. Indicates the next state S t+1 The maximum Q value among all possible actions.

[0032] Based on the above technical solution, preferably, in step S35, the convergence condition is to meet any of the following conditions:

[0033] (1) The number of training times reaches the preset maximum number of iterations N;

[0034] (2) In K consecutive iterations, the maximum update amplitude of the Q value is less than the preset threshold of the Q value update amplitude;

[0035] (3) In M consecutive iterations, the average value of the Q-value change of all state-action pairs is less than the preset threshold of the Q-value change.

[0036] Based on the above technical solution, preferably, step S4 includes:

[0037] S41, obtaining real-time operating status data of the current flue gas denitrification system, inputting the real-time operating status data into the trained reinforcement learning model, predicting multiple possible ammonia water injection amount adjustment actions, and obtaining an initial ammonia water injection amount;

[0038] S42, constructing a search interval near the initial ammonia injection amount, and calculating the comprehensive objective function value J for each candidate ammonia injection amount within the search interval i ;

[0039] S43, if there is J i <J i-1 , then update the ammonia injection amount and return to step S42 to continue optimization; if there is J i >J i-1 or n consecutive iterations J i ≈J i-1 , the optimization is terminated and the model predicts the ammonia injection amount.

[0040] On the basis of the above technical solution, preferably, step S5 specifically includes: introducing a reaction rate parameter based on the flue gas denitration reaction model, establishing a correction relationship between the reaction rate parameter and the model predicted ammonia injection amount, correcting the model predicted ammonia injection amount predicted by the reinforcement learning model based on the correction relationship, and calculating the optimal ammonia injection amount. The correction relationship between the reaction rate parameter and the model predicted ammonia injection amount is:

[0041]

[0042] in, Indicates the optimal ammonia injection amount, [NO x ] represents the concentration of nitrogen oxides, T represents the temperature, r represents the reaction rate parameter, represents the model's predicted ammonia injection amount, represents the historical training data set, A represents the constant coefficient, and B, C, D, E and F represent the weight coefficients.

[0043] On the basis of the above technical solution, preferably, in step S3, the historical data operation data includes flue gas temperature, flue gas flow, ammonia injection amount, ammonia concentration, nitrogen oxide concentration before denitrification, nitrogen oxide concentration after denitrification, denitrification efficiency and ammonia escape concentration.

[0044] The method for accurately controlling the amount of ammonia water injection for a flue gas denitrification system of the present invention has the following beneficial effects compared with the prior art:

[0045] (1) By building a reinforcement learning model, the ability to accurately adjust the amount of ammonia injection in the flue gas denitrification system is significantly improved; by building a flue gas denitrification reaction model and a functional relationship between the amount of ammonia injection and the denitrification efficiency, combined with real-time operating data and historical operation data, the system can dynamically adjust the amount of ammonia injection to ensure the maximization of the denitrification efficiency, while effectively controlling ammonia escape. The intelligent adjustment method overcomes the limitations of the traditional system that relies on static empirical models and can adapt to complex and changeable industrial site environments;

[0046] (2) Based on the functional relationship between the amount of ammonia water injection and the denitrification efficiency, combined with the constraints of the system, a comprehensive objective function is constructed, so that the system can find the best balance point in multi-objective optimization, which not only improves the denitrification efficiency, but also effectively controls ammonia escape, ensuring the economy and environmental protection of the system operation;

[0047] (3) By building a reinforcement learning model, defining the environmental state and action space, and constructing a reward function based on denitrification efficiency and ammonia escape, it is possible to perform adaptive learning and optimization in a complex industrial environment. Through experience replay and Q-value update mechanism, the model can continuously improve prediction accuracy and decision-making ability, and ultimately achieve precise control of ammonia injection amount and reduce excessive consumption of chemical agents;

[0048] (4) The reaction rate parameter is introduced and a correction relationship is constructed. The model prediction results are corrected by considering the reaction kinetic characteristics, which overcomes the shortcomings of the traditional system in dynamic response, improves the system's ability to respond quickly to operating condition fluctuations, and makes the adjustment of the ammonia injection amount more accurate and timely. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0050] Figure 1 This is a flow chart of a method for accurately controlling the amount of ammonia water injection for a flue gas denitration system according to the present invention;

[0051] Figure 2 This is a real-time prediction optimization flow chart of the present invention. DETAILED DESCRIPTION

[0052] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] like Figure 1 As shown, the present invention provides a method for accurately controlling the injection amount of ammonia water for a flue gas denitrification system, comprising the following steps:

[0054] S1. Constructing a flue gas denitration reaction model, and constructing a functional relationship between the ammonia injection amount and the denitration efficiency based on the flue gas denitration reaction model;

[0055] S2. Based on the functional relationship between the ammonia injection amount and the denitrification efficiency and the constraints of the flue gas denitrification system, a comprehensive objective function is constructed;

[0056] S3. Collect historical operation data of the flue gas denitrification system and build a reinforcement learning model, in which the environmental state includes flue gas flow, nitrogen oxide concentration and ammonia escape, the action space is the adjustment amount of ammonia injection, and a reward function is built based on denitrification efficiency and ammonia escape; use historical operation data to iteratively train the reinforcement learning model, and maximize the reward function to make the model learn the optimal strategy, so that the comprehensive objective function value on the historical operation data is minimized;

[0057] S4, obtaining the real-time operating status data of the current flue gas denitrification system, inputting the real-time operating status data into the trained reinforcement learning model, and outputting the model-predicted ammonia injection amount;

[0058] S5. Construct a correction relationship, correct the ammonia injection amount predicted by the model based on the correction relationship, and calculate the optimal ammonia injection amount.

[0059] The invention significantly improves the ability to precisely adjust the amount of ammonia water injection in the flue gas denitrification system by constructing a reinforcement learning model; by constructing a flue gas denitrification reaction model and a functional relationship between the amount of ammonia water injection and the denitrification efficiency, combined with real-time operating data and historical operation data, the system can dynamically adjust the amount of ammonia water injection to ensure the maximization of the denitrification efficiency and effectively control ammonia escape; the intelligent adjustment method overcomes the limitations of the traditional system that relies on static empirical models and can adapt to complex and changeable industrial field environments

[0060] Furthermore, step S1 specifically includes:

[0061] In the process of flue gas denitrification, ammonia water reacts with nitrogen oxides (NO x ) reacts to generate nitrogen (N2) and water (H2O), and a flue gas denitrification reaction model is constructed:

[0062] NO x +NH3→N2+H2O

[0063] According to the analysis of the flue gas denitrification reaction model, the amount of ammonia water injection, nitrogen oxide concentration, temperature and flue gas flow rate all affect the denitrification efficiency. Among them, the amount of ammonia water injection directly affects the denitrification efficiency. Too much injection may cause ammonia escape, while too little injection will lead to poor denitrification effect. Based on the analysis results of the flue gas denitrification reaction model, the functional relationship between the amount of ammonia water injection and the denitrification efficiency can be expressed as:

[0064]

[0065] in, Indicates the amount of ammonia injection, [NO x ] represents the concentration of nitrogen oxides, T represents the temperature, and v represents the flue gas flow rate.

[0066] Specifically:

[0067] (1) Influence of ammonia injection amount: Ammonia is used as a reducing agent in the denitration reaction, and its flow rate (i.e. injection amount) directly affects the denitration efficiency. Within a certain range, increasing the ammonia injection amount can improve the denitration efficiency, but excessive ammonia may cause ammonia escape. The influence of ammonia injection amount on denitration efficiency can be expressed as:

[0068]

[0069] Among them, a1, b1 and c1 are all unknown coefficients.

[0070] (2) Effect of nitrogen oxide concentration: The higher the nitrogen oxide concentration, the more ammonia is needed to react to achieve the same denitrification efficiency. At the same time, high concentrations of nitrogen oxides may inhibit the denitrification reaction. The effect of nitrogen oxide concentration on denitrification efficiency is expressed as:

[0071]

[0072] Among them, a2, b2 and c2 are all unknown coefficients.

[0073] (2) Effect of temperature: Generally, within a certain temperature range, increasing the temperature can accelerate the reaction rate and improve the denitration efficiency; however, too high a temperature may cause problems such as catalyst deactivation. The effect of temperature on denitration efficiency is expressed as:

[0074] f T (T) = a3T 2 +b3T+c3

[0075] Among them, a3, b3 and c3 are all unknown coefficients.

[0076] (3) Effect of flue gas flow rate: Flue gas flow rate affects the residence time of flue gas in the denitrification reactor, and thus affects the denitrification efficiency. Too fast a flow rate may lead to incomplete reaction, while too slow a flow rate may cause blockage and other problems. The effect of flue gas flow rate on denitrification efficiency is expressed as:

[0077] f v (v) = a4v 2 +b4v+c4

[0078] Among them, a4, b4 and c4 are all unknown coefficients.

[0079] Taking into account the influence of the above factors, the functional relationship between the ammonia injection amount and the denitrification efficiency is finally expressed as:

[0080]

[0081] Among them, η denox It represents the denitrification efficiency. Indicates the effect of ammonia injection amount, represents the influence of nitrogen oxide concentration, f T (T) represents the effect of temperature, f v (V) represents the influence of flue gas flow rate, k1, k2, k3 and k4 are weight coefficients, reflecting the relative importance of different factors on denitrification efficiency; k5 is a constant term, which represents the basic denitrification efficiency under standard operating conditions.

[0082] By constructing a flue gas denitrification reaction model and the functional relationship between the ammonia injection amount and the denitrification efficiency, the influence of various process parameters on the denitrification efficiency can be accurately described, providing a reliable theoretical basis for the subsequent reinforcement learning control strategy.

[0083] Furthermore, in step S2, with the goal of minimizing ammonia slip and maximizing denitrification efficiency, a comprehensive objective function is constructed based on the functional relationship between ammonia injection amount and denitrification efficiency and system constraints. First, the objective function needs to simultaneously satisfy the following specific constraints:

[0084] ① Constraints on the amount of ammonia injection: The amount of ammonia injection must be within the system design range 1. The concentration of ammonia water must meet the process requirements; 2. Environmental protection index constraints: the denitrification efficiency must meet the minimum environmental protection standards, the nitrogen oxide emission concentration must meet the emission standards, and the ammonia escape must be controlled within the allowable range; 3. Operating parameter constraints: the reaction temperature must be maintained within the catalyst activity temperature window, the flue gas flow rate must be within the design load range, and the system pressure loss must be within the allowable range.

[0085] Based on the above constraints, a comprehensive objective function is constructed, and the calculation formula of the comprehensive objective function is:

[0086]

[0087] Among them, η denox is the denitrification efficiency, which reflects the effectiveness of the system in removing nitrogen oxides. denox ), which can ensure the efficient operation of the system in reducing pollutant emissions; Q max is the maximum allowable ammonia injection amount, Represents the ratio of the amount of ammonia water injection to its maximum allowable value. This item is used to measure the efficiency of ammonia water use and ensure that the injection amount is within a reasonable range to avoid ammonia escape caused by excessive injection; [NO x ] in is the concentration of nitrogen oxides in flue gas before denitrification, [NO x ] out The concentration of nitrogen oxides in flue gas after denitrification, It is the ratio of the concentration of nitrogen oxides in the flue gas after denitrification to the concentration before denitrification, which reflects the purification effect of the denitrification process. By minimizing this ratio, the purification capacity of the system can be further optimized. w1, w2, and w3 are weight coefficients, w1+w2+w3=1.

[0088] By minimizing the comprehensive objective function J, the overall optimization of multiple objectives such as denitrification efficiency, ammonia consumption and emission compliance can be achieved under the premise of meeting various constraints.

[0089] Further, step S3 specifically includes: after completing the construction of the comprehensive objective function, collecting historical operation data of the flue gas denitrification system, wherein the historical operation data includes flue gas temperature, flue gas flow, ammonia injection amount, ammonia concentration, nitrogen oxide concentration before denitrification, nitrogen oxide concentration after denitrification, denitrification efficiency and ammonia escape concentration. Perform data preprocessing on the historical operation data, including outlier processing, data normalization processing and time series alignment processing, to ensure that the data quality meets the model training requirements.

[0090] Based on the above preprocessed historical data, a reinforcement learning model is constructed. The specific process of the reinforcement learning model is as follows:

[0091] S31. Define the environment state and action space based on historical operation data.

[0092] The environmental status includes flue gas flow (unit: cubic meter / hour), nitrogen oxide concentration (unit: mg / cubic meter) and ammonia escape (unit: mg / cubic meter). These parameters directly reflect the operating status of the system; the action space is defined as the adjustment amount of ammonia injection (unit: kg / hour), and its value range is [-δ max ,δ max ], where δ max is the maximum adjustment allowed.

[0093] S32. Construct a reward function based on maximizing denitration efficiency and minimizing ammonia escape, wherein the calculation expression of the reward function is:

[0094]

[0095] Among them [NO x ] in is the concentration of nitrogen oxides before denitrification, [NO x ] out is the concentration of nitrogen oxides after denitrification, NH3 is the amount of ammonia escape, is the maximum allowed ammonia escape, ω1 and ω2 are weight coefficients.

[0096] S33. Construct an evaluation network Q(S,A;θ) and a target network Q(S,A;θ'), and establish an experience replay pool for storing state transition samples (S,A,r,S'), where S represents the environment state, A represents the action space, θ represents the parameter set of the evaluation network, θ' represents the parameter set of the target network, r represents the immediate reward, and the reward value is calculated once per time step; S' represents the next environment state after the action is executed;

[0097] Specifically, the evaluation network is used to estimate the Q value corresponding to the current state-action, where the input layer receives the environment state S, the output layer corresponds to the Q value of the action adjusted by different ammonia injection amounts, and the network parameter θ is continuously updated and optimized through back propagation. The target network is used to calculate the target Q value and provide a more stable learning target. The parameter θ' is regularly copied from the evaluation network to reduce training instability.

[0098] S34. Execute model training:

[0099] S341, initialization phase: randomly initialize the evaluation network parameters θ, copy the evaluation network parameters to the target network θ'=θ; clear the experience replay pool;

[0100] S342, training iteration process: adopt ε-greedy strategy to select actions, randomly select actions for exploration with probability ε, and select the maximum action considered by the current evaluation network with probability 1-ε. After executing the selected action, calculate the immediate reward r according to the denitrification efficiency and ammonia escape, and obtain the immediate reward r and the next state S'. Store the state transition sample (S, A, r, S') in the experience replay pool, randomly sample batch data from the experience replay pool for training, and update the parameters of the evaluation network through the Q value update formula.

[0101] Among them, the calculation expression of the Q value update formula is:

[0102]

[0103] Among them, Q(S t , A t ) indicates that in state S t Next, execute action A t The expected cumulative reward; α represents the learning rate, which controls the step size of parameter update; r t Indicates execution of action A t The immediate reward obtained after the environment is obtained; γ represents the discount factor, which is used to balance the importance of immediate rewards and future rewards; A′ represents the possible actions for the next environment state, Indicates the next state S t+1 The maximum Q value among all possible actions.

[0104] S35. Determine whether the convergence condition is met based on the Q value. If not, return to step S34 to continue training. If it is met, obtain the final reinforcement learning model. The model can automatically decide the optimal ammonia injection adjustment strategy according to the current system state to achieve balanced control of denitration efficiency and ammonia escape. At the same time, it can continue to learn and optimize in actual applications, and continuously update and improve the control strategy through real-time data.

[0105] The convergence conditions include:

[0106] (1) The number of training iterations reaches the preset maximum number of iterations N. This is a basic termination condition to ensure that the training process does not proceed indefinitely. When the number of training iterations reaches the preset maximum value N, the training will stop even if other convergence conditions are not met.

[0107] (2) In K consecutive iterations, the maximum update amplitude of the Q value is less than the preset threshold of the Q value update amplitude, that is, the difference between the single Q value before and after the update max|Q(S t+1 ,A t+1 )-Q(S t ,A t )| is less than the preset threshold of Q value update amplitude. This item focuses on the most significant change in a single iteration. When the maximum update amplitude in K consecutive iterations is less than the preset threshold, it indicates that the model has achieved relative stability in the main state-action pairs.

[0108] (3) In M consecutive iterations, the average value of the Q-value change of all state-action pairs is less than the preset threshold of the Q-value change, that is, the average of the Q-value changes of all state-action pairs reflects the global convergence degree of the entire model. When the average change is less than the preset threshold for M consecutive times, it indicates that the model has reached a stable state as a whole.

[0109] Specifically, the convergence condition is to meet any of the following conditions:

[0110] a. After each iteration of training, first determine whether the maximum number of iterations N has been reached;

[0111] b. If the maximum number of iterations N is not reached, the maximum update amplitude of the Q value of this iteration is calculated and compared with the Q value of the previous iteration;

[0112] c. Calculate the average Q value change of all state-action pairs in this iteration and compare it with the result of the Q value change of all state-action pairs in the previous iteration;

[0113] d. If any of the convergence conditions of the Q-value update amplitude and the Q-value change are met, the training is stopped to obtain the final reinforcement learning model;

[0114] e. If none of the conditions are met, return to step S34 to continue training.

[0115] Furthermore, if Figure 2 As shown, in step S4, after completing the training of the reinforcement learning model, the model needs to be applied to the actual flue gas denitrification system control. This step obtains real-time operating status data, combines the trained reinforcement learning model, and uses an iterative optimization method to determine the optimal ammonia injection amount. Specifically, it includes:

[0116] S41. Acquire real-time operating status data of the current flue gas denitrification system, input the real-time operating status data into the trained reinforcement learning model, predict multiple possible ammonia water injection amount adjustment actions, and obtain an initial ammonia water injection amount.

[0117] Specifically, the real-time operating status data includes flue gas temperature, flue gas flow, current ammonia injection amount, ammonia concentration, nitrogen oxide concentration before denitrification, nitrogen oxide concentration after denitrification, real-time denitrification efficiency and ammonia escape concentration. The collected real-time operating status data is preprocessed, including outlier detection, data normalization and other operations to ensure that the data quality meets the model input requirements. The preprocessed real-time operating status data is input into the trained reinforcement learning model, and the model predicts multiple possible ammonia injection amount adjustment actions based on the current state. Specifically, the model outputs a set of candidate adjustment amounts, which act on the current injection amount to obtain multiple possible initial ammonia injection amounts. Usually the top k actions with the highest Q value are selected to obtain k initial candidate injection amounts.

[0118] S42, constructing a search interval near the initial ammonia water injection amount, and calculating a comprehensive objective function value J for each candidate ammonia water injection amount within the search interval.

[0119] Specifically, a search interval is constructed with the initial ammonia injection amount as the center, and the range of the search interval is usually set to ±10% of the initial injection amount. M points are uniformly sampled within the interval as candidate ammonia injection amounts. For each candidate injection amount, the corresponding comprehensive objective function value J is calculated. i .

[0120] S43, if there is J i <J i-1 , then update the ammonia injection amount and return to step S42 to continue optimization; if there is J i >J i-1 or n consecutive iterations J i ≈J i-1 , the optimization is terminated and the model predicts the ammonia injection amount.

[0121] If there is J i <J i-1 (That is, the objective function value of the current iteration is less than the objective function value of the previous iteration), indicating that the optimization direction is correct. At this time, the initial optimal ammonia injection amount is updated, and the process returns to step S42 to continue searching for a better solution in the new search interval.

[0122] If any of the following situations occurs, the optimization process is terminated and the current model-predicted ammonia injection amount is output: i >J i-1 , indicating that continued optimization may lead to worse results; or, in n consecutive iterations, J i ≈J i-1, indicating that the optimization has reached a stable state.

[0123] The above iterative optimization strategy can ensure convergence while making full use of the knowledge learned by the reinforcement learning model to achieve precise adjustment of the ammonia injection amount. By setting reasonable termination conditions, the quality of the optimization results is ensured and the waste of computing resources caused by over-optimization is avoided. The final output model predicts the ammonia injection amount, which will serve as the basic data for the next correction.

[0124] Furthermore, in step S5, based on the prediction of the reinforcement learning model, a reaction kinetics correction mechanism is introduced to optimize the prediction results by constructing a correction relationship to obtain a more accurate ammonia injection amount control strategy. Specifically, it includes:

[0125] S51. Introduce reaction rate parameters based on the flue gas denitrification reaction model.

[0126] The reaction rate parameter reflects the speed of the reaction under specific temperature and concentration conditions, and has an important influence on the precise control of the amount of ammonia injection. The introduction of the reaction rate parameter enables the correction process to better adapt to the actual reaction kinetics. The reaction rate (r) is usually related to the amount of ammonia injection, temperature and NOx concentration. The reaction rate formula can be expressed as:

[0127]

[0128] k represents the reaction rate constant, [NO x ] represents the concentration of nitrogen oxides, [NH3] represents the concentration of ammonia water; E a represents activation energy, R represents gas constant, and T represents reaction temperature.

[0129] S52, establishing a correction relationship between the reaction rate parameter and the model-predicted ammonia injection amount, and correcting the model-predicted ammonia injection amount predicted by the reinforcement learning model based on the correction relationship, and calculating the optimal ammonia injection amount. The correction relationship between the reaction rate parameter and the model-predicted ammonia injection amount is:

[0130]

[0131] in, Indicates the optimal ammonia injection amount, that is, the corrected ammonia injection amount, [NO x ] represents the concentration of nitrogen oxides, T represents the temperature, r represents the reaction rate parameter, represents the model's predicted ammonia injection amount, represents the historical training data set, A represents the constant coefficient, and B, C, D, E and F represent the weight coefficients.

[0132] In the present invention, by introducing reaction rate parameters and historical data influencing factors, the model prediction value is dynamically adjusted to make it more suitable for actual working conditions, ensuring that the best denitrification effect and ammonia utilization rate can be achieved under different operating conditions. The optimal ammonia injection amount outputted in the end not only takes into account the current environmental status and model prediction, but also integrates historical experience and reaction kinetic characteristics, and has high accuracy and robustness.

[0133] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for accurately controlling the amount of ammonia water injection for a flue gas denitrification system, characterized in that: The following steps are involved: S1. Constructing a flue gas denitration reaction model, and constructing a functional relationship between the ammonia injection amount and the denitration efficiency based on the flue gas denitration reaction model; S2. Based on the functional relationship between the ammonia injection amount and the denitrification efficiency and the constraints of the flue gas denitrification system, a comprehensive objective function is constructed; S3. Collect historical operation data of the flue gas denitrification system and build a reinforcement learning model, in which the environmental state includes flue gas flow, nitrogen oxide concentration and ammonia escape, the action space is the adjustment amount of ammonia injection, and a reward function is built based on denitrification efficiency and ammonia escape; use historical operation data to iteratively train the reinforcement learning model, and maximize the reward function to make the model learn the optimal strategy, so that the comprehensive objective function value on the historical operation data is minimized; S4, obtaining the real-time operating status data of the current flue gas denitrification system, inputting the real-time operating status data into the trained reinforcement learning model, and outputting the model-predicted ammonia injection amount; S5. Construct a correction relationship, correct the ammonia injection amount predicted by the model based on the correction relationship, and calculate the optimal ammonia injection amount.

2. The method for accurately controlling the amount of ammonia water injection for a flue gas denitration system according to claim 1, characterized in that: Step S1 includes: Constructing flue gas denitrification reaction model: NO x +NH3→N2+H2O According to the analysis of the flue gas denitrification reaction model, the ammonia injection amount, nitrogen oxide concentration, temperature and flue gas flow rate all affect the denitrification efficiency; based on the analysis results of the flue gas denitrification reaction model, the functional relationship between the ammonia injection amount and the denitrification efficiency is constructed: Among them, η denox It represents the denitrification efficiency. Indicates the effect of ammonia injection amount, represents the influence of nitrogen oxide concentration, f T (T) represents the effect of temperature, f v (V) represents the influence of flue gas flow rate, k1, k2, k3 and k4 are weight coefficients, and k5 is a constant term, which represents the basic denitrification efficiency.

3. The method for accurately controlling the amount of ammonia water injection for a flue gas denitration system according to claim 1, characterized in that: In step S2, the calculation formula of the comprehensive objective function is: Among them, η denox is the denitrification efficiency, Q max is the maximum ammonia injection amount, [NO x ] in is the concentration of nitrogen oxides in flue gas before denitrification, [NO x ] out The concentration of nitrogen oxides in the flue gas after denitrification, w1, w2, w3 are weight coefficients.

4. The method for accurately controlling the amount of ammonia water injection for a flue gas denitration system according to claim 1, characterized in that: In step S3, a reward function is constructed based on the denitration efficiency and the ammonia escape amount, and the calculation expression of the reward function is: Among them [NO x ] in is the concentration of nitrogen oxides before denitrification, [NO x ] out is the concentration of nitrogen oxides after denitrification, NH3 is the amount of ammonia escape, is the maximum allowed ammonia escape, ω1 and ω2 are weight coefficients.

5. The method for accurately controlling the amount of ammonia water injection for a flue gas denitration system according to claim 4, characterized in that: In step S3, building a reinforcement learning model specifically includes: S31, defining the environment state and action space based on historical operation data; S32. Constructing a reward function based on denitration efficiency and ammonia escape; S33. Construct an evaluation network Q(S,A;θ) and a target network Q(S,A;θ'), and establish an experience replay pool for storing state transition samples (S,A,r,S'), where S represents the environment state, A represents the action space, θ represents the parameter set of the evaluation network, θ' represents the parameter set of the target network, r represents the immediate reward, and S' represents the next environment state after the action is executed; S34, execute model training: use the ε-greedy strategy to select actions, execute the selected actions and obtain immediate rewards r, store state transition samples in the experience replay pool, randomly sample batch data from the experience replay pool, and update the Q value through the Q value update formula; S35. Based on the Q value, determine whether the convergence condition is met. If not, return to S34 to continue training. If so, obtain the final reinforcement learning model.

6. A method for accurately controlling the amount of ammonia water injection for a flue gas denitration system according to claim 5, characterized in that: The calculation expression of the Q value update formula is: Among them, Q(S t , A t ) indicates that in state S t Next, execute action A t The expected cumulative reward; α represents the learning rate; r t Indicates execution of action A t The immediate reward after γ represents the discount factor; A′ represents the possible action for the next environment state. Indicates the next state S t+1 The maximum Q value among all possible actions.

7. A method for accurately controlling the amount of ammonia water injection for a flue gas denitration system according to claim 6, characterized in that: In step S35, the convergence condition is to meet any of the following conditions: (1) The number of training times reaches the preset maximum number of iterations N; (2) In K consecutive iterations, the maximum update amplitude of the Q value is less than the preset threshold of the Q value update amplitude; (3) In M consecutive iterations, the average value of the Q-value change of all state-action pairs is less than the preset threshold of the Q-value change.

8. The method for accurately controlling the amount of ammonia water injection for a flue gas denitration system according to claim 1, characterized in that: Step S4 includes: S41, obtaining real-time operating status data of the current flue gas denitrification system, inputting the real-time operating status data into the trained reinforcement learning model, predicting multiple possible ammonia water injection amount adjustment actions, and obtaining an initial ammonia water injection amount; S42, constructing a search interval near the initial ammonia injection amount, and calculating the comprehensive objective function value J for each candidate ammonia injection amount within the search interval i ; S43, if there is J i <J i-1 , then update the ammonia injection amount and return to step S42 to continue optimization; if there is J i >J i-1 or n consecutive iterations J i ≈J i-1 , the optimization is terminated and the model predicts the ammonia injection amount.

9. The method for accurately controlling the amount of ammonia water injection for a flue gas denitration system according to claim 1, characterized in that: Step S5 specifically includes: introducing a reaction rate parameter based on the flue gas denitration reaction model, establishing a correction relationship between the reaction rate parameter and the model predicted ammonia injection amount, correcting the model predicted ammonia injection amount predicted by the reinforcement learning model based on the correction relationship, and calculating the optimal ammonia injection amount. The correction relationship between the reaction rate parameter and the model predicted ammonia injection amount is: in, Indicates the optimal ammonia injection amount, [NO x ] represents the concentration of nitrogen oxides, T represents the temperature, r represents the reaction rate parameter, The model is used to predict the amount of ammonia injection. represents the historical training data set, A represents the constant coefficient, and B, C, D, E and F represent the weight coefficients.

10. The method for accurately controlling the amount of ammonia water injection for a flue gas denitration system according to claim 1, characterized in that: In step S3, the historical data operation data includes flue gas temperature, flue gas flow, ammonia injection amount, ammonia concentration, nitrogen oxide concentration before denitrification, nitrogen oxide concentration after denitrification, denitrification efficiency and ammonia escape concentration.

Images