Intelligent denitration control method and system based on PID (Proportion Integration Differentiation) and Transformer

Abstract

The invention discloses an intelligent denitration control method and system based on PID (Proportion Integration Differentiation) and Transform. The method comprises the following steps: acquiring an operation parameter sequence of an SCR (Selective Catalytic Reduction) denitration system in real time; a time sequence input into the Transform model and the feature dimension are exchanged to form a reverse Transform model, the predicted outlet NOx concentration is obtained, an ammonia injection amount adjusting value is calculated according to the deviation between the predicted outlet NOx concentration and the target outlet NOx concentration, the opening degree of an ammonia injection valve is adjusted in real time, and the reverse Transform model is adjusted on line based on the deviation between the actual outlet NOx concentration and the predicted outlet NOx concentration. The method is applied to the field of denitration control, the change trend of the NOx concentration in a period of time in the future can be predicted more accurately, so that the injection amount of ammonia is adjusted in advance, and the lag defect of an original PID controller is effectively overcome.

Description

Technical Field

[0001] The present invention relates to the technical field of denitration control, and in particular to a denitration intelligent control method and system based on PID and inverse transformer. Background Art

[0002] Selective catalytic reduction (SCR) denitrification technology is the mainstream solution for reducing nitrogen oxide (NOx) emissions in industrial flue gas treatment. Its principle is to inject ammonia (NH3) as a reducing agent into the flue under the action of a catalyst (such as V2O5-WO3 / TiO2) to convert NOx into harmless N2 and H2O. However, the precise control of ammonia injection amount in actual operation faces many problems. For example, from ammonia injection to NOx concentration detection, it needs to go through mixing, catalytic reaction and sensor response, with a total lag time of 30-60 seconds, and traditional feedback control is prone to overshoot or oscillation due to the lag effect.

[0003] Traditional PID controllers that rely on error integration and differential feedback are prone to integral saturation and phase delay in hysteresis systems. Although long short-term memory networks (LSTM) and recurrent neural networks (RNN) can capture time series dependencies, their gradient vanishing / exploding problems lead to the accumulation of prediction errors in very long sequences (>60 seconds). In addition, although the self-attention mechanism (Self-Attention) of the Transformer architecture can model global dependencies, fixed position encoding is difficult to adapt to the dynamic timing offset of the SCR system, and multi-head attention is prone to ignoring local key features (such as transient responses to sudden working conditions) in long sequences. Summary of the invention

[0004] In view of the problem that the ammonia injection amount is difficult to accurately control in the existing SCR denitrification process, the present invention provides a denitrification intelligent control method and system based on PID and inverse Transformer, which can more accurately predict the NOx concentration change trend in the future, so as to adjust the ammonia injection amount in advance, thereby effectively overcoming the hysteresis defect of the original PID controller.

[0005] To achieve the above object, the present invention provides a denitrification intelligent control method based on PID and inverse Transformer, comprising the following steps: Step 1: collect the operating parameter sequence of the SCR denitration system in real time and pre-process it; Step 2, swapping the time series and feature dimensions input into the Transformer model to form an inverse Transformer model, and inputting the operating parameter sequence into the inverse Transformer model to obtain a predicted outlet NOx concentration within a prediction time window; Step 3, calculating an ammonia injection amount adjustment value based on a PID controller based on a deviation between the predicted outlet NOx concentration and the target outlet NOx concentration; Step 4: adjusting the ammonia injection valve opening in real time based on the ammonia injection amount adjustment value, collecting the actual outlet NOx concentration, and adjusting the inverse Transformer model and the PID controller online based on the deviation between the actual outlet NOx concentration and the predicted outlet NOx concentration.

[0006] In one embodiment, the operating parameter sequence has multiple characteristic data sequences, including but not limited to inlet NOx concentration, outlet NOx concentration, ammonia escape rate, reactor temperature, and ammonia injection valve opening.

[0007] In one embodiment, in step 1, the preprocessing includes performing the following operations on each of the feature data sequences: First, calculate the mean μ and standard deviation σ of the feature data in the feature data sequence, and Characteristic data Mark as outliers; The abnormal points in the feature data sequence are interpolated and replaced, and the interpolated feature data sequence is normalized to: ; in, For feature data The corresponding feature data after normalization.

[0008] In one embodiment, the loss function of the inverted Transformer model during the training phase is for: ; in, To reverse the Transformer model training process The weight coefficient of the round iteration; Represents the Huber loss, specifically: ; ; in, The number of samples contained in a training batch for the inverted Transformer model, For the batch The Huber loss corresponding to samples is: For the batch The predicted value of outlet NOx concentration corresponding to each sample is: For the batch The actual value of outlet NOx concentration corresponding to the sample, is a hyperparameter; Represents the MSE loss, specifically: ; ; in, For the batch The MSE loss corresponding to each sample.

[0009] In one embodiment, the weight coefficient Specifically: ; in, , are the initial weight and final value weight respectively, and , is the attenuation coefficient, is the number of training rounds.

[0010] In one embodiment, in step 3, the ammonia injection amount adjustment value is specifically: ; ; in, is the ammonia injection amount adjustment value, is the deviation at the current moment, is the target outlet NOx concentration, To improve the predicted outlet NOx concentration T seconds into the future output by the Transformer model, is the accumulated deviation from the initial moment to the current moment, To use the interval, , , They are proportional gain, integral gain and differential gain respectively.

[0011] In one embodiment, in step 4, the real-time adjustment of the ammonia injection valve opening based on the ammonia injection amount adjustment value is specifically: like , then increase the opening of the ammonia injection valve, and the adjustment amount is ,in, is the maximum valve opening, is the adjustment coefficient; like , then reduce the opening of the ammonia injection valve, the adjustment amount is ; like , the current opening of the ammonia injection valve remains unchanged.

[0012] In one embodiment, in step 4, online adjustment of the inverse Transformer model and the PID controller specifically includes: Step 401, setting a monitoring time window, and calculating the mean absolute error between the actual outlet NOx concentration and the predicted outlet NOx concentration within each monitoring time window; Step 402, determine whether the mean absolute error of N consecutive monitoring time windows exceeds a set threshold: If yes, proceed to step 403; Otherwise, wait for calculating the mean absolute error of the next monitoring time window and then proceed to step 402 again; Step 403, after updating the learning rate of the inverse Transformer model, calculate the mean absolute error of the next monitoring time window and determine whether it is less than a set threshold: If yes, after saving the learning rate of the inverted Transformer model, proceed to step 404; Otherwise, proceed to step 403 again; Step 404: dynamically adjust the proportional gain and the integral time of the PID controller based on the mean absolute error of the current monitoring time window and the mean absolute error of the previous monitoring time window.

[0013] In one embodiment, step 404 is specifically: ; ; in, , is the adjusted proportional gain and integral time, , is the proportional gain and integral time before adjustment, is the adjustment factor, is the mean absolute error of the current monitoring time window, is the mean absolute error of the previous monitoring time window.

[0014] To achieve the above object, the present invention also provides a denitration intelligent control system based on PID and inverse transformer, which uses the above method to perform denitration intelligent control. The denitration intelligent control system includes: The data acquisition unit is used to collect the operating parameter sequence of the SCR denitration system in real time and pre-process it; A data prediction unit, used for swapping the time series and feature dimensions inputted in the Transformer model to form an inverse Transformer model, and inputting the operating parameter sequence into the inverse Transformer model to obtain a predicted outlet NOx concentration within a prediction time window; A PID calculation unit, configured to calculate an ammonia injection amount adjustment value based on a PID controller according to a deviation between the predicted outlet NOx concentration and the target outlet NOx concentration; The online adjustment unit is used to adjust the opening of the ammonia injection valve in real time according to the ammonia injection amount adjustment value, collect the actual outlet NOx concentration, and make online adjustments to the inverse Transformer model and the PID controller based on the deviation between the actual outlet NOx concentration and the predicted outlet NOx concentration.

[0015] Compared with the prior art, the present invention has the following beneficial technical effects: The present invention transfers the self-attention mechanism from the time dimension to the feature dimension by swapping the time series input in the Transformer model with the feature dimension, so as to better capture the feature information in the long sequence, and then more accurately predict the trend of NOx concentration changes in the future, so as to cooperate with the PID controller to adjust the ammonia injection amount in advance, thereby effectively overcoming the hysteresis defect of the original PID controller. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] 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 the structures shown in these drawings without paying creative work.

[0017] Figure 1 This is a flow chart of the intelligent denitration control method based on PID and inverse Transformer in Example 1 of the present invention; Figure 2 This is a flow chart of online adjustment of the inverse Transformer model and the PID controller in Embodiment 1 of the present invention; Figure 3 It is a structural block diagram of the intelligent control of denitration based on PID and inverse Transformer in Example 2 of the present invention; Figure 4 This is a structural block diagram of the terminal device in Example 3 of the present invention.

[0018] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in 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.

[0020] In addition, the technical solutions between the various embodiments of the present invention can be combined with each other, but it must be based on the fact that ordinary technicians in the field can implement it. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0021] Example 1 like Figure 1 The figure shows a denitrification intelligent control method based on PID and inverse transformer disclosed in this embodiment, which mainly includes the following steps: Step 1: collect the operating parameter sequence of the SCR denitration system in real time and pre-process it; Step 2: swap the time series and feature dimensions input into the Transformer model to form an inverse Transformer model, and input the operating parameter sequence into the inverse Transformer model to obtain the predicted outlet NOx concentration within the prediction time window, where NOx refers to nitrogen oxides such as nitric oxide, nitrogen dioxide, and nitrous oxide; Step 3, calculating the ammonia injection amount adjustment value based on the PID controller according to the deviation between the predicted outlet NOx concentration and the target outlet NOx concentration; Step 4: adjust the ammonia injection valve opening in real time based on the ammonia injection amount adjustment value, collect the actual outlet NOx concentration, and adjust the inverse Transformer model and PID controller online based on the deviation between the actual outlet NOx concentration and the predicted outlet NOx concentration.

[0022] In this embodiment, the operating parameter sequence has multiple characteristic data sequences, including inlet NOx concentration, outlet NOx concentration, ammonia escape rate, reactor temperature, ammonia injection valve opening, etc., and the frequency can be set to 1 second, 2 seconds, etc. It is worth noting that in the specific application process, the characteristic data sequence is not limited to including the aforementioned characteristic data, but can also include other characteristic data such as the amount of coal entering the furnace, its weight, and the amount of flue gas.

[0023] In the specific implementation process of step 1, the preprocessing of the operating parameter sequence includes the removal of abnormal points, supplementation of proposed points and normalization of each characteristic data sequence in the operating parameter sequence. The specific implementation process is as follows: First, use box plots, Z-scores, or other methods to identify outliers in the feature data sequence and eliminate extreme values ​​caused by noise or sensor errors. For example, calculate the mean μ and standard deviation σ of the feature data in the feature data sequence and list the features in the feature data sequence that meet the Characteristic data Mark as outliers; Then, the abnormal points in the characteristic data sequence are interpolated and replaced to ensure the continuity of the characteristic data sequence and avoid the model misjudging the periodicity due to missing data; Finally, the interpolated feature data sequence is normalized to improve the convergence speed of the data in the optimization process such as gradient descent. The normalization process is as follows: ; in, For feature data The corresponding feature data after normalization.

[0024] In the specific implementation process of step 2, this embodiment forms an inverted Transformer model by swapping the time series input with the feature dimension in the Transformer model, so that the Transformer transfers the self-attention mechanism from the time dimension to the feature dimension, thereby better capturing the feature information in the long sequence, and further more accurately predicting the NOx concentration change trend in the future.

[0025] Specifically, the input structure of the original Transformer model is (B, T, D), where B is the batch size, T is the number of time steps, and D is the feature dimension. In this embodiment, the input structure of the inverted Transformer model is (B, D, T), that is, the input time series and feature dimensions are swapped. For example, in the SCR denitrification system, the operating parameter sequence selects valve opening, ammonia concentration, inlet NOx concentration, outlet NOx concentration, and reactor temperature respectively. The original Transformer model analyzes the dependency of the time dimension T (such as the correlation between NOx concentrations at previous and subsequent moments) through the self-attention mechanism, but the interactive modeling of the feature dimension D (such as the relationship between multiple parameters such as valve opening and temperature) is insufficient. In the SCR denitrification system, the outlet NOx concentration is affected by the cross-feature synergy of multiple operating parameters (for example, ammonia concentration and reactor temperature jointly determine the reaction efficiency), and relying solely on time attention may ignore the complex coupling relationship between features. The inverted Transformer model in this embodiment shifts the focus of the self-attention mechanism from the time dimension T to the feature dimension D, and regards each feature as an independent sequence. The model focuses on analyzing the global correlation between different features (such as how valve opening affects ammonia concentration and temperature) rather than the time fluctuation of a single feature, thereby more accurately predicting the trend of NOx concentration changes in the future.

[0026] In the specific application process, the construction process of the inverted Transformer model is: first, use the tensor transpose operation to convert the original input (B, T, D) to (B, D, T); second, retain the encoder-decoder structure of the Transformer, but adjust the multi-head attention mechanism to adapt to the inverted input dimension.

[0027] The inverse Transformer model in this embodiment can be pre-trained offline and then trained online, or it can be trained directly online. For example, offline training uses historical data to train the model, and online training can be performed in the form of updating sliding window data. This embodiment constructs the loss function of the inverse Transformer model training phase by combining Huber loss and MSE loss. , specifically: ; in, To reverse the Transformer model training process The weight coefficient of the round iteration; Represents the Huber loss, specifically: ; ; in, The number of samples contained in a training batch for the inverted Transformer model, For the batch The Huber loss corresponding to samples is: For the batch The predicted value of outlet NOx concentration corresponding to each sample is: For the batch The actual value of outlet NOx concentration corresponding to the sample, is a hyperparameter used to control the transition point of the loss function from quadratic to linear. For example, it can be set to; Represents the MSE loss, specifically: ; ; in, For the batch The MSE loss corresponding to each sample.

[0028] Huber loss is highly robust to outliers (such as sensor noise and sudden changes in working conditions). When the absolute value of the prediction error exceeds the hyperparameter When , the loss function degenerates into linear growth, thus effectively avoiding abnormal samples from dominating the gradient update. However, the MSE loss is sensitive to small deviations of normal data, and the square term can be used to enhance the model's fitting accuracy for steady-state conditions. Therefore, this embodiment combines the Huber loss with the MSE loss to construct the loss function of the inverted Transformer model training phase. , which can effectively balance the robustness and accuracy of outlet NOx concentration.

[0029] As a preferred implementation, this embodiment also discloses a method for dynamically adjusting the weight coefficient in the loss function, and its specific implementation process is: ; in, , are the initial weight and final value weight respectively, and , is the attenuation coefficient used to control the transition smoothness, which can be set to , is the number of training rounds. In specific applications, it can be set to , That is, at the beginning of training, Huber Loss is given a high weight to suppress the interference of outliers. As the training progresses, it gradually transitions to MSE dominance to improve the model's fitting accuracy for regular data.

[0030] In this embodiment, the process of the PID controller calculating the ammonia injection amount adjustment value is: ; ; in, is the ammonia injection amount adjustment value, is the deviation at the current moment, is the target outlet NOx concentration, To improve the predicted outlet NOx concentration T seconds into the future output by the Transformer model, is the accumulated deviation from the initial moment to the current moment, To use the interval, , , They are proportional gain, integral gain and differential gain respectively.

[0031] The ammonia injection amount adjustment value is calculated , the ammonia injection valve opening can be adjusted in real time based on the ammonia injection amount adjustment value, and the specific implementation process is as follows: like , then increase the opening of the ammonia injection valve, and the adjustment amount is ,in, is the maximum valve opening, is the adjustment coefficient. For example, it can be set to ; like , then reduce the opening of the ammonia injection valve, the adjustment amount is ; like , the current opening of the ammonia injection valve remains unchanged.

[0032] In the specific implementation process of step 4, this embodiment adjusts the inverse Transformer model and the PID controller online according to the deviation between the actual outlet NOx concentration and the predicted outlet NOx concentration, thereby forming a closed-loop process of continuous optimization. Figure 2 ,The specific implementation process of online adjustment of the inverse Transformer model and PID controller includes the following steps: Step 401, setting a monitoring time window, and calculating the mean absolute error between the actual outlet NOx concentration and the predicted outlet NOx concentration in each monitoring time window, wherein the monitoring time window can be set to 10min~30min; Step 402, determine whether the mean absolute error of N consecutive (for example, 3) monitoring time windows exceeds a set threshold: If yes, proceed to step 403; Otherwise, wait for calculating the mean absolute error of the next monitoring time window and then proceed to step 402 again; Step 403, after updating the learning rate of the inverted Transformer model (for example, reducing the learning rate to 10% of the original value, etc.), calculate the mean absolute error of the next monitoring time window and determine whether it is less than the set threshold: If yes, save the learning rate of the inverted Transformer model and proceed to step 404; Otherwise, proceed to step 403 again; Step 404, dynamically adjust the proportional gain and integral time of the PID controller based on the mean absolute error of the current monitoring time window and the mean absolute error of the previous monitoring time window, so that the PID controller always matches the current working conditions. For example, when the model prediction error decreases, increasing the proportional gain can speed up the response speed, and shortening the integral time can enhance the integral effect and accelerate the elimination of steady-state errors, specifically: ; ; in, , is the adjusted proportional gain and integral time, , is the proportional gain and integral time before adjustment, is the mean absolute error of the current monitoring time window, is the mean absolute error of the previous monitoring time window, To prevent the adjustment coefficient from sudden changes in parameters, for example, it can be set to .

[0033] In the specific implementation process of step 403, the adjustment is not limited to the method of updating the learning rate. You can also choose to update the operating parameter sequence time window, for example, increase the existing operating parameter sequence time window by 2 hours or shorten it by 1 hour, etc. You can also freeze the embedding layer of the improved Transformer model, only train the self-attention layer and the fully connected layer, or reset the weight coefficient of the loss function, etc.

[0034] It is worth noting that although this embodiment Figure 1 , Figure 2 The steps in the process are shown in sequence as indicated by the arrows, but the steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of the steps, and the steps can be executed in other orders. Figure 1 , Figure 2At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0035] Example 2 Based on the denitration intelligent control method based on PID and inverse transformer in Example 1, this embodiment discloses a denitration intelligent control system based on PID and inverse transformer, referring to Figure 3 The denitration intelligent control system includes a data acquisition unit, a data prediction unit, a PID calculation unit and an online adjustment unit. Specifically: The data acquisition unit is used to collect the operating parameter sequence of the SCR denitration system in real time and pre-process it; The data prediction unit is used to swap the time series and feature dimensions input in the Transformer model to form an inverse Transformer model, and input the operating parameter sequence into the inverse Transformer model to obtain the predicted outlet NOx concentration within the prediction time window; The PID calculation unit is used to calculate the ammonia injection amount adjustment value based on the PID controller according to the deviation between the predicted outlet NOx concentration and the target outlet NOx concentration; The online adjustment unit is used to adjust the ammonia injection valve opening in real time according to the ammonia injection amount adjustment value, collect the actual outlet NOx concentration, and make online adjustments to the inverse Transformer model and PID controller based on the deviation between the actual outlet NOx concentration and the predicted outlet NOx concentration.

[0036] In this embodiment, the specific working process and working principle of the installation relationship calculation unit, the transfer relationship calculation unit and the alignment unit are the same as those in the method of embodiment 1, so they are not described in detail in this embodiment. Each unit module can be implemented in whole or in part by software, hardware and a combination thereof, and each unit module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above unit modules.

[0037] Example 3 like Figure 4The terminal device disclosed in this embodiment includes a transmitter, a receiver, a memory, and a processor. The transmitter is used to send instructions and data, the receiver is used to receive instructions and data, the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions stored in the memory to implement the method in the above-mentioned embodiment 1.

[0038] It should be noted that the above memory can be independent or integrated with the processor. When the memory is independently provided, the terminal device further includes a bus for connecting the memory and the processor.

[0039] Example 4 This embodiment discloses a computer-readable storage medium, in which computer-executable instructions are stored. When a processor executes the computer-executable instructions, the method in the above-mentioned embodiment 1 is implemented.

[0040] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0041] The above description is only a preferred embodiment of the present invention, and does not limit the protection scope of the present invention. All equivalent structural changes made by using the contents of the present invention specification and drawings under the inventive concept of the present invention, or directly / indirectly applied in other related technical fields are included in the protection scope of the present invention.

Claims

1. A denitrification intelligent control method based on PID and inverse Transformer, characterized in that: The steps include: Step 1: collect the operating parameter sequence of the SCR denitration system in real time and pre-process it; Step 2, swapping the time series and feature dimensions input into the Transformer model to form an inverse Transformer model, and inputting the operating parameter sequence into the inverse Transformer model to obtain a predicted outlet NOx concentration within a prediction time window; Step 3, calculating an ammonia injection amount adjustment value based on a PID controller according to a deviation between the predicted outlet NOx concentration and the target outlet NOx concentration; Step 4: adjusting the ammonia injection valve opening in real time based on the ammonia injection amount adjustment value, collecting the actual outlet NOx concentration, and adjusting the inverse Transformer model and the PID controller online based on the deviation between the actual outlet NOx concentration and the predicted outlet NOx concentration.

2. The denitrification intelligent control method based on PID and inverse Transformer according to claim 1 is characterized in that: The operating parameter sequence has a variety of characteristic data sequences, including but not limited to inlet NOx concentration, outlet NOx concentration, ammonia escape rate, reactor temperature, and ammonia injection valve opening.

3. The denitrification intelligent control method based on PID and inverse transformer according to claim 2 is characterized in that: In step 1, the preprocessing includes performing the following operations on each of the feature data sequences: First, calculate the mean μ and standard deviation σ of the feature data in the feature data sequence, and Characteristic data Mark as an outlier point; The abnormal points in the feature data sequence are interpolated and replaced, and the interpolated feature data sequence is normalized to: in, For feature data The corresponding feature data after normalization.

4. The denitrification intelligent control method based on PID and inverse transformer according to claim 1, 2 or 3, characterized in that: The loss function of the inverted Transformer model during the training phase for: in, To reverse the Transformer model training process The weight coefficient of the round iteration; Represents the Huber loss, specifically: in, The number of samples contained in a training batch for the inverted Transformer model, For the batch The Huber loss corresponding to samples is: For the batch The predicted value of outlet NOx concentration corresponding to each sample is: For the batch The actual value of outlet NOx concentration corresponding to the sample, is a hyperparameter; Represents the MSE loss, specifically: in, For the batch The MSE loss corresponding to each sample.

5. The denitrification intelligent control method based on PID and inverse Transformer according to claim 4 is characterized in that: The weight coefficient Specifically: in, , are the initial weight and final value weight respectively, and , is the attenuation coefficient, is the number of training rounds.

6. The denitrification intelligent control method based on PID and inverse transformer according to claim 1, 2 or 3, characterized in that: In step 3, the ammonia injection amount adjustment value is specifically: in, is the ammonia injection amount adjustment value, is the deviation at the current moment, is the target outlet NOx concentration, To improve the predicted outlet NOx concentration T seconds into the future output by the Transformer model, is the accumulated deviation from the initial moment to the current moment, To use the interval, , , They are proportional gain, integral gain and differential gain respectively.

7. The denitrification intelligent control method based on PID and inverse transformer according to claim 6 is characterized in that: In step 4, the real-time adjustment of the ammonia injection valve opening based on the ammonia injection amount adjustment value is specifically as follows: like , then increase the opening of the ammonia injection valve, and the adjustment amount is ,in, is the maximum valve opening, is the adjustment coefficient; like , then reduce the opening of the ammonia injection valve, the adjustment amount is ; like , the current opening of the ammonia injection valve remains unchanged.

8. The denitrification intelligent control method based on PID and inverse transformer according to claim 1, 2 or 3, characterized in that: In step 4, online adjustment of the inverse Transformer model and the PID controller specifically includes: Step 401, setting a monitoring time window, and calculating the mean absolute error between the actual outlet NOx concentration and the predicted outlet NOx concentration within each monitoring time window; Step 402, determine whether the mean absolute error of N consecutive monitoring time windows exceeds a set threshold: If yes, proceed to step 403; Otherwise, wait for calculating the mean absolute error of the next monitoring time window and then proceed to step 402 again; Step 403, after updating the learning rate of the inverse Transformer model, calculate the mean absolute error of the next monitoring time window and determine whether it is less than a set threshold: If yes, after saving the learning rate of the inverted Transformer model, proceed to step 404; Otherwise, proceed to step 403 again; Step 404: dynamically adjust the proportional gain and the integral time of the PID controller based on the mean absolute error of the current monitoring time window and the mean absolute error of the previous monitoring time window.

9. The denitrification intelligent control method based on PID and inverse transformer according to claim 8 is characterized in that: Step 404 is specifically as follows: in, , is the adjusted proportional gain and integral time, , is the proportional gain and integral time before adjustment, is the adjustment factor, is the mean absolute error of the current monitoring time window, is the mean absolute error of the previous monitoring time window.

10. A denitrification intelligent control system based on PID and inverse transformer, characterized in that: The method according to any one of claims 1 to 9 is used to perform intelligent denitration control, wherein the intelligent denitration control system comprises: The data acquisition unit is used to collect the operating parameter sequence of the SCR denitration system in real time and pre-process it; A data prediction unit, used for swapping the time series and feature dimensions inputted in the Transformer model to form an inverse Transformer model, and inputting the operating parameter sequence into the inverse Transformer model to obtain a predicted outlet NOx concentration within a prediction time window; A PID calculation unit, configured to calculate an ammonia injection amount adjustment value based on a PID controller according to a deviation between the predicted outlet NOx concentration and the target outlet NOx concentration; The online adjustment unit is used to adjust the opening of the ammonia injection valve in real time according to the ammonia injection amount adjustment value, collect the actual outlet NOx concentration, and make online adjustments to the inverse Transformer model and the PID controller based on the deviation between the actual outlet NOx concentration and the predicted outlet NOx concentration.

Images