Multi-step temperature prediction method for low-temperature catalytic desulfurization and denitrification integrated device

Abstract

The invention provides a multi-step temperature prediction method for a low-temperature catalytic desulfurization and denitrification integrated device, and belongs to the technical field of temperature prediction based on deep learning. The method comprises the following steps: firstly, collecting nine kinds of data including reaction conditions, equipment states and an optimal reaction temperature sequence; secondly, a designed data processing module is adopted to carry out noise reduction decomposition on original time series data, key features are reserved, and data quality is improved; inputting a designed and trained multi-step prediction model fusing the hierarchical ESN reserve pool and the self-attention mechanism, and generating a temperature multi-step prediction result to guide the device to regulate and control in advance; during model training, model parameters are optimized by using an improved sparrow algorithm, and efficient search of a parameter space is realized by dynamically updating a strategy through population classification. According to the method, through the synergistic effect of data enhancement, multi-scale modeling and intelligent optimization, the reaction efficiency and the operation stability are remarkably improved, and the bottleneck of traditional single-model prediction precision is broken through.

Description

Technical Field

[0001] The present invention belongs to the technical field of temperature prediction based on deep learning, and particularly relates to a multi-step temperature prediction method for a low-temperature catalytic desulfurization and denitrification integrated device. Background Art

[0002] The low-temperature catalytic desulfurization and denitrification integrated device is an innovative technology for industrial flue gas treatment. Its core lies in the synergistic removal of sulfur oxides and nitrogen oxides through low-temperature catalytic reactions. Such devices usually adopt an ammonia-free catalytic system or organic catalysts to efficiently convert pollutants within a relatively low temperature range. For example, sulfoxide catalysts are used to capture acidic gases and promote oxidation reactions, ultimately generating recoverable by-products such as ammonium sulfate and ammonium nitrate. This not only avoids the risks of ammonia escape and secondary pollution in traditional processes but also realizes resource utilization. The device design integrates a modular structure and intelligent control. For example, the segmented catalyst layout can adapt to the dynamic changes in the flue gas temperature field, while the multi-stage cyclone dust removal and demisting system can simultaneously remove particulate matter and aerosols to meet the ultra-low emission requirements. In addition, this technology breaks through the temperature limitation of catalysts in traditional technologies and can operate stably without additional heat supply, significantly reducing energy consumption and operation and maintenance costs. It is particularly suitable for low-temperature flue gas treatment scenarios in non-electric industries such as coking and steelmaking, and has become a key solution to promote the green transformation of industry.

[0003] During the actual use of the low-temperature catalytic desulfurization and denitrification integrated device, the reaction temperature of the device is a key parameter affecting the desulfurization and denitrification efficiency. In actual experiments, it is necessary to adjust the reaction temperature of the device in combination with the actual reaction state and reaction conditions of the device to enhance the desulfurization and denitrification efficiency of the device and reduce the overall energy consumption of the device. Therefore, multi-step prediction of the optimal reaction temperature based on the reaction conditions and state of the device is an essential technical link in the automatic adjustment and control of temperature.

[0004] Existing temperature prediction technologies include the following:

[0005] Temperature prediction method based on physical models: By establishing a system of nonlinear partial differential equations containing chemical reaction kinetics and the coupling mechanism of gas flow heat transfer and mass transfer, combined with the device geometric structure, catalyst activity distribution, and inlet gas parameters (such as flow rate, component concentration), and using finite element technology, the temperature prediction is realized. However, this method has too high a calculation cost in actual use, and the collected physical data is too sensitive to dynamic interference, resulting in a reduced prediction rate under actual working conditions;

[0006] Machine learning-based temperature prediction method: The machine learning-based temperature prediction method usually constructs a model through historical data and time series features, uses models such as linear regression and random forest to learn the non-linear relationship between variables, combines feature engineering to extract periodic, trend, and spatial correlation information, and improves the generalization ability of the model through cross-validation and hyperparameter optimization, and finally realizes the prediction of temperature. However, when performing multi-step temperature prediction, this method is difficult to extract the features of long time series, resulting in a low prediction accuracy;

[0007] Deep learning-based temperature prediction method: When performing multi-step temperature prediction through deep learning, a large amount of sample data needs to be collected and a deep learning neural network model needs to be constructed to obtain the temperature prediction result. However, the existing neural network models lack the processing of data noise, resulting in the existence of data noise during the actual operation of the model, which will reduce the prediction accuracy of the model. At the same time, when predicting for a relatively long time step, there is still a problem of low accuracy.

[0008] In addition, the existing methods cannot perform multi-step time prediction according to the device state conditions within a reaction period, so that more refined temperature adjustment cannot be achieved, resulting in the device being unable to perform adaptive temperature adjustment according to the actual operating state, resulting in limited reaction efficiency. Summary of the Invention

[0009] In view of the above problems, the present invention proposes a multi-step temperature prediction method for a low-temperature catalytic desulfurization and denitrification integrated device, which can automatically predict the optimal reaction temperature of the device during the temperature adjustment interval, and realize adaptive temperature adjustment throughout the operation process, which can not only reduce the consumption of human resources, but also improve the efficiency of desulfurization and denitrification operations and reduce energy consumption.

[0010] The present invention provides a multi-step temperature prediction method for a low-temperature catalytic desulfurization and denitrification integrated device, including the following processes:

[0011] S1, data is collected at time intervals of T as time steps, and M consecutive data collections are used as a set of temperature characteristic time series data. The collected data includes flue gas flow rate Yl, sulfur dioxide concentration Yn, nitride concentration Yd, dust content Yf, ammonia injection amount Yp, space velocity Yk, induced draft fan frequency Ye, heater power Yw, and device temperature Ya;

[0012] S2, the temperature characteristic time series data is denoised based on the data processing module; the IMF component set is obtained through improved empirical mode decomposition, and then combined with the energy density and frequency derivative analysis of the short-time Fourier transform, the key IMF components are dynamically screened;

[0013] S3. Input the key IMF components obtained in S2 into the trained temperature multi-step prediction model. The temperature multi-step prediction model is designed based on the self-attention mechanism and the echo state network, which fully extracts the recent features, medium-term features, and long-term features of the key IMF components and outputs the temperature prediction values at the next m time points.

[0014] Preferably, the construction method of the data set for training the temperature multi-step prediction model includes:

[0015] Carry out data collection operations in multiple low-temperature catalytic desulfurization and denitrification integrated devices. The time interval T is used as the time step, and M consecutive data are collected as a set of temperature feature time series data Datain. Based on the obtained temperature feature time series data Datain, the reaction temperature range Dya ∈ [Dyamin, Dyamax] of one device is obtained, that is, the reaction temperature of the device is between Dyamin and Dyamax at this time;

[0016] The temperature range is evenly refined into p set values and randomly assigned as the control temperature at each time point during the reaction process, so as to generate a large number of reaction temperature sequences. At the same time, based on these temperature sequences, experimental reaction tests are carried out, and the quality of these temperature sequences is evaluated according to the reaction efficiency and energy consumption after the reaction. The best reaction temperature sequence Sdya is used as the optimal reaction temperature sequence and integrated with the data Datain to obtain a complete data set Datay.

[0017] Preferably, the experimental reaction test based on these temperature sequences is for the temperature prediction target. By adjusting the temperature values at each time point within the reaction temperature range [Dyamin, Dyamax], a sulfur dioxide analyzer and a nitride analyzer are deployed to detect the sulfur dioxide concentration and nitride concentration in the flue gas before and after treatment to calculate the reaction efficiency of the device, and the optimal reaction temperature sequence at m time points is selected by using the calculated reaction efficiency and the energy consumption of the entire reaction process obtained by reading the meter.

[0018] Preferably, the process of obtaining the IMF component set by the improved empirical mode decomposition is as follows:

[0019] S21. A complete set of data Datain contains 9 specific input feature time series data Syn, Syd, Syl, Syk, Syp, Syf, Sye, Syw, Sya. The above 9 input feature time series data are spliced into an input matrix S(t) containing all input feature time series data;

[0020] Based on the total time series data S(t) obtained by splicing, the mean value of this time series data S(t) is calculated to be m, and then Gaussian white noise with an amplitude of 0.01m is added to the data S(t) to obtain the total time series data Sg(t) after noise addition; all the maximum and minimum points in this time series are calculated, and the extracted maximum points are fitted using the cubic spline interpolation method to obtain the upper envelope S upper (t), the minimum value points extracted are fitted to obtain the lower envelope S lower (t);

[0021] S22, then calculate and obtain the mean signal m(t) of the total time series data;

[0022] S23, based on the obtained mean signal m(t), subtracting the mean signal m(t) from the signal Sg(t) to obtain a preliminary modal signal h(t) of the total time series data, and calculating a standard deviation of the preliminary modal signal h(t), recorded as SD;

[0023] S24, using the obtained preliminary modal signal h(t) as a new time series signal to repeat steps S21 to S23, when its standard deviation SD is less than 0.3 three times in a row, using the preliminary modal signal h(t) calculated for the third time as the first IMF component c(t) obtained by decomposing the total time series data S(t);

[0024] S25, removing the first IMF component c(t) from the signal Sg(t) to obtain a residual signal r(t), and using the residual signal r(t) as new input data;

[0025] Iterate process S21 to process S25 until the decomposition iteration number K is reached, and the IMF component set HIMF obtained by decomposing this total time series data is obtained.

[0026] Preferably, the energy density and frequency derivative analysis combined with the short-time Fourier transform is used to dynamically screen the key IMF components, and the specific process is as follows:

[0027] First, a Hanning window with a window length of 256 points is used to perform short-time Fourier transform on the IMF components in the set HIMF to obtain its corresponding time-frequency matrix Sp(t, f), and calculate the energy density E(t, f) and frequency derivative operator of its time-frequency matrix

[0028] Then, the energy density and frequency derivative operator corresponding to each IMF component of the total time series data IMF component set HIMF are calculated, and they are recombined to obtain the IMF component frequency domain characteristic component set PIMF;

[0029] Calculate the weight factor α corresponding to each IMF component in turn k :

[0030]

[0031] Among them, sqrt(*) represents the standard deviation solving function, and its weight factor α k The larger it is, the higher the energy proportion occupied by this IMF component in the entire set of IMF components, and the smaller the frequency fluctuation of this IMF component, indicating that this component is an excellent IMF component that can better express the characteristics of this time series data;

[0032] According to the above weight factor calculation method, calculate the weight factor corresponding to each IMF component in the IMF component set;

[0033] Sort according to the weight factors corresponding to each calculated IMF component, and retain the IMF components with the top weights, and the corresponding data indices are q1, q2,..., q l , and combine them to obtain the final feature representation Ps(t) of this total time series data.

[0034] Preferably, for the temperature multi-step prediction model, Ps(t) is divided into recent data according to the time index t medium-term data and long-term data Three different time period data; for the data in different time periods, after being processed by the fully connected layer and the corresponding time memory pool, and then integrated through the fully connected layer to obtain the final optimal response temperature multi-step prediction result;

[0035] Among them, the primary features extracted from the long-term data through the fully connected layer are input into the long-term memory pool network to obtain the long-term temperature time series correlation feature LFea. The primary features extracted from the medium-term data through the fully connected layer and LFea are added and then sent into the medium-term memory pool network to obtain the medium-term temperature time series correlation feature MFea. The primary features extracted from the recent data CPs(t) through the fully connected layer and MFea are added and then input into the recent memory pool network to obtain the recent temperature time series correlation feature CFea.

[0036] Preferably, the three time memory pool networks are all composed of three ESN network units and the self-attention mechanism. The input features of the memory pool network are linearly processed by the three ESN network units in sequence, and the output features of each ESN network are concatenated to obtain the concatenated feature; and the concatenated feature is input into the self-attention mechanism for the fusion of the output features of the three ESN network units; in the self-attention mechanism, three learnable parameter matrices W q , W k , W vMap the features output by the three ESN network units in the splicing feature to three different feature spaces to obtain the feature matrices q, k, and v; subsequently, use the Softmax normalization function to fuse the features q, k, and v to obtain the output features of the temporal memory pool network;

[0037] In the designed memory pool network, the leakage rate of the ESN network unit in the short-term memory pool network is 0.9, the leakage rate of the ESN network unit in the medium-term memory pool network is 0.5, and the leakage rate of the ESN network unit in the long-term memory pool network is 0.2. The number of neurons in each ESN network unit is set to 500.

[0038] Preferably, the trained temperature multi-step prediction model is a model optimization strategy using an improved sparrow algorithm to optimize the model parameters in the temperature multi-step prediction model to find the specific model parameters corresponding to the best performance; the specific process includes:

[0039] S31, Initialize the sparrow population; optimize the spectral radius and sparsity parameters qsr, qsd of the ESN network unit in the short-term memory pool network, the spectral radius and sparsity parameters zsr, zsd of the ESN network unit in the middle-layer memory pool network, and the spectral radius and sparsity parameters ssr, ssd of the ESN network unit in the long-term memory pool network; Initialize the sparrow population, and each solution in the population contains random values of the six parameters [qsr, zsr, ssr, qsd, zsd, ssd] within the value range; complete the initialization of N solutions to obtain a sparrow population with an initialized population size of N;

[0040] S32, Use the parameters included in each population as the actual parameters used in the temperature multi-step prediction model, use the dataset to expand the model training, and use the loss function as the objective function to evaluate whether the parameters of the population are excellent;

[0041] Sort according to the calculated values of the objective function from smallest to largest; divide into excellent solutions Zmqy, intermediate solutions Zmqm, and backward solutions Zmqb; in addition, define the solution corresponding to the smallest objective function value as the best solution Zmq best ;

[0042] S33, Design different position update methods for different solution types to accelerate the optimization speed of the parameters; for the backward solutions, perform a discard operation and discard them from the total population;

[0043] S34, Iterate the processes of S32 and S33 to the target number of times, and / use the parameter set corresponding to the best population in the last iteration as the actual parameters used by the ESN network unit in the temperature multi-step prediction model to obtain the finally optimized device best reaction temperature multi-step prediction model.

[0044] Preferably, the loss function calculates the loss value Los for each prediction using the mean square error loss function based on the result Yout predicted by the model and the correct temperature sequence data Sdya in the dataset; it represents the error between the predicted temperature sequence value and the true temperature sequence value. The larger the loss value, the greater the error, and vice versa, the smaller the error. This loss function is also used as the objective function for evaluating the sparrow population parameters.

[0045] Preferably, in S31, the specific process of initializing the sparrow population is as follows:

[0046] Initialize the sparrow population. Each solution in the population contains random values of six parameters [qsr, zsr, ssr, qsd, zsd, ssd] within the value range. For the three different spectral radius parameters qsr, zsr, and ssr, the lower limit value parameter rdown = 0.3, and the upper limit value parameter rtop = 1.6. For the three different sparsity parameters qsd, zsd, and ssd, the lower limit value parameter qdown = 0.01, and the upper limit value parameter qtop = 0.3. Subsequently, design a random initialization factor xk o The calculation method is as follows:

[0047]

[0048] where Rand[0, 1] represents a random number from 0 to 1, o ∈ [1, N], and N represents the population size, that is, N solutions are set up;

[0049] where Rand[0, 1] represents a random number from 0 to 1, o ∈ [1, N], and N represents the population size, that is, N solutions are initialized;

[0050] The o-th solution Zmq obtained by initialization o represents qsr o , zer o , ssr o , qsd o , zsd o , ssd o a combination of six parameters; the calculation methods of the three spectral radius parameters qsr1, zsr1, and ssr1 are as follows: First, calculate the difference between the upper limit value parameter rtop and the lower limit value parameter rdown, and multiply the obtained difference by the random initialization factor xk o Then add this product to the lower limit value parameter rdown to obtain the calculated parameter value; qsd o , zsd o , ssd oThe calculation methods of the three sparsity parameters are as follows: First, calculate the difference between the upper limit value parameter qtop and the lower limit value parameter qdown, and multiply the obtained difference by the randomly initialized factor xk o and then add this product to the lower limit value parameter qdown to obtain the calculated parameter value.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] (1) Design of the data processing module based on the adaptive frequency-domain modal joint reconstruction algorithm: The innovation of this point lies in proposing an adaptive data processing method based on the combination of improved empirical mode decomposition (EMD) and frequency-domain feature joint screening. By introducing Gaussian white noise-assisted decomposition, iterative standard deviation termination conditions, and an IMF component screening mechanism that combines energy density and frequency derivative operators, accurate filtering of noise and efficient retention of features in time-series data are achieved. In the multi-step temperature prediction of the low-temperature catalytic desulfurization and denitrification integrated device, this method significantly improves the signal-to-noise ratio and feature purity of the training data by removing redundant components and strengthening key modal features, enabling the prediction model to more accurately capture the dynamic characteristics under complex working conditions, thereby enhancing the robustness and long-term prediction accuracy of the model and providing reliable data support for the intelligent control and optimal operation of industrial equipment.

[0053] (2) Construction of the multi-step temperature prediction model based on the self-attention mechanism and the echo state network (ESN): The innovation of this point lies in proposing a multi-step prediction model based on a time-distributed architecture. By decomposing the prediction task into independent time steps and recursively fusing historical prediction results, and combining the collaborative optimization of the multi-level ESN memory pool and the self-attention mechanism, efficient extraction of multi-scale dynamic features and accurate capture of long-term dependence relationships are achieved. In the multi-step temperature prediction of the low-temperature catalytic desulfurization and denitrification integrated device, the combination of the self-attention mechanism to strengthen the temporal correlation between features significantly improves the stability and accuracy of multi-step temperature prediction under complex working conditions.

[0054] (3) Design of the model optimization module based on the improved sparrow algorithm: The innovation of this point lies in proposing improved population initialization and update strategies, enabling the model parameters to be searched and optimized in a wider space, avoiding falling into local optimal solutions, thereby improving the optimization efficiency of the model parameters and ensuring the effect of model parameter optimization. The generalization ability of the model is improved, ensuring that when performing multi-step temperature prediction under extreme working conditions and rare working conditions, a relatively good prediction accuracy can still be maintained.

[0055] (4) The present invention designs a brand-new data input in this scenario, using flue gas flow rate, sulfur dioxide concentration, nitride concentration, dust content, ammonia injection volume, air velocity, induced draft fan frequency, heater power, and device temperature as inputs, covering the key parameters in the three stages of mass transfer, heat transfer, and reaction, and serving as the data characteristics for the interpretability of multi-step prediction of the temperature of the low-temperature catalytic desulfurization and denitrification integrated device. Description of the Drawings

[0056] Figure 1 It is a flow chart of the overall technical route of the present invention.

[0057] Figure 2 It is a flow chart of the operation of the data processing module of the present invention.

[0058] Figure 3 It is a schematic diagram of the network structure of the multi-step temperature prediction model of the present invention.

[0059] Figure 4 It is a diagram of the memory pool network architecture of the present invention.

[0060] Figure 5 It is a diagram of the comparison results of the best temperature prediction experiment in the embodiment of the present invention.

[0061] Figure 6 It is a diagram of the comparison results of the desulfurization and denitrification efficiency experiment in the embodiment of the present invention. Detailed Embodiment

[0062] The overall technical route of the present invention is as Figure 1 shown:

[0063] Construction of a multi-step temperature prediction dataset for a low-temperature catalytic desulfurization and denitrification integrated device: During the construction of the dataset, nine input features, namely flue gas flow rate (Yl), sulfur dioxide concentration (Yn), nitride concentration (Yd), dust content (Yf), ammonia injection rate (Yp), space velocity (Yk), induced draft fan speed (Ye), heater power (Yw), and device temperature (Ya), were calibrated first, covering the key parameters in the three stages of mass transfer, heat transfer, and reaction. Time-series data at 100 time points were collected at time intervals T through sensors such as sulfur dioxide analyzers, ultrasonic flow meters, and β-ray dust meters to form the input data Datain. For the temperature prediction target, by adjusting the temperature value at each time point within the reaction temperature range [Dyamin, Dyamax] (step size (Dyamax - Dyamin) / 30), a sulfur dioxide analyzer and a nitride analyzer were deployed to detect the sulfur dioxide concentration and nitride concentration in the flue gas before and after treatment to calculate the reaction efficiency of the device. The best reaction temperature sequence Sdya at 10 time points was selected by using the calculated reaction efficiency and the energy consumption of the entire reaction process obtained from reading the meter. Finally, N groups of samples containing the input data Datain and the corresponding optimal temperature label Sdya were integrated to complete the construction of the multi-step temperature prediction dataset Data;

[0064] Design of a data processing module based on an adaptive frequency-domain modal joint reconstruction algorithm: Based on the initial multi-step temperature prediction dataset collected by S1, the present invention designs an adaptive frequency-domain modal joint reconstruction algorithm to further decompose and process the time-series data in the initial dataset to enhance data expression and provide higher-quality training data for model training;

[0065] Construction of a multi-step temperature prediction model based on self-attention mechanism and echo state network (ESN): The present invention designs an echo state network based on a hierarchical ESN reservoir structure to ensure the ability to extract time-series data features at multiple time scales, and integrates a self-attention mechanism into the model to further improve the overall prediction ability of the model;

[0066] Design of a model optimization module based on an improved sparrow algorithm: To further optimize the overall performance of the multi-step temperature prediction model, the present invention designs a model optimization strategy based on an improved sparrow algorithm to optimize the model parameters in the multi-step temperature prediction model and find the specific model parameters corresponding to the best performance, achieving the effect of improving the temperature prediction accuracy;

[0067] Actual deployment and use of the model; Deploy relevant data sensors into the integrated low-temperature catalytic desulfurization and denitrification device to achieve real-time collection of various status data of the device, deploy the data processing module based on the adaptive frequency-domain modal joint reconstruction algorithm and the optimized multi-step temperature prediction model into relevant computer terminals, and send the data collected by the sensors in real time into the multi-step temperature prediction model through the operation of the data processing module to achieve multi-step temperature prediction, and adjust the temperature of the device in advance according to the obtained temperature prediction results to improve the overall reaction efficiency of the device.

[0068] The following combines specific embodiments to further illustrate the specific implementation process of the present invention.

[0069] I. Dataset construction

[0070] In order to achieve multi-step prediction of the optimal working temperature of the integrated low-temperature catalytic desulfurization and denitrification device in the present invention, it is necessary to obtain the optimal temperature sequence of the device corresponding to a certain device reaction condition during dataset construction. Using multiple integrated low-temperature catalytic desulfurization and denitrification devices as data collection carriers, and performing high-frequency temperature adjustment and testing within a reasonable reaction temperature range, and finally analyzing from the desulfurization and denitrification results and the energy consumption of the whole process to obtain the optimal temperature sequence of the device corresponding to a certain device reaction condition. The specific dataset construction process is as follows:

[0071] 1. Calibration of input data for the dataset: During the actual operation of the integrated low-temperature catalytic desulfurization and denitrification device, different flue gas conditions, reaction control parameters, and equipment status parameters will affect the optimal working temperature of the device. In the present invention, the flue gas flow rate Yl, sulfur dioxide concentration Yn, nitride concentration Yd, dust content Yf, ammonia injection amount Yp, space velocity Yk, induced draft fan frequency Ye, heater power Yw, and device temperature Ya are selected. The above selection of input features covers the mass transfer, heat transfer, and reaction stages during the desulfurization and denitrification operation, and has a high correlation with the temperature during desulfurization and denitrification operation; in addition, the difficulty of data collection during actual data collection is low, ensuring the generalization of the data and the convenience of collection;

[0072] 2. Initial data collection of input features: Based on the calibrated input features, data collection is carried out for nine characteristic data of flue gas flow rate Yl, sulfur dioxide concentration Yn, nitride concentration Yd, dust content Yf, ammonia injection volume Yp, space velocity Yk, induced draft fan speed Ye, heater power Yw, and device temperature Ya. Deploy a sulfur dioxide analyzer and a nitride analyzer to collect the sulfur dioxide concentration Yn and nitride concentration Yd data of the desulfurization and denitrification integrated device. Deploy an ultrasonic flowmeter to collect its flue gas flow rate Yl and space velocity Yk data; deploy an ammonia flowmeter to collect the ammonia injection volume Yp data; deploy a β-ray dust meter to collect the dust content Yf data; deploy an infrared sensor to collect the device temperature data Ya; the induced draft fan speed Ye and heater power Yw can be directly obtained by reading the device parameters. When carrying out data collection, the collection time interval is T, that is, sample data is collected every time interval T. Each piece of time series data collected contains sample data at 100 time points;

[0073] Syn = [Syn1, Syn2,..., Syn 100

[0074] Syd = [Syd1, Syd2,..., Syd 100

[0075] Syl = [Syl1, Syl2,..., Syl 100

[0076] Syk = [Syk1, Syk2,..., Syk 100

[0077] Syp = [Syp1, Syp2,..., Syp 100

[0078] Syf = [Syf1, Syf2,..., Syf 100

[0079] Sye = [Sye1, Sye2,..., Sye 100

[0080] Syw = [Syw1, Syw2,..., Syw 100

[0081] Sya = [Sya1, Sya2,..., Sya 100

[0082] ​​​​​​​​​Among them, Syn, Syd, Syl, Syk, Syp, Syf, Sye, Syw, and Sya respectively represent the time-series data of each group of sulfur dioxide concentration Yn, nitride concentration Yd, dust content Yf, ammonia injection amount Yp, air velocity Yk, induced draft fan speed Ye, heater power Yw, and device temperature Ya obtained by collection.

[0083] 3. Construction of the initial dataset for multi-step temperature prediction: Based on the described method for collecting initial data of input features, data collection operations are carried out in multiple actual low-temperature catalytic desulfurization and denitrification integrated devices, and an overall input data Datain is obtained.

[0084] Datain = [Syn, Syd, Syl, Syk, Syp, Syf, Sye, Syw, Sya]

[0085] In the present invention, the time adjustment interval of the reaction temperature Dya is the sampling interval time T of the input data, and the obtained reaction temperature sequence contains reaction temperature data at 10 time points. To ensure that the obtained reaction temperature sequence data is optimal, the present invention has carried out a large number of temperature tests based on the actual low-temperature catalytic desulfurization and denitrification integrated device. Based on the obtained device input feature data Datain, first, a reaction temperature range Dya ∈ [Dyamin, Dyamax] of the device is obtained, that is, the reaction temperature of the device should be between Dyamin and Dyamax at this time.

[0086] To obtain the accurate optimal reaction temperature, this temperature range is evenly refined into 30 set values and randomly assigned as the control temperature for each time point during the reaction process, thereby generating a large number of reaction temperature sequences. At the same time, based on these temperature sequences, experimental reaction tests are carried out, and the quality of these temperature sequences is evaluated according to the reaction efficiency and energy consumption after the reaction ends. Finally, after a large number of reaction tests are completed, the group of temperature sequences Sdya with the best reaction effect is used as the optimal reaction temperature sequence and integrated with the data Datain to obtain a complete dataset Datay;

[0087] Sdya = [Dya1, Dya2,..., Dya 10

[0088] Datay = [Datain, Sdya]

[0089] Based on the described dataset construction method, the optimal reaction temperature sequence is obtained under various different device state conditions Datain to complete data collection. After collecting N groups of data, the construction of the multi-step temperature prediction dataset Data is completed.

[0090] Data = [Datay1, Datay2, ..., Datay N

[0091] II. Design of Data Processing Module

[0092] The present invention designs a data processing module based on an adaptive frequency-domain modal joint reconstruction algorithm to further process the input time-series data Datain in the data set Data, strengthen the feature expression of the data, and provide more accurate and robust training data for subsequent model training to improve the accuracy of subsequent multi-step temperature prediction. The specific process is as Figure 2 shown.

[0093] S21. Improved empirical mode decomposition operation:

[0094] (1) A set of overall input data Datain contains 9 specific input feature time-series data Syn, Syd, Syl, Syk, Syp, Syf, Sye, Syw, Sya. The above 9 input feature time-series data are concatenated into an input matrix S(t) containing all input feature time-series data:

[0095] S(t) = Contact[Syn, Syd, Syl, Syk, Syp, Syf, Sye, Syw, Sya], where Contact[*] represents a data concatenation operation function;

[0096] Based on the concatenated total time-series data S(t), the mean value m of this time-series data S(t) is calculated. Subsequently, Gaussian white noise with an amplitude of 0.01m is added to the data S(t) to obtain the noisy total time-series data Sg(t). All maximum and minimum points in the current time-series data are calculated, and the extracted maximum points are fitted using cubic spline interpolation to obtain the upper envelope S upper (t), and the extracted minimum points are fitted to obtain the lower envelope S lower (t);

[0097] (2) Subsequently, the mean signal m(t) of this total time-series data is calculated;

[0098]

[0099] (3) Based on the obtained mean signal m(t), this mean signal m(t) is subtracted from the signal Sg(t) to obtain the preliminary modal signal h(t) of this total time-series data, and the standard deviation of this preliminary modal signal h(t) is calculated and denoted as SD;

[0100] h(t) = Sg(t) - m(t)

[0101] ​(4) Repeat steps (1) to (3) using the obtained preliminary modal signal h(t) as the new time series signal. When its standard deviation SD is less than 0.3 for three consecutive times, use the preliminary modal signal h(t) calculated for the third time as the first IMF component c(t) decomposed from this total time series data S(t);

[0102] (5) Remove the first IMF component c(t) from the signal Sg(t) to obtain the residual signal r(t), and use this residual signal r(t) as the new input data. Repeat the iterative process (1) to (5) until the decomposition iteration times K are reached to obtain the set of IMF components HIMC decomposed from this total time series data;

[0103] r(t) = Sg(t) - c(t)

[0104] HIMC = {IMF1(t), IMF2(t),..., IMF K (t)}

[0105] S22. Based on the set of IMF components HIMC of the total time series data obtained in the S21 process, the present invention eliminates redundant components in the HIMC set, further optimizes the component expression of the data, reduces the computational amount required in the actual use of the model, and improves the speed of multi-step temperature prediction to adapt to the continuously changing internal environment of the device. The present invention designs an IMF component screening unit based on the short-time Fourier transform and the frequency derivative operator to optimize the IMF component expression of the data. First, perform a short-time Fourier transform on the IMF components in the set HIMC using a Hanning window with a window length of 256 points to obtain its corresponding time-frequency matrix Sp(t, f), and calculate the energy density E(t, f) of the time-frequency matrix and the frequency derivative operator That is:

[0106] Sp(t,f) k = STFT(IMF k )

[0107] E(t, f) k = |Sp(t, f) k | 2

[0108]

[0109] where k ∈ [1, K], STFT(*) represents the short-time Fourier transform operation function, |*| represents the modulus operation function, f t (*) represents the first-order partial derivative function with respect to the variable t, Sp(t, f) k , E(t,f) k , They are all binary functions, where t represents the independent variable time and f represents the independent variable frequency;

[0110] S23. Based on the energy density and frequency derivative operator solving method described in S22, calculate the energy density and frequency derivative operator corresponding to each IMF component in the total time series data IMF component set HIMF, and recombine them to obtain the IMF component frequency domain feature component set PIMF:

[0111]

[0112] Among them, the first element in {*, *} represents the energy density corresponding to the IMF component, and the second element represents the frequency derivative operator corresponding to the IMF component;

[0113] S24. Calculate the weight factor α corresponding to each IMF component in turn k :

[0114]

[0115] Among them, sqrt(*) represents the standard deviation solving function, and its weight factor α k The larger it is, the higher the energy proportion occupied by this IMF component in the entire IMF component set, and the smaller the frequency fluctuation of this IMF component, indicating that this component is an excellent IMF component that can better express the characteristics of this time series data;

[0116] S25. According to the weight factor calculation method described in S24, calculate the weight factor corresponding to each IMF component in the IMF component set. According to experience, 8 IMF components can achieve a relatively accurate representation of the time series data characteristics. Therefore, in the present invention, the weight factors corresponding to each calculated IMF component are sorted, and only the first 8 IMF components with the weight factors are retained, and the corresponding data indices are q1, q2,..., q8, and they are combined to obtain the final feature representation Ps(t) of this total time series data;

[0117] Ps(t) = [IMFq1, IMFq2,..., IMFq8]

[0118] IMFq1, IMFq2,..., IMFq8 represent the IMF components with the first 8 weight factors;

[0119] S26. Perform the data processing operations on each input data Datain in the data set Data as shown in steps S21 to S25 to strengthen the feature expression of the input data Datain, and define the high-standard training data set obtained after the data processing as Datag.

[0120] III. Construction of Temperature Multi-Step Prediction Model

[0121] Since the ESN network has a low complexity during training and has good non-linear fitting ability, combined with the self-attention mechanism, it can capture various temporal dependence relationships. And when conducting actual temperature prediction, it is necessary to ensure the prediction robustness of the model under various noises and interferences, reduce the training cost and actual deployment cost of the model, and improve the adaptability of the model to various low-temperature catalytic desulfurization and denitrification integrated devices. Therefore, the present invention constructs a temperature multi-step prediction model based on the self-attention mechanism and the echo state network. Its overall architecture is as Figure 3 shown.

[0122] 1. Architecture design of the multi-step prediction model for the optimal reaction temperature: The multi-step prediction model for the optimal reaction temperature designed by the present invention adopts a hierarchical processing architecture based on the data collection time. Based on the high-standard training dataset Datag obtained after the data processing operations of the data processing module, according to the data collection time, the data in the dataset Datag is divided into data in three different time periods: recent data, medium-term data, and long-term data. For the data in different time periods, after being processed by the fully connected layer and the corresponding time memory pool, it is further integrated through the fully connected layer to obtain the final multi-step prediction result of the optimal reaction temperature. It should be noted that in order to further capture the dependence relationship between data at different times, the input data of the medium-term memory pool is the superposition of the medium-term data and the output data of the long-term memory pool, and the input data of the recent memory pool is the superposition of the recent data and the output data of the medium-term memory pool.

[0123] 2. Design of the memory pool network: The architecture of the memory pool network is as Figure 4 shown:[[]]

[0124] Based on the designed architecture of the temperature multi-step prediction model, which includes memory pool networks at three different time scales of long-term, medium-term, and recent to process the long-term data, medium-term data, and recent data divided from Datag according to the data collection time respectively. Specifically, the length of the dataset Datag is TM, and according to the time series data index, Datag is divided into three segments, namely recent data Datac, medium-term data Datam, and long-term data Datal, and the length of each segment of data is TS / 3. The following describes the design process of the memory pool network: Each time memory pool network consists of three ESN network units and the self-attention mechanism. The input data of the memory pool network is linearly processed by three ESN networks in sequence, and the output features of each ESN network are added together to obtain the concatenated feature Feadz. And Feadz is sent into the self-attention mechanism for further processing. In the self-attention mechanism, first, three learnable parameter matrices W q , W k , W vMap the feature Feadz to three different feature spaces to obtain the feature matrices q, k, and v.

[0125] q = W q * Feadz

[0126] k = W k * Feadz

[0127] v = W v * Feadz

[0128] Subsequently, use the Softmax normalization function to fuse the features q, k, and v to obtain the output feature Feaout of the data processing pool.

[0129]

[0130] Among them, Softmax(*) represents the normalization operation function, d represents the scaling factor, and k T represents the transpose of the feature matrix k.

[0131] The leakage rate parameter of the ESN network unit in the reservoir determines the memory ability of the reservoir for historical data. The higher the reservoir leakage rate, the stronger its ability to extract the features of recent data. The lower the reservoir leakage rate, the stronger its ability to extract the features of long-term data. In the memory pool network designed in the present invention, the leakage rate of the ESN network unit in the short-term memory pool is 0.9, the leakage rate of the ESN network unit in the medium-term memory pool is 0.5, and the leakage rate of the ESN network unit in the long-term memory pool is 0.2. The number of neurons in each ESN network unit is set to 500. In addition, the spectral radius and sparsity parameters of the ESN network unit in its shallow memory pool are qsr and qsd respectively, the spectral radius and sparsity parameters of the ESN network unit in the middle memory pool are zsr and zsd respectively, and the spectral radius and sparsity parameters of the ESN network unit in its long-term memory pool are ssr and ssd respectively.

[0132] 3. Implementation of the optimal reaction temperature multi-step prediction model: Based on the designed memory pool, complete the design of the optimal reaction temperature multi-step prediction model. In the temperature multi-step prediction model, the high-standard training data set Datag obtained after data operation and processing is divided into three segments: recent data, medium-term data, and long-term data according to the data collection time. Send the long-term data into the long-term data processing pool, add the output features of the medium-term data and the long-term data processing pool and send them into the medium-term data processing pool, and add the output features of the recent data and the medium-term data processing pool and send them into the recent data processing pool. Ensure that when performing temperature prediction, the input data can maintain a high prediction accuracy regardless of whether there is a time-scale change at the minute level or the hour level. Subsequently, send the outputs obtained from the three data processing pools into a fully connected layer for final processing to obtain the final temperature multi-step prediction result Yout.

[0133] IV. Model Optimization Strategy Design Based on Improved Sparrow Algorithm

[0134] Loss Function Design: To evaluate the accuracy of the model in multi-step temperature prediction and reduce the error of multi-step temperature prediction, in the present invention, according to the result Yout predicted by the model and the correct temperature sequence data Sdya in the dataset, the mean square error loss function is used to solve the loss value Los of each prediction. It represents the error between the predicted temperature sequence value and the true temperature sequence value. The larger the loss value, the greater the error; conversely, the smaller the error. This loss function is simultaneously used as the objective function for evaluating the parameters of the subsequent sparrow population.

[0135] Los = MSE(Sdya, Yout)

[0136] Where, MSE(*) represents the mean square error loss calculation function.

[0137] (1) Initialization of Sparrow Population: After completing the design of the loss function, it is also necessary to define the initialization method of the sparrow population. The optimization module designed in the present invention aims to optimize the parameters qsr, zsr, ssr, qsd, zsd, ssd of the ESN network unit. Since there are many parameters to be optimized, in order to ensure that each parameter has a larger parameter space coverage after the initialization operation, the present invention proposes an improved parameter initialization method. Compared with the general initialization method, this method not only satisfies the traversal of the initialized sparrow population within the parameter range space but also takes into account a larger population exploration space, so as to ensure that the optimized ESN network unit has better temperature prediction ability and improve the accuracy of the model in multi-step temperature prediction.

[0138] Initialize the sparrow population. Each solution in the population contains random values within the value range of the six parameters [qsr, zsr, ssr, qsd, zsd, ssd]. For the three different spectral radius parameters qsr, zsr, ssr, the lower limit value parameter rdown = 0.3, and the upper limit value parameter rtop = 1.6. For the three different sparsity parameters qsd, zsd, ssd, the lower limit value parameter qdown = 0.01, and the upper limit value parameter qtop is 0.3. Subsequently, design the calculation method of the random initialization factor xk o of.

[0139]

[0140] Where, Rand[0, 1] represents a random number from 0 to 1, o ∈ [1, N], and N represents the population size, that is, a total of N solutions are set in the initialization of the present invention.

[0141] Zmq1 = [qsr1, zer1, ssr1, qsd1, zsd1, ssd1]

[0142] In the above formula, the first solution Zmq1 obtained by initialization represents

[0143] A combination of six parameters qsr1, zer1, ssr1, qsd1, zsd1, and ssd1. Then, the randomly initialized parameters of the o-th group of the initialized population are as follows:

[0144] qsr o = rdown + (rtop - rdown) * xk o

[0145] zsr o = rdown + (rtop - rdown) * xk o

[0146] ssr o = rdown + (rtop - rdown) * xk o

[0147] qsd o = qdown + (qtop - qdown) * xk o

[0148] qsd o = qdown + (qtop - qdown) * xk o

[0149] qsd o = qdown + (qtop - qdown) * xk o

[0150] The above formula is the specific calculation method for each parameter in the solution. It should be noted that although the calculation methods are similar, the initialization factor xk corresponding to each group of parameters o is a random number between 0 and 1, that is, the initialization factor xk of each group of parameters o are all different to ensure that similar parameters have different values during the initialization process of the entire population. Initialize N solutions in the above manner to obtain a sparrow population with an initialized population size of N:

[0151] ZZmq = [Zmq1, Zmq2,..., Zmq N

[0152] ​(2) Using the sparrow population ZZmq obtained by initialization, the parameters included in each population are respectively used as the actual parameters in the temperature multi-step prediction model, and the model training is carried out using the optimized dataset, and the designed loss function is used as the objective function to evaluate whether the parameters of the population are excellent; specifically, the larger the calculation result of the loss value, the worse the performance of the corresponding population parameters, and vice versa, the better the performance. Based on this principle, the advantages and disadvantages of N groups of sparrow population parameters are calculated in turn, and they are sorted in ascending order according to the calculated value of the objective function; the solutions in the top 20% are used as excellent solutions Zmqy, the solutions in the range of 20% to 80% are used as intermediate solutions Zmqm, and the solutions in the bottom 20% are used as backward solutions Zmqb. In addition, the solution corresponding to the minimum objective function value is defined as the optimal solution Zmq best .

[0153] (3) According to the solution types finally divided in (2), the present invention designs different position update methods for different solution types to accelerate the parameter optimization speed. First, for the excellent solution Zmqy, it needs to move slightly towards the optimal solution and introduce a certain amount of randomness to further explore the optimal position.

[0154]

[0155] Among them, represents the specific parameter value of the u-th parameter in the solution after the update of the excellent population, Zmqyu represents the specific parameter value of the u-th parameter before the update of the excellent solution, represents the parameter value of the u-th parameter in the optimal solution, Rand[0, 1] represents a random number from 0 to 1, and γ is a scaling factor with a value of 0.2. For the intermediate solution, the position update criterion designed by the present invention is based on moving towards the optimal solution.

[0156]

[0157] Among them, represents the specific parameter value of the u-th parameter in the solution after the update of the intermediate solution, Zmqm u represents the specific parameter value of the u-th parameter before the update of the intermediate solution, represents the parameter value of the u-th parameter in the optimal solution, and δ is a scaling factor with a value of 0.8.

[0158] For the backward solution, a discard operation is performed to discard it from the total population.

[0159] (4) Based on the population type definition method and the sparrow population position update method described in (2) and (3), iterate the technical processes described in (2) to (3) for 6 times. And use the parameter set corresponding to the best population in the last iteration as the parameters actually used by the ESN network unit in the temperature multi-step prediction model, so as to obtain the finally optimized device's best reaction temperature multi-step prediction model.

[0160] V. Model Deployment and Experiment

[0161] Hardware Sensor Deployment: Arrange data acquisition sensors in the actually operating integrated low-temperature catalytic desulfurization and denitrification device to collect the actual operating status data of the device, including sulfur dioxide analyzers, nitride analyzers, ultrasonic flow meters, ammonia flow meters, β-ray dust meters, and infrared sensors;

[0162] Data Processing Module Deployment: Deploy the data processing module to the corresponding computer terminal to perform data processing operations on the data collected from the hardware sensors and the device parameter data read;

[0163] Temperature Prediction Model Deployment: Also deploy the optimized temperature multi-step prediction model to the computer terminal and debug this prediction model to ensure that the data processed by the data processing module can be normally sent into the temperature prediction model, and ensure that the model can successfully obtain multi-step temperature prediction results;

[0164] Temperature Multi-step Prediction Implementation: Based on the above technical links, when the integrated low-temperature catalytic desulfurization and denitrification device is actually operating, collect the state parameter data of each system of the device in real time and obtain the optimal temperature multi-step prediction results in real time;

[0165] Intelligent Regulation: According to the previously obtained temperature multi-step prediction results, adaptively optimize and adjust the actual operating temperature of the device to ensure that the device temperature and the device state are in the best matching state, thereby optimizing the operation efficiency of the equipment and reducing the energy consumption in the whole processing process.

[0166] Since the Echo State Network (ESN) and the Long Short-Term Memory Recurrent Neural Network (LSTM) have good performance in solving prediction problems, the present invention takes the Mean Absolute Percentage Error (MAPE) as the evaluation index and compares the algorithm of the present invention with these two algorithms in terms of the accuracy of the best reaction temperature. The closer the predicted best temperature is to the true best temperature (the smaller the MAPE), the higher the accuracy of this prediction result. The final experimental results are as Figure 5 shown.

[0167] Specifically, the present invention refines each reaction cycle into 10 time points for comparing the prediction accuracy of the best temperature.

[0168] According to the experimental results, it can be found that the prediction accuracy of the best reaction temperature of the algorithm of the present invention is the best throughout the entire cycle. In addition, as the time point passes, the later the temperature data, the more difficult it is to predict, so the overall prediction accuracy of the three algorithms is decreasing. However, since the present invention adopts a hierarchical processing architecture for time series data, data from different time periods can enter the corresponding memory pool network for feature extraction, so that the prediction accuracy of the algorithm of the present invention does not decrease significantly as the time point passes; in contrast, LSTM and ESN have obvious problems of decreased prediction accuracy. The experimental results verify the effectiveness of the temperature prediction architecture proposed by the present invention.

[0169] In order to further verify the efficiency improvement of the model in actual desulfurization and denitrification tasks, the present invention detects the sulfide and nitride content of the reaction gas before the desulfurization and denitrification reaction and at each reaction time point, and uses this as an evaluation indicator of the desulfurization and denitrification efficiency. Specifically, the present invention compares the efficiency of three methods: fixed temperature method, ESN predicted temperature, and LSTM predicted temperature in desulfurization and denitrification tasks. In addition, the artificial fixed temperature remains consistent throughout the entire experimental cycle, while ESN, LSTM and the present invention method adopt the multi-step refined temperature prediction strategy proposed in the present invention to dynamically predict and adjust the reaction temperature at ten time points. The final experimental results are as follows. Figure 6 shown.

[0170] According to the experimental results, it can be found that the overall sulfide and nitride content continues to decrease over time; however, since the optimal temperature in the entire reaction process changes with the environment, it is difficult for the fixed temperature method to maintain the optimal reaction efficiency throughout the entire cycle, so its denitrification and desulfurization rates are slow. ESN, LSTM and the method of the present invention adopt a multi-step refinement temperature prediction strategy to ensure that the optimal temperature can be adjusted in real time during the reaction cycle, so that the desulfurization and denitrification efficiency is significantly higher than the fixed temperature method. In addition, the present invention combines the adaptive frequency domain modal joint reconstruction algorithm and the sparrow optimization algorithm, which not only enhances the model's ability to deeply mine data features, but also further optimizes the model architecture, thereby having a higher desulfurization and denitrification efficiency than ESN and LSTM. The experimental results are consistent with the optimal temperature prediction experimental results, verifying the practicality of the algorithm of the present invention in actual desulfurization and denitrification tasks.

[0171] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0172] Although the specific implementation manners of the present invention are described above, they are not limitations on the protection scope of the present invention. Those skilled in the art should understand that, based on the technical solutions of the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present invention.

Claims

1. A multi-step temperature prediction method for an integrated low-temperature catalytic desulfurization and denitrification device, characterized in that It includes the following processes: S1. Data collection is performed with a time interval T as the time step. Data is continuously collected M times as a set of temperature feature time series data. The collected data includes flue gas flow rate Yl, sulfur dioxide concentration Yn, nitride concentration Yd, dust content Yf, ammonia injection amount Yp, space velocity Yk, induced draft fan frequency Ye, heater power Yw, and device temperature Ya; S2. Denoise the temperature feature time series data based on the data processing module; obtain the IMF component set through improved empirical mode decomposition, and then dynamically screen the key IMF components by combining the energy density and frequency derivative analysis of the short-time Fourier transform; S3. Input the key IMF components obtained in S2 into the trained temperature multi-step prediction model. The temperature multi-step prediction model is designed based on the self-attention mechanism and the echo state network, fully extracts the recent features, medium-term features, and long-term features of the key IMF components, and outputs the temperature prediction values at the next m time points.

2. The temperature multi-step prediction method for an integrated low-temperature catalytic desulfurization and denitrification device according to claim 1, characterized in that, The construction method of the data set for training the temperature multi-step prediction model includes: Data collection operations are carried out in multiple low-temperature catalytic desulfurization and denitrification integrated devices. With the time interval T as the time step, data is continuously collected M times as a set of temperature feature time series data Datain. Based on the obtained temperature feature time series data Datain, the reaction temperature range Dya ∈ [Dyamin, Dyamax] of one device is obtained, that is, the reaction temperature of the device is between Dyamin and Dyamax at this time; The temperature range is evenly refined into p set values and randomly assigned as the control temperature at each time point during the reaction process, thereby generating a large number of reaction temperature sequences; at the same time, experimental reaction tests are carried out based on these temperature sequences, and the advantages and disadvantages of these temperature sequences are evaluated according to the reaction efficiency and energy consumption after the reaction ends. The set of temperature sequences Sdya with the best reaction effect is used as the best reaction temperature sequence and integrated with the data Datain to obtain a complete data set Datay.

3. A method for multi-step prediction of the temperature of a low-temperature catalytic desulfurization and denitrification integrated device according to claim 2, characterized in that: The experimental reaction test based on these temperature sequences is for the temperature prediction target. By adjusting the temperature value at each time point within the reaction temperature range [Dyamin, Dyamax], a sulfur dioxide analyzer and a nitride analyzer are deployed to detect the sulfur dioxide concentration and nitride concentration in the flue gas before and after treatment to calculate the reaction efficiency of the device, and the best reaction temperature sequence at m time points is selected by using the calculated reaction efficiency and the energy consumption of the entire reaction process obtained from the meter reading.

4. A multi-step temperature prediction method for a low-temperature catalytic desulfurization and denitrification integrated device as described in claim 1, characterized in that: The specific process of obtaining the IMF component set through improved empirical mode decomposition is as follows: S21. A complete set of data Datain contains 9 specific input feature time series data Syn, Syd, Syl, Syk, Syp, Syf, Sye, Syw, Sya. The above 9 input feature time series data are spliced into an input matrix S(t) containing all input feature time series data; Based on the total time series data \(S(t)\) obtained by splicing, the mean value of this time series data \(S(t)\) is calculated as \(m\). Subsequently, Gaussian white noise with an amplitude of \(0.01m\) is added to the data \(S(t)\) to obtain the noise-added total time series data \(S_g(t)\); all the maximum and minimum points in this time series are calculated, and the cubic spline interpolation method is used to fit the extracted maximum points to obtain the upper envelope \(S\) upper (t), and the extracted minimum points are fitted to obtain the lower envelope \(S\) lower (t); S22, then calculate and obtain the mean signal m(t) of the total time series data; S23, based on the obtained mean signal m(t), subtracting the mean signal m(t) from the signal Sg(t) to obtain a preliminary modal signal h(t) of the total time series data, and calculating a standard deviation of the preliminary modal signal h(t), recorded as SD; S24, using the obtained preliminary modal signal h(t) as a new time series signal to repeat steps S21 to S23, when its standard deviation SD is less than 0.3 three times in a row, using the preliminary modal signal h(t) calculated for the third time as the first IMF component c(t) obtained by decomposing the total time series data S(t); S25, removing the first IMF component c(t) from the signal Sg(t) to obtain a residual signal r(t), and using the residual signal r(t) as new input data; Iterate process S21 to process S25 until the decomposition iteration number K is reached, and the IMF component set HIMF obtained by decomposing this total time series data is obtained.

5. A multi-step temperature prediction method for an integrated low-temperature catalytic desulfurization and denitrification device according to claim 4, characterized in that: The energy density and frequency derivative analysis combined with the short-time Fourier transform is used to dynamically screen the key IMF components. The specific process is as follows: First, perform a short-time Fourier transform on the IMF components in the set HIMF using a Hanning window with a window length of 256 points to obtain the corresponding time-frequency matrix Sp(t, f), and calculate the energy density E(t, f) and the frequency derivative operator of its time-frequency matrix Then, the energy density and frequency derivative operator corresponding to each IMF component of the total time series data IMF component set HIMF are calculated, and they are recombined to obtain the IMF component frequency domain characteristic component set PIMF; Calculate the weight factor α corresponding to each IMF component in sequence k : where sqrt(*) represents the standard deviation solving function, and its weighting factor α k The larger it is, the higher the energy proportion occupied by this IMF component in the entire set of IMF components, and the smaller the frequency fluctuation of this IMF component, indicating that this component is an excellent IMF component that can better express the characteristics of this time series data; According to the above weight factor calculation method, calculate the weight factor corresponding to each IMF component in the IMF component set; Sort the weight factors corresponding to each IMF component obtained by calculation, and retain the IMF components with the first \(m\) weight factors. The corresponding data indices are \(q_1, q_2, \cdots, q_m\). l Combine them to obtain the final feature representation \(P_s(t)\) of this total time series data.

6. A multi-step temperature prediction method for a low-temperature catalytic desulfurization and denitrification integrated device according to claim 1, characterized in that: The temperature multi-step prediction model divides Ps(t) into recent data according to the time index t medium-term data and long-term data Data for three different time periods; for data in different time periods, after being processed by a fully connected layer and the corresponding time memory pool, it is then integrated through a fully connected layer to obtain the final optimal reaction temperature multi-step prediction result; Among them, the primary features of the long-term data extracted by the fully connected layer are input into the long-term memory pool network to obtain the long-term temperature time series association features LFea, the primary features of the medium-term data extracted by the fully connected layer are added to LFea and then sent to the medium-term memory pool network to obtain the medium-term temperature time series association features MFea, and the primary features of the recent data CPs(t) extracted by the fully connected layer are added to MFea and then input into the recent memory pool network to obtain the recent temperature time series association features CFea.

7. A multi-step temperature prediction method for a low-temperature catalytic desulfurization and denitrification integrated device as claimed in claim 6, characterized in that: All three temporal memory pool networks are composed of three ESN network units and a self-attention mechanism. The input features of the memory pool network are linearly processed by the three ESN network units in sequence, and the output features of each ESN network are concatenated to obtain the concatenated features; and the concatenated features are input into the self-attention mechanism for fusing the output features of the three ESN network units; in the self-attention mechanism, three learnable parameter matrices W q , W k , W v are first used to map the features output by the three ESN network units in the concatenated features to three different feature spaces to obtain the feature matrices q, k, v; then the Softmax normalization function is used to fuse the features q, k, v to obtain the output features of the temporal memory pool network; In the designed memory pool network, the leakage rate of the ESN network unit in the short-term memory pool network is 0.9, the leakage rate of the ESN network unit in the medium-term memory pool network is 0.5, and the leakage rate of the ESN network unit in the long-term memory pool network is 0.

2. The number of neurons in each ESN network unit is set to 500.

8. A multi-step temperature prediction method for an integrated low-temperature catalytic desulfurization and denitrification device according to claim 7, characterized in that: The trained temperature multi-step prediction model adopts the model optimization strategy of the improved sparrow algorithm to optimize the model parameters in the temperature multi-step prediction model and find the specific model parameters corresponding to the best performance; the specific process includes: S31, Sparrow population initialization; optimize the spectral radius and sparsity parameters qsr, qsd of the ESN network units in the short-term memory pool network, the spectral radius and sparsity parameters zsr, zsd of the ESN network units in the middle-term memory pool network, and the spectral radius and sparsity parameters ssr, ssd of the ESN network units in the long-term memory pool network; initialize the sparrow population, where each solution in the population contains random values of the six parameters [qsr, zsr, ssr, qsd, zsd, ssd] within the value range; complete the initialization of N solutions to obtain a sparrow population with an initialized population size of N. S32, Use the parameters included in each population as the actual parameters used in the temperature multi-step prediction model respectively, expand the model training using the dataset, and use the loss function as the objective function to evaluate whether the parameters of the population are excellent. Sort in ascending order according to the calculated values of the objective function; divide them into excellent solutions Zmqy, intermediate solutions Zmqm, and lagging solutions Zmqb; in addition, define the solution corresponding to the minimum objective function value as the optimal solution Zmq best ; S33, Design different position update methods for different solution types to accelerate the parameter optimization speed; for backward solutions, perform a discard operation and discard them from the total population. S34, Iterate the processes of S32 and S33 to the target number of times, and / use the parameter set corresponding to the best population in the last iteration as the actual parameters used by the ESN network units in the temperature multi-step prediction model to obtain the finally optimized device's optimal reaction temperature multi-step prediction model.

9. A multi-step temperature prediction method for an integrated low-temperature catalytic desulfurization and denitration device according to claim 8, characterized in that: The loss function is to solve the loss value Los of each prediction using the mean square error loss function based on the result Yout predicted by the model and the correct temperature sequence data Sdya in the dataset; it represents the error between the predicted temperature sequence value and the true temperature sequence value. The larger the loss value, the larger the error, and vice versa, the smaller the error; this loss function is also used as the objective function for evaluating the sparrow population parameters.

10. A multi-step temperature prediction method for an integrated low-temperature catalytic desulfurization and denitrification device according to claim 8, characterized in that: In S31, the specific process of sparrow population initialization is as follows: Initialize the sparrow population. Each solution in the population contains random values of six parameters [qsr, zsr, ssr, qsd, zsd, ssd] within their value ranges. For the three different spectral radius parameters qsr, zsr, and ssr, the lower limit value parameter rdown = 0.3 and the upper limit value parameter rtop = 1.

6. For the three different sparsity parameters qsd, zsd, and ssd, the lower limit value parameter qdown = 0.01 and the upper limit value parameter qtop is 0.

3. Subsequently, design the random initialization factor xk o The calculation method is as follows: Among them, Rand[0, 1] represents a random number from 0 to 1, o ∈ [1, N], and N represents the population size, that is, N solutions are initialized. The o-th solution Zmq obtained by initialization o represents qsr o , zer o , sSr o , qsd o , zsd o , ssd o A combination of six parameters; among them, the calculation methods of the three spectral radius parameters qsr1, zsr1, and ssr1 are as follows: First, calculate the difference between the upper limit value parameter rtop and the lower limit value parameter rdown, and multiply the obtained difference by the randomly initialized factor xk o and then add the product to the lower limit value parameter rdown to obtain the calculated parameter value; qsd o , zsd o , ssd o The calculation methods of the three sparsity parameters are as follows: First, calculate the difference between the upper limit value parameter qtop and the lower limit value parameter qdown, and multiply the obtained difference by the randomly initialized factor xk o and then add the product to the lower limit value parameter qdown to obtain the calculated parameter value.

Images