A dynamic prediction and energy-saving control system for VOCs emission of a foundry plant catalytic combustion device and medium

Abstract

The application discloses a kind of casting plant catalytic combustion equipment VOCs emission dynamic prediction and energy-saving control system and medium, system includes: high-frequency VOCs concentration monitoring unit, using TDLAS and FTIR spectrometer and by data fusion algorithm obtains high-precision concentration data, wherein the secondary differential signal of TDLAS analyzer is separately extracted as key feature;Data preprocessing and feature engineering module, process multi-source data;Dynamic emission prediction module, built-in attention mechanism time series convolution network ATCN model, by fusing the secondary differential signal and historical data, realize early emission concentration prediction;Multi-objective optimization control module, using model predictive control MPC framework, with the lowest integrated operating cost as target, wherein dynamic virtual carbon penalty coefficient β can be dynamically adjusted according to workshop production plan, and rolling optimization generates advance control instruction, adjusts equipment power and fan frequency;The application realizes the early VOCs treatment, forward control and energy saving of prediction.

Description

Technical Field

[0001] This invention belongs to the field of industrial waste gas purification technology, specifically relating to a dynamic prediction and energy-saving control system and medium for VOCs emissions of catalytic combustion equipment in a foundry workshop. Background Technology

[0002] The foundry industry is a crucial basic industrial sector in my country. During its production processes, smelting, casting, core making, and sand treatment generate large amounts of volatile organic compounds (VOCs), primarily including benzene compounds, alkanes, and alcohols. These VOCs not only pollute the environment but also pose serious health risks to operators. Currently, catalytic combustion technology (RCO / RTO) is the mainstream technology for treating VOCs in foundry workshops. Through the action of a catalyst, VOCs are oxidized and decomposed into CO2 and H2O at relatively low temperatures, thereby achieving purification.

[0003] However, existing catalytic combustion equipment still faces numerous technical bottlenecks in actual operation. Firstly, in terms of monitoring, traditional systems often employ single-point sensors, such as PID or FID sensors, which are ill-suited to the harsh environment of foundries—high temperature, high humidity, and high dust levels. Measurement data is easily affected by interference and exhibits poor stability. More importantly, these sensors typically only provide instantaneous concentration values ​​and cannot reflect the changing trends of VOCs concentrations, causing the control system to remain in a passive response state.

[0004] In terms of control strategies, existing equipment generally adopts simple control based on fixed thresholds or traditional PID control. This control method has significant lag: the equipment only starts to increase the temperature and airflow when an increase in concentration is detected, and by the time the control action takes effect, the peak VOC concentration has often passed. This delayed control not only fails to effectively prevent instantaneous exceedances but also results in huge energy waste. To ensure treatment effectiveness, the equipment usually has to maintain high-power operation year-round, leading to high operating costs.

[0005] At the application level of advanced monitoring technologies, even if existing systems adopt advanced spectral technologies such as TDLAS, they are limited to using the final calculated concentration values ​​for monitoring and control. They fail to deeply explore the original physical signals, such as the dynamic trend information contained in the second harmonic signal that changes before the concentration value itself. This results in the system being unable to capture early signs of impending sudden changes in VOC concentration from the perspective of physical principles, and the essence of predictive ability has not been broken through.

[0006] At the optimization control level, the key parameters in the objective function of existing optimization algorithms, such as the penalty coefficient for exceeding emission standards, are usually fixed values. This static setting cannot adapt to the characteristics of drastic fluctuations in casting production load, and cannot achieve proactive and smooth increase in control intensity before high load arrives, or maximize energy saving effect during low load periods. This restricts the system from achieving the upper limit of global optimal performance in terms of both environmental protection and energy saving.

[0007] Furthermore, existing systems lack intelligent predictive capabilities. Casting production is inherently intermittent; for example, the pouring of large castings can instantly generate large amounts of VOCs, but traditional systems cannot anticipate such changes in operating conditions. Simultaneously, production planning information, such as pouring time and casting material, is completely isolated from the waste gas treatment system, failing to provide decision support for optimized control.

[0008] Another prominent problem is the system's lack of adaptability. Catalysts gradually deactivate during use, and equipment performance declines over time, but the parameters of traditional control systems remain constant, causing the treatment effect to gradually deteriorate over operating time and requiring frequent manual intervention and adjustments.

[0009] The existing system also falls short in handling special operating conditions. During planned shutdowns and maintenance periods, equipment often continues to operate at high power; and for sudden concentration surges, there is a lack of effective emergency control measures. This one-size-fits-all operating mode clearly cannot meet the needs of refined management.

[0010] In summary, existing catalytic combustion systems generally suffer from problems such as low monitoring accuracy, control lag, high energy consumption, and poor adaptability, and cannot meet increasingly stringent environmental protection requirements and the urgent need for energy conservation and emission reduction. Summary of the Invention

[0011] In order to overcome the problems of low monitoring accuracy, control lag, high energy consumption and poor adaptability that are common in existing catalytic combustion systems, this invention discloses a dynamic prediction and energy-saving control system for VOCs emissions of catalytic combustion equipment in a foundry workshop and a medium that can effectively solve the above-mentioned technical problems.

[0012] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0013] A dynamic prediction and energy-saving control system for VOCs emissions from a catalytic combustion device in a foundry workshop includes:

[0014] The high-frequency VOCs concentration monitoring unit includes a tunable diode laser absorption spectrometer (TDLAS) installed on the intake pipe of the catalytic combustion equipment and a Fourier transform infrared (FTIR) spectrometer installed on the exhaust pipe. It is used to collect the total hydrocarbon (THC) concentration value of the intake air and the non-methane total hydrocarbon (NMHC) concentration value of the exhaust air in real time at a frequency of not less than 1 Hz.

[0015] The equipment operating condition parameter acquisition unit is connected to the programmable logic controller (PLC) of the catalytic combustion equipment to acquire the operating condition parameters of the equipment in real time. The operating condition parameters include at least the inlet air flow, catalytic bed temperature, electric heater power and induced draft fan frequency.

[0016] The data preprocessing and feature engineering module is connected to the high-frequency VOCs concentration monitoring unit and the equipment operating parameter acquisition unit. It is used to perform wavelet transform denoising, Z-score-based standardization on the received time series data, and construct statistical features based on a sliding time window. The statistical features include at least the mean, variance and slope of the VOCs concentration within the window.

[0017] The dynamic emission prediction module is connected to the output of the data preprocessing and feature engineering module. It embeds an attention-enhanced temporal convolutional network (ATCN) model. This model takes the current and historical time-series data after preprocessing and feature engineering as input and outputs the predicted value of the exhaust non-methane total hydrocarbon (NMHC) concentration after a specific time window in the future. The time-series data includes the intake total hydrocarbon (THC) concentration, intake flow rate, and catalyst bed temperature.

[0018] The multi-objective optimization control module, communicating with the dynamic emission prediction module and the programmable logic controller (PLC), is configured to compare the predicted NMHC concentration of exhaust gas with a preset emission concentration threshold. With the optimization objectives of minimizing overall operating cost and achieving emission standards, it employs a model predictive control (MPC) framework, performing rolling optimization based on the process model of the controlled object to generate proactive control commands. These commands are used to adjust the heating power of the electric heater and the operating frequency of the induced draft fan. The overall operating cost includes the cost of electricity consumption and the virtual carbon penalty cost incurred due to excessive emissions.

[0019] Preferably, the system establishes a data buffer area to receive real-time data streams from the TDLAS analyzer and FTIR spectrometer in the high-frequency VOCs concentration monitoring unit, and processes them through a data fusion algorithm;

[0020] The data fusion algorithm is as follows: an adaptive weighted fusion algorithm is adopted, using the high-frequency response data of the TDLAS analyzer as the main body and the multi-component concentration quantitative analysis results provided by the FTIR spectrometer as the correction factor to calibrate the measured values ​​of the TDLAS analyzer online and output the final high-precision and high-reliability VOCs concentration values.

[0021] The adaptive weighted fusion algorithm is used to generate dynamic calibration coefficients based on the long-term trend deviation between the measurements of the TDLAS analyzer and the FTIR spectrometer, and to perform online calibration of the measurements of the TDLAS analyzer.

[0022] Preferably, the construction, training, and online prediction process of the attention-enhanced temporal convolutional network (ATCN) model is deeply coupled with the physical measurement characteristics of the TDLAS analyzer in the high-frequency VOCs concentration monitoring unit, specifically including:

[0023] A unique input feature of the ATCN model is the second derivative signal of the laser scanning absorption spectrum provided by the TDLAS analyzer, which is extremely sensitive to the rate of change of VOCs gas concentration.

[0024] During the model training phase, the second derivative signal and VOCs concentration time series data are used together as training inputs, enabling the ATCN model, especially its multi-head self-attention mechanism layer, to learn the intrinsic physical relationship between the rate of change of VOCs concentration and the final emission peak.

[0025] During the online prediction phase, the ATCN model utilizes real-time acquired second-differential signals to prioritize capturing the trend of impending abrupt changes in VOC concentrations. Combined with historical time-series data, it achieves ultra-early and high-precision prediction of exhaust NMHC concentrations, especially their inflection points and sudden peaks.

[0026] Preferably, the multi-objective optimization control module uses a model predictive control (MPC) framework for rolling optimization solutions, and the specific process is as follows:

[0027] Establish a finite-time domain optimization problem with a time step of minutes;

[0028] The objective function is to minimize the overall operating cost in the prediction time domain, and its mathematical expression is:

[0029]

[0030] Where J represents the overall operating cost, and N represents the prediction time-domain step size. For the first The total power of the equipment in the step, where Δt is the time step size. For the first The predicted emission concentration of the step, The emission standard limit is α, where α is the unit electricity price coefficient. For the first The dynamic virtual carbon penalty coefficient of the step;

[0031] Dynamic virtual carbon penalty coefficient The value selection strategy is as follows:

[0032] Future production plan data is obtained from the Manufacturing Execution System (MES) in the foundry through the OPC UA protocol. The future production plan data includes at least the material of the castings to be produced.

[0033] If the castings planned for production in a specific future period are made of materials with high VOC emissions, then the VOC emissions will dynamically increase within the corresponding prediction time domain. The value of is chosen to increase the penalty weight of exceeding emission standards in the optimization objective function, forcing the optimization algorithm to adopt a more conservative strategy of increasing heating power before this period arrives;

[0034] If the castings planned for production in a specific future period are made of materials with low strength due to VOCs, then the production capacity will be reduced accordingly. The value of is chosen to give the optimization algorithm more room for energy-saving adjustment;

[0035] The constraints include the operating range of the catalyst bed temperature, the rate of increase and decrease of the electric heater power, and the range of frequency adjustment of the induced draft fan.

[0036] The Particle Swarm Optimization (PSO) algorithm is used to solve the optimization problem online, obtain the optimal control command sequence, and apply the first control command in the sequence to the controlled device.

[0037] Preferably, the multi-objective optimization control module is also configured with a multi-mode switching strategy:

[0038] When a planned shutdown and maintenance signal is received from the Manufacturing Execution System (MES) in the foundry, the system automatically switches to sleep mode. The control command will enable the catalytic combustion equipment to operate at the lowest power while maintaining the lowest catalyst activity temperature.

[0039] When the dynamic emission prediction module predicts that the intake THC concentration will increase sharply within the next 5 minutes, it automatically switches to the pre-response mode and increases the power of the electric heater in advance based on the prediction results to avoid exceeding the emission peak.

[0040] Preferably, the system further includes:

[0041] The model performance monitoring and online update module continuously compares the predicted and actual measured values ​​of exhaust non-methane total hydrocarbon (NMHC) concentration.

[0042] When the prediction error exceeds the preset tolerance for multiple consecutive cycles, or when a systematic drift in the operating conditions caused by catalyst deactivation or equipment performance degradation is detected, the model update process is automatically triggered.

[0043] The update process uses the latest data to perform online incremental learning on the last convolutional layer and subsequent layers of the temporal convolutional network ATCN model to adapt to slowly changing operating conditions.

[0044] Preferably, the data preprocessing and feature engineering module also receives data from multi-channel photoionization detectors (PID sensors) arranged in the sand processing, core making, and pouring areas of the foundry workshop, for monitoring the VOCs concentration distribution in the workshop environment. The environmental VOCs concentration data, after preprocessing, is input as an auxiliary feature to the dynamic emission prediction module.

[0045] Preferably, the TDLAS analyzer in the high-frequency VOCs concentration monitoring unit selects a specific laser wavelength for scanning and measuring the characteristic absorption spectra of typical benzene series compounds and alkane gases in the VOCs composition of the foundry.

[0046] Preferably, a computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the aforementioned VOCs emission dynamic prediction and energy-saving control system.

[0047] The beneficial effects of this invention are as follows:

[0048] This invention achieves ultra-early and high-precision prediction of VOCs concentrations. It combines a tunable diode laser absorption spectroscopy (TDLAS) analyzer with a Fourier transform infrared (FTIR) spectrometer, and innovatively employs an adaptive weighted fusion algorithm to process the measurement data from both. This fully leverages the advantages of TDLAS's high-frequency response and FTIR's multi-component quantitative analysis. Crucially, this invention breaks away from the traditional approach of using only concentration values, innovatively extracting and utilizing the second-order differential signal from the TDLAS analyzer as a unique input feature of the ATCN model. This signal is extremely sensitive to the rate of change in gas concentration, enabling the prediction model to capture trends prior to changes in the concentration value itself. This achieves ultra-early and high-precision prediction of inflection points and sudden peaks in emission concentration changes, providing crucial early warning time for control systems and fundamentally solving the prediction lag problem.

[0049] This invention achieves intelligent and forward-looking multi-objective optimization control, employing a Model Predictive Control (MPC) framework. With the optimization objectives of minimizing energy consumption cost and virtual carbon penalty cost, this invention innovatively makes the dynamic virtual carbon penalty coefficient β in the objective function a dynamic variable. This coefficient can adaptively adjust based on future production plans obtained from the Manufacturing Execution System (MES), such as casting materials: when facing production plans for high-strength materials, it automatically increases the β value, forcing the system to adopt a conservative strategy of enhanced combustion in advance to eliminate the risk of exceeding standards; when facing low-strength materials, it decreases the β value, giving the algorithm more room for energy-saving optimization, realizing a leap from passive response to proactive optimization based on production situation awareness, and balancing the contradiction between environmental protection and energy conservation.

[0050] Breaking through the limitations of traditional control modes, it adopts the Model Predictive Control (MPC) framework, with the optimization objective of minimizing power consumption cost and virtual carbon penalty cost. Under the premise of strictly ensuring emission compliance, the optimal control command is obtained through rolling optimization, realizing the transformation from passive response to active optimization, and fundamentally solving the contradiction between control lag and high energy consumption.

[0051] The system's intelligence level has been improved. Through multi-mode switching strategies, the system can intelligently identify special operating conditions, such as planned shutdowns and sudden increases in concentration, and automatically adjust the control strategy to ensure treatment effectiveness while maximizing energy-saving potential, demonstrating the system's high level of intelligence.

[0052] This ensures the long-term stability of the system. Through model performance monitoring and online update mechanisms, the system can automatically detect slow changes such as catalyst deactivation and equipment performance degradation, and trigger incremental learning processes to fine-tune the prediction model. This enables the system to adapt to changes in operating conditions, maintain long-term operational stability, and reduce the need for manual intervention.

[0053] This invention has developed a complete intelligent manufacturing solution. It deeply integrates advanced sensing technology, industrial IoT, artificial intelligence algorithms and advanced process control technology to realize a closed-loop intelligent system from accurate monitoring and intelligent prediction to optimized control. It provides an efficient, reliable and energy-saving overall solution for VOCs treatment in the casting industry and has important engineering application value.

[0054] With promising economic and environmental benefits, the system can effectively reduce energy consumption and prevent emissions from exceeding standards while ensuring that emissions meet standards through accurate prediction and optimized control, thus achieving good economic and environmental benefits.

[0055] In summary, this invention, through multiple collaborative innovations in the sensing layer, prediction layer, and control layer, effectively solves the problems of low monitoring accuracy, control lag, high energy consumption, and poor adaptability in existing catalytic combustion systems, providing an effective technical means to achieve intelligent, refined, and efficient VOCs treatment in the casting industry. Attached Figure Description

[0056] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other embodiments can be derived from the provided drawings without creative effort.

[0057] Figure 1The diagram below shows the structural block diagram of a VOCs emission dynamic prediction and energy-saving control system for a catalytic combustion equipment in a foundry workshop, as provided in Embodiment 1 of the present invention. Detailed Implementation

[0058] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, exemplary embodiments will be described in detail below, examples of which are illustrated in the accompanying drawings. In the following description relating to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of methods and systems consistent with some aspects of this application as detailed in the appended claims.

[0059] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0060] The following detailed description of the specific implementation methods, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided in detail.

[0061] Example 1

[0062] Please see Figure 1 A dynamic prediction and energy-saving control system for VOCs emissions from a catalytic combustion equipment in a foundry workshop includes the following steps:

[0063] The high-frequency VOCs concentration monitoring unit includes a tunable diode laser absorption spectrometer (TDLAS) installed on the intake pipe of the catalytic combustion equipment and a Fourier transform infrared (FTIR) spectrometer installed on the exhaust pipe. It is used to collect the total hydrocarbon (THC) concentration value of the intake air and the non-methane total hydrocarbon (NMHC) concentration value of the exhaust air in real time at a frequency of not less than 1 Hz.

[0064] The equipment operating condition parameter acquisition unit is connected to the programmable logic controller (PLC) of the catalytic combustion equipment to acquire the operating condition parameters of the equipment in real time. The operating condition parameters include at least the inlet air flow, catalytic bed temperature, electric heater power and induced draft fan frequency.

[0065] The data preprocessing and feature engineering module is connected to the high-frequency VOCs concentration monitoring unit and the equipment operating parameter acquisition unit. It is used to perform wavelet transform denoising, Z-score-based standardization on the received time series data, and construct statistical features based on a sliding time window. The statistical features include at least the mean, variance and slope of the VOCs concentration within the window.

[0066] The dynamic emission prediction module is connected to the output of the data preprocessing and feature engineering module. It embeds an attention-enhanced temporal convolutional network (ATCN) model. This model takes the current and historical time-series data after preprocessing and feature engineering as input and outputs the predicted value of the exhaust non-methane total hydrocarbon (NMHC) concentration after a specific time window in the future. The time-series data includes the intake total hydrocarbon (THC) concentration, intake flow rate, and catalyst bed temperature.

[0067] The multi-objective optimization control module, communicating with the dynamic emission prediction module and the programmable logic controller (PLC), is configured to compare the predicted NMHC concentration of exhaust gas with a preset emission concentration threshold. With the optimization objectives of minimizing overall operating cost and achieving emission standards, it employs a model predictive control (MPC) framework, performing rolling optimization based on the process model of the controlled object to generate proactive control commands. These commands are used to adjust the heating power of the electric heater and the operating frequency of the induced draft fan. The overall operating cost includes the cost of electricity consumption and the virtual carbon penalty cost incurred due to excessive emissions.

[0068] Furthermore, the system establishes a data buffer area to receive real-time data streams from the TDLAS analyzer and FTIR spectrometer in the high-frequency VOCs concentration monitoring unit, and processes them through a data fusion algorithm.

[0069] Since the response times and analysis cycles of the two instruments are different, the system first performs time-series alignment of the two instruments based on the data timestamps to ensure that the data used for fusion corresponds to the flue gas samples at the same sampling time.

[0070] The data fusion algorithm runs online according to the following logical steps:

[0071] An adaptive weighted fusion algorithm is adopted, which uses the high-frequency response data of the TDLAS analyzer as the main body of fusion. Because the TDLAS analyzer has the advantage of extremely fast response speed, down to the second level, it can capture instantaneous changes in concentration and meet the needs of real-time control.

[0072] Using the multi-component concentration quantitative analysis results provided by the FTIR spectrometer as a correction factor, the FTIR spectrometer can accurately identify and quantify multiple VOCs components. Its measurement results are highly accurate, but the analysis cycle is usually long (minutes).

[0073] One implementation of the adaptive weighted fusion algorithm is to use the high-precision quantitative results of the FTIR spectrometer to calibrate the possible drift or error of the TDLAS analyzer's measurements;

[0074] The data processed by the above fusion algorithm is the final high-precision and high-reliability VOCs concentration value output, which retains the high-frequency characteristics of TDLAS and has the quantitative accuracy of FTIR.

[0075] It is worth mentioning that the specific calibration logic for calibrating the measured values ​​of the TDLAS analyzer can be as follows:

[0076] The system continuously monitors the long-term trend deviation between the measurements taken by the TDLAS analyzer and the FTIR spectrometer.

[0077] When new data from the FTIR spectrometer becomes available, the algorithm calculates the difference between the TDLAS analyzer value and the FTIR spectrometer value at that moment.

[0078] Based on this difference, a dynamic calibration coefficient is generated, such as an offset or a scaling factor, and applied to the TDLAS analyzer high-frequency data over a subsequent period until the next FTIR spectrometer data point arrives.

[0079] Furthermore, the construction, training, and online prediction processes of the attention-enhanced temporal convolutional network (ATCN) model are deeply coupled with the physical measurement characteristics of the TDLAS analyzer in the high-frequency VOCs concentration monitoring unit, specifically including:

[0080] The TDLAS analyzer uses wavelength modulation spectroscopy technology. Its laser output wavelength is modulated by the center frequency of the absorption spectrum of the target gas, while it is modulated by a high-frequency sinusoidal signal. The signal received by the detector is demodulated by the lock-in amplifier, which can simultaneously obtain the first harmonic, second harmonic and even higher harmonic signals.

[0081] This invention innovatively selects the second harmonic signal as a unique input feature of the ATCN model. It is the second derivative signal of the laser scanning absorption spectrum provided by the TDLAS analyzer. This second derivative signal has a peak at the center of the absorption spectral line, and its peak amplitude is proportional to the gas concentration. At the same time, due to its sharper waveform and better symmetry, it is extremely sensitive to the rate of change of VOCs gas concentration. When the VOCs concentration begins to rise slightly but rapidly, the concentration value itself may not have reached the high threshold, but the amplitude of the second derivative signal has already changed significantly, thus becoming an early detection of concentration abrupt changes.

[0082] The system is equipped with a high-precision data acquisition card to ensure that the concentration value output by the TDLAS analyzer comes from the second harmonic peak amplitude and the second differential signal, that is, the original waveform data of the second harmonic, and has the same timestamp.

[0083] Feature extraction is performed on the second-differential signal, and statistical characteristics such as root mean square value, peak-to-peak value and waveform factor are calculated within a sliding time window to quantify its intensity and quality.

[0084] The extracted second-order differential signal features are normalized and time-aligned with the VOCs concentration time series data, air flow rate, catalyst bed temperature and other operating parameters in the same time period, and then concatenated into a multi-dimensional feature vector, which is used as the input of the ATCN model.

[0085] The multi-head self-attention mechanism layer of the ATCN model automatically learns and assigns higher attention weights to the features of the second-differential signal by calculating the correlation between the query, key, and value vectors. This means that the model is trained to focus on these weak physical signals that indicate an impending drastic change in concentration.

[0086] During the model training phase, a historical dataset containing a large number of typical working conditions, especially those that cause a sudden increase in concentration, such as pouring operations, is used. The model learns specific patterns of the second differential signal through training, such as the inherent physical correlation and mathematical mapping relationship between the rapid jump in amplitude and the VOCs concentration peak that appears after 5-15 sampling periods.

[0087] During the online prediction phase, the system inputs the latest second-order differential signal features along with historical data into the pre-trained ATCN model in real time.

[0088] Because the second derivative signal is sensitive to the rate of concentration change before the concentration value itself changes, the ATCN model can capture the trend of an impending sudden change in VOCs concentration. Its output is no longer a simple concentration prediction, but a super early prediction result with higher confidence that incorporates physical precursor signals.

[0089] This method enables ultra-early and high-precision prediction of exhaust non-methane total hydrocarbon (NMHC) concentrations, especially their inflection points and sudden peaks. Compared with models that only use historical concentration values, the early warning time provided by this invention can be tens of seconds to several minutes in advance, providing a crucial decision window for subsequent multi-objective optimization control modules and fundamentally avoiding the lag in control actions.

[0090] Furthermore, the multi-objective optimization control module employs a model predictive control (MPC) framework for rolling optimization solutions. The specific process is as follows:

[0091] Establish a finite-time domain optimization problem with a time step of minutes;

[0092] The objective function is to minimize the overall operating cost in the prediction time domain, and its mathematical expression is:

[0093]

[0094] Where J represents the overall operating cost, and N represents the prediction time-domain step size. For the first The total power of the catalytic combustion equipment in the first step mainly consists of the total electrical power of the electric heater and the induced draft fan, where Δt is the time step. For the first The predicted emission concentrations for this step are output by the ATCN model. The emission standard limit is α, where α is the unit electricity price coefficient. For the first The dynamic virtual carbon penalty coefficient of the step;

[0095] When the system starts, it loads the MPC optimization parameters from the configuration file, including: prediction time step N, time step Δt, unit electricity price coefficient α (yuan / kWh), and upper and lower limits and rise and fall rate limits of equipment constraints such as catalyst bed temperature, electric heater power, and induced draft fan frequency;

[0096] The MPC objective function J is designed to minimize the overall operating cost over a future period, which consists of two parts:

[0097] Cost of electricity consumption: Virtual carbon penalty cost: ;

[0098] Dynamic virtual carbon penalty coefficient The value selection strategy is as follows:

[0099] Using the OPC UA protocol, future production plan data is obtained in real time from the Manufacturing Execution System (MES) in the foundry, with a focus on analyzing the material field of the castings to be produced.

[0100] The system has a pre-stored strength mapping table for casting materials and VOCs. For example: ductile iron QT series -> high strength; gray cast iron HT series -> medium strength; metal mold castings -> low strength. This module compares the materials in the MES plan with the mapping table and executes the following strategy:

[0101] If, during a specific future period, the material of the castings planned for production in steps k1 to k2 of the prediction time domain is determined to be a material with high VOCs production intensity, then during that period, the dynamic increase... The value of ;

[0102] By significantly increasing the penalty weight of emission exceedance items in the objective function, the MPC optimization algorithm tends to adopt a conservative strategy of increasing heating power and air volume to avoid high penalties in the future. This makes full preparations for treating the expected high concentration of exhaust gas and greatly reduces the risk of exceeding the standard.

[0103] If the material of the castings planned for production in a specific future period is determined to be a material with low VOCs production strength, then the corresponding dynamic reduction will be implemented. The value of is chosen to give the optimization algorithm more room for energy-saving adjustment;

[0104] This is equivalent to relaxing emission constraints in the optimization objective, giving the optimization algorithm greater freedom to find energy-saving strategies, such as reducing power and reducing air volume, thereby deeply exploring energy-saving potential and reducing operating costs under absolutely safe operating conditions;

[0105] At the beginning of each control cycle, i.e., at each time step Δt, the following closed-loop process is executed:

[0106] The system reads the current system status from the on-site PLC, such as the catalyst bed temperature and current power, and obtains the ultra-early predicted concentration values ​​for the next N steps from the dynamic emission prediction module.

[0107] Based on the current state, predicted values, and data determined by MES. The sequence is used to construct a complete optimization problem, and the particle swarm optimization (PSO) algorithm is used to solve the problem online to obtain a sequence of optimal control commands.

[0108] The electric heater power and induced draft fan frequency of the first control command in the sequence are sent to the PLC for execution;

[0109] Entering the next control cycle, the above process is repeated, and a new round of prediction and optimization is carried out based on the latest system state, thereby achieving the advanced control effect of rolling optimization and feedback correction.

[0110] Furthermore, the multi-objective optimization control module is also configured with a multi-mode switching strategy:

[0111] In the multi-objective optimization control module, a mode switching logic unit is set up. This logic unit continuously monitors the status from two key signal sources: the MES interface of the foundry workshop, which subscribes to the planned shutdown and maintenance signal via the OPC UA protocol; and the dynamic emission prediction module, which provides the predicted trend of intake THC concentration for the next 5 minutes.

[0112] When the mode switching logic unit receives a high-priority planned shutdown maintenance signal from the foundry workshop manufacturing execution system (MES), which is typically a Boolean value of TRUE, it immediately triggers a mode switch.

[0113] The system automatically switches from the default MPC optimization mode to hibernation mode;

[0114] In hibernation mode, the optimization objective function J of MPC is temporarily modified, and the virtual carbon penalty term is set to zero. Since production is about to stop, there is no risk of exceeding the limit. The optimization objective is simplified to the lowest energy consumption cost under the premise of meeting the minimum constraints.

[0115] MPC will solve for a series of control commands, which will gradually reduce the catalytic combustion equipment to a level that can only maintain the minimum active temperature of the catalyst, such as 250°C, and significantly reduce the frequency of the induced draft fan to the minimum value that meets the basic ventilation requirements of the workshop, so as to operate at the lowest power.

[0116] Upon receiving the production resumption signal from the Manufacturing Execution System (MES) in the foundry workshop, the system exits sleep mode, the controller begins to heat up in advance, and switches back to MPC optimization mode after the operating conditions stabilize.

[0117] The mode switching logic unit analyzes the prediction data provided by the dynamic emission prediction module in real time. When the dynamic emission prediction module predicts that the intake THC concentration will increase sharply within the next 5 minutes, if the intake THC concentration will rise from the current value to more than 50% of its average value or other configurable thresholds, it is determined to be a sharp increase in concentration and immediately triggers mode switching.

[0118] The system automatically switches from MPC optimization mode to pre-response mode;

[0119] In the pre-response mode, MPC no longer waits for the optimization solution, but instead adopts the feedforward control principle based on the predicted concentration change magnitude.

[0120] Based on a preset concentration-power mapping table, a step command to significantly increase the power of the electric heater is directly generated. For example, the power is immediately increased to 80% of the rated power to store heat for the upcoming high-concentration exhaust gas in advance. The power of the electric heater is increased in advance based on the prediction results to avoid exceeding the emission peak.

[0121] Once the actual intake air concentration begins to decrease and the prediction model indicates that the peak risk has been eliminated, the system exits the pre-response mode and reverts to the MPC optimization mode for fine-tuning.

[0122] Furthermore, the VOCs emission dynamic prediction and energy-saving control system also includes:

[0123] The model performance monitoring and online update module continuously receives exhaust non-methane total hydrocarbon (NMHC) concentration measurements from the high-frequency VOCs concentration monitoring unit and actual predicted values ​​from the dynamic emission prediction module.

[0124] It calculates the prediction absolute error and rolling root mean square error (RMSE) for each control cycle in real time, such as 10 seconds, based on data from the most recent hour.

[0125] The module contains two parallel trigger judgment logics:

[0126] Error exceeding the limit judgment: If the absolute prediction error exceeds 5% of the preset tolerance for 30 consecutive periods, the model performance is judged to have degraded, and the update process is triggered.

[0127] Systematic drift detection employs the Statistical Process Control (SPC) method to plot a control chart of the prediction error. If the error sequence shows a continuous upward trend or exceeds the upper / lower limit of the control chart, it indicates that there is systematic operating condition drift caused by catalyst deactivation or equipment performance degradation, which also triggers the update process.

[0128] Once the update process is triggered, the system automatically starts a background learning task;

[0129] Collect high-quality time-series data from a recent period, such as the past 7 days, after data preprocessing and feature engineering, as a new training dataset;

[0130] Employ fine-tuning strategies from transfer learning;

[0131] First, freeze all the preceding layers of the already trained attention-enhanced temporal convolutional network (ATCN) model. These layers learn general temporal features, such as trends and periodic patterns, which usually do not need to be changed.

[0132] Then, only the last convolutional layer and subsequent layers of the attention-enhanced temporal convolutional network ATCN model are unlocked. These layers are responsible for mapping abstract features to specific concentration prediction values ​​and are most sensitive to changes in working conditions.

[0133] Using a new dataset, the unlocked layers are trained incrementally with a smaller learning rate, with the training objective being to minimize the prediction error on the new data.

[0134] After training, the performance of the fine-tuned model is evaluated using a validation set. If the performance improvement is significant, such as a reduction of more than 20% in the root mean square error (RMSE) or a return to an acceptable range, the old model parameters currently in use online are seamlessly replaced with the new model parameters to complete the online update.

[0135] Furthermore, the data preprocessing and feature engineering module also receives data from multi-channel photoionization detectors (PID sensors) arranged in the sand processing, core making, and pouring areas of the foundry, for monitoring the VOCs concentration distribution in the workshop environment;

[0136] Each area's PID sensor collects the environmental VOCs concentration value in real time at a set sampling frequency, for example, once per minute. Multi-channel means that one host can connect to multiple probes, which are arranged in different parts to achieve distributed measurement.

[0137] The data preprocessing and feature engineering module receives measurement data from all the aforementioned PID sensors in real time via industrial bus or industrial Ethernet protocol.

[0138] The transmitted data packet contains a sensor ID, which identifies the specific process, timestamp, and concentration measurement value;

[0139] The received environmental VOCs concentration data must undergo the same preprocessing procedure as the main process data to ensure data quality and consistency.

[0140] The Raida criterion is used to eliminate outliers caused by transient sensor interference.

[0141] For temporary data loss caused by communication interruption, linear interpolation is used to fill in the missing data.

[0142] The moving average method is used to smooth the data in order to eliminate random fluctuation noise;

[0143] All PID sensor data timestamps are strictly aligned with the data timestamps of equipment such as the intake manifold TDLAS analyzer to ensure that data points at the same time correspond to the same operating conditions.

[0144] After preprocessing, the data from each PID sensor is treated as an independent auxiliary feature;

[0145] These features are combined with key features such as total hydrocarbon (THC) concentration in the intake air, intake air flow rate, and catalyst bed temperature to form an enhanced feature vector with higher dimensions and integrated workshop environmental information after preprocessing.

[0146] Ultimately, this enhanced feature vector is input as a whole into the dynamic emissions prediction module for model training and real-time prediction.

[0147] Furthermore, the TDLAS analyzer in the high-frequency VOCs concentration monitoring unit selects a specific laser wavelength for scanning and measuring the characteristic absorption spectra of typical benzene series compounds and alkane gases in the VOCs composition of the foundry.

[0148] First, gas chromatography-mass spectrometry analysis was performed on VOCs exhaust gas from typical foundry workshops, such as iron casting and steel casting workshops, to identify typical benzene series compounds, such as benzene, toluene, xylene, and alkane gases, such as methane, ethane, and propane, as the main components and their approximate concentration ranges.

[0149] Consult the HITRAN spectral database or relevant standards to identify the characteristic absorption lines of the aforementioned typical benzene series compounds and alkane gases in the mid-infrared band;

[0150] The principle for selecting absorption lines is to ensure sufficient detection sensitivity, to select lines that interfere less with other background gases such as water vapor and carbon dioxide absorption lines, and to ensure that the center wavelength of the selected line is within the tunable range of a commercially available semiconductor laser.

[0151] Based on the selected characteristic absorption spectral lines, determine their center wavelength, and configure a distributed feedback DFB laser with an emission wavelength that can cover the specific laser wavelength of the center wavelength for the TDLAS analyzer;

[0152] In the software of the TDLAS analyzer, set the scanning control parameters of the laser so that its output wavelength can accurately and at high frequency scan through the characteristic absorption spectral line;

[0153] The TDLAS analyzer detects the absorption intensity of a laser at a specific laser wavelength after the laser passes through the flue gas, and calculates the concentration of the gas being measured in real time based on the Lambert-Beer law.

[0154] This concentration value, namely the total hydrocarbon (THC) concentration in the intake air, is output to the data preprocessing and feature engineering module in real time.

[0155] Furthermore, this system can be implemented through a computer-readable storage medium containing a computer program. When the computer program is executed by a processor, it implements the aforementioned VOCs emission dynamic prediction and energy-saving control system.

Claims

1. A dynamic prediction and energy-saving control system for VOCs emission of a foundry plant catalytic combustion equipment, characterized in that, The system comprises: a high-frequency VOCs concentration monitoring unit, comprising a tunable diode laser absorption spectroscopy (TDLAS) analyzer arranged on an air inlet pipeline of the catalytic combustion device and a Fourier transform infrared (FTIR) spectrometer arranged on an exhaust pipeline, for collecting total hydrocarbon (THC) concentration values of the inlet air and non-methane hydrocarbon (NMHC) concentration values of the exhaust air in real time at a frequency of no less than 1 Hz; a device working condition parameter acquisition unit, which is in communication connection with a programmable logic controller (PLC) of the catalytic combustion device, for acquiring running working condition parameters of the device in real time, the running working condition parameters at least including inlet air flow, catalytic bed temperature, electric heater power and induced draft fan frequency; a data preprocessing and feature engineering module, which is in signal connection with the high-frequency VOCs concentration monitoring unit and the device working condition parameter acquisition unit, for performing wavelet transform denoising, Z-score-based standardization processing on the received time series data, and constructing statistical features based on a sliding time window, the statistical features at least including mean, variance and slope of VOCs concentration within the window; a dynamic emission prediction module, which is connected to an output end of the data preprocessing and feature engineering module, and in which an attention mechanism enhanced time series convolution network (ATCN) model is embedded, the model taking current and historical time series data after preprocessing and feature engineering as input, and outputting non-methane hydrocarbon (NMHC) concentration prediction values after a future time window; the time series data including total hydrocarbon (THC) concentration values of the inlet air, inlet air flow, catalytic bed temperature; a multi-objective optimization control module, which is in communication connection with the dynamic emission prediction module and the programmable logic controller (PLC), and is configured to compare the non-methane hydrocarbon (NMHC) concentration prediction values with a preset emission concentration threshold value; taking the lowest comprehensive operation cost and emission standard compliance as optimization objectives, using a model predictive control (MPC) framework, performing rolling optimization solution based on a process model of the controlled object, and generating a lead control instruction; the control instruction is used for adjusting heating power of the electric heater and operating frequency of the induced draft fan; the comprehensive operation cost includes electric energy consumption cost and virtual carbon penalty cost due to excessive emission; wherein the system sets up a data buffer area, respectively receives real-time data streams from the TDLAS analyzer and the FTIR spectrometer in the high-frequency VOCs concentration monitoring unit, and processes the data streams through a data fusion algorithm; the data fusion algorithm specifically comprises: using an adaptive weighted fusion algorithm, taking high-frequency response data of the TDLAS analyzer as the main body, taking multi-component concentration quantitative analysis results provided by the FTIR spectrometer as a correction factor, performing online calibration on the measurement values of the TDLAS analyzer, and outputting VOCs concentration values with high accuracy and high reliability; the adaptive weighted fusion algorithm is used for generating a dynamic calibration coefficient to perform online calibration on the measurement values of the TDLAS analyzer according to long-term trend deviation of the measurement values of the TDLAS analyzer and the FTIR spectrometer. The construction, training and online prediction process of the attention mechanism enhanced time sequence convolution network ATCN model are deeply coupled with the physical measurement characteristics of the TDLAS analyzer in the high-frequency VOCs concentration monitoring unit, and specifically include: An exclusive input feature of the ATCN model is a second derivative signal of a laser scanning absorption spectrum provided by the TDLAS analyzer, which is extremely sensitive to the rate of change of VOCs gas concentration; In the model training stage, the second derivative signal is used as a training input together with VOCs concentration time sequence data, so that the multi-head self-attention mechanism layer of the ATCN model can learn the internal physical correlation between the VOCs concentration change rate and the final emission peak value; In the online prediction stage, the ATCN model uses the real-time collected second derivative signal to preferentially capture the trend that VOCs concentration is about to mutate, and then combines historical time sequence data to realize the ultra-early and high-precision prediction of the change inflection point and burst peak value of the exhaust NMHC concentration.

2. The VOCs emission dynamic prediction and energy saving control system according to claim 1, characterized in that, The multi-objective optimization control module adopts a model predictive control MPC framework for rolling optimization solution, and the specific process is as follows: A finite time domain optimization problem is established with a minute level as a time step; The objective function is to minimize the comprehensive operation cost in the prediction time domain, which is mathematically expressed as: ； where J is the comprehensive operation cost, N is the prediction time domain step, is the total power of the device at the step, and Δt is the time step, is the predicted emission concentration at the step, is the emission standard limit value, and α is the unit electricity price coefficient, is the dynamic virtual carbon penalty coefficient at the step. The dynamic virtual carbon penalty coefficient The value strategy is as follows: Future production plan data is obtained from the casting plant manufacturing execution system MES through the OPC UA protocol, and the future production plan data at least includes the planned production of castings materials; If the castings planned for production in the future are made of materials with high VOCs production strength, then the dynamic increase will occur within the corresponding prediction time domain for that period. The value of is chosen to increase the penalty weight of exceeding emission standards in the optimization objective function, forcing the optimization algorithm to adopt a more conservative strategy of increasing heating power before this period arrives; If the casting material planned to be produced in the future period is a material with low VOCs generation intensity, the value of the VOCs generation intensity is reduced accordingly to give the optimization algorithm a larger energy-saving adjustment space. The constraint conditions include the catalytic bed temperature working range constraint, the electric heater power rise rate constraint, and the induced draft fan frequency regulation range constraint; The particle swarm optimization PSO algorithm is used to solve the optimization problem online to obtain an optimal control instruction sequence, and the first control instruction in the sequence is applied to the controlled device.

3. The VOCs emission dynamic prediction and energy saving control system according to claim 2, wherein, The multi-objective optimization control module is also configured with a multi-mode switching strategy: When a planned shutdown and maintenance signal is received from the casting plant manufacturing execution system MES, the system automatically switches to the dormant mode, and the control instruction will make the catalytic combustion device run at the lowest power while maintaining the minimum activity temperature of the catalyst; When the dynamic emission prediction module predicts that the intake THC concentration will suddenly increase within the next 5 minutes, the system automatically switches to the pre-response mode, and the electric heater power is increased in advance based on the prediction result to avoid the emission peak exceeding the standard.

4. The VOCs emission dynamic prediction and energy saving control system according to claim 2, wherein, The system also includes: A model performance monitoring and online updating module that continuously compares the predicted value of the exhaust non-methane hydrocarbon NMHC concentration with the actual measured value; When the prediction error exceeds the preset tolerance for consecutive multiple cycles, or when a systematic drift in working conditions caused by catalyst deactivation or device performance degradation is detected, the model updating process is automatically triggered; The updating process uses the latest data to perform online incremental learning on the last convolution layer and subsequent layers of the time sequence convolution network ATCN model to adapt to slowly changing working conditions.

5. The VOCs emission dynamic prediction and energy saving control system according to claim 1, wherein, The data preprocessing and feature engineering module also receives multi-channel photoionization detector (PID) sensor data from the casting plant sand handling, core making, and pouring areas to monitor the concentration distribution of VOCs in the plant environment. The preprocessed environmental VOCs concentration data are used as auxiliary features input to the dynamic emission prediction module.

6. The VOCs emission dynamic prediction and energy saving control system according to claim 1, wherein, The TDLAS analyzer in the high-frequency VOCs concentration monitoring unit selects laser wavelengths for scanning measurements based on the characteristic absorption spectral lines of typical benzene series and alkane gases in the VOCs composition of the casting plant.

7. A computer readable storage medium characterized in that, The computer readable storage medium stores a computer program that, when executed by a processor, implements the VOCs emission dynamic prediction and energy-saving control system of any one of claims 1 to 6.

Images