A VOCs carbon can running state prediction method and system based on deep machine learning

Abstract

The application discloses a VOCs carbon tank running state prediction method and system based on deep machine learning, and relates to the technical field of data processing. The VOCs carbon tank running state prediction method based on deep machine learning extracts an adsorption time sequence parameter set by acquiring running state time sequence data; extracts a corrected saturation and a pressure difference change rate based on a multi-factor adsorption state deduction of an activated carbon adsorption capacity model; generates a breakthrough warning signal and constructs a candidate time sequence window when a breakthrough condition is met; predicts a regeneration performance index set and a corresponding confidence interval through attention perception of a long short-term memory network; extracts current working condition information to infer and predict a regeneration dynamic optimization weight set through a large language model, the application evaluates the regeneration performance index set through dynamic optimization weights, takes the confidence interval as a risk evaluation factor and incorporates it into the evaluation process, and outputs a corresponding running state prediction result, so as to realize comprehensive prediction of the activated carbon adsorption state and the regeneration running state.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for predicting the operating status of VOCs carbon canisters based on deep machine learning. Background Technology

[0002] In activated carbon adsorption technology, when waste gas passes through an activated carbon bed, the VOCs in it are adsorbed and retained by the activated carbon, and the purified gas meets the emission standards. As the adsorption process continues, the adsorption sites of the activated carbon are gradually occupied by VOCs molecules, and its adsorption capacity continues to decrease. When the activated carbon is close to the adsorption saturation state, it needs to be regenerated. The adsorbed VOCs are released from the activated carbon through thermal desorption, restoring its adsorption capacity and realizing the recycling of activated carbon. In the actual operation of activated carbon adsorption and regeneration, the adsorption stage and the regeneration stage constitute a reciprocating cycle. The two stages influence each other and jointly determine the overall operating state.

[0003] The limitations of existing technologies include at least the following problems: the actual adsorption state of activated carbon in existing technologies is affected by the coupling of multiple factors, and the saturation varies under different operating conditions for the same running time. The time-periodic inference method is difficult to accurately describe the true adsorption state of the carbon canister, resulting in uncertainty in the judgment of the adsorption endpoint. On this basis, when entering the regeneration stage, there is a lack of means to perceive the degree of performance recovery after regeneration in advance. At the same time, it is difficult to evaluate the comprehensive effect under different regeneration operating parameters for carbon canisters with different operating histories and different decay states. This makes it difficult to effectively predict the actual effect of the regeneration operation before operation, thus affecting the advance judgment of the treatment effect. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for predicting the operating status of VOCs activated carbon canisters based on deep machine learning, which solves the problems of inaccurate characterization of adsorption status and lack of advance prediction and difference assessment of regeneration effect in existing technologies.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the operating status of VOCs activated carbon canisters based on deep machine learning, comprising the following steps: acquiring time-series data of the operating status of VOCs activated carbon canisters and extracting a set of adsorption time-series parameters for VOCs activated carbon canisters; based on a pre-established activated carbon adsorption capacity model and combined with the adsorption time-series parameter set, extracting the corrected saturation and differential pressure change rate of VOCs activated carbon canisters through multi-factor adsorption state deduction; when the corrected saturation and differential pressure change rate meet preset breakthrough conditions, issuing a breakthrough warning signal, and constructing multiple sets of candidate conditions based on the operating status time-series data. Select a time series window; input each group of candidate time series windows into a pre-trained long short-term memory network, and after adjusting for saturation and pressure difference rate of change weighted attention perception, predict the regeneration performance index set and corresponding confidence interval for each group of candidate time series windows; extract current operating condition information from the operating status time series data, and input it into a pre-trained large language model for semantic reasoning, predicting and outputting the regeneration dynamic optimization weight set; evaluate the regeneration performance index set corresponding to each group of candidate time series windows according to the dynamic optimization weight, incorporate the confidence interval as a risk penalty term into the evaluation, deduce the VOCs carbon tank operating status prediction result and output it.

[0006] Furthermore, the specific steps for extracting the corrected saturation of the VOCs activated carbon canister are as follows: Extract the inlet and outlet VOCs concentration sequences, airflow sequences, temperature and humidity sequences, and activated carbon canister pressure difference sequences from the adsorption time series parameter set; input the inlet and outlet VOCs concentration sequences and airflow sequences into the activated carbon adsorption capacity model to extract the adsorption increment per unit time. : ;

[0007] in, Let be the air volume at the i-th sampling time. Let VOCs concentration at the inlet be the concentration at the i-th sampling time. Let VOCs concentration at the outlet be the concentration at the i-th sampling time. The sampling time interval, The adsorption increment per unit time corresponds to the i-th sampling time; the cumulative adsorption amount is obtained by summing all adsorption increments per unit time along the time axis. : ;

[0008] Obtain the actual mass of activated carbon filled in the VOCs activated carbon canister and maximum adsorption capacity per unit mass And combined with the cumulative adsorption amount After comprehensive processing, the basic saturation is obtained. : Substituting the temperature and humidity sequence and the carbon canister pressure difference sequence into the preset corresponding functions, the temperature and humidity correction factors are obtained. and differential pressure correction factor and the basic saturation Correction processing is performed to obtain the corrected saturation of VOCs in the carbon canister. : .

[0009] Furthermore, when the adsorption time series parameter set lacks inlet VOCs concentration, the corrected saturation is extracted through the following steps: obtaining the initial removal efficiency and performing time decay processing to obtain the removal efficiency sequence; estimating the inlet VOCs concentration sequence based on the outlet VOCs concentration sequence and the removal efficiency sequence; comprehensively processing the inlet and outlet VOCs concentration sequences and the air volume sequence to obtain the basic saturation; and extracting the current canister pressure difference and the rate of change of pressure difference as dynamic correction factors based on the canister pressure difference sequence to correct the basic saturation and obtain the corrected saturation.

[0010] Furthermore, the specific steps for predicting the regeneration performance index set corresponding to each group of candidate time-series windows are as follows: Each group of candidate time-series windows is input into a pre-trained Long Short-Term Memory (LSTM) network, and the hidden states at each time step are extracted via forward propagation; the corrected saturation and pressure change rate are concatenated to generate an attention query vector, and the hidden states at each time step are used as key and value vectors, respectively, to map them to the attention weights at each time step; the attention weights at each time step are dynamically reconstructed based on the corrected saturation; the corresponding value vectors are weighted and summed using the dynamically reconstructed attention weights at each time step to obtain weighted attention features, which are then input into the fully connected layer of the LSM network. After mapping by the fully connected layer, the regeneration performance index set corresponding to each candidate time-series window is output.

[0011] Furthermore, the specific steps for predicting the confidence intervals corresponding to the regeneration performance index set for each candidate time window are as follows: embed a dropout layer before the fully connected layer of the long short-term memory network; perform multiple independent forward propagations for the same candidate time window, with the dropout layer randomly blocking different neurons in each forward propagation, and output multiple sets of prediction results corresponding to the candidate time window respectively; calculate the mean and standard deviation of each index in the multiple sets of prediction results, and jointly determine the confidence intervals of each index in the regeneration performance index set.

[0012] Furthermore, a training sample set containing multiple historical time-series windows and corresponding real regeneration performance index labels is constructed, and each historical time-series window is divided into historical segments and future segments. Each historical time-series window is sequentially input into a Long Short-Term Memory (LSTM) network, and after forward propagation, the corresponding predicted values ​​of regeneration performance indexes, predicted values ​​of outlet VOCs concentration, and estimated values ​​of adsorption increment are output in each branch of the network. The mean square error between the predicted values ​​of regeneration performance indexes and the real regeneration performance index labels is used as the data fitting term. Based on the future segments, the predicted values ​​of outlet VOCs concentration and the estimated values ​​of adsorption increment are used to calculate the mass conservation constraint term. Based on the data fitting term and the mass conservation constraint term, a bi-objective loss function is constructed. The parameters of the LSM network are updated through backpropagation with the minimization of the bi-objective loss function as the optimization objective, thus completing the pre-training.

[0013] Furthermore, the specific steps for predicting the output regeneration dynamic optimization weight set are as follows: convert the values ​​and states of each working condition parameter in the current working condition information into natural language description text; take each indicator in the regeneration performance index set as the optimization objective and each parameter in the current working condition information as the constraint condition, and input the natural language description text into the pre-trained large language model.

[0014] Semantic reasoning is performed on the coupling relationship between optimization objectives and constraints using a large language model, outputting the priority ranking results of each optimization objective; the priority ranking results are then mapped to the weight coefficients corresponding to each optimization objective. : Where j represents the number of the current optimization objective, and k represents the numbers of all optimization objectives participating in the weight calculation. Prioritize the j-th optimization objective ( Indicates the highest priority. (Indicates the lowest priority) This is the attenuation coefficient (with a value of 0.5). Let the weight coefficients be those for the j-th optimization objective. The priority ranking values ​​are the corresponding to the k optimization objectives.

[0015] Furthermore, the specific steps for outputting the predicted operating status of the VOCs charcoal canister are as follows:

[0016] For each group of candidate time series windows, the regeneration dynamic optimization weight set is used as the evaluation coefficient to comprehensively evaluate the corresponding regeneration performance index set, and the evaluation score of each group of candidate time series windows is obtained; the interval width of the confidence interval is extracted, and the risk penalty value of each group of candidate time series windows is determined; based on the evaluation score and risk penalty value of each group of candidate time series windows, the comprehensive evaluation score of each group of candidate time series windows is obtained; the candidate time series window with the highest comprehensive evaluation score is selected as the optimal operating state prediction result of the VOCs carbon tank and output.

[0017] Furthermore, when the adsorption analysis parameter set lacks inlet VOCs concentration, after outputting the corrected saturation, the following steps are also included: real-time monitoring of outlet VOCs concentration and carbon canister pressure difference, and determining whether self-calibration is triggered; in response to the triggering of self-calibration, obtaining the removal efficiency based on the increase in carbon canister pressure difference and outlet VOCs concentration through a preset mapping relationship; comparing the removal efficiency with the initial removal efficiency, and updating the decay coefficient in the time decay model based on the comparison result; recalculating the removal efficiency estimate with the updated decay coefficient, and completing the self-calibration of the saturation estimation result.

[0018] A deep machine learning-based VOCs activated carbon canister operation status prediction system includes: a data acquisition unit for extracting the adsorption time series parameter set of the VOCs activated carbon canister; an adsorption status inference unit for inferring the corrected saturation and differential pressure change rate of the VOCs activated carbon canister based on a pre-established activated carbon adsorption capacity model; a penetration warning unit for determining whether a penetration warning signal is issued; a candidate time series construction unit for constructing multiple sets of candidate time series windows; a deep prediction unit for inputting each set of candidate time series windows into a pre-trained long short-term memory network, and predicting the regeneration performance index set and corresponding confidence intervals for each set of candidate time series windows through weighted attention perception based on corrected saturation and differential pressure change rate; a weight generation unit for extracting current operating condition information and inputting it into a pre-trained large language model to predict and output a regeneration dynamic optimization weight set; and an evaluation prediction unit for evaluating the regeneration performance index set of each candidate window using the regeneration dynamic optimization weight set, incorporating the confidence interval as a risk penalty term in the evaluation, and inferring and outputting the VOCs activated carbon canister operation status prediction result.

[0019] The present invention has the following beneficial effects:

[0020] (1) The VOCs carbon canister operation status prediction method based on deep machine learning calculates the adsorption increment per unit time by using the difference between the inlet and outlet VOCs concentrations and accumulates them to obtain the cumulative adsorption amount. The basic saturation is converted by the filling amount and the maximum adsorption capacity. Then, temperature and humidity correction factors and pressure difference correction factors are introduced to correct the basic saturation. Finally, the corrected saturation reflecting the true state of the carbon canister is obtained. This method can comprehensively consider the influence of temperature, humidity, pressure difference and adsorption process changes on the state of the carbon canister under different operating conditions. It also provides an inversion estimation method for the inlet concentration for operating conditions with missing inlet concentration monitoring data. The inlet concentration is calculated by the outlet concentration and time decay model, and the saturation estimation process is corrected by combining the pressure difference change rate.

[0021] (2) The VOCs activated carbon canister operation status prediction method based on deep machine learning uses a long short-term memory network based on historical time series data. By modifying the saturation and pressure difference change rate weighted attention perception, the model selectively focuses on historical information, so that when predicting regeneration effect, waste gas treatment efficiency, energy consumption level and activated carbon decay rate, it pays more attention to the time period related to the current adsorption state. The same candidate window is forward propagated multiple times to obtain multiple sets of prediction results. The confidence interval of each index is determined by the mean and standard deviation. In the pre-training stage, the mass conservation constraint term and the data fitting term are introduced to form a dual-objective loss function, so that the network prediction process can comprehensively utilize historical time series information, current adsorption state information and prediction uncertainty information.

[0022] (3) The VOCs carbon canister operation status prediction method based on deep machine learning converts the current operating condition parameters into natural language descriptions and inputs them into a large language model for semantic reasoning. It outputs the priority ranking of each optimization objective and maps it into weight coefficients, so that the weight of the optimization objective can be automatically adapted according to the current operating condition. The confidence interval width is included as a risk penalty item in the comprehensive evaluation, so that both the prediction effect and the prediction reliability are considered when predicting the operation status. Thus, the evaluation and screening of regeneration status prediction under different operating conditions are realized.

[0023] (4) The VOCs activated carbon canister operation status prediction system based on deep machine learning extracts the adsorption time series parameter set through the data acquisition unit, the adsorption status inference unit infers and corrects the saturation and pressure difference change rate based on the activated carbon adsorption capacity model, the penetration early warning unit triggers the early warning signal according to the preset conditions, the candidate time series construction unit generates multiple sets of candidate time series windows, the deep prediction unit predicts the regeneration performance index set and confidence interval corresponding to each set of windows with a long short-term memory network, the weight generation unit infers and outputs the dynamically optimized weight set with a large language model, and the evaluation prediction unit evaluates each set of prediction results with the weight set and incorporates the confidence interval as a risk penalty item into the score, and finally infers the optimal operation status prediction result, thereby improving the accurate perception of saturation in the adsorption stage and covering the reliability of the operation effect prediction in the regeneration stage.

[0024] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0025] Figure 1 This is a flowchart of a method for predicting the operating status of VOCs carbon canisters based on deep machine learning, according to the present invention.

[0026] Figure 2 This is a flowchart illustrating the specific steps of predicting the regeneration dynamic optimization weight set in the VOCs carbon canister operation status prediction method based on deep machine learning of the present invention.

[0027] Figure 3 This is a block diagram of a VOCs carbon canister operation status prediction system based on deep machine learning according to the present invention. Detailed Implementation

[0028] Please see Figure 1 The present invention provides a technical solution: a method for predicting the operating status of VOCs carbon canisters based on deep machine learning, comprising the following steps: acquiring time series data of the operating status of VOCs carbon canisters and extracting the adsorption time series parameter set of VOCs carbon canisters;

[0029] Based on a pre-established activated carbon adsorption capacity model and combined with an adsorption time series parameter set, the corrected saturation and differential pressure change rate of VOCs in the activated carbon canister are extracted through multi-factor adsorption state deduction.

[0030] If the correction saturation and differential pressure change rate meet the preset penetration conditions, a penetration warning signal will be issued;

[0031] In response to the penetration warning signal, based on the time series data of the operating status, the future time data of the three dimensions of temperature, air volume and duration are used as the variables to be predicted, the future time data of the other dimensions are set to be the same as the current time, and multiple sets of candidate time series windows are constructed, each set of candidate time series windows corresponds to different combinations of values ​​of the variables to be predicted;

[0032] Each candidate time series window is input into a pre-trained long short-term memory network. After adjusting the saturation and pressure difference change rate weighted attention perception, the regeneration performance index set and corresponding confidence interval of each candidate time series window are predicted. The regeneration performance index set includes regeneration efficiency, waste gas treatment efficiency, energy consumption level and activated carbon decay rate.

[0033] Extract current operating condition information from the time series data of the operating status, input it into the pre-trained large language model for semantic reasoning, and predict the output to regenerate the dynamically optimized weight set;

[0034] The regeneration performance index set corresponding to each candidate time series window is evaluated based on the dynamic optimization weight. The confidence interval is included in the evaluation as a risk penalty term. The candidate time series window that makes the evaluation result optimal is deduced and output as the operation status prediction result of the VOCs carbon tank.

[0035] Among them, the operating status time series data are multi-dimensional time series data collected by sensing devices deployed in each process link during the operation of VOCs carbon canister (activated carbon adsorption canister) and mobile regeneration equipment according to a preset sampling period. The sampling period is determined according to the sensor response time and the VOCs carbon canister status change period. In one embodiment, the sampling period can be set to 10 seconds.

[0036] The time-series data of the operating status includes the following core operating parameters: temperature and humidity of adsorbed gas, temperature and humidity of regenerated gas, adsorption air volume, regenerated gas air volume, pressure difference of carbon canister, multi-point temperature of catalyst bed, running time of adsorption section, running time of regeneration section, regeneration cycle and switching frequency, inlet VOCs concentration and outlet VOCs concentration, and VOCs component type. The above parameters are stored uniformly after being aligned by timestamps, forming a full-process multi-dimensional structured time-series database covering adsorption process, mobile regeneration process and catalytic oxidation process.

[0037] Extract the inlet VOCs concentration sequence required for subsequent calculations from this database. Export VOCs concentration sequence Air volume sequence , temperature and humidity sequence and the differential pressure sequence of the carbon canister This constitutes an adsorption time series parameter set, where t is the time index, representing the sampling time of each parameter.

[0038] The specific steps for constructing multiple candidate time series windows are as follows:

[0039] Each candidate time series window contains a length of... A historical timeline and a length of The future time series segment, and the historical time series segment, are all operational status time series data within a fixed duration before the triggering of the penetration warning. In this embodiment... The value is 128 sampling points, and this historical segment remains unchanged when constructing each group of candidate windows;

[0040] In the future time series segment, the data of the three dimensions of regenerated gas temperature, regenerated gas volume and duration are set as variables to be predicted, and all combinations of their respective value ranges are filled in according to the preset step size.

[0041] In one embodiment, the effective range of the regeneration gas temperature is 150℃~300℃ (not exceeding the maximum tolerance temperature of activated carbon), with a step size of 25℃, for a total of 7 values; the range of the regeneration gas flow rate is 2000~6000m³ / h, with a step size of 500m³ / h, for a total of 9 values; and the range of the regeneration time is 1.5h~4.0h (not less than the shortest effective desorption time), with a step size of 0.5h, for a total of 6 values.

[0042] Among them, the regeneration duration, as a scalar control parameter, is copied and expanded to a length of... The constant sequence is filled into the input vector at each sampling time of the future segment, so that the model can infer the final performance under different regeneration times based on this feature;

[0043] In one implementation, The value is 12, while the minimum range of the regeneration time is 1.5h. Therefore, all moments in the future segment are in the regeneration operation stage, and the regeneration gas temperature and air volume are kept constant. As for the inlet VOCs concentration, considering that the introduction of exhaust gas is usually stopped during the regeneration stage, it is filled to 0 in the future segment.

[0044] For the pressure differential in the charcoal canister, data is supplemented based on the start / stop status of the fan during the regeneration phase. When the fan is in bypass operation, since the charcoal canister is in a non-adsorption state, the pressure differential baseline value determined by historical stable operating data under bypass operation is used for filling. The pressure differential baseline value is obtained by collecting pressure differential data over multiple consecutive sampling periods under fan bypass operation and based on statistical results. When the fan stops operating and no new pressure differential data is acquired, the pressure differential value at the previous valid sampling time is used for maintenance. Other non-control dimensions, such as temperature and humidity, can copy the current value and extend it to a length of [length missing]. The sequence is filled;

[0045] For the VOCs concentration dimension at the exit, the actual exit concentration at the last moment of the historical segment is used as the input for the first moment of the future segment. The exit concentration input at each subsequent moment is predicted by the second branch based on the network output at the previous moment, in order to maintain consistency with the autoregressive prediction method in the pre-training stage.

[0046] All data from each dimension underwent Z-score standardization before being input into the Long Short-Term Memory network. If there are missing inlet VOCs concentration monitoring data for the current operating conditions, the inlet VOCs concentration dimension uses the inlet concentration value estimated at the current time by removing efficiency sequence inversion. Fill in the blanks.

[0047] Specifically, the steps for extracting the corrected saturation of VOCs from the carbon canister are as follows:

[0048] The inlet and outlet VOCs concentration sequences, air volume sequences, temperature and humidity sequences, and carbon canister pressure difference sequences are extracted from the adsorption time series parameters. The length of each sequence is N, which means there are N sampling points from the historical start time to the current time.

[0049] The inlet and outlet VOCs concentration sequences and air volume sequences were input into the activated carbon adsorption capacity model to extract the VOCs adsorption increment per unit time of the activated carbon canister. :

[0050] ;

[0051] in, Let be the air volume at the i-th sampling time. Let VOCs concentration at the inlet be the concentration at the i-th sampling time. Let VOCs concentration at the outlet be the concentration at the i-th sampling time. The sampling time interval, This represents the adsorption increment per unit time corresponding to the i-th sampling moment;

[0052] The cumulative adsorption amount up to the current moment is obtained by summing up all the adsorption increments per unit time from the historical start time to the current time along the time axis. :

[0053] ;

[0054] Cumulative adsorption amount Characterizes the total mass of VOCs adsorbed by activated carbon from the start of its use to the present moment;

[0055] Obtain the actual mass of activated carbon filled in the VOCs activated carbon canister and maximum adsorption capacity per unit mass (The maximum adsorption capacity per unit mass is obtained from the parameters provided by the manufacturer based on the type of activated carbon material used), and is combined with the cumulative adsorption capacity. Comprehensive processing was performed to obtain the basic saturation level of VOCs in the carbon canister. :

[0056] ;

[0057] Basic saturation This is a dimensionless ratio, ranging from 0 to 1, representing the proportion of the current adsorption capacity of activated carbon to its maximum theoretical adsorption capacity. When the value approaches 1, it indicates that the activated carbon is close to saturation.

[0058] The activated carbon filling amount is obtained based on the activated carbon canister design parameters or activated carbon replacement and maintenance records; when activated carbon is replaced, it is updated according to the actual quality of the replaced activated carbon.

[0059] Substituting the temperature and humidity sequence and the carbon canister pressure difference sequence into the preset corresponding functions, the temperature and humidity correction factors are obtained. and differential pressure correction factor and the basic saturation The corrected saturation of VOCs in the carbon canister is obtained by performing a correction process, specifically:

[0060] Temperature and humidity correction factor Used to correct the effect of changes in ambient temperature and humidity on the adsorption capacity of activated carbon;

[0061] In one implementation, the temperature correction factor is calculated using the following formula: The formula for calculating the humidity correction factor is: ;

[0062] Final temperature and humidity correction factor That is, if the product of temperature correction and humidity correction is less than 0.85, then the value is 0.85; otherwise, the actual product value is used.

[0063] Under baseline conditions (temperature not exceeding 25°C and humidity not exceeding 60%RH). For 1.0, It is 1.0. It is 1.0;

[0064] Differential pressure correction function The pressure difference in the carbon canister is used to correct the effect of changes in bed resistance on the estimation of adsorption capacity. The rate of change of pressure differential reflects the bed resistance level. Reflecting the rate of change of bed resistance, the rate of change of pressure difference It is obtained by dividing the pressure difference of the carbon canister between two adjacent samples by the sampling interval;

[0065] The differential pressure correction function is defined in bivariate form:

[0066] ;

[0067] in, The reference pressure difference is 1.0 kPa. The maximum permissible pressure difference (value is 3.0 kPa). The coefficient for the rate of change (with a value of 0.15) The scaling factor for the rate of change (valued at 2.0 h / kPa) is used to truncate the output of this function so that it is not lower than the preset lower limit of 0.1. When the rate of change of pressure difference is 0, the second factor is approximately 0.925. As the rate of change of pressure difference decreases negatively, this factor approaches 1. As the rate of change of pressure difference increases positively, this factor gradually decreases, effectively reducing the correction factor value when the pressure difference rises rapidly.

[0068] Correct saturation The calculation method is as follows:

[0069] ;

[0070] If the calculation yields If so, it is truncated to 1 to conform to the physical upper limit of saturation, thus correcting the saturation. The saturation value is the result of joint verification by environmental correction and bed resistance correction. The corrected saturation and the pressure difference change rate together constitute the core input parameters for subsequent penetration early warning judgment, long short-term memory network attention mechanism and multi-factor state inference.

[0071] When the adsorption time series parameter set lacks the inlet VOCs concentration, the corrected saturation of the VOCs in the carbon canister is extracted through the following steps:

[0072] The initial removal efficiency is obtained and subjected to time decay processing to obtain the removal efficiency sequence, as follows:

[0073] Removal efficiency The proportion of VOCs adsorbed and removed after passing through the activated carbon bed is represented by the value of VOCs decreasing over time, which characterizes the degradation process of the activated carbon's adsorption capacity.

[0074] Initial removal efficiency The measured removal efficiency value is taken during the first run after fresh activated carbon is loaded, and the value ranges from 0.90 to 0.99.

[0075] In one implementation, the initial removal efficiency The value is 0.95, and the time decay treatment uses an exponential decay form to describe the change in removal efficiency with running time:

[0076] ;

[0077] in, This refers to the cumulative operating time since the activated carbon was loaded and put into use. The attenuation coefficient characterizes the rate at which the adsorption capacity of activated carbon degrades. Based on historical operating data, calibration is performed. In one implementation method, for granular activated carbon used to treat spray painting exhaust gas, Values ;

[0078] Based on the outlet VOCs concentration sequence and removal efficiency sequence, the inlet VOCs concentration sequence is estimated by inversion, specifically as follows:

[0079] ;

[0080] By comprehensively processing the inlet and outlet VOCs concentration sequences and air volume sequences, the basic saturation of the VOCs activated carbon canister is obtained, which is the cumulative adsorption capacity calculated according to the aforementioned activated carbon adsorption capacity model. With base saturation ;

[0081] Based on the pressure difference sequence of the carbon canister, the current pressure difference and the rate of change of pressure difference are extracted as dynamic correction factors to correct the baseline saturation, resulting in the corrected saturation of the VOCs carbon canister, specifically:

[0082] At the same time, extract the current pressure difference of the carbon canister. and its rate of change As dynamic correction factors, both factors work together to correct the base saturation:

[0083] ;

[0084] in, This is the differential pressure correction function. Reflecting the current bed resistance level, Reflecting the rate of change in bed resistance, the two factors are combined as a dynamic correction factor to characterize the impact of increased adsorption load and changes in bed resistance on saturation estimation.

[0085] If the corrected saturation and differential pressure change rate meet the preset penetration conditions, a penetration warning signal is generated, specifically as follows:

[0086] The penetration condition is that the corrected saturation reaches or exceeds the preset saturation threshold, and the differential pressure change rate is greater than zero. The preset saturation threshold is determined according to the activated carbon material and design conditions, and the value range is 0.80 to 0.95.

[0087] In one embodiment, the activated carbon loading for treating the spraying exhaust gas is 200 kg, and the maximum adsorption capacity per unit mass is 800 mg / g. When the cumulative adsorption reaches about 144 kg, the outlet concentration begins to show an upward trend, and at this time the saturation threshold value is 0.90.

[0088] If the condition is met, a penetration warning signal is generated and stored. This signal includes the warning trigger timestamp, the current corrected saturation, the differential pressure change rate, and the warning level.

[0089] The warning level is determined based on the value of the correction saturation: when 0.80 ≤ current correction saturation < 0.90, it is the "prompt" level; when 0.90 ≤ current correction saturation < 0.95, it is the "warning" level; and when the current correction saturation ≥ 0.95, it is the "emergency" level.

[0090] Specifically, the steps for predicting the set of regeneration performance indicators corresponding to each group of candidate time series windows are as follows:

[0091] Each candidate time window is input into the pre-trained Long Short-Term Memory network, and the hidden state of each time step is extracted through forward propagation. The hidden state contains a compressed representation of all input information before that time step.

[0092] The current corrected saturation and pressure difference change rate are concatenated to generate an attention query vector. This attention query vector represents the instantaneous operating conditions under the current adsorption state. The attention query vector is then linearly transformed through a trainable fully connected layer to obtain the transformed attention query vector, specifically:

[0093] The attention query vector is generated by concatenating the current corrected saturation and the rate of change of pressure difference. :

[0094] ;

[0095] in, This is a vector concatenation operation that concatenates the corrected saturation (scalar) and the rate of change of pressure (scalar) into a two-dimensional vector. ;

[0096] Attention query vector A linear transformation is performed through a trainable fully connected layer, mapping it to the same dimensional space as the hidden state:

[0097] ;

[0098] in, This is the weight matrix. For bias vectors, This is the transformed attention query vector;

[0099] The hidden states at each time step are used as key and value vectors, respectively. The transformed attention query vector, key vector, and value vector at each time step are processed through the corresponding dropout layer to obtain the processed attention query vector, key vector, and value vector. The dot product of the processed attention query vector and the key vector at each time step is calculated and mapped to the attention weight at each time step using a normalized exponential function. Specifically:

[0100] Hidden states at each time step The key vector is obtained through linear transformation. Sum value vector :

[0101] ;

[0102] in, This is the weight matrix. It is the bias vector;

[0103] Will Each is obtained through the corresponding discard layer. Calculate the transformed query vector With key vectors at each time step The attention weights are obtained by taking the dot product of the two numbers and normalizing it using Softmax.

[0104] ;

[0105] The attention weights at each time step are dynamically reconstructed based on the corrected saturation, specifically as follows:

[0106] Attention weights at each time step based on modified saturation Dynamic reconstruction is performed when the correction saturation falls below the first-state gating threshold. At this time, the first-state gating threshold is used as an optimizable parameter in the attention weight dynamic reconstruction network. Its initial value can be 0.50, and it is updated through backpropagation of the prediction error in the training samples, thus adjusting the attention weights at each time step. Reconstructed into a uniform distribution:

[0107] ;

[0108] When the correction saturation is higher than the second-state gating threshold At that time, the second-state gating threshold is used as an optimization parameter during network training. Its initial value can be 0.80, and it is adjusted according to the principle of minimizing the model loss function, thus affecting the attention weights. Reconstructed into a distribution oriented toward recent time steps:

[0109] ;

[0110] in, The concentration adjustment coefficient is based on relative position. Replace the absolute position index i, so that when N changes The value range remains stable to avoid numerical overflow;

[0111] This parameter sets a candidate value range before model training and selects the value based on the prediction error of the validation set. By changing the value of the concentration adjustment coefficient, the prediction loss of the corresponding model on the validation set is calculated, and the value corresponding to the minimum validation loss is taken as the final parameter.

[0112] In one implementation, the candidate range for the concentration adjustment coefficient is set to 1.0 to 2.0, the initial test value is 1.5, and the final value is determined after comparing the prediction error of the validation set.

[0113] When the corrected saturation is between and During this period, the original attention weights remain unchanged, i.e. ;

[0114] Attention weights at each time step after dynamic reconstruction For the corresponding value vector We perform a weighted summation to obtain the weighted attention features, namely:

[0115] ;

[0116] The weighted attention features are input into the fully connected layer of the Long Short-Term Memory network. After mapping by the fully connected layer, the regeneration performance index set corresponding to each candidate time window is output, specifically:

[0117] Will By discarding the layers before the final fully connected layer, we obtain Then, the fully connected layer is input, and after mapping by the fully connected layer, the set of regeneration performance indicators corresponding to each candidate time series window is output. The mapping relationship of the fully connected layer is as follows:

[0118] ;

[0119] in, This is the weight matrix of the fully connected layer. For bias vectors, The output vector consists of four indicators: regeneration efficiency, waste gas treatment efficiency, energy consumption level, and activated carbon decay rate, forming a set of regeneration performance indicators. ;

[0120] Among them, regeneration efficiency refers to the ratio of the mass of VOCs desorbed and removed during the regeneration process to the cumulative adsorption amount of activated carbon before regeneration, reflecting the degree of recovery of adsorption capacity by the regeneration operation.

[0121] Waste gas treatment efficiency refers to the VOCs removal rate measured when the carbon canister is put back into adsorption operation after regeneration; energy consumption level refers to the total energy consumed to complete a single regeneration operation.

[0122] The activated carbon decay rate refers to the percentage decrease in the maximum adsorption capacity of activated carbon relative to its initial capacity after each regeneration.

[0123] The specific steps for predicting the confidence intervals corresponding to the regeneration performance index sets for each candidate time series window are as follows:

[0124] Dropout layers are embedded before each fully connected layer in the Long Short-Term Memory (LSTM) network. Each dropout layer randomly masks the output of its neurons according to a dropout ratio determined during training. The value is determined by optimization based on the prediction error of the validation set during the model training phase and the model's generalization performance, and its range is from 0.1 to 0.3.

[0125] In one implementation, the model validation error is compared under different discard ratios. The value is set to 0.2. In different forward propagation rounds, the location of the neurons in the discard layer shield is different.

[0126] Multiple independent forward propagations are performed on the same candidate time window. In each forward propagation, different neurons in the dropout layer are randomly masked, so that each forward propagation corresponds to a different subnetwork. A set of predicted values ​​for the regeneration performance index corresponding to the candidate time window is output, resulting in multiple sets of prediction results. Specifically:

[0127] Suppose that T independent forward propagations are performed (T is 50 in this embodiment), and each sub-network outputs a set of predicted values ​​of regeneration performance indicators corresponding to the candidate time series window. ,in A total of T sets of prediction results were obtained;

[0128] Calculate the mean of each indicator in multiple sets of prediction results. with standard deviation ( (Corresponding to four indicators), the standard deviation is used as a measure of prediction uncertainty. The larger the standard deviation, the higher the degree of dispersion between the predicted values ​​of each sub-network output. That is, the lower the stability of the output of the Long Short-Term Memory Network for the current candidate time window and the lower the prediction reliability.

[0129] The confidence intervals for each index in the regeneration performance index set are determined using both the mean and standard deviation, specifically as follows:

[0130] Determine the corresponding critical value of the standard normal distribution based on the set confidence level, and use the mean as the basis. with standard deviation Calculate the upper and lower bounds of the confidence intervals for each indicator:

[0131] ;

[0132] in, The critical value corresponding to the standard normal distribution at the confidence level, and the interval width. The wider the width, the higher the uncertainty of the indicator's prediction result;

[0133] In one implementation, a 95% confidence level is used, corresponding to z=1.96.

[0134] The pre-training steps for Long Short-Term Memory (LSTM) networks are as follows:

[0135] A training sample set is constructed, containing multiple historical time-series windows, corresponding real regeneration performance index labels, and corresponding real outlet VOCs concentrations, specifically as follows:

[0136] Before training, the labels of the real regeneration performance indicators are Z-score standardized to make their mean 0 and standard deviation 1, so as to eliminate the influence of the difference in dimensions on the loss function. During inference, the predicted values ​​output by the network are de-standardized to restore them to the original dimensions.

[0137] In one implementation, the training sample set contains 2,000 historical time series windows, each with a duration of 24 hours (corresponding to 288 sampling points with a sampling interval of 5 minutes).

[0138] The historical time series windows are all taken from the complete operating cycle that includes the regeneration operation, covering the entire process of pre-regeneration adsorption, regeneration process and post-regeneration adsorption initial stage. Therefore, future segments may include the regeneration period when the inlet concentration is zero.

[0139] Divide each historical time series window into front Historical fragments and their aftermath The future segments of each time step (the remaining intermediate time steps are generated into multiple sets of samples through sliding sampling) are used to ensure that the sequence length during the training phase is consistent with that during the inference phase, thus ensuring the matching of model parameters and receptive fields.

[0140] The first 80% of the samples were used for training, and the last 20% were used for validation. During training, the sliding step size was 30 sampling points to expand the sample size.

[0141] Each historical time series window is sequentially input into the Long Short-Term Memory (LSTM) network. After forward propagation, the corresponding predicted values ​​of regeneration performance indicators, predicted values ​​of outlet VOCs concentration, and estimated values ​​of adsorption increment are output in each branch of the network. Specifically:

[0142] The historical time series windows in the training samples are divided into historical segments (previous and future). (the moment) and future fragments (later) At any given moment, during the forward propagation phase of the network, for historical fragments ( The network receives the complete input vector (including inlet concentration, outlet concentration, air volume and other dimensions) in a conventional manner, and uses a teacher-forced method (using real historical data as input) to extract temporal features and generate the hidden state at each time step.

[0143] For future fragments ( The network's input strategy is switched to: inlet concentration and air volume still use real values; for the outlet concentration dimension, the first moment of the future segment ( ) Always use the true value as input; from the second time point ( The planned sampling strategy will be implemented from the beginning, with the planned sampling probability. Using the actual exit concentration at that moment as input, with probability... Switch the input for this dimension from the true value to the exit concentration value predicted by the second branch of the network at the previous time step. ;

[0144] The second branch uses this information to gradually predict the complete export concentration sequence of future segments. Planned sampling probability The value gradually decreases with each training cycle, starting at 0.9 and decreasing by 0.05 every 10 training cycles until it reaches 0.1, in order to gradually approximate the fully autoregressive mode during inference.

[0145] Long Short-Term Memory (LSTM) networks are multi-task output structures that include an attention-aware module. In the linear transformation layer of the attention module (… After that, before the final output fully connected layer, and before the fully connected layer for predicting the outlet concentration of the second branch and the fully connected layer for extrapolating the adsorption increment of the third branch, discard layers are embedded respectively, with the discard probability set to 1. ;

[0146] During the pre-training phase, each dropout layer randomly deactivates neurons with the aforementioned probability; during the inference phase, the dropout layer remains active to support the Monte Carlo dropout method for uncertainty estimation. During pre-training, the same dynamic attention weighting module as in the inference phase is added above the LSTM layer, and the weighted attention features are input into the fully connected layer and output in three ways.

[0147] The first branch is the performance prediction branch, which is used to map the weighted attention features to predicted values ​​of the regeneration performance index. ;

[0148] The second branch is the export concentration prediction branch, used to predict the hidden states at each time step. The mapping yields the predicted VOCs concentration at the outlet at the corresponding time. ;

[0149] The third branch is the adsorption increment calculation branch, which receives the hidden state at each time step. Output scalar adsorption increment estimate ,in, This is the weight matrix. It is the bias vector;

[0150] Query vector generation matrix of the attention module Key value transformation matrix The parameters of the fully connected layer and the three branches all participate in the backpropagation update during pre-training. During the training phase, the three branches jointly participate in the forward propagation and loss function calculation.

[0151] During the inference phase, the first branch outputs a set of regeneration performance indicators, the second branch is used to generate the outlet concentration sequence for future time moments to maintain the integrity of the input sequence and the autoregressive structure, and the third branch stops the calculation.

[0152] The mean square error between the predicted and actual regeneration performance index labels is used as the data fitting term. Specifically:

[0153] ;

[0154] in, Let j be the predicted value of the j-th indicator. This represents the true label value of the j-th indicator;

[0155] Based on the actual inlet concentration and actual air volume in the future segment, the predicted value of outlet VOCs concentration and the estimated value of adsorption increment are used to calculate the mass conservation constraint term:

[0156] Real entry concentration based on future fragments Actual air volume and predicted export concentration Calculate the mass conservation residual of the future segment:

[0157] ;

[0158] in This represents the adsorption increment value estimated at the corresponding future time for the third branch (adsorption increment estimation branch);

[0159] Mass conservation constraint Defined as the mean square value of the residuals of future segments:

[0160] ;

[0161] Based on the data fitting term and the mass conservation constraint term, a dual-objective loss function is constructed. Minimizing this dual-objective loss function is the optimization objective. The parameters of the Long Short-Term Memory network are updated through backpropagation to complete pre-training. Specifically:

[0162] Bi-objective loss function:

[0163] ;

[0164] in, The mass conservation constraint weighting coefficient has a value ranging from 0.01 to 0.1, and in one embodiment, it is set to 0.05.

[0165] With the goal of minimizing the bi-objective loss function, the parameters of the long short-term memory network (such as various weight matrices, bias vectors, etc.) are updated through backpropagation. The training process continues until the loss function value on the validation set no longer decreases for 10 consecutive training cycles, at which point the model is considered to have converged and the pre-training is completed.

[0166] It should be noted that during training, for each historical time window, the attention query vector is constructed by concatenating the corrected saturation and pressure change rate corresponding to the last sampling moment of the historical segment, which is completely consistent with the construction method in the inference stage.

[0167] During the LSTM network pre-training stage, in order to improve the model's adaptability to missing inlet VOCs concentrations, some inlet VOCs concentration data in the training samples are replaced with estimated values ​​obtained through the aforementioned inversion method according to the data replacement probability.

[0168] The data replacement probability is used to simulate scenarios where the inlet VOCs concentration sensor malfunctions or data is missing during actual operation. It is determined based on the statistical proportion of missing inlet VOCs concentration events in historical operating data.

[0169] The number of valid sampling points and the number of missing sampling points for VOCs concentration at the inlet were counted within multiple complete operating cycles. The actual missing ratio was calculated and used as the initial setting value for the data replacement probability. When the missing ratio of historical data was insufficient to cover the model training requirements, perturbation tests were conducted near the initial setting value. The final value was determined by comparing the model prediction error under different replacement probabilities.

[0170] In one implementation, the data replacement probability is set to 30%.

[0171] Specifically, such as Figure 2 As shown, the specific steps for predicting the output regeneration and dynamically optimizing the weight set are as follows:

[0172] The values ​​and statuses of each operating condition parameter in the current operating condition information are converted into natural language description text, specifically:

[0173] The values ​​and statuses of each operating condition parameter in the current operating condition information are converted into natural language description text. The current operating condition information includes the following parameters: number of times activated carbon has been regenerated. Current ambient temperature Current ambient humidity Current electricity price time period type (Peak / Normal / Row) Production Scheduling Urgency (Emergency / Normal / Relaxed) and charcoal canister pressure differential status (Normal / High / Severely Blocked);

[0174] Among them, the number of times the activated carbon has been regenerated is obtained from the historical operating cycle data; the current ambient temperature and humidity are collected in real time by temperature and humidity sensors; the current electricity price period type is determined according to the peak, flat and valley electricity price division rules published by the local power grid; the urgency of the production schedule is input externally or preset by the system; and the pressure difference status of the carbon canister is determined according to the current pressure difference value.

[0175] Differential pressure status of charcoal canister Determine based on the current differential pressure value: when This is normal; when The price was slightly higher at that time; It was a time of severe congestion;

[0176] The natural language description text is generated by comparing each operating condition parameter with its preset threshold to generate a state description statement;

[0177] In one implementation, the current operating conditions are as follows: the activated carbon has been regenerated 8 times (recommended lifespan is 10 times), the ambient temperature is 38°C (above the normal threshold of 25°C), the ambient humidity is 65%RH (above the normal threshold of 60%RH), the electricity price is during peak hours (1.2 yuan / kWh), the production schedule urgency level is "urgent", and the carbon canister pressure difference is 1.8 kPa (belonging to a relatively high state).

[0178] The converted natural language description text is as follows: "The activated carbon has been regenerated 8 times and is nearing the end of its life (recommended life is 10 times); the ambient temperature is 38℃, which is a high temperature condition in summer; the ambient humidity is 65%, which is high; the current electricity price is peak (1.2 yuan / kWh); the production schedule is tight, and regeneration and production must be completed within 3 hours; the pressure difference of the carbon tank has reached 1.8 kPa and is still rising, and there is slight blockage in the bed;

[0179] Using the indicators in the regeneration performance index set as optimization objectives and the parameters in the current working condition information as constraints, natural language description text is input into the pre-trained large language model.

[0180] Semantic reasoning is performed on the coupling relationship between optimization objectives and constraints using a large language model, and the priority ranking of each optimization objective under the current working condition is output, specifically:

[0181] The large language model employs a pre-trained model based on the Transformer architecture (such as LLaMA), and uses a sample set containing working condition descriptions and corresponding weight ranking pairs for instruction fine-tuning. The fine-tuning sample pair format is as follows: the input text is the working condition description text, and the expected output is a priority ranking in JSON format. Example: Input: "The activated carbon has been regenerated 8 times... the pressure difference in the carbon canister has reached 1.8 kPa and is still rising." Expected output: {"priority": [regeneration efficiency, energy consumption level, activated carbon decay rate, waste gas treatment efficiency]};

[0182] During inference, the prefix constraint is: "Please output priority sorting in JSON format:" to ensure that the model output conforms to the expected format. During inference, the current working condition information is formatted into natural language description text according to the pre-built working condition information description template, and the prefix constraint is added to force the output to be a JSON structured string to ensure stable generation of sorting results.

[0183] The operating condition information description template is established based on the model input variable type, including operating condition fields such as the number of times activated carbon has been regenerated, current ambient temperature and humidity, current electricity price time period type, production scheduling urgency, and carbon tank pressure differential status, and specifies the arrangement order and description format of each field;

[0184] If a large language model does not have the ability to directly output structured data, a rule parser can be added after the model output to extract priority ranking information from unstructured text.

[0185] Based on the various indicators in the regeneration performance index set (regeneration efficiency) Waste gas treatment efficiency Energy consumption level Activated carbon decay rate Using the parameters in the current working condition information as constraints, the natural language description text is input into the pre-trained large language model as the optimization objective;

[0186] The output format of the large language model is preset to a JSON structured string, which contains a list of indicators sorted by priority from high to low and their corresponding reasoning.

[0187] In one implementation, the reasoning process is as follows: if “tight production schedule” is identified, it is determined that “regeneration efficiency” should be prioritized and is the main contradiction at the current stage; if “activated carbon has been regenerated 8 times” is identified, it is determined that “activated carbon decay rate” should be considered as an important factor, but its priority is lower than the current production demand; if “peak electricity price” is identified, it is determined that “energy consumption level” should be strictly controlled.

[0188] The above reasoning results are converted into a priority ranking: regeneration efficiency > energy consumption level > activated carbon decay rate > waste gas treatment efficiency.

[0189] The priority ranking results are mapped to the weight coefficients corresponding to each optimization objective, forming a regenerative dynamic optimization weight set, specifically:

[0190] The weights are calculated using an exponential decay formula based on the ranking position:

[0191] ;

[0192] Where j represents the number of the current optimization objective, and k represents the numbers of all optimization objectives involved in the weight calculation. Prioritize the j-th optimization objective ( Indicates the highest priority. (Indicates the lowest priority) This is the attenuation coefficient (with a value of 0.5). Let the weight coefficients be those for the j-th optimization objective. The priority ranking values ​​are the corresponding to the k optimization objectives.

[0193] Specifically, the steps for outputting the predicted operating status of the VOCs canister are as follows:

[0194] For each group of candidate time series windows, the regeneration dynamic optimization weight set is used as the evaluation coefficient to comprehensively evaluate the corresponding regeneration performance index set of each group of candidate time series windows, and the evaluation score of each group of candidate time series windows is obtained, as follows:

[0195] The regeneration performance index set for each candidate window is normalized: For positive indices (regeneration efficiency, waste gas treatment efficiency), the normalization formula is as follows: For negative indicators (energy consumption level, activated carbon decay rate), the normalization formula is: ;

[0196] All after normalization All values ​​range from 0 to 1, with larger values ​​being better. Reflects the optimal goal of regeneration efficiency. This reflects the stable target of treatment effect (i.e., waste gas treatment efficiency). Reflecting the target of minimum energy consumption, The target value that reflects the longest lifespan of activated carbon (i.e., the larger the normalized rate of activated carbon decay, the slower the decay and the longer the lifespan).

[0197] For each group of candidate time series windows, the weight set is dynamically optimized using regeneration. As evaluation coefficients, the normalized regeneration performance index set corresponding to each group of candidate time windows... A comprehensive evaluation was conducted, and the weighted evaluation scores of each group of candidate time series windows were calculated. The calculation method is as follows:

[0198] ;

[0199] Extract the confidence interval width corresponding to each group of candidate time series windows, and determine the risk penalty value for each group of candidate time series windows, specifically as follows:

[0200] In one implementation, the range width of the energy consumption level index that is directly related to operating costs is selected. As a measure of forecast uncertainty, the risk penalty term is calculated using a scale-consistent approach:

[0201] ;

[0202] in, This represents the maximum width of the confidence interval for energy consumption levels across all candidate windows. This is a preset risk trade-off coefficient used to adjust the degree of influence of forecast uncertainty in the comprehensive evaluation;

[0203] It can be determined by comparing the deviation between the comprehensive evaluation results and the actual operating effect under different α values ​​during the model validation phase.

[0204] In one implementation, by verifying the impact of different risk trade-off coefficient values ​​on the prediction results of the optimal operating state, when the comprehensive evaluation results match the actual operating effects to a high degree, α is set to 0.1.

[0205] In another implementation, a weighted sum of the range widths of four indicators can be used as a risk penalty:

[0206] ;

[0207] in, Assign weights to the contribution of each indicator's uncertainty to operational risk (assigned according to the importance of the indicator; for example, energy consumption level and activated carbon decay rate can be given higher weights). This represents the maximum confidence interval width of the j-th indicator across all candidate windows;

[0208] Based on the evaluation scores and risk penalty values ​​of each group of candidate time series windows, the comprehensive evaluation score of each group of candidate time series windows is obtained: ;

[0209] The candidate time series windows with the highest comprehensive evaluation scores are selected as the optimal operating state prediction results for the VOCs carbon tank and output, specifically:

[0210] In all candidate timing windows, find the one that makes Maximized candidate window index :

[0211] ;

[0212] Where M is the total number of candidate time series windows;

[0213] The first The values ​​of the variables to be predicted corresponding to the candidate time series windows (i.e., the target state value of regeneration temperature) Target state value of regenerated gas air volume Regeneration time target state value The output is the predicted operating status of the VOCs carbon canister.

[0214] Specifically, when the adsorption analysis parameter set lacks inlet VOCs concentration, the following steps are included after outputting the corrected saturation:

[0215] Real-time monitoring of the current VOCs concentration at the outlet and the differential pressure value of the charcoal canister And determine whether to trigger self-calibration, specifically:

[0216] When the pressure difference of the carbon canister is monitored at three consecutive sampling times and the outlet VOCs concentration shows a synchronous and continuous upward trend, the self-calibration of the saturation estimation results is triggered. The criterion for determining the continuous upward trend is: the pressure difference at three consecutive sampling times meets the following conditions. And the outlet concentration at three consecutive sampling times satisfies ;

[0217] In response to the self-calibration trigger, the removal efficiency is obtained based on the pressure difference in the activated carbon canister and the increase in the outlet VOCs concentration, through a preset mapping relationship, specifically:

[0218] The mapping relationship is established using a two-dimensional lookup table, where the input variable is the current differential pressure deviation. and export concentration growth rate The output variable is the actual removal efficiency correction coefficient. ;

[0219] The two-dimensional lookup table is obtained by calibration using historical operating data. Specifically, it collects data on the differential pressure, inlet VOCs concentration, outlet VOCs concentration, and actual removal efficiency of the VOCs canister under different operating cycles. Based on the historical operating data, it calculates the differential pressure deviation and outlet concentration growth rate corresponding to each operating state. Using the differential pressure deviation and outlet concentration growth rate as two-dimensional input coordinates and the actual removal efficiency correction coefficient as the output value, it establishes a correspondence between the input state and the removal efficiency correction coefficient.

[0220] In a simplified implementation, this mapping relationship can be approximated by the following formula:

[0221] ;

[0222] in To reflect the impact of pressure differential deviation and outlet concentration change rate on actual removal efficiency, correction coefficients obtained by fitting historical operating data are used in this embodiment. , The rate of change of outlet concentration is approximately calculated by dividing the concentration difference between two consecutive sampling times by the sampling interval.

[0223] The removal efficiency is compared with the initial removal efficiency, and the decay coefficient in the time decay model is updated based on the comparison result, i.e.:

[0224] Will If the difference between the initial removal efficiency and the current efficiency exceeds the efficiency change threshold, the decay coefficient is updated using the following formula. :

[0225] ;

[0226] Recalculate the estimated removal efficiency and inlet VOCs concentration at subsequent time points using the updated attenuation coefficient (following the steps above), and recalculate the cumulative adsorption amount, basic saturation and corrected saturation to complete the self-calibration of the saturation estimation results.

[0227] The efficiency change judgment threshold is obtained based on the historical removal efficiency data during the stable operation phase of activated carbon, and is used to determine whether the current removal efficiency change exceeds the normal operating fluctuation range.

[0228] In one implementation, the efficiency change discrimination threshold is set to ±5% by statistically analyzing the range of change in removal efficiency relative to the initial removal efficiency over multiple normal operating cycles.

[0229] It should be noted that this step is performed before each penetration condition determination.

[0230] Please see Figure 3 This invention provides a technical solution: a VOCs activated carbon canister operation status prediction system based on deep machine learning, comprising: a data acquisition unit for acquiring time-series data of the VOCs activated carbon canister operation status and extracting an adsorption time-series parameter set of the VOCs activated carbon canister; an adsorption status deduction unit for extracting the corrected saturation and differential pressure change rate of the VOCs activated carbon canister based on a pre-established activated carbon adsorption capacity model and combined with the adsorption time-series parameter set, through multi-factor adsorption status deduction; a penetration warning unit for issuing a penetration warning signal when the corrected saturation and differential pressure change rate meet preset penetration conditions; a candidate time-series construction unit for constructing multiple sets of candidate time-series windows based on the operation status time-series data in response to the penetration warning signal; and a deep prediction unit for combining the candidate time-series windows into a single window. The time-series window is input into a pre-trained long short-term memory network. After being modified by saturation and pressure difference change rate weighted attention perception, it predicts the regeneration performance index set and corresponding confidence interval for each candidate time-series window. The regeneration performance index set includes regeneration efficiency, waste gas treatment efficiency, energy consumption level, and activated carbon decay rate. The weight generation unit is used to extract the current operating condition information from the operating status time-series data and input it into a pre-trained large language model for semantic reasoning to predict and output the regeneration dynamic optimization weight set. The evaluation prediction unit is used to evaluate the regeneration performance index set corresponding to each candidate time-series window according to the dynamic optimization weight, incorporate the confidence interval as a risk penalty term into the evaluation, deduce the candidate time-series window that makes the evaluation result optimal, and output it as the operating status prediction result of the VOCs carbon canister.

[0231] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0232] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting the operating status of VOCs charcoal canisters based on deep machine learning, characterized in that, Includes the following steps: Obtain the time-series data of the VOCs activated carbon canister's operating status and extract the adsorption time-series parameter set of the VOCs activated carbon canister; Based on a pre-established activated carbon adsorption capacity model and combined with an adsorption time series parameter set, the corrected saturation and differential pressure change rate of VOCs in the activated carbon canister are extracted through multi-factor adsorption state deduction. When the correction saturation and differential pressure change rate meet the preset penetration conditions, a penetration warning signal is issued, and multiple candidate time series windows are constructed based on the operating status time series data; Each candidate time series window is input into a pre-trained long short-term memory network. After adjusting the saturation and pressure difference change rate weighted attention perception, the regeneration performance index set and corresponding confidence interval of each candidate time series window are predicted. Extract current operating condition information from the time series data of the operating status, input it into the pre-trained large language model for semantic reasoning, and predict the output to regenerate the dynamically optimized weight set; The regeneration performance index set corresponding to each candidate time series window is evaluated based on the dynamic optimization weight. The confidence interval is included as a risk penalty term in the evaluation. The prediction results of the VOCs carbon canister's operating status are deduced and output.

2. The method for predicting the operating status of VOCs carbon canisters based on deep machine learning according to claim 1, characterized in that, The specific steps for correcting the saturation of VOCs extracted from the carbon canister are as follows: The inlet and outlet VOCs concentration sequences, air volume sequences, temperature and humidity sequences, and carbon canister pressure difference sequences were extracted from the adsorption time series parameters. The inlet and outlet VOCs concentration sequences and air volume sequences were input into the activated carbon adsorption capacity model to extract the adsorption increment per unit time. : ； in, Let be the air volume at the i-th sampling time. Let VOCs concentration at the inlet be the concentration at the i-th sampling time. Let VOCs concentration at the outlet be the concentration at the i-th sampling time. The sampling time interval, This represents the adsorption increment per unit time corresponding to the i-th sampling moment; The cumulative adsorption amount is obtained by summing up all adsorption increments per unit time along the time axis. : ； Obtain the actual mass of activated carbon filled in the VOCs activated carbon canister and maximum adsorption capacity per unit mass And combined with the cumulative adsorption amount After comprehensive processing, the basic saturation is obtained. : ； Substituting the temperature and humidity sequence and the carbon canister pressure difference sequence into the preset corresponding functions, the temperature and humidity correction factors are obtained. and differential pressure correction factor and the basic saturation Correction processing is performed to obtain the corrected saturation of VOCs in the carbon canister. : 。 3. The method for predicting the operating status of VOCs carbon canisters based on deep machine learning according to claim 2, characterized in that, When the adsorption time series parameter set lacks inlet VOCs concentration, the corrected saturation is extracted through the following steps: The initial removal efficiency is obtained and then subjected to time decay processing to obtain the removal efficiency sequence; Based on the effluent VOCs concentration sequence and removal efficiency sequence, the inlet VOCs concentration sequence is estimated by inversion. The inlet and outlet VOCs concentration sequences and air volume sequences were comprehensively processed to obtain the baseline saturation. Based on the carbon canister pressure difference sequence, the current carbon canister pressure difference and the rate of change of pressure difference are extracted as dynamic correction factors to correct the basic saturation and obtain the corrected saturation.

4. The method for predicting the operating status of VOCs carbon canisters based on deep machine learning according to claim 2, characterized in that, The specific steps for predicting the set of regeneration performance indicators corresponding to each group of candidate time series windows are as follows: Each candidate time series window is input into the pre-trained Long Short-Term Memory network, and the hidden state at each time step is extracted through forward propagation. The modified saturation and the rate of change of pressure are concatenated to generate an attention query vector, and the hidden state at each time step is used as the key vector and value vector respectively, so as to map the attention weights at each time step. The attention weights at each time step are dynamically reconstructed based on the corrected saturation. The corresponding value vectors are weighted and summed using the attention weights of each time step after dynamic reconstruction to obtain weighted attention features, which are then input into the fully connected layer of the long short-term memory network. After mapping by the fully connected layer, the set of regeneration performance indicators corresponding to each candidate time window is output.

5. The method for predicting the operating status of VOCs carbon canisters based on deep machine learning according to claim 4, characterized in that, The specific steps for predicting the confidence intervals corresponding to the regeneration performance index sets for each candidate time series window are as follows: A dropout layer is embedded before the fully connected layer of the Long Short-Term Memory network; Multiple independent forward propagations are performed on the same candidate time window. In each forward propagation, different neurons in the discard layer are randomly masked, and multiple sets of prediction results corresponding to the candidate time window are output respectively. Calculate the mean and standard deviation of each indicator in multiple sets of prediction results, and jointly determine the confidence interval of each indicator in the regeneration performance indicator set.

6. The method for predicting the operating status of VOCs carbon canisters based on deep machine learning according to claim 4, characterized in that, The pre-training steps for Long Short-Term Memory (LSTM) networks are as follows: Construct a training sample set containing multiple historical time series windows and corresponding real regeneration performance index labels, and divide each historical time series window into historical segments and future segments; Each historical time series window is sequentially input into the Long Short-Term Memory Network. After forward propagation, the corresponding predicted values ​​of regeneration performance indicators, predicted values ​​of outlet VOCs concentration, and estimated values ​​of adsorption increment are output in each branch of the network. The mean square error between the predicted value and the actual label of the regeneration performance index is used as the data fitting term. Based on future segments, the mass conservation constraint terms are calculated using the predicted value of VOCs concentration at the outlet and the estimated value of adsorption increment. Based on the data fitting term and the mass conservation constraint term, a dual-objective loss function is constructed. The optimization objective is to minimize the dual-objective loss function. The parameters of the long short-term memory network are updated through backpropagation to complete the pre-training.

7. The method for predicting the operating status of VOCs carbon canisters based on deep machine learning according to claim 1, characterized in that, The specific steps for predicting the output and regenerating the dynamic optimization weight set are as follows: Convert the values ​​and status of each working condition parameter in the current working condition information into natural language description text; Using the indicators in the regeneration performance index set as optimization objectives and the parameters in the current working condition information as constraints, natural language description text is input into the pre-trained large language model. Semantic reasoning is performed on the coupling relationship between optimization objectives and constraints using a large language model, and the priority ranking results of each optimization objective are output. Map the priority ranking results to the weight coefficients corresponding to each optimization objective. : ； Where j represents the number of the current optimization objective, and k represents the numbers of all optimization objectives involved in the weight calculation. Prioritize the j-th optimization objective ( Indicates the highest priority. (Indicates the lowest priority) This is the attenuation coefficient (with a value of 0.5). Let the weight coefficients be those for the j-th optimization objective. The priority ranking values ​​are the corresponding to the k optimization objectives.

8. The method for predicting the operating status of VOCs carbon canisters based on deep machine learning according to claim 7, characterized in that, The specific steps for outputting the predicted operating status of the VOCs charcoal canister are as follows: For each group of candidate time series windows, the regeneration dynamic optimization weight set is used as the evaluation coefficient to comprehensively evaluate the corresponding regeneration performance index set and obtain the evaluation score of each group of candidate time series windows. Extract the width of the confidence interval and determine the risk penalty value for each group of candidate time series windows; Based on the evaluation scores and risk penalty values ​​of each group of candidate time series windows, the comprehensive evaluation score of each group of candidate time series windows is obtained. The candidate time series window with the highest comprehensive evaluation score is selected as the optimal operating state prediction result of the VOCs carbon canister and output.

9. The method for predicting the operating status of VOCs carbon canisters based on deep machine learning according to claim 3, characterized in that, When the adsorption analysis parameter set lacks inlet VOCs concentration, the following steps are included after outputting the corrected saturation: Real-time monitoring of outlet VOCs concentration and carbon canister pressure differential, and determination of whether self-calibration is triggered; In response to the self-calibration trigger, the removal efficiency is obtained through a preset mapping relationship based on the pressure difference of the carbon canister and the increase in the VOCs concentration at the outlet. The removal efficiency is compared with the initial removal efficiency, and the decay coefficient in the time decay model is updated based on the comparison results. The removal efficiency estimate is recalculated using the updated attenuation coefficient, thus completing the self-calibration of the saturation estimation results.

10. A VOCs charcoal canister operation status prediction system based on deep machine learning, employing the VOCs charcoal canister operation status prediction method based on deep machine learning as described in any one of claims 1-9, characterized in that, include: The data acquisition unit is used to extract the adsorption time series parameter set of VOCs carbon canisters; The adsorption state extrapolation unit is used to extrapolate the corrected saturation and differential pressure change rate of VOCs in the activated carbon canister based on a pre-established activated carbon adsorption capacity model. Penetration warning unit is used to determine whether the conditions for issuing a penetration warning signal are met; Candidate time series construction unit, used to construct multiple sets of candidate time series windows; The deep prediction unit is used to input each group of candidate time windows into a pre-trained long short-term memory network, and after correcting the saturation and pressure change rate weighted attention perception, predict the regeneration performance index set and corresponding confidence interval of each group of candidate time windows. The weight generation unit is used to extract current working condition information and input it into a pre-trained large language model to predict and regenerate the dynamically optimized weight set. The evaluation and prediction unit is used to evaluate the regeneration performance index set of each candidate window using the regeneration dynamic optimization weight set, incorporate the confidence interval as a risk penalty term into the evaluation, and deduce and output the prediction results of the VOCs carbon canister's operating status.

Images