A method for reducing the emission of electrolyte solvent

Abstract

The application discloses a kind of electrolyte solvent emission reduction methods, belong to lithium battery electrolyte solvent production emission reduction and quality control field.This method is applied to dimethyl carbonate continuous production device, collects reaction, recovery, rectification, pollution and on-line offline composition data, and is adjusted by directed flow topological graph, space-time graph neural network, time series Transform, material balance correction, conformal prediction interval and constraint model prediction control output each control variable adjustment amount.

Description

Technical Field

[0001] This invention relates to the field of emission reduction and quality control in lithium battery electrolyte solvent production, and particularly to a method for reducing electrolyte solvent emissions. Background Technology

[0002] The production of lithium-ion battery electrolyte solvents such as dimethyl carbonate typically involves continuous units including transesterification, methanol recovery, reactive distillation, and distillation of dimethyl carbonate products. In continuous dimethyl carbonate production units, on-site control relies heavily on reaction load, tower temperature and pressure, reflux ratio, yield, and limited online composition data for judgment. However, byproducts such as dipropylene glycol, organic waste liquid load, and battery-grade quality indicators often require offline detection or hysteresis analysis to obtain the necessary information.

[0003] Existing methods often struggle to promptly distinguish the impact of byproduct sources, changes in the recovery stream, and changes in the distillation cut-off point on waste liquid load and solvent loss when dealing with coupled changes between the reaction, recovery, and rectification sections. There are time lags and residence time differences between offline byproduct data and online sensor sequences, which can easily lead to soft measurement results deviating from the current control cycle. At the same time, simply adjusting the reflux ratio, reboiler load, or cut-off point according to fixed empirical thresholds may introduce battery-grade product quality risks while reducing waste liquid discharge.

[0004] Therefore, there is a need for an electrolyte solvent emission reduction method that can address the shortcomings of the existing technologies. Summary of the Invention

[0005] One objective of this invention is to propose a method for reducing electrolyte solvent emissions. Addressing the issue that the generation of byproducts and organic waste liquid in continuous dimethyl carbonate production units is affected by the reaction load, recovery flow, and distillation cut-off point, and that detection is often delayed, this invention proposes a technical solution based on directed process topology graphs, spatiotemporal neural networks, time-series Transformers, material balance correction, conformal prediction intervals, and constraint model predictive control. This invention achieves the technical effects of reducing byproduct generation and organic waste liquid load, reducing dimethyl carbonate loss, and maintaining battery-level quality control stability.

[0006] This invention provides a method for reducing electrolyte solvent emissions, applied to a continuous dimethyl carbonate production unit, and includes the following steps: S1. Collect the operating variables, composition variables, by-product detection variables, and waste liquid organic matter variables of the continuous dimethyl carbonate production unit, and generate a process time series dataset according to a unified sampling period; S2. Construct a directed process topology graph according to the reaction equipment, recovery equipment, distillation equipment, material flow, recovery flow, and discharge flow, setting the equipment as nodes, the material flow, recovery flow, and discharge flow as edges, and writing the multi-time window process features into the corresponding nodes or edges; S3. Process the directed process topology graph using a spatiotemporal graph neural network to obtain the coupled representation of the reaction section, recovery section, and distillation section; S4. Input the coupled representation and the process time series dataset into a time series Transformer model to obtain the by-product generation rate, organic waste liquid load, dimethyl carbonate loss, and battery-grade product quality. S5. The output of step S4 is corrected using material balance constraints, tower segment cutting constraints, and conformal prediction intervals to obtain the corrected soft measurement results, uncertainty level, correction decision field, and safety state set; S6. The constrained model predictive control is executed with the corrected soft measurement results, uncertainty level, correction decision field, and safety state set as inputs, and the reaction load adjustment, methanol recovery flow rate adjustment, dimethyl carbonate recovery flow rate adjustment, reflux ratio adjustment, reboiler load adjustment, and distillation cut point adjustment are output. When the uncertainty level reaches the preset conservative control level, the input is invalid, or the candidate set is empty, the conservative adjustment of the above six types of control variables is output according to the conservative control branch.

[0007] Optionally, S1 includes: Collect reaction load, feed flow rate, reaction temperature, reaction pressure, temperature and pressure of reactive distillation column section, temperature and pressure of dimethyl carbonate distillation column section, reflux ratio, reboiler load, top product, side stream, bottom product, methanol recovery flow rate, dimethyl carbonate recovery flow rate, wastewater flow rate, online composition data, offline by-product data, and organic matter index of waste liquid; Align online data with sampling periods of one to ten minutes, write offline by-product data into the corresponding historical window according to the sampling timestamp, and write a calibration flag for missing offline values. Short window, medium window and long window features are generated. The short window covers the current control cycle and the two control cycles before it and is used to calculate the mean, slope and extreme values. The medium window covers the statistics of material residence time. The long window covers the historical distribution features of offline detection lag.

[0008] Optionally, S2 includes: Establish a node record for each equipment node, the node record including equipment type, material retention amount, temperature and pressure statistics, load status and composition estimate; An edge record is established for each edge, which includes flow rate, composition, flow direction, recovery flag, emission flag, and initial edge weight; The material residuals of each equipment node are calculated based on the feed flow rate, recovery flow rate, output flow rate, sewage discharge flow rate and composition estimate. The edge weights between nodes are then corrected based on the synchronous changes of the material residuals with the by-product generation rate, organic waste liquid load and dimethyl carbonate loss in the historical calibration period, so as to obtain the dynamic edge weights and waste liquid source contribution coefficient. Furthermore, the dynamic edge weights are determined based on the material residuals, specifically including: normalizing the signed material residuals according to the material throughput of the equipment nodes to obtain the signed residual ratio; The absolute value of the signed residual ratio is taken as the residual intensity, and the correlation between the residual intensity and the changes in the by-product generation rate, the organic waste liquid load and the dimethyl carbonate loss during the historical calibration period is calculated to obtain the residual correlation coefficient. The residual correlation coefficient is weighted with the initial edge weight to synthesize a dynamic edge weight, and the dynamic edge weight is truncated within the range of zero to one. The contribution coefficient of the waste liquid source is obtained by normalizing the dynamic edge weights of the upstream path of the same discharge stream. Based on the online finite difference model within the most recent stable operating window, the corresponding recovery flow column in the linearized prediction model frozen in the previous effective period, or historical disturbance test records, the sensitivity records of the methanol recovery flow increment and dimethyl carbonate recovery flow increment to the upper limit of the organic waste liquid load and the upper limit of the quality risk in the previous effective period are generated. Based on the waste liquid source contribution coefficient and the sensitivity record, a candidate record for the recovery flow direction is generated, and the candidate record for the recovery flow direction is output to S6, so that S6 applies directional constraints or freezes the methanol recovery flow adjustment amount and the dimethyl carbonate recovery flow adjustment amount. Optionally, S3 includes: Input node records, edge records, multi-time-window process features, and dynamic edge weights into the spatiotemporal graph neural network; The spatiotemporal graph neural network includes a graph message passing layer, a time gating layer, and an inter-segment attention layer. The graph message passing layer aggregates the logistics impact of adjacent device nodes according to dynamic edge weights. The time gating layer extracts lag changes according to short, medium, and long windows. The inter-segment attention layer outputs the coupling weights between the reaction segment, the recovery segment, and the distillation segment. The coupling weights are combined with the hidden states of the nodes to form the coupling representation.

[0009] Optionally, S4 includes: Input the coupling representation, process time series dataset, offline detection timestamp, and material residence time parameters into the time series Transformer model; The time-series Transformer model sets a detection lag mask and a dwell time position code to make the online variables of the current control cycle correspond to the by-product detection variables that arrive late. The time-series Transformer model outputs the byproduct generation rate of dipropylene glycol, the organic waste liquid load, the loss of dimethyl carbonate, the quality risk of battery-grade products, and the risk component of heavy components entrained in the quality risk of battery-grade products, and writes the same control cycle identifier for each output.

[0010] Optionally, S5 includes: Based on the material residuals of each equipment node, the by-product generation rate, organic waste liquid load and dimethyl carbonate loss are non-negatively conserved and corrected. After the quality risk and the risk component of its heavy components are clipped to the range of zero to one, the current cycle correction soft measurement results are obtained. Based on the adjustment amount of each candidate distillation cut point and the control candidate sequence generated before the start of step S5 by the control boundary table, the executable control quantity of the previous cycle, the action rate limit, and the preset enumeration step size, a set of candidate action prediction results is generated. Combined with the column top temperature, sensitive plate temperature, column bottom temperature, column section pressure, reflux ratio, and production rate, column section cutting constraints are formed, and candidate action prediction results that do not meet the battery-grade composition limit are deleted. The set of candidate action prediction results is only used for candidate cutting feasibility flag generation, constraint screening, feasibility judgment, and objective function evaluation. When the candidate action prediction result set is empty, retain the previous valid current cycle correction soft measurement result, set to conservative control level and enable conservative control state. The conformal prediction quantile is determined based on the non-consistency score of the calibration sample. The upper and lower bound intervals of each output and the upper bound of the risk of the recombinant component are generated using the conformal prediction quantile and the calibration scale of the corresponding output. The uncertainty level is determined according to the interval width and the preset interval in which the quality risk upper bound is located. Generate and output the current cycle correction soft measurement result, candidate action prediction result set, correction decision field, safety status set and manual review flag. The correction decision field includes candidate cutting feasibility flag, uncertainty level, risk score, specification status, residual constrained status, invalid input status, model confidence threshold trigger status and manual review status. Furthermore, the conformal prediction interval is used for hierarchical control, specifically including: normalizing the absolute values ​​of the prediction errors of by-product generation rate, organic waste liquid load, dimethyl carbonate loss, quality risk and heavy component entrainment risk components according to the corresponding calibration scale and using them as non-consistency scores. In a calibrated sample set with a target coverage of 90% to 99%, quantiles are determined using finite sample order statistics. Multiply the quantile value by the corresponding output calibration scale, and then add or subtract it from the aforementioned corresponding output to obtain the upper and lower bounds; When the upper bound exceeds the corresponding preset control threshold, the upper bound is input into the model to predict the risk penalty term of the control objective function. When the upper bound does not exceed the recovery threshold and the interval width does not exceed the width threshold, output the conservative control recovery criterion.

[0011] Optionally, S6 includes: Read the current cycle correction soft measurement results, candidate action prediction result set, correction decision field, safety status set and manual review flag generated by the correction process, and select the control branch in the priority order of invalid input, empty candidate action prediction result set, threshold version mismatch, manual review flag, model confidence threshold trigger, no feasible solution, ordinary conservative level and normal level; The control state vector is composed of the upper limit of the by-product generation rate, the upper limit of the organic waste liquid load, the upper limit of the dimethyl carbonate loss, the upper limit of the quality risk, the upper limit of the heavy component entrainment risk, the reaction load, the methanol recovery flow rate, the dimethyl carbonate recovery flow rate, the reflux ratio, the reboiler load, and the cut point position corresponding to the current cycle correction soft measurement result. The control state vector does not use the candidate prediction value in the candidate action prediction result set as the current control state. The control vector consists of the reaction load adjustment, methanol recovery flow rate adjustment, dimethyl carbonate recovery flow rate adjustment, reflux ratio adjustment, reboiler load adjustment, and distillation cut-off point adjustment. The prediction model is identified after the control state vector, control vector and disturbance vector are normalized according to the corresponding calibration scale matrix. The candidate action prediction results set is used for constraint screening, feasibility judgment and model predictive control objective function evaluation. The model predictive control objective function is solved under the limits of battery-level composition, equipment load, recovery flow rate, tower section temperature and pressure, emission load non-negative limit and risk upper limit. The objective function includes by-product penalty term, waste liquid load penalty term, solvent loss penalty term, quality risk penalty term, heavy component entrainment risk penalty term and action change penalty term after threshold normalization. When the uncertainty level reaches the preset conservative control level, the emission load limit and quality risk limit are tightened, and the allowable range of the response load adjustment is narrowed. Furthermore, the adjustment amount of the distillation cut-off point is determined based on the column section sensitivity matrix, specifically including: a candidate input vector composed of the column top temperature gradient, sensitive plate temperature gradient, column bottom temperature gradient, column section pressure gradient, reflux ratio disturbance and candidate cut-off point adjustment amount, and the sensitivity matrix output of the change in dimethyl carbonate loss, the change in the risk of heavy component entrainment and the change in the load of waste liquid containing organic matter. When the sensitivity matrix output is a change normalized according to the corresponding upper bound calibration scale, it is first inversely normalized according to the corresponding upper bound calibration scale to become the change in engineering quantity; The upper bound of the risk carried by the recombinant component is the upper bound obtained by the conformal prediction interval of the risk component carried by the recombinant component. For each candidate cutting point adjustment, calculate the candidate-level cutting feasibility indicator and the emission reduction direction consistency indicator. The emission reduction direction consistency indicator is set to allow when the upper limit of dimethyl carbonate loss after candidate adjustment does not increase, the upper limit of heavy component entrainment risk does not increase, and the upper limit of organic waste liquid load decreases. Only the candidate cut point adjustment amount that meets the candidate-level cut feasibility indicator, the emission reduction direction consistency indicator, and the battery-level composition limit is used as the candidate set of distillation cut point adjustment amount and input into the model predictive control. The model predictive control determines the distillation cut point adjustment amount within the candidate set.

[0012] The beneficial effects of this invention are: 1. By constructing a directed process topology diagram based on equipment, logistics, recycling flow, and emission flow, and expressing the synchronous changes in material residuals, by-products, waste liquid load, and solvent loss with dynamic edge weights, the contribution of waste liquid sources can be explicitly input into subsequent control links, thereby improving the targeting of emission reduction adjustments.

[0013] 2. By extracting the coupling relationship between the reaction section, recovery section and rectification section through a spatiotemporal graph neural network, and characterizing the detection lag and material residence time through a time-series Transformer, the by-product generation rate, organic waste liquid load, dimethyl carbonate loss and battery-grade product quality risk can be evaluated within the same control period.

[0014] 3. By correcting the soft measurement results through material balance, tower section cutting constraints, and conformal prediction intervals, and inputting the normalized upper bound risk into the constraint model for predictive control, conservative control can be implemented when uncertainty exceeds the limit, reducing quality risks caused by detection lag or model deviation. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This invention describes the overall process from data acquisition to model predictive control and conservative branch judgment.

[0016] Figure 2 This describes the processing flow in S5 that outputs the correction results and safety status after candidate set judgment, conformal prediction classification, and recovery condition judgment. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1 An electrolyte solvent emission reduction method, applied to a continuous dimethyl carbonate production unit, includes: S1. Collect the operating variables, composition variables, by-product detection variables, and waste liquid organic matter variables of the continuous dimethyl carbonate production unit, and generate a process time series dataset according to a unified sampling period; S2. Construct a directed process topology graph according to the reaction equipment, recovery equipment, distillation equipment, material flow, recovery flow, and discharge flow, setting the equipment as nodes, the material flow, recovery flow, and discharge flow as edges, and writing the multi-time window process features into the corresponding nodes or edges; S3. Process the directed process topology graph using a spatiotemporal graph neural network to obtain the coupled representation of the reaction section, recovery section, and distillation section; S4. Input the coupled representation and the process time series dataset into a time series Transformer model to obtain the by-product generation rate, organic waste liquid load, dimethyl carbonate loss, and battery-grade product quality. S5. The output of step S4 is corrected using material balance constraints, tower segment cutting constraints, and conformal prediction intervals to obtain the corrected soft measurement results, uncertainty level, correction decision field, and safety state set; S6. The constrained model predictive control is executed with the corrected soft measurement results, uncertainty level, correction decision field, and safety state set as inputs, and the reaction load adjustment, methanol recovery flow rate adjustment, dimethyl carbonate recovery flow rate adjustment, reflux ratio adjustment, reboiler load adjustment, and distillation cut point adjustment are output. When the uncertainty level reaches the preset conservative control level, the input is invalid, or the candidate set is empty, the conservative adjustment of the above six types of control variables is output according to the conservative control branch.

[0019] In this specific embodiment, S1 includes: The electrolyte solvent production emission reduction control system is deployed on the production control server of the dimethyl carbonate continuous production unit. This production control server receives raw production records through a distributed control system interface, an online chromatography interface, a wastewater total organic carbon detection interface, and a laboratory information system interface, and assigns them a control cycle number. Time series dataset used in the primary key establishment process , The record fields include , , , , , and ,in Used to distinguish between online variables, offline byproduct variables, and waste liquid organic matter variables. Used to mark missing, out-of-bounds, and calibration-needed states; The acquisition module in each control cycle Reading reaction load Raw material flow rate Reaction temperature Reaction pressure Temperature of the section of the reactive distillation column and pressure Temperature of the dimethyl carbonate distillation column section and pressure Reflux ratio Reboiler load Production volume at the top of the tower Lateral extraction volume , Tower bottom output Methanol recovery flow rate Dimethyl carbonate recovery flow rate and sewage flow And write the engineering units of each variable into Field; Online composition data is provided by the online chromatography system. The acquisition module writes the molar fractions of methanol, dimethyl carbonate, light components, heavy components, and water into the composition vector. First, remove components whose instrument status is faulty or less than zero; then, when the components participating in normalization are... satisfy And each At that time, the acquisition module is in Press at time Perform normalization, where The effective threshold is defined and given by the online chromatographic calibration table; when If out-of-bounds components exist, the acquisition module does not generate a normalized component vector, but instead writes it into the database. And then move the sample to the exception cache; Offline byproduct data is written to the laboratory information system after the test is completed, and the acquisition module reads the dipropylene glycol concentration. Recombination peak area Internal peak area and sample collection time The offline byproduct data is written to the corresponding historical window according to the sampling timestamp. The historical window matching threshold is set to a sampling timestamp that is no more than half a sampling period from the center of the window; whereby... The unit is milligrams per liter, and the sample source field is... The dimethyl carbonate stream or the sludge buffer tank stream is indicated by the dilution factor in the numerical segment. Recorded by the laboratory information system; the data acquisition module is set to... Generate offline byproduct generation rate labels, where The mass flow rate of the sample's logistics at the sampling time is expressed in kilograms per hour. This refers to the density of the logistics, expressed in kilograms per liter. The conversion factor from milligrams to kilograms is obtained. The unit is kilograms per hour; The acquisition module also converts the peak area of ​​the recombinant fraction into a risk label for the recombinant fraction, first calculating the normalized peak area of ​​the internal standard. Then, according to the calibration curve obtained before deployment. The estimated concentrations of the recombinant components were obtained, where , and the same unit zero protection quantity All data are from the chromatographic peak area calibration table; calculations were then performed. ,in This refers to the upper limit of the heavy component concentration in the battery grade specification sheet. As a risk assessment metric for recombinant components, The risk label is a dimensionless value in the range of zero to one; the acquisition module writes the detection method, calibration curve version, sample reference, and spectrum batch number into the data. Field; When the current control cycle If the corresponding offline value has not yet been received, Write in The online variables are retained for subsequent lag alignment in the time series model. When offline records are returned, the acquisition module relocates the historical window according to the sample acquisition timestamp without overwriting the current period's online value, and writes the historical window number, detection completion time, and recalculation flag into the data. Field; The organic matter index of the waste liquid is generated by the sewage flow meter and the total organic carbon detection interface. The acquisition module calculates the observed organic matter load of the waste liquid. ,in The value is the waste liquid mass flow rate, expressed in kilograms per hour. Total organic carbon concentration, expressed in milligrams per liter. The density of the waste liquid is expressed in kilograms per liter. The conversion factor from milligrams to kilograms is obtained. The unit is kilograms per hour; the system will Write The waste liquid density version number, conversion factor source, and detection interface number are written into the database. Field; The alignment module sets the sampling grid with a control cycle of one to ten minutes. In this specific embodiment, the sampling cycle is set to five minutes. When online variables fall into the same sampling grid, the time-weighted average is taken. Offline variables are entered into the history window according to the sampling timestamp. Missing online values ​​are continued by the previous valid value for no more than two control cycles and written into the history window. If the variable is invalidated after two control cycles, a missing mask will be written into the subsequent model input. ; The feature generation module revolves around the current cycle. Construct a short window , middle window and long window In this specific implementation method Covering three control cycles, The coverage is based on twelve control cycles determined by the material residence time. Covering the most recent 24 hours, window features include mean. Slope ,extremum and effective sample ratio ,in Obtained and written from the slope of the least squares line of Field; When conflicts exist between online chromatography, wastewater discharge detection, or offline detection, the acquisition module generates conflict flags according to the data source priority. Online chromatography is used to determine current composition trends, offline byproduct values ​​are used for historical window calibration, and waste liquid organic matter indicators are used for emission load observation. If two valid online values ​​for the same material component appear in the same period, the record with the timestamp closer to the center of the sampling grid is retained, and the other record is written to the exception cache. For manual review; After data acquisition and alignment are completed, the production control server outputs the process time series dataset. Window feature set Missing mask and data quality indicators And write these objects into the time series database. The table is used for reading during step S2 when constructing the directed process topology graph.

[0020] In this specific embodiment, S2 includes: The topology building module reads the result obtained in step S1 and And establish a directed process topology graph according to the device flowchart. ,in This includes an ester exchange reactor, a reactive distillation column, a methanol recovery column, a dimethyl carbonate distillation column, a recovery tank, a sludge buffer tank, and a product tank. This includes the feed side, reactant side, top outlet side, side stream outlet side, bottom outlet side, recovery stream side, and discharge stream side, with the side direction consistent with the actual flow direction. Each device node Corresponding to one node record Fields include , , , , , and ,in Temperatures at the top of the storage column, the sensitive plate, and the bottom of the column. Store component vectors estimated jointly by online chromatography and material balance; Each logistics edge Record a corresponding edge Fields include , , , , , and ,in The data was obtained from the device design logistics relationships and historical steady-state sample statistics, showing the recovery flow edge. Set as one, discharge flow side Set to one; The topology building module writes the short window, medium window, and long window features from step S1 into nodes or edges; temperature, pressure, load, and composition features into the corresponding device nodes; and flow rate, recovery flow rate, extraction flow rate, and sewage discharge flow rate features into the corresponding edges. The writing rules are determined by variable labels. Mapping table between node number or edge number Confirmed, not in Variables that are hit are written to the exception cache and do not participate in the current cycle's topology update; To calculate material residuals, the topology building module performs calculations on each device node. Establish a closed-loop quality management system ,in For the edge In the cycle mass flow rate The rate of change of liquid holdup at the equipment node. The unit is kilograms per hour; the system further calculates the signed residual ratio. and residual strength ,in This represents the total throughput of the node. For a zero protection level of one kilogram per hour, all subsequent threshold triggers will use non-negative values. ; The dynamic edge weight correction module in the historical calibration period set Internal calculation of residual strength Changes in the rate of by-product formation Waste liquid load variation and the change in the amount of dimethyl carbonate loss The synchronous change relationship; before calculating the synchronous change relationship, the module will adjust each change quantity according to the historical calibration scale of the same type of output. , and Convert to dimensionless change , and And comparisons are only made under the same control cycle, the same sampling window, and the same calibration version; residual correlation coefficient The result is obtained by weighting the residual strength with the synchronization coefficients of three dimensionless variables, all of which are limited to the range of -1 to 1, and then weighted according to the calibration weight table. , and Weighted synthesis, where the weights sum to one, yields... It is a dimensionless quantity, and the larger the value, the stronger the contribution of the node residual to the same source as emission load, solvent loss and by-product changes; The edge weight update rule is ,in For the edge The initial edge weights, For the edge The upstream node, This is a correction factor for the edge type. This means truncating the results to the range of zero to one. If the effective traffic field of a node is less than 80% of the number of fields required for that node, then it is preserved. And write ; The waste liquid source contribution coefficient is obtained by normalizing the dynamic edge weights of the upstream path of the emission flow, and the emission edge... upstream path set Topology building block calculation ,in For path The contribution coefficient to the emission flow, if the denominator is zero, is written according to the contribution ratio under historical steady-state operating conditions. And mark it as pending calibration; The topology building module also generates candidate records for the recovery flow direction based on the contribution coefficient of the waste liquid source. This record includes the contribution coefficient of the methanol recovery path. Contribution coefficient of dimethyl carbonate recovery pathway Methanol contribution threshold Dimethyl carbonate contribution threshold The process involves retrieving the flow sensitivity field, the field to be calibrated, and the field that allows adjustment direction. The retrieved flow sensitivity field is derived from the online finite difference model within the most recently stable operating window and the linearized prediction model frozen in the previous cycle. The corresponding recycled traffic column, or generated from historical disturbance test records after version freezing, includes the following fields: , , and , respectively, represent the dimensionless change direction of the upper bound of waste liquid load and the upper bound of quality risk in the previous effective period after the methanol or dimethyl carbonate recovery flow rate increases by one action scale. Positive values ​​indicate that the upper bound increases, and non-positive values ​​indicate that the upper bound does not increase. When the number of effective samples is lower than the sensitivity sample threshold, the disturbance direction is inconsistent with the current control boundary, or the sensitivity version is not frozen before the start of the current period, the direction field is written as uncertain and manual review is triggered. The recovery flow adjustment constraint is not triggered solely by the contribution coefficient, but only when , and ,and When not marked as requiring calibration, the direction field records the methanol recovery flow rate increase direction allow flag and writes the candidate direction. Only when , and ,and When not marked as requiring calibration, the direction field records the dimethyl carbonate recovery flow rate increase direction allow flag and writes the candidate direction. When the sensitivity direction is uncertain, the contribution coefficient is backfilled according to historical stable operating conditions, the contribution coefficient needs to be calibrated, or the recovery equipment is close to the load limit, the direction field is written as uncertain or manually reviewed; step S2 does not freeze the recovery flow adjustment amount for this cycle. Follow They are output together to step S6 as candidate constraints for model predictive control; After the topology is constructed, the system outputs a directed flow topology graph. Node feature matrix Edge feature matrix Dynamic edge weight matrix Contribution coefficient of waste liquid source and will Version number Write it to the graph cache for the spatiotemporal graph neural network to read in step S3.

[0021] In this specific embodiment, S3 includes: The soft measurement server reads the directed flow topology graph output in step S2. Node feature matrix Edge feature matrix and dynamic edge weight matrix And call the spatiotemporal graph neural network The coupling relationships between the reaction section, recovery section, and rectification section are encoded. The input retains the device node number, edge direction, recycle flow flag, and discharge flow flag to ensure that graph messages are transmitted along the actual logistics direction; The graph message passing layer communicates with each node. Compute neighborhood messages ,in For the first Hidden state of upstream nodes in the layer For edge recording features, The dynamic edge weights obtained in step S2 The message summation result is written to the node's hidden state cache as a combination of linear transformations with layered normalization and nonlinear activations. Node status updates adopt ,in For nodes Window feature concatenation vector, , and For parameter matrices, The hyperbolic tangent activation function outputs... This indicates the combined impact of logistics and status of a node in the current control cycle; The time-gating layer reads the features of short, medium, and long windows respectively, and then... , and Calculate the three gating coefficients and generate the time-fused state. All three gating coefficients are dimensionless numbers in the range of zero to one; The inter-segment attention layer groups the device nodes into reaction segments, recovery segments, and distillation segments, forming segment vectors for each segment. , and First, calculate the original attention logic quantity. ,in and For the projection matrix, For the projection dimension; for a fixed source segment Inter-segment attention layer in all target segments Normalization yields , Used to indicate source segment For the target segment The direction of coupling contribution; When the dynamic edge weight of a certain recycling flow edge is marked as in step S2 In this case, the spatiotemporal graph neural network does not delete the edge, but instead sets the data quality component in its edge features to one and adds a quality penalty term to the attention layer. The coupling weights after penalty are calculated as follows: Renormalization, where the normalization dimension remains the same as the source segment. All target segments below, As given in the quality penalty parameter table, the quality penalty only applies to the original logic quantity and not directly to the normalized quantity. Execute softmax again; The spatiotemporal graph neural network was trained using historical batch data before deployment, with training samples... As input, the calibration results of offline byproduct generation rate, waste liquid organic load, and dimethyl carbonate loss are used as supervision signals. The loss function consists of a prediction error term, a side weight smoothing term, and a material residual consistency term. The model parameters are frozen and passed during deployment. Field records; The soft measurement server will display the final hidden state of each node. Inter-segment coupling weights Waste liquid source contribution coefficient and data quality indicators Combined into a coupled characterization and will Write Cache; if any critical segment vector is missing, use the segment vector from the most recent valid period and... Set to pending review; Coupling characterization output from step S3 It includes node hidden states, inter-segment coupling weights, path contributions, and quality penalty information. The coupling representation has the same control cycle number as the process time series dataset in step S1. And it serves as the graphical input to the temporal Transformer model in step S4.

[0022] In this specific embodiment, S4 includes: The timing prediction module reads the coupling characterization output from step S3. and the process time series dataset formed in step S1 And construct the input sequence ,in In this specific implementation, the sequence length is set to twenty-four control cycles. It is obtained by splicing together online process variables, window features, missing mask, offline detection timestamps, dwell time encoding, and coupled representations; Dwell time location coding is calculated based on the logistics path and equipment liquid holding capacity. The time series prediction module analyzes the path... calculate ,in For node liquid holdup, This represents the total throughput of the node. For zero protection quantity in the same unit, The unit is hours, and the model will The mapping is performed as a sine / cosine position code and added to the input sequence; The detection lag mask is used to align offline byproduct detection values ​​with the actual sampling time. The time-series prediction module creates a mask for each offline sample. When the sample collection time Belongs to the input sequence window and the detection result is within the period When the previous has already returned ,otherwise ; for the position of the input sequence The module maps sampling time. Obtain and only The location is included in the offline label calibration channel; Temporal Transformer Model It includes a process encoding layer, a delayed attention layer, and a multi-task output layer. The process encoding layer will... Mapped to and the current control cycle Online variables, coupled representations, and quality indicators are mapped to query vectors. The delayed attention layer first forms an effective location set. ,when When empty, directly use the current period code containing only online variables and... Write to the output record when Press when not empty Calculate attention weights, where For a matrix, For sequence position Corresponding logistics routes The dwell time bias varies with candidate location; when Direct order This prevents the blocked offline tags from receiving non-zero attention weights; The multi-task output layer simultaneously outputs the byproduct generation rate. Organic waste liquid load Dimethyl carbonate loss Battery-grade product quality risks and the risk component of recombinant components ,in , and The units are all kilograms per hour. and All are dimensionless risk values ​​in the range of zero to one; To ensure consistent output direction, both training and deployment specify... The larger the value, the higher the quality risk of the battery-grade product; the quality risk is caused by the risk carried by the recombining components. Risk of excessive moisture Risk of deviation from dimethyl carbonate purity Weighted formation, in which , , ; This is the moisture content reading. The purity of dimethyl carbonate, and From the battery-grade specifications sheet, and From the risk assessment scale table, , and The weights are derived from the quality risk weight table and are non-negative and sum to one. The weight table is frozen in version according to product specifications before the start of each control cycle. When the offline byproduct labels in the input sequence have not yet been returned, the time series prediction module uses online composition changes, material residuals, and coupling characterization to predict the current cycle, while simultaneously... Set to 1. After the offline tags are returned, the historical error samples are updated through the calibration cache, but the control quantities already issued in the current control cycle are not written back, so as to avoid the use of future information for prediction and control in the same cycle. The temporal Transformer model was trained using historical batches before deployment, with the training objective being to normalize the multi-task loss. ,in , , , and For historical calibration values, , , , and For the corresponding output calibration scale parameters, to The weights are dimensionless and derived from the training configuration table. Used to limit output abrupt changes between adjacent cycles; After inference is completed, the time series prediction module will identify the five prediction outputs and the same control cycle. Model version Hysteresis mask summary Input quality flags are written into the soft measurement raw output table. This table is then used as input for step S5 conservation correction, tower segment cutting constraints, and conformal prediction processing.

[0023] In this specific embodiment, S5 includes: The calibration evaluation module reads the raw soft measurement output from step S4. , , , and Simultaneously read the node material residuals from step S2. Dynamic edge weights Contribution coefficient of waste liquid source And generate correction task records. , The fields include , , , and ; Material balance correction first applies conservation constraints to the byproduct formation rate, waste liquid load, and dimethyl carbonate loss. The correction method is as follows: ,in , For nodes For output The residual contribution coefficients are derived from regression results of historical calibration samples; in this specific implementation, and All are measured in kilograms per hour And the coefficients are dimensionless; if the output... If the measurement reference comes from the volumetric flow rate or concentration label, then Provided and written from the unit conversion table frozen before the start of this period. ; And according to each output satisfy ,when When a node's input side has a positive open-circuit relative to its output side, the deduction term is used to reduce the overestimation of the output caused by the node's residual. When there is a negative open-circuit quantity on the output side relative to the input side, the sign of the deduction term causes the correction output to increase accordingly, but it is still subject to non-negative clipping; in terms of the anomaly handling sequence, the correction evaluation module first removes nodes whose residual strength exceeds the upper limit and then... Set the values ​​to zero, then allocate the contribution coefficients of nodes sharing the same residual source according to the dynamic edge weight ratio. Next, perform a non-negative closed projection on the remaining nodes and re-normalize them. If the remaining nodes are empty, the projection fails, or the unit conversion table version does not match, freeze the previous valid correction result and write it to the database. Quality risk does not participate in the quality flow conservation deduction; the correction and evaluation module defines it as... And define the risk component carried by the recombinant component as Both are dimensionless correction outputs in the range of zero to one; When the residual strength of any device node Exceeding the upper limit of the calibrated residual at this node At that time, the correction and evaluation module does not directly use the node residual to correct the soft measurement results, but instead uses the node contribution coefficient. Set to zero and write If at least one node does not exceed the limit, the correction evaluation module renormalizes the remaining nodes according to the non-negative closed projection rule described in the previous paragraph. If no node exceeds the limit, the previous valid correction soft measurement result is retained and... Enabled; at the same time, the contribution coefficient of the waste liquid source in the upstream path of this node is only retained for the constraint direction judgment of the model predictive control, and will no longer participate in the current cycle conservation deduction.

[0024] Candidate action prediction result set Generated by the candidate generation module, the candidate sources include the adjustment amount of each candidate cut point in the preset distillation cut point candidate library. and the constraint control candidate sequence generated independently before the start of step S5. The From the control boundary table Executable adjustment amount in the previous period The generation of action rate limits, preset enumeration step sizes, and boundary version numbers is independent of the solver output in step S6 of this cycle. The candidate record fields must include at least the candidate number, six types of control variable adjustments, boundary version, action rate hit flag, and validity flag. The candidate generation module uses the original output of step S4 as a benchmark, and combines the predicted impact of candidate cut point adjustment and candidate control vector on the light components at the top of the column, the temperature gradient of the sensitive plate, the heavy components at the bottom of the column, and the composition limits of the battery level to generate candidate-level cut feasibility indicators. and candidate action prediction output When any of the candidate adjusted battery level composition limits, tower section temperature and pressure limits, or heavy component upper limits are not met, And delete the corresponding candidate. If after deletion If empty, retain the previous valid correction result as the frozen output. Set to conservative control level and enable it. Step S6 is processed as a branch with no feasible solution; Candidate set elements Including conservation correction output and candidate-level cutting feasibility indicator. Specification hit flag and candidate level output range width; when When not empty, the correction and evaluation module performs a check on each candidate. Predicted output based on candidate actions Conformal quantiles frozen before the start of the current period and calibration scale Generate upper and lower bounds for candidate levels, where hour and , hour and Then the dimensionless width was calculated. and ,in The waste liquid load control threshold, As a quality risk threshold, and To correspond to the calibration scale; then according to the evaluation value Sort the candidate elements. to The weights are non-negative dimensionless and sum to one, and are derived from the candidate ranking weight table and the versions are frozen by product brand before the start of the control period. This is the benefit evaluation value. An increase in the value indicates that the candidate result is more suitable as a constraint screening and objective function evaluation object. The sorting direction is by... The first candidate in descending order is selected as the priority candidate in the candidate action prediction result set; if two candidate elements... If they are the same, then choose the upper bound of quality risk. Candidate elements with lower values ​​are treated as parallel results; Battery class composition limits are specified in the specification sheet. This table provides fields including lower limit of dimethyl carbonate purity, upper limit of moisture content, upper limit of methanol residue, upper limit of heavy components, applicable product grade, testing method, and version number. If the current product grade does not match the specification table, the calibration and evaluation module will retain the previous valid specification version and... Set to pending review; Conformal prediction intervals are based on the calibration sample set Establishment, frozen before the start of each control cycle. The version numbers of the risk threshold table and control weight table are used; newly added offline error samples in the current period are only written to the next calibration version; the calibration evaluation module processes each output... Calculate the non-consistent score ,in For historical calibration values, This is the correction output for the corresponding period. For calibration scale parameters of the same unit or dimension, fractions Dimensionless values ​​and stored surface; In target coverage Set within the range of 90% to 99%, this specific implementation method takes... The calibration evaluation module will freeze the version. The non-consistent scores are sorted in ascending order and their order is taken. corresponding As quantiles of a finite sample; subsequently, for generate and ,right generate and ,in and Output respectively The lower and upper bounds, That is, the upper bound of the risk associated with the recombinant component; The risk penalty score for conformal intervals is defined as follows: , , , , and To correspond to the preset control threshold, to The risk weights are non-negative and sum to one. As a penalty term, an increase in the value indicates a higher risk of emissions and air quality in the current cycle, which will be addressed through subsequent model predictive control to reduce emissions. Set risk penalties for directions; Risk threshold table Record , , , , and The unit, applicable product brand, calibration sample range, and version number; among which , and The units are consistent with the upper bound of the corresponding mass flow rate. and A dimensionless risk threshold in the range of zero to one. This is a dimensionless risk penalty threshold; when the threshold table version does not match the current product brand, the correction evaluation module does not output a normal level, but instead... Set to conservative control level and write This causes step S6 to be executed according to the set of safe states; The uncertainty level is determined by the interval width and the upper bound of the quality risk, and the correction evaluation module calculates it. When any Exceeding the preset interval width threshold Upper limit of quality risk Exceeding the quality risk threshold Upper limit of risks associated with recombinant components Exceeding the risk threshold of recombinant components or risk score Exceeding the risk score threshold At that time, Set to conservative control level, threshold , , and All are from the calibration threshold table; The classification results also generate correction decision fields for step S6 to read. ,in Indicates whether candidate cut adjustments can be used for control. Indicates the level of uncertainty. Represents the monotonic score of risk; set of safe states include , , , and Fields; the correction and evaluation module will be in step S1 The following are summarized as follows: key online values ​​exceeding the hold period, necessary traffic or missing component fields, and missing necessary topology fields in S2. In step S3 In step S4 Model confidence is below the threshold and Summarized as The following are summarized as follows: contribution coefficients to be calibrated, specification versions to be reviewed, uncertainties in the direction of recovered flow sensitivity, and versions of the cutting sensitivity matrix not frozen. The anomaly priority is determined in the following order: invalid input, empty candidate action prediction result set, threshold version mismatch, manual review flag, low confidence model flag, and normal conservative level. Any high-priority anomaly will trigger an exception. Enable it and disable step S6 to exit the conservative boundary; The conservative control recovery criterion uses the upper bound and interval width as comparison benchmarks, and the correction evaluation module reads the recovery threshold table. This table records , , , , , Number of consecutive periods Applicable product brand, control level, and version number; among which , and The units are respectively consistent with the upper bound of the corresponding mass flow rate and satisfy the condition of not exceeding the corresponding control threshold. and A dimensionless threshold between zero and one, and not exceeding [a certain value]. and , The dimensionless width threshold. The historically conservative control mechanism exited sample calibration and the version was frozen before the start of the current cycle; the calibration and evaluation module was continuously... Each control cycle simultaneously satisfies , , , , And all At that time, Set to pending exit state; if any condition is not met, remain in the pending exit state. Enable and output conservative control recovery flag in step S6 New recovery samples added in the current cycle are only written to the next recovery threshold version; After calibration is completed, the calibration evaluation module outputs the calibration soft measurement results for the current period. Upper and lower bound intervals Candidate action prediction result set Uncertainty level Feasibility indicators for candidate-level cutting Correcting decision fields and safe state set and write these objects to The table provides the model predictive control call for step S6; the current periodic correction soft measurement results are used to construct the control state vector, and the candidate action prediction result set is only used for constraint screening, feasibility judgment and objective function evaluation.

[0025] In this specific embodiment, S6 includes: The model predictive control module first reads the current periodic calibration soft measurement result output in step S5. Upper and lower bound intervals Candidate action prediction result set Uncertainty level Correcting decision fields Set of safe states and manual review mark And read the waste liquid source contribution coefficient from step S2. The control module selects the control branch in a unique order: invalid input, empty candidate action prediction result set, threshold version mismatch, manual review flag, low confidence model, no feasible solution, ordinary conservative level and normal level. When any high-priority anomaly occurs, the ordinary model prediction control solution result shall not be issued. The control state vector is defined as follows: ,in , , , and These are the upper bounds for byproduct formation rate, waste liquid load, dimethyl carbonate loss, quality risk, and risk of entrainment of heavy components, respectively. and These are the methanol recovery flow rate and the dimethyl carbonate recovery flow rate, respectively. The distillation cut point position is encoded; the control state vector is generated only using the current cycle correction soft measurement result and its upper and lower bound intervals, and does not use the candidate prediction value in the candidate action prediction result set as the current control state. Control vector is defined as Each component represents the adjustment amount of the reaction load, methanol recovery flow rate, dimethyl carbonate recovery flow rate, reflux ratio, reboiler load, and distillation cut-off point, respectively. The control module sets the unit, upper and lower limits, and action rate limits of each component from the control boundary table. Read from; The prediction model uses linearized state updates. ,in It is an eleven-dimensional state vector. It is a six-dimensional control vector. The perturbation vector; matrix , and From the most recently stable running window It can be identified. The requirements are: no high-priority anomalies within the window, the load of major equipment not reaching the upper or lower limits, and the number of valid samples not less than [number missing]. The pre-identification control module reads the state scale matrix. Action scale matrix and perturbation scale matrix The state residuals, control variables, and disturbance variables are converted into dimensionless coordinates to identify the target. ;in The diagonal elements are the control threshold or risk scale for the five categories of upper bounds, the reaction load scale, the recovery flow rate scale, the reflux ratio scale, the reboiler load scale, and the cut-off point location scale, respectively. All scales are from the control scale table frozen before the start of the current cycle. When the matrix dimensions do not match, the number of samples is insufficient, the scale table is missing, or the identification residual exceeds the threshold, the control module freezes the previous effective model version and switches to the conservative control branch. The matrix version number is frozen before the start of each control cycle. The model predictive control objective function is ,in , , , and For the corresponding control threshold, The motion scale is a diagonal matrix. to For dimensionless risk penalty weights, It is a dimensionless motion change penalty matrix, with all weights derived from the control weight table and maintained as a fixed version under the same product brand and control level; The constraint set includes the upper bound nonnegativity constraint on mass flow rate. , and Risk upper bound constraint and Battery class composition constraints Recombination component with constraint Equipment load constraints Methanol recovery flow rate constraint Dimethyl carbonate recovery flow rate constraint Tower section temperature and pressure constraints and emission load constraints Each threshold is read from the corresponding equipment and product specification table and written into the solution record for this cycle. The control module reads the candidate records of the recovery flow direction output in step S2 and the upper and lower bound intervals output in step S5, and sets the waste liquid source contribution coefficient. The mapping is used to adjust the direction of the recycled flow; the recycled flow sensitivity record is generated by the online finite difference, linearized prediction model within the most recent stable operating window. The methanol recovery flow rate series and dimethyl carbonate recovery flow rate series were obtained, or from historical perturbation test records. All sensitivities were set before calculation. and Normalize and freeze the version before the start of this cycle. , , and The positive direction is defined as an increase corresponding to the upper bound; when the contribution coefficient of the methanol recovery path leading to the emission stream... Exceeding the contribution threshold Furthermore, the sensitivity of the methanol recovery flow rate increment to the upper limit of waste liquid load and the upper limit of quality risk is... and When, add Directional constraints are used to prohibit reducing the methanol recovery flow rate; when the dimethyl carbonate recovery path contribution coefficient Exceeding the contribution threshold ,and and When, add Directional constraints are used to prohibit reducing the dimethyl carbonate recovery flow rate; if the number of effective sensitivity samples is insufficient, the direction is uncertain, or the recovery tower is close to its load limit, or If it is marked as pending calibration, the corresponding recovery flow adjustment amount will be frozen according to the executable adjustment amount of the previous cycle and written into the manual review reason field; The candidate set for distillation cut-off point adjustment is filtered by the column segment sensitivity matrix, and the control module constructs the matrix. The input column includes the candidate cut point adjustment amount. Reflux ratio disturbance Temperature gradient at the top of the tower Sensitive plate temperature gradient Temperature gradient of the tower bottom and tower section pressure gradient The output rows include changes in dimethyl carbonate loss, changes in the risk of heavy component entrainment, and changes in waste liquid load; for candidate elements structure , and according to Obtain the normalized candidate influence values; Identification was achieved through a combination of historical disturbance tests and online finite-difference spectroscopy. Input quantities were normalized according to the control boundary table, and output quantities were normalized according to the corresponding upper bound calibration scale. In this specific embodiment, the positive direction is defined as the movement of the cutting point towards the top of the tower; the control module then presses... , and Generate changes in engineering quantities, among which , and These are the upper bound calibration scales for dimethyl carbonate loss, heavy component entrainment risk, and waste liquid load, respectively; when the cutting point actuator is missing, the sensitive plate temperature is unavailable, or the reverse normalization scale is missing, or... When the version is not frozen, the adjustment amount of the candidate cut point is set to zero and a manual review flag is written; Adjustment amount for each candidate cutting point The control module calculates the feasibility indicators for candidate-level cutting. Consistency mark with emission reduction direction When the candidate adjusted battery level composition limits, tower section temperature and pressure limits, and heavy component upper limits are all met, ,otherwise ;when , and At the same time ,otherwise ,only and The candidate quantities are used as the candidate set of distillation cut point adjustment quantities and input into the solver. The model predicts and controls the determination of the distillation cut point adjustment quantities within this candidate set. All the above comparisons are performed on the same scale of the engineering quantity or the risk upper bound of zero to one. The normalized influence quantity is not directly added to the engineering quantity upper bound.

[0026] when When the conservative control level is reached, the control module tightens the emission load threshold to Tighten the quality risk threshold to And the allowable range for reaction load adjustment was changed from Narrowed to ,in Units and Same and satisfy , For dimensionless risk tightening quantity and satisfying , and From the conservative control parameter table; if there is no feasible solution after tightening, then output the conservative adjustment amount of the six types of control variables according to the safety control table; No feasible solution branch is determined by the solver state. The candidate action prediction result set is empty or a high-priority exception is triggered; the control command table stores the executable adjustment amount from the previous cycle. Execution goals of the previous cycle Use when conservative branches need to freeze adjustment amounts. Use when the conservative branch needs to freeze the absolute execution target. ,in These are the currently measured control variables. For action rate limits; when At that time, the control module will Set to zero, , , and Output according to the above freeze adjustment rules or safety control table values, and according to... Output safe load reduction adjustment amount; when When, the adjustment amounts for the six types of control variables are adopted from the safety control table values, the executable adjustment amounts of the previous cycle, or the adjustment amounts calculated from the execution targets of the previous cycle, and marked as pending review; when When there are no invalid inputs, the control module tightens the threshold and solves the conservative problem. If the solution fails, a safe control table value is used. When other infeasible conditions are triggered, the control module will... Set to zero, and Output according to the freeze adjustment rules, and Set to the freeze adjustment amount or a non-negative safety value allowed by the direction constraint, and then... Output safety response load adjustment amount; The positive load reduction recorded in the safety control table, with units consistent with the reaction load, is ultimately... It is also subject to the action rate limit and the lower limit of equipment load; all six types of adjustment quantities mentioned above are written into the control command table and marked as conservative branch outputs; If no feasible solution occurs for two consecutive control cycles, or If any of the following conditions are met: threshold version mismatch, low confidence model flag, etc., the control module will control the command table. The logo remains the same, while maintaining Enable and send uncertainty exceedance records to the production control system; in this state, even if the ordinary solver returns a feasible solution, the ordinary solution must not be issued. Only when a feasible solution reappears in a subsequent cycle will step S5 be output. Furthermore, the control module is only allowed to restore normal control boundaries when the manual verification flag is cleared; The model predictive control solver outputs the control vector for the previous cycle in each control period. The adjustment amounts for reaction load, methanol recovery flow rate, dimethyl carbonate recovery flow rate, reflux ratio, reboiler load, and distillation cut-off point are written into the control command table. The production control system generates the execution target value based on the current measured value and the adjustment amount, and records the solution status, constraint margin and conservative control flag for feedback and update in the next control cycle.

[0027] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0028] This invention connects process topology, time lag modeling, conservation correction, and constraint control into a continuous technology chain, enabling a calculable correspondence between the adjustment amounts of reaction load, methanol recovery flow rate, dimethyl carbonate recovery flow rate, reflux ratio, reboiler load, and distillation cut-off point and by-products, waste liquid load, solvent loss, and product quality risks.

[0029] This invention further utilizes a dynamic edge weight correction mechanism driven by material balance residuals, a conformal confidence interval hierarchical control mechanism, and a tower segment sensitivity matrix to screen and adjust the cutting point, so that the control output can simultaneously meet the requirements of quality constraints and emission reduction direction consistency.

Claims

1. A method for reducing electrolyte solvent emissions, characterized in that, Applications in continuous dimethyl carbonate production facilities include: S1. Collect operating variables, composition variables, by-product detection variables, and organic matter variables in waste liquid from the continuous dimethyl carbonate production unit, and generate a process time-series dataset according to a unified sampling period. S2. Construct a directed process topology graph based on reaction equipment, recovery equipment, distillation equipment, material flow, recovery flow, and discharge flow, setting equipment as nodes and material flow, recovery flow, and discharge flow as edges, and writing multi-time-window process features into the corresponding nodes or edges. S3. Process the directed process topology graph using a spatiotemporal graph neural network to obtain coupled representations of the reaction section, recovery section, and distillation section. S4. Input the coupled representations and process time-series dataset into a time-series Transformer model to obtain by-product generation rate, organic waste liquid load, dimethyl carbonate loss, and battery-grade product quality risk. S5. The output of step S4 is corrected using material balance constraints, tower section cutting constraints, and conformal prediction intervals to obtain the corrected soft measurement results, uncertainty level, corrected decision field, and safety state set; S6. The constrained model predictive control is executed with the corrected soft measurement results, uncertainty level, corrected decision field, and safety state set as inputs, and the reaction load adjustment, methanol recovery flow rate adjustment, dimethyl carbonate recovery flow rate adjustment, reflux ratio adjustment, reboiler load adjustment, and distillation cut point adjustment are output. When the uncertainty level reaches the preset conservative control level, the input is invalid, or the candidate set is empty, the conservative adjustment of the above six types of control variables is output according to the conservative control branch.

2. The electrolyte solvent emission reduction method according to claim 1, characterized in that, S1 includes: Data collection includes reaction load, feed flow rate, reaction temperature, reaction pressure, temperature and pressure of the reactive distillation column section, temperature and pressure of the dimethyl carbonate distillation column section, reflux ratio, reboiler load, top product output, side stream output, bottom product output, methanol recovery flow rate, dimethyl carbonate recovery flow rate, wastewater discharge flow rate, online composition data, offline by-product data, and organic matter indicators in the waste liquid. Online data is aligned to a sampling period of one to ten minutes. Offline by-product data is written to the corresponding historical window according to the sampling timestamp, and a calibration flag is added for missing offline values. Short, medium, and long window features are generated. The short window covers the current control cycle and the two preceding control cycles and is used to calculate the mean, slope, and extreme values. The medium window covers the statistics of material residence time, and the long window covers the historical distribution characteristics of offline detection lag.

3. The electrolyte solvent emission reduction method according to claim 1, characterized in that, S2 includes: A node record is established for each equipment node, including equipment type, material retention, temperature and pressure statistics, load status, and composition estimate. An edge record is established for each edge, including flow rate, composition, flow direction, recovery flag, discharge flag, and initial edge weight. The material residual of each equipment node is calculated based on the feed flow rate, recovery flow rate, output flow rate, wastewater discharge flow rate, and composition estimate. The edge weights between nodes are then corrected based on the synchronous changes of the material residuals with the byproduct generation rate, organic wastewater load, and dimethyl carbonate loss during the historical calibration period, resulting in dynamic edge weights and wastewater source contribution coefficients.

4. The electrolyte solvent emission reduction method according to claim 1, characterized in that, S3 includes: inputting node records, edge records, multi-time-window process features, and dynamic edge weights into a spatiotemporal graph neural network; the spatiotemporal graph neural network includes a graph message passing layer, a time gating layer, and an inter-segment attention layer; the graph message passing layer aggregates the material flow impact of adjacent equipment nodes according to dynamic edge weights; the time gating layer extracts lag changes according to short, medium, and long windows; the inter-segment attention layer outputs the coupling weights between the reaction segment, the recovery segment, and the distillation segment; the coupling weights are combined with the hidden states of nodes to form the coupling representation.

5. The electrolyte solvent emission reduction method according to claim 1, characterized in that, S4 includes: inputting the coupling characterization, process time series dataset, offline detection timestamp, and material residence time parameters into a time series Transformer model; setting a detection lag mask and residence time position encoding in the time series Transformer model so that the online variables of the current control cycle correspond to the by-product detection variables that arrive later; the time series Transformer model outputs by-product generation rate of dipropylene glycol, organic waste liquid load, dimethyl carbonate loss, battery-grade product quality risk, and the risk component of heavy components entrained in the battery-grade product quality risk, and writing the same control cycle identifier for each output.

6. The electrolyte solvent emission reduction method according to claim 1, characterized in that, S5 includes: Based on the material residuals at each equipment node, non-negative conservation corrections are applied to the byproduct generation rate, organic waste liquid load, and dimethyl carbonate loss. After trimming the quality risk and its associated risk components within a range of zero to one, the current cycle's corrected soft measurement results are obtained. Based on the adjustment amount for each candidate distillation cut-off point and the control candidate sequence generated before step S5 using the control boundary table, the executable control quantity of the previous cycle, the action rate limit, and the preset enumeration step size, a candidate action prediction result set is generated. This set, combined with the column top temperature, sensitive plate temperature, column bottom temperature, column section pressure, reflux ratio, and production rate, forms column section cutting constraints. Candidate action prediction results that do not meet the battery-grade composition limits are deleted. This candidate action prediction result set is only used for generating candidate cutting feasibility indicators, constraint screening, feasibility judgment, and objective function evaluation; it is not used as the control state vector in step S6. The current cycle correction soft measurement result; when the candidate action prediction result set is empty, the previous valid current cycle correction soft measurement result is retained, the conservative control level is set, and the conservative control state is enabled; the conformal prediction quantile value is determined according to the inconsistency score of the calibration sample, and the upper and lower bounds of the current cycle correction soft measurement result and the upper bound of the risk of the recombinant component are generated with the conformal prediction quantile value and the corresponding output calibration scale, and the uncertainty level is determined according to the preset interval where the interval width and the quality risk upper bound are located; the current cycle correction soft measurement result, the candidate action prediction result set, the correction decision field, the safety status set, and the manual review flag are generated and output, and the correction decision field includes the candidate cutting feasibility flag, uncertainty level, risk score, specification status, residual limited status, invalid input status, model confidence threshold trigger status, and manual review status.

7. The electrolyte solvent emission reduction method according to claim 6, characterized in that, S6 include: The current cycle calibration soft measurement result, candidate action prediction result set, calibration decision field, safety state set, and manual review flag generated by the calibration process are read, and the control branch is selected in the following priority order: invalid input, empty candidate action prediction result set, threshold version mismatch, manual review flag, model confidence threshold trigger, no feasible solution, ordinary conservative level, and normal level. The upper limit of the by-product generation rate, the upper limit of the organic waste liquid load, the upper limit of the dimethyl carbonate loss, the upper limit of the quality risk, the upper limit of the heavy component entrainment risk, the reaction load, the methanol recovery flow rate, the dimethyl carbonate recovery flow rate, the reflux ratio, the reboiler load, and the cut-off point position corresponding to the current cycle calibration soft measurement result are used to form a control state vector, and the control state vector does not use the candidate prediction values ​​in the candidate action prediction result set as the current control state. The reaction load adjustment amount, the methanol recovery flow rate adjustment amount, and the dimethyl carbonate adjustment amount are used as the control state. The control vector is composed of the methyl ester recovery flow rate adjustment, reflux ratio adjustment, reboiler load adjustment, and distillation cut-off point adjustment. The prediction model is identified after normalizing the control state vector, control vector, and disturbance vector according to the corresponding calibration scale matrix. The set of candidate action prediction results is used for constraint screening, feasibility judgment, and evaluation of the model predictive control objective function. The model predictive control objective function is solved under battery-level composition limits, equipment load limits, recovery flow rate limits, column section temperature and pressure limits, non-negative emission load limits, and risk upper limits. The objective function includes by-product penalty terms, waste liquid load penalty terms, solvent loss penalty terms, quality risk penalty terms, heavy component entrainment risk penalty terms, and action change penalty terms after threshold normalization. When the uncertainty level reaches the preset conservative control level, the emission load limits and quality risk limits are tightened, and the allowable range of the reaction load adjustment is narrowed.

8. The electrolyte solvent emission reduction method according to claim 3, characterized in that, The dynamic edge weights in step S2 are determined based on the material residuals, specifically including: normalizing the signed material residuals according to the material throughput of the equipment nodes to obtain a signed residual ratio; using the absolute value of the signed residual ratio as the residual strength, and calculating the correlation between the residual strength and the changes in the by-product generation rate, the organic waste liquid load, and the dimethyl carbonate loss in the historical calibration period to obtain a residual correlation coefficient; weighting the residual correlation coefficient with the initial edge weights to synthesize dynamic edge weights, and truncating the dynamic edge weights within the range of zero to one; normalizing the dynamic edge weights of the upstream path of the same discharge flow to obtain the waste liquid source contribution coefficient; and generating the corresponding recovery flow column or historical disturbance test records from the online finite difference model within the most recent stable operating window, the linearized prediction model frozen in the previous effective period, etc. Sensitivity records are generated for the incremental methanol recovery flow and the incremental dimethyl carbonate recovery flow, respectively, to the upper limit of the organic waste liquid load and the upper limit of the quality risk in the previous effective period. Based on the waste liquid source contribution coefficient and the sensitivity records, candidate recovery flow direction records are generated. These candidate records include a methanol recovery path contribution coefficient, a dimethyl carbonate recovery path contribution coefficient, a contribution threshold, a sensitivity direction field, a field to be calibrated, a field of allowed adjustment direction, and a version field. The candidate recovery flow direction records are output to step S6, whereby step S6 reads the candidate recovery flow direction records and applies directional constraints or freezes the methanol recovery flow adjustment and the dimethyl carbonate recovery flow adjustment. Step S2 does not determine the methanol recovery flow adjustment or the dimethyl carbonate recovery flow adjustment for the current period.

9. The electrolyte solvent emission reduction method according to claim 6, characterized in that, The conformal prediction interval in step S5 is used for hierarchical control, specifically including: normalizing the absolute values ​​of prediction errors for by-product generation rate, organic waste liquid load, dimethyl carbonate loss, quality risk, and risk components carried by heavy components according to the corresponding calibration scale, and using them as inconsistency scores; determining quantile values ​​in a calibration sample set with a target coverage rate of 90% to 99% according to finite sample order statistics; multiplying the quantile values ​​by the corresponding output calibration scale, and then adding and subtracting them from the aforementioned corresponding outputs to obtain upper and lower bounds; when the upper bound exceeds the corresponding preset control threshold, inputting the upper bound into the risk penalty term of the model's predictive control objective function; and outputting a conservative control recovery criterion when the upper bound continuously does not exceed the recovery threshold and the interval width continuously does not exceed the width threshold.

10. The electrolyte solvent emission reduction method according to claim 7, characterized in that, The distillation cut-off point adjustment in step S6 is determined based on the column section sensitivity matrix, including: using the column top temperature gradient, sensitive plate temperature gradient, column bottom temperature gradient, column section pressure gradient, reflux ratio disturbance, and candidate cut-off point adjustment to form a candidate input vector; and using the changes in dimethyl carbonate loss, heavy component entrainment risk, and organic waste liquid load as the sensitivity matrix output; when the sensitivity matrix output is a change normalized according to the corresponding upper bound calibration scale, it is first inversely normalized to an engineering quantity change according to the corresponding upper bound calibration scale; the upper bound of the heavy component entrainment risk is calculated by the heavy component entrainment risk component through a common... The upper bound of the predicted interval is obtained; for each candidate cut point adjustment amount, the candidate-level cut feasibility indicator and the emission reduction direction consistency indicator are calculated. The emission reduction direction consistency indicator is set to allow when the upper bound of the dimethyl carbonate loss amount after the candidate adjustment does not increase, the upper bound of the risk of heavy component entrainment does not increase, and the upper bound of the organic waste liquid load decreases; only the candidate cut point adjustment amount that meets the candidate-level cut feasibility indicator, the emission reduction direction consistency indicator, and the battery-level composition limit is used as the candidate set of distillation cut point adjustment amount and input into the model prediction control. The model prediction control determines the distillation cut point adjustment amount within the candidate set.

Images