A method and system for intelligent parameter adjustment of a papermaking process oriented to energy consumption optimization

Abstract

The present application relates to the field of intelligent manufacturing, and more particularly to a papermaking process intelligent parameter adjustment method and system for energy consumption optimization, comprising the following steps: S1: obtaining historical production and energy consumption report, product quality standard, production capacity plan, equipment capacity curve and safety red line, energy price and carbon emission coefficient, defining multi-objective optimization function and constraint; S2: obtaining production process data and preprocessing to obtain time series sample data; S3: performing feature construction based on the time series sample data to obtain a feature set, and using NOTEARS to construct a process-energy consumption directed acyclic graph to obtain a causal graph; S4: constructing a short-term energy consumption prediction model and a quality prediction model, and based on the feature set, the causal graph, the target KPI and the constraint, performing short-term energy consumption prediction and quality prediction; S5: according to the target function and the constraint, the short-term energy consumption prediction and the quality prediction result, performing multi-objective strategy optimization. The present application can effectively reduce the energy consumption per unit product while ensuring product quality and production safety.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing, and in particular to a method and system for intelligent parameter adjustment in papermaking processes aimed at energy consumption optimization. Background Technology

[0002] As a traditional high-energy-consuming industry, the paper industry accounts for a significant proportion of the total energy consumption of the manufacturing sector. Statistics show that my country's paper industry consumes over 50 million tons of standard coal equivalent annually, with energy consumption per unit product far exceeding international advanced levels. The paper production process is complex, involving multiple stages such as pulping, washing, bleaching, papermaking, and drying, with intricate coupling relationships between material, energy, and information flows. Traditional experience-based operation and segmented independent control models struggle to achieve coordinated optimization across the entire process chain, resulting in low energy efficiency, significant quality fluctuations, and persistently high production costs.

[0003] With the introduction of "dual carbon" targets and the continuous rise in energy prices, paper manufacturing enterprises are facing pressure to conserve energy and reduce emissions. How to minimize energy consumption while ensuring product quality and production safety has become a core technical problem that the industry urgently needs to solve. However, paper manufacturing processes involve numerous parameters (hundreds of variables such as temperature, pressure, concentration, and flow rate), complex mechanisms (involving multi-physical field coupling such as heat and mass transfer, chemical reactions, and mechanical actions), and varied disturbance factors (raw material characteristics, environmental conditions, equipment status, etc.). Traditional single-point optimization and experience-based adjustment methods can no longer meet the needs of refined management in modern production. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide a method and system for intelligent parameter adjustment in papermaking processes aimed at energy consumption optimization, which can effectively reduce energy consumption per unit product while ensuring product quality and production safety.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for intelligent parameter adjustment in papermaking processes aimed at energy consumption optimization includes the following steps:

[0007] S1: Obtain historical production and energy consumption reports, product quality standards, capacity plans, equipment capacity curves and safety limits, energy prices and carbon emission coefficients, and define multi-objective optimization functions and constraints;

[0008] S2: Acquire production process data and preprocess it to obtain time-series sample data;

[0009] S3: Based on time-series sample data, feature sets are constructed and a directed acyclic graph of process-energy consumption is constructed using NOTEARS to obtain a causal graph.

[0010] S4: Construct short-term energy consumption prediction models and quality prediction models, and perform short-term energy consumption prediction and quality prediction based on feature sets, causal graphs, target KPIs and constraints;

[0011] S5: Based on the objective function and constraints, short-term energy consumption prediction and quality prediction results, perform multi-objective strategy optimization.

[0012] Furthermore, obtain historical production and energy consumption reports, product quality standards, capacity plans, equipment capacity curves and safety limits, energy prices and carbon emission coefficients, and define multi-objective optimization functions and constraints, as follows:

[0013] Obtain basic data from paper manufacturing enterprises, including historical production reports, detailed energy consumption statistics reports, and quality standard requirements for each product; at the same time, obtain capacity plans, capacity curves in equipment technical files, equipment safety operation red lines, local energy price policies, and carbon emission coefficients in the national carbon emission trading system;

[0014] Construct a multi-objective optimization function:

[0015] ;

[0016] ;

[0017] ;

[0018] ;

[0019] ;

[0020] Where x is the decision variable vector, containing all adjustable process parameters; t is time; J energy (x) represents the unit product energy consumption cost; J quality (x) represents the product quality deviation; J stability (x) represents the system's production stability index; w1(t), w2(t), and w3(t) are the dynamic weight coefficients corresponding to each objective; w i,base S represents the baseline weight for the i-th objective; j (t) represents the j-th dynamic influencing factor; α ij (S j (t) represents the weight adjustment factor based on the impact factors; m is the total number of impact factors; i is the target category number; Q product K represents the product output per unit time; C represents the total number of energy types; k (t) represents the real-time unit price of the k-th energy type; E k (x) represents the consumption of the k-th type of energy under the decision variable vector x; P k Carbon emission price; ξ kThe carbon emission factor for energy type k; w q,j Let be the weight of the j-th quality indicator;

[0021] L j (q j (x)) is the loss function for the j-th indicator, reflecting the penalty for quality deviation; q j (x) represents the predicted value of the j-th quality indicator; r represents the number of quality grade penalty items λ. k P represents the weight of the k-th penalty term; k (x) represents the penalty level; p represents the number of selected critical process parameters; ω s,i The stability weight for the i-th process parameter; CV i (x) is the coefficient of variation of the i-th process parameter; σ i (x) is the standard deviation of the i-th parameter; μ i (x) is the mean of the i-th parameter;

[0022] The constraints include process constraints, quality constraints, and equipment constraints, and the penalty function method is used to transform the constraints into penalty terms of the objective function:

[0023] ;

[0024] Where, n h For the constraint quantity; M i g is the penalty coefficient for the i-th constraint; i (x) represents the i-th inequality constraint;

[0025] n e N represents the number of equality constraints. j h is the penalty coefficient for the j-th equation; j (x) is the j-th equality constraint.

[0026] Furthermore, the production process data, including DCS / PLC real-time data, energy consumption metering, online quality, laboratory offline testing, equipment status, maintenance records, and work orders / change logs, undergoes the following preprocessing:

[0027] Data from different systems is uniformly converted into standard JSON format, and data source identifiers, timestamps, and quality marker metadata are added. Linear interpolation is used to align data with different sampling frequencies onto a unified time grid, while handling time deviations caused by clock drift and network latency.

[0028] Different processing strategies are selected based on the missing data pattern: linear interpolation is used for short-term missing data, spline interpolation is used for medium-term missing data, and KNN algorithm based on similar operating conditions is used for estimation for long-term missing data. Finally, Butterworth low-pass filter is used to remove high-frequency noise, and the cutoff frequency is determined according to the dynamic characteristics of the process parameters.

[0029] Furthermore, feature sets are obtained by constructing features based on time-series sample data, as detailed below:

[0030] First, regarding the core process variables, including the cooking temperature T... cook Steam pressure P steam Internet speed V wire Pressing pressure P press C-concentration of pre-net pulp consistency Drying cylinder temperature T cylinder It calculates multi-scale statistical and dynamic characteristics, including sliding window mean, variance, extreme values, skewness, kurtosis, quantiles, linear trend slope and fit R2, with the window covering short-term, medium-term and long-term, to capture slow drift and fast disturbances.

[0031] Secondly, the coding process lag and response inertia are used to construct multi-order lag terms, differences, and rates of change to characterize the time dependence of setting changes, process response, and energy consumption / quality. Thirdly, physical coupling and efficiency-related interaction features are added. Subsequently, contextual and periodic features are incorporated, including order urgency, product type, raw material batch, work group, equipment health, time-of-use electricity price, and time sine and cosine coding.

[0032] Data quality and stability filtering is performed on the original feature pool: features with missing rates greater than preset values, outlier rates greater than preset values, and those that are significantly non-stationary in unit root tests and unsuitable for lag modeling are removed, and logarithmic transformation is performed on heavy-tailed distributions; then redundancy control and multicollinearity management are performed: redundancy is removed based on Pearson / Spearman correlation and hierarchical clustering, and representative features are retained for highly correlated features in the same cluster; VIF is calculated, and variables with VIF greater than preset values ​​are removed or their influence is absorbed by a regularization model, and finally, a standardized feature matrix X is output.

[0033] Furthermore, a directed acyclic graph of process and energy consumption was constructed using NOTEARS to obtain a causal graph, as detailed below:

[0034] Let X be the standardized feature matrix, and the optimization objective be least squares + L1 sparse terms, constrained by smoothness and acyclicity:

[0035] ;

[0036] Where n is the number of samples; It is the Frobenius norm (the square root of the sum of squares of the matrix elements); λ is the L1 norm, used for adding sparsity regularization; λ is the L1 regularization hyperparameter; h(W) is the acyclic constraint smoothing function; The Hadamard product is a matrix formed by squaring its elements. The index of the matrix elements; This is the matrix trace operation; p is the eigenvalue, used as the baseline value for acyclic constraints;

[0037] Using the augmented Lagrange method:

[0038] ;

[0039] in, Let be the objective function with Lagrange multipliers and a penalty term; α is the Lagrange multiplier; ρ is the augmented Lagrange penalty coefficient;

[0040] W is updated iteratively by L-BFGS, and ρ is gradually increased and the Lagrange multiplier α is updated until convergence. To enhance the credibility of the temporal causal direction, reasonable lag terms are included in the input features and control variables are introduced to address potential confusion.

[0041] A directed graph G=(V,E) and weights are constructed for subsequent critical path intensity ranking. Kernel mapping feature expansion is performed before NOTEARS. After obtaining the directed graph, a visualized causal graph is generated, and the critical coupling paths of energy consumption-quality-process are systematically identified.

[0042] Furthermore, after obtaining the calibrated DAG, a visualized causal graph is generated, and the key coupling paths of energy consumption, quality, and process are systematically identified, as follows:

[0043] First, the links are sorted according to edge weight strength, betweenness centrality, and causal importance, and the Top-N critical paths, their influence directions, and resilience coefficients are extracted. Then, for each critical path, the intervention sensitivity and local resilience are calculated, the quantitative impact of increasing / decreasing the setpoint on energy consumption and quality is given, and strategy suggestions and trade-off hints within the operationally feasible domain are output.

[0044] Furthermore, the short-term energy consumption prediction model uses a shared encoder and task head to jointly predict total energy consumption and energy consumption by energy source and by work segment, as detailed below:

[0045] The shared encoder is built on LSTM:

[0046] ;

[0047] in, The time-series input feature window used for energy consumption prediction has a length of w; h represents the hidden state of the previous time step.t The hidden representation of the LSTM output at the current time;

[0048] Task Header:

[0049] ;

[0050] ;

[0051] in, This represents the predicted energy consumption for the k-th type of work section at a lead time of h; h is the predicted lead time; W (k) b (k) These are the linear mapping weights and bias parameters for the k-th prediction head, respectively; k set description: elec is the power consumption, steam is the steam consumption, gas is the gas consumption, pull is the pulping section energy consumption, paper is the papermaking section energy consumption, and coat is the coating section energy consumption. This represents the predicted total energy consumption at time h ahead.

[0052] Equipment health correction should consider efficiency degradation and load nonlinearity.

[0053] ;

[0054] ;

[0055] in, H represents the k-th type of energy consumption prediction after being converted to the actual state of the equipment; t For equipment health; Age is the equipment's equivalent service time or cumulative runtime; Load t Current load factor; This is the device state correction function; h, a, l correspond to H respectively. t ,Aget,Load t ;α 1k ,α 2k ,α 3k The identification coefficient for the k-th type of energy consumption; a design Reference life; l rated This is the rated load.

[0056] Furthermore, the quality prediction model employs multi-task learning with shared representations for multiple quality indicators, and constructs task-specific inputs by combining causal paths, as detailed below:

[0057] Shared representation:

[0058] ;

[0059] in, For the time-series input feature window used for quality prediction; For the shared feature extractor, multivariate time series are compressed into a low-dimensional representation; zt is the shared representation vector at time t; index prediction:

[0060] ;

[0061] in, This is a quantitative prediction value leading h; The predicted moisture content is h ahead of time; The intensity prediction value is h ahead of time; , for the regression header of each task, from the shared representation z t Output the point prediction for the corresponding indicator; quality tolerance weighted loss:

[0062] ;

[0063] ;

[0064] Where j is the quality index; This is the predicted value of indicator j leading h. δ represents the true value of index j; j γ represents the tolerance band for index j; j The weighting coefficient for out-of-band penalty controls the amplification of out-of-tolerance deviations; w q,j L represents the weight of index j in the total mass loss. j L is the segmented weighted loss for index j. quality This represents the total loss for all quality tasks.

[0065] Furthermore, multi-objective strategy optimization, including workshop-level global optimization, unit-level predictive control, and equipment-level rapid control, yields hierarchical optimal setpoint trajectories and operational strategies within feasible regions, as follows: The workshop-level global optimization uses NSGA-II to generate the Pareto front and outputs the target setpoint intervals for each section; the unit-level predictive control uses MPC to consider the uncertainty of quality prediction as a chance constraint; and the equipment-level rapid control tracks the MPC setpoint through PID control.

[0066] A papermaking process intelligent parameter adjustment system for energy consumption optimization includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps of the papermaking process intelligent parameter adjustment method for energy consumption optimization as described above:

[0067] 1. This invention employs the NOTEARS algorithm to construct a process-energy consumption causal graph, transforming the traditional black-box model into an intelligent system with interpretable process mechanisms. This automatic discovery of causal relationships not only enhances the model's credibility and robustness but also provides operators with clear process adjustment logic;

[0068] 2. This invention constructs a dual-model system for short-term energy consumption and quality prediction, adopts a shared encoder and multi-task learning architecture, which ensures both prediction accuracy and computational efficiency. By integrating feature engineering, causal constraints and physical corrections, the model output is closer to the actual achievable range of the process. On this basis, a hierarchical multi-objective optimization framework is established to dynamically balance the weights of energy consumption minimization, quality stabilization and production smoothing.

[0069] 3. This invention realizes intelligent and flexible production scheduling. The weight of the objective function can be automatically adjusted according to external conditions such as energy price fluctuations, order urgency, and equipment health status. The constraints can be dynamically switched according to the process status, so that production decisions are always optimally matched with the market environment and equipment status. Attached Figure Description

[0070] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0071] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0072] refer to Figure 1 In this embodiment, a method for intelligent parameter adjustment of papermaking processes for energy consumption optimization is provided, including the following steps:

[0073] S1: Obtain historical production and energy consumption reports, product quality standards, capacity plans, equipment capacity curves and safety limits, energy prices and carbon emission coefficients, and define multi-objective optimization functions and constraints;

[0074] S2: Acquire production process data and preprocess it to obtain time-series sample data;

[0075] S3: Based on time-series sample data, feature sets are constructed and a directed acyclic graph of process-energy consumption is constructed using NOTEARS to obtain a causal graph.

[0076] S4: Construct short-term energy consumption prediction models and quality prediction models, and perform short-term energy consumption prediction and quality prediction based on feature sets, causal graphs, target KPIs and constraints;

[0077] S5: Based on the objective function and constraints, short-term energy consumption prediction and quality prediction results, perform multi-objective strategy optimization, including workshop-level global optimization, unit-level predictive control and equipment-level rapid control, to obtain the hierarchical optimal setpoint trajectory and the operation strategy within the feasible domain.

[0078] In this embodiment, historical production and energy consumption reports, product quality standards, production capacity plans, equipment capacity curves and safety limits, energy prices and carbon emission coefficients are obtained. A multi-objective optimization function and constraints are defined, as follows:

[0079] Acquire basic data from paper manufacturing enterprises, including historical production reports (covering monthly output, product mix, production efficiency, and other indicators), detailed energy consumption statistics (electricity consumption, steam consumption, natural gas consumption, circulating water consumption, etc. by work section), and quality standard requirements for each product (technical indicators such as basis weight, moisture content, thickness, strength, and whiteness, and their allowable deviation ranges); at the same time, obtain capacity plans, capacity curves from equipment technical files (such as maximum throughput of the digester, maximum speed of the paper machine, maximum steam flow rate of the drying cylinder, etc.), equipment safety operation red lines (temperature upper limit, pressure extreme value, vibration threshold, etc.), local energy price policies (electricity time-of-use rates, steam price, natural gas price), and carbon emission coefficients from the national carbon emission trading system;

[0080] A multi-objective optimization function is constructed. The first objective is to minimize the comprehensive energy consumption cost per unit product; the second objective is to minimize product quality deviation; and the third objective is to maximize the stability of the production system, which is measured by the reciprocal of the coefficient of variation of process parameters.

[0081] ;

[0082] ;

[0083] ;

[0084] ;

[0085] ;

[0086] Where x is the decision variable vector, containing all adjustable process parameters; t is time; J energy (x) represents the unit product energy consumption cost; J quality (x) represents the product quality deviation; J stability (x) represents the system's production stability index; w1(t), w2(t), and w3(t) are the dynamic weight coefficients corresponding to each objective; w i,base S represents the baseline weight for the i-th objective; j (t) represents the j-th dynamic influencing factor (such as energy price, order urgency, equipment status, etc.); αij (S j (t) represents the weight adjustment factor based on the impact factors; m is the total number of impact factors; i is the target category number; Q product K represents the product output per unit time; K represents the total number of energy types (such as electricity, steam, natural gas, etc.); C represents the product output per unit time. k (t) represents the real-time unit price of the k-th energy type; E k (x) represents the consumption of the k-th type of energy under the decision variable vector x; P k Carbon emission price; ξ k The carbon emission factor for energy type k; w q,j Let be the weight of the j-th quality indicator;

[0087] L j (q j (x)) is the loss function for the j-th indicator, reflecting the penalty for quality deviation; q j (x) represents the predicted value of the j-th quality indicator; r represents the number of quality grade penalty items λ. k P represents the weight of the k-th penalty term; k (x) represents the penalty level (e.g., second-class product, scrap loss, etc.); p represents the number of selected critical process parameters; ω s,i The stability weight for the i-th process parameter; CV i (x) is the coefficient of variation (ratio of standard deviation to mean) of the i-th process parameter; σ i (x) is the standard deviation of the i-th parameter; μ i (x) is the mean of the i-th parameter.

[0088] The constraints include process constraints, quality constraints, and equipment constraints, and the penalty function method is used to transform the constraints into penalty terms of the objective function:

[0089] ;

[0090] Where, n h For the constraint quantity; M i g is the penalty coefficient for the i-th constraint; i (x) represents the i-th inequality constraint;

[0091] n e N represents the number of equality constraints. j h is the penalty coefficient for the j-th equation; j (x) is the j-th equality constraint.

[0092] Specifically, in this embodiment, process constraints include process window limitations such as cooking temperature range of 150-170°C, pressure range of 0.6-0.8 MPa, alkali concentration of 16-20%, pulp concentration of 0.6-1.2%, and screen speed of 800-2000 m / min; quality constraints ensure that the quantitative deviation of the product is controlled within ±2%, the moisture deviation within ±0.5%, and the thickness deviation within ±3%; equipment constraints ensure that the load rate of each piece of equipment does not exceed 90% of its design capacity, the vibration value of key equipment does not exceed the safety threshold, and the temperature rise does not exceed the design limit. The initial weights of each objective are determined using the analytic hierarchy process: energy cost accounts for 40%, quality stability accounts for 35%, and production efficiency accounts for 25%. These weight coefficients will be dynamically adjusted according to actual production conditions and market demand.

[0093] In this embodiment, the production process data, including DCS / PLC real-time data, energy consumption metering (electricity / steam / gas / water), online quality control (QCS), laboratory offline testing, equipment status (vibration, temperature rise), maintenance records, and work order / change logs, are preprocessed as follows:

[0094] Data from different systems is uniformly converted into standard JSON format, and data source identifiers, timestamps, and quality marker metadata are added. Linear interpolation is used to align data with different sampling frequencies onto a unified time grid, while handling time deviations caused by clock drift and network latency.

[0095] Different processing strategies are selected based on the missing data pattern: linear interpolation is used for short-term missing data (<5 sampling points), spline interpolation is used for medium-term missing data (5-20 sampling points), and KNN algorithm based on similar operating conditions is used for estimation for long-term missing data. Finally, Butterworth low-pass filters are used to remove high-frequency noise, and the cutoff frequency is determined according to the dynamic characteristics of the process parameters: 0.1Hz for temperature parameters, 0.2Hz for flow parameters, and 0.5Hz for pressure parameters.

[0096] In this embodiment, a feature set is obtained by constructing features based on time-series sample data, as detailed below:

[0097] First, regarding the core process variables, including the cooking temperature T... cook Steam pressure P steam Internet speed V wire Pressing pressure P press C-concentration of pre-net pulp consistency Drying cylinder temperature T cylinderCalculate multi-scale statistical and dynamic characteristics, including sliding window mean, variance, extreme values, skewness, kurtosis, quantiles, linear trend slope and fit R2, with the window covering short-term (10–30 min), medium-term (1–4 h) and long-term (8–24 h) to capture slow drift and rapid disturbances.

[0098] Secondly, the coding process lag and response inertia are used to construct multi-order lag terms, differences, and rates of change to characterize the time dependence of setting changes, process response, and energy consumption / quality. Thirdly, physical coupling and efficiency-related interactive features are added, such as temperature-pressure coupling T*P, evaporation load Q*(M_in-M_out)*h_fg, power consumption / output, steam / evaporation rate, vacuum / network speed, etc., to enhance the characterization of the mechanism. Subsequently, contextual and periodic features are incorporated, including order urgency, product type, raw material batch, work group, equipment health (vibration / temperature rise comprehensive score), time-of-use electricity price, and time-of-use sine and cosine coding (hours / weeks / months) to reflect the impact of external conditions and operating strategies on energy consumption and quality.

[0099] Data quality and stability filtering is performed on the original feature pool: features with missing rates greater than preset values, outlier rates greater than preset values, and those that are significantly non-stationary in unit root tests and unsuitable for lag modeling are removed, and logarithmic transformation is performed on heavy-tailed distributions; then redundancy control and multicollinearity management are performed: redundancy is removed based on Pearson / Spearman correlation and hierarchical clustering, and representative features are retained for highly correlated features in the same cluster; VIF (variance inflation factor) is calculated, and variables with VIF greater than preset values ​​are removed or their influence is absorbed by a regularization model, and finally, a standardized feature matrix X is output.

[0100] In this embodiment, NOTEARS is used to construct a directed acyclic graph of process-energy consumption to obtain a causal graph, as detailed below:

[0101] Let X be the standardized feature matrix, and the optimization objective be least squares + L1 sparse terms, constrained by smoothness and acyclicity:

[0102] ;

[0103] Where n is the number of samples; It is the Frobenius norm (the square root of the sum of squares of the matrix elements); λ is the L1 norm, used for adding sparsity regularization; λ is the L1 regularization hyperparameter; h(W) is the acyclic constraint smoothing function; The Hadamard product is a matrix formed by squaring its elements. The index of the matrix elements; This is the matrix trace operation; p is the eigenvalue, used as the baseline value for acyclic constraints;

[0104] Using the augmented Lagrange method:

[0105] ;

[0106] in, Let be the objective function with Lagrange multipliers and a penalty term; α is the Lagrange multiplier; ρ is the augmented Lagrange penalty coefficient;

[0107] W is updated iteratively using L-BFGS, and ρ is gradually increased while the Lagrange multipliers α are updated until convergence. To enhance the credibility of the temporal causal direction, reasonable lag terms (such as...) are incorporated into the input features. And introduce control variables to address potential confusion;

[0108] A directed graph G=(V,E) and weights are constructed for subsequent critical path intensity ranking. Kernel mapping feature expansion is performed before NOTEARS. After obtaining the directed graph, a visualized causal graph is generated, and the critical coupling paths of energy consumption-quality-process are systematically identified.

[0109] In this embodiment, after obtaining the calibrated DAG, a visualized causal graph is generated, and the key coupling paths of energy consumption, quality, and process are systematically identified, as follows:

[0110] First, the links are sorted according to edge weight strength, betweenness centrality, and causal importance (SHAP on SEM), and the Top-N critical paths and their influence directions and elasticity coefficients are extracted. For example, in the papermaking drying section, a typical high-impact link is: wire speed Vwire → drying load DryLoad → steam flow rate S_steam → steam energy consumption Esteam; in the press-drying coupling section: press pressure Ppress → press dryness SRout → drying load → energy consumption; in the pulping-papering section: cooking temperature / alkali concentration → fiber characteristics (soft measurement) → press dewatering efficiency → drying energy consumption; quality coupling paths include: drying cylinder temperature / steam pressure → paper web moisture → quality deviation / rework rate. Subsequently, for each critical path, intervention sensitivity and local resilience are calculated, providing a quantitative impact of increasing / decreasing setpoints on energy consumption and quality. Strategy suggestions and trade-off hints within the operationally feasible domain are output (e.g., increasing P_press to a certain range can reduce steam energy consumption by X% while ensuring moisture deviation ≤ 0.5%, but V_wire needs to be adjusted simultaneously to maintain capacity and stability). The final results include: a causal adjacency matrix and visualization, a critical path list (impact direction, intensity, and credibility), intervention simulation results (point estimation and intervals), strategy guidance (priority, expected benefits, and risk points), and retraining / re-verification thresholds, used to drive subsequent predictive modeling, MPC feedforward, and shop floor RTO optimization.

[0111] In this embodiment, the energy consumption prediction model uses a shared encoder and task head to jointly predict total energy consumption and energy consumption by energy source and by work segment, as detailed below:

[0112] The shared encoder is built on LSTM:

[0113] ;

[0114] in, The time-series input feature window used for energy consumption prediction has a length of w; it includes process settings, process measurements, environmental variables, context (such as shift / product), etc., and only retains the causal parent set and key confounding factors; h represents the hidden state of the previous time step. t The hidden representation of the LSTM output at the current time;

[0115] Task Header:

[0116] ;

[0117] ;

[0118] in, This represents the predicted energy consumption for the k-th type of work section at a lead time of h; h is the predicted lead time; W (k) b (k) These are the linear mapping weights and bias parameters for the k-th prediction head, respectively; k set description: elec is the power consumption, steam is the steam consumption, gas is the gas consumption, pull is the pulping section energy consumption, paper is the papermaking section energy consumption, and coat is the coating section energy consumption. This represents the total energy consumption forecast (sum of energy sources / work sections) at time h ahead.

[0119] Equipment health correction should consider efficiency degradation and load nonlinearity.

[0120] ;

[0121] ;

[0122] in, H represents the k-th type of energy consumption prediction after being converted to the actual state of the equipment; t For equipment health; Age is the equipment's equivalent service time or cumulative runtime; Load t Current load factor; This is the device state correction function; h, a, l correspond to H respectively. t ,Aget,Load t ;α 1k ,α 2k ,α 3k The identification coefficient for the k-th type of energy consumption (determining the sensitivity of health, aging, and load to energy consumption); a design Reference life; l rated This is the rated load.

[0123] In this embodiment, the quality prediction model employs multi-task learning with shared representations for multiple quality indicators, and constructs task-specific inputs by combining causal paths, as detailed below:

[0124] Shared representation:

[0125] ;

[0126] in, For the time-series input feature window used for quality prediction; For the shared feature extractor, multivariate time series are compressed into a low-dimensional representation; zt is the shared representation vector at time t; index prediction (examples: quantitative, moisture content, intensity):

[0127] ;

[0128] in, This is a quantitative prediction value leading h; The predicted moisture content is h ahead of time; The intensity prediction value is h ahead of time; , for the regression header of each task, from the shared representation z t Output point predictions for the corresponding indicators; quality tolerance weighted loss (piecewise quadratic or asymmetric penalty):

[0129] ;

[0130] ;

[0131] Where j is a quality index (such as basis, moisture, strength, thickness, brightness, etc.). This is the predicted value of indicator j leading h. δ represents the true value of index j; j γ represents the tolerance band for index j; j The weighting coefficient for out-of-band penalty controls the amplification of out-of-tolerance deviations; w q,j L represents the weight of index j in the total mass loss. j L is the piecewise weighted loss of index j (different penalties within / outside the tolerance). quality This represents the total loss for all quality tasks.

[0132] In this embodiment, multi-objective strategy optimization includes workshop-level global optimization, unit-level predictive control, and equipment-level fast control, resulting in hierarchical optimal setpoint trajectories and operational strategies within feasible regions, as follows: The workshop-level global optimization (15–30 min step size) uses NSGA-II to generate the Pareto front, with adaptive weights (energy price / order priority), outputting the target setpoint range for each work section; the unit-level predictive control (1–5 min step size) uses MPC to consider the uncertainty of quality prediction as a chance constraint; and the equipment-level fast control (<100 ms) tracks the MPC setpoint through PID control.

[0133] A smart parameter adjustment system for papermaking processes aimed at energy consumption optimization is characterized by comprising a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps of the smart parameter adjustment method for papermaking processes aimed at energy consumption optimization as described above.

[0134] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0135] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0136] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0137] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0138] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent parameter adjustment in papermaking processes aimed at energy consumption optimization, characterized in that, Includes the following steps: S1: Obtain historical production and energy consumption reports, product quality standards, capacity plans, equipment capacity curves and safety limits, energy prices and carbon emission coefficients, and define multi-objective optimization functions and constraints; S2: Acquire production process data and preprocess it to obtain time-series sample data; S3: Based on time-series sample data, feature sets are constructed and a directed acyclic graph of process-energy consumption is constructed using NOTEARS to obtain a causal graph. S4: Construct short-term energy consumption prediction models and quality prediction models, and perform short-term energy consumption prediction and quality prediction based on feature sets, causal graphs, target KPIs and constraints; S5: Based on the objective function and constraints, short-term energy consumption prediction and quality prediction results, perform multi-objective strategy optimization; The feature set is obtained by constructing features based on time-series sample data, as detailed below: First, regarding the core process variables, including the cooking temperature T... cook Steam pressure P steam Internet speed V wire Pressing pressure P press C-concentration of pre-net pulp consistency Drying cylinder temperature T cylinder It calculates multi-scale statistical and dynamic characteristics, including sliding window mean, variance, extreme values, skewness, kurtosis, quantiles, linear trend slope and fit R2, with the window covering short-term, medium-term and long-term, to capture slow drift and fast disturbances. Secondly, the coding process lag and response inertia are used to construct multi-order lag terms, differences, and rates of change to characterize the time dependence of setting changes, process response, and energy consumption / quality. Thirdly, physical coupling and efficiency-related interaction features are added. Subsequently, contextual and periodic features are incorporated, including order urgency, product type, raw material batch, work group, equipment health, time-of-use electricity price, and time sine and cosine coding. Data quality and stability filtering is performed on the original feature pool: features with missing rates greater than preset values, outlier rates greater than preset values, and those that are significantly non-stationary in unit root tests and are not suitable for lag modeling are removed, and logarithmic transformation is performed on the heavy-tailed distribution. Subsequently, redundancy control and multicollinearity management are carried out: redundancy is removed based on Pearson / Spearman correlation and hierarchical clustering, and representative features are retained for highly correlated features in the same cluster. Calculate VIF, remove variables with VIF greater than the preset value or absorb their influence with a regularization model, and finally output the standardized feature matrix X; The NOTEARS process-energy consumption directed acyclic graph is used to construct a causal graph, which is then used to obtain the causal spectrum, as detailed below: Let X be the standardized feature matrix, and the optimization objective be least squares + L1 sparse terms, constrained by smoothness and acyclicity: ； Where n is the number of samples; It is the Frobenius norm; λ is the L1 norm, used for adding sparsity regularization; λ is the L1 regularization hyperparameter; h(W) is the acyclic constraint smoothing function; For Hadamard product; is the matrix element exponent; tr(·) is the matrix trace operation; p is the eigenvalue, used as the baseline value for acyclic constraints; Using the augmented Lagrange method: ； in, Let be the objective function with Lagrange multipliers and a penalty term; α is the Lagrange multiplier; ρ is the augmented Lagrange penalty coefficient; W is updated iteratively by L-BFGS, and ρ is gradually increased and the Lagrange multiplier α is updated until convergence. To enhance the credibility of the temporal causal direction, lag terms are included in the input features and control variables are introduced to address potential confusion. A directed graph G=(V,E) and weights are constructed for subsequent critical path intensity ranking. Kernel mapping feature expansion is performed before NOTEARS. After obtaining the directed graph, a visualized causal graph is generated, and the critical coupling paths of energy consumption-quality-process are systematically identified.

2. The intelligent parameter adjustment method for papermaking process oriented towards energy consumption optimization according to claim 1, characterized in that, The process involves acquiring historical production and energy consumption reports, product quality standards, production capacity plans, equipment capacity curves and safety limits, energy prices and carbon emission coefficients, and defining a multi-objective optimization function and constraints, as detailed below: Obtain basic data from paper manufacturing enterprises, including historical production reports, detailed energy consumption statistics reports, and quality standard requirements for each product; at the same time, obtain capacity plans, capacity curves in equipment technical files, equipment safety operation red lines, local energy price policies, and carbon emission coefficients in the national carbon emission trading system; Construct a multi-objective optimization function: ； ； ； ； ； Where x is the decision variable vector, containing all adjustable process parameters; t is time; J energy (x) represents the unit product energy consumption cost; J quality (x) represents the product quality deviation; J stability (x) represents the system's production stability index; These are the dynamic weighting coefficients corresponding to each objective; S represents the baseline weight for the i-th objective; j (t) represents the j-th dynamic influencing factor; α ij (S j (t) represents the weight adjustment factor based on the impact factors; m is the total number of impact factors; i is the target category number; Q product K represents the product output per unit time; K represents the total number of energy types. E represents the real-time unit price of energy type k′; k′ (x) represents the consumption of the k′th type of energy under the decision variable vector x; Carbon emission price; ξ k′ The carbon emission factor for energy type k′; w q,j L represents the weight of the j-th quality indicator; j (q j (x)) is the loss function for the j-th indicator, reflecting the penalty for quality deviation; q j (x) represents the predicted value of the j-th quality indicator; r represents the number of quality grade penalty items λ. k P represents the weight of the k-th penalty term; k (x) represents the penalty level; p represents the number of selected critical process parameters; ω s,i The stability weight for the i-th process parameter; CV i (x) is the coefficient of variation of the i-th process parameter; σ i (x) is the standard deviation of the i-th parameter; μ i (x) is the mean of the i-th parameter; The constraints include process constraints, quality constraints, and equipment constraints, and the penalty function method is used to transform the constraints into penalty terms of the objective function: ； Where, n h For the constraint quantity; M i g is the penalty coefficient for the i-th constraint; i (x) represents the i-th inequality constraint; n e N represents the number of equality constraints. j h is the penalty coefficient for the j-th equation; j (x) represents the j-th equality constraint.

3. The intelligent parameter adjustment method for papermaking process oriented towards energy consumption optimization according to claim 1, characterized in that, The production process data includes real-time DCS / PLC data, energy consumption metering, online quality data, laboratory offline testing data, equipment status, maintenance records, and work order / change logs. The preprocessing details are as follows: Data from different systems is uniformly converted into standard JSON format, and data source identifiers, timestamps, and quality marker metadata are added. Linear interpolation is used to align data with different sampling frequencies onto a unified time grid, while handling time deviations caused by clock drift and network latency. Different processing strategies are selected based on the missing data pattern: linear interpolation is used for short-term missing data, spline interpolation is used for medium-term missing data, and KNN algorithm based on similar operating conditions is used for estimation for long-term missing data. Finally, Butterworth low-pass filter is used to remove high-frequency noise, and the cutoff frequency is determined according to the dynamic characteristics of the process parameters.

4. The intelligent parameter adjustment method for papermaking process oriented towards energy consumption optimization according to claim 1, characterized in that, After obtaining the calibrated directed graph G, a visualized causal graph is generated, and the key coupling paths of energy consumption, quality, and process are systematically identified, as follows: First, the links are sorted according to edge weight strength, betweenness centrality, and causal importance, and the Top-N key paths and their influence directions and resilience coefficients are extracted; then, for each key path, the intervention sensitivity and local resilience are calculated, the quantitative impact of increasing / decreasing the set value on energy consumption and quality is given, and strategy suggestions and trade-off hints within the operationally feasible domain are output.

5. The intelligent parameter adjustment method for papermaking process oriented towards energy consumption optimization according to claim 1, characterized in that, The short-term energy consumption prediction model uses a shared encoder and task head to jointly predict total energy consumption and segmented energy consumption, as detailed below: The shared encoder is built on LSTM: ； in, The time-series input feature window used for energy consumption prediction has a length of w; h t-1 h represents the hidden state of the previous time step. t The hidden representation of the LSTM output at the current time; Task Header: ； ； in, This represents the predicted energy consumption for the k-th type of work section at a lead time of h; h is the predicted lead time; W (k) b (k) These are the linear mapping weights and bias parameters for the k-th prediction head, respectively; k set description: pullp is the energy consumption of the pulping section, paper is the energy consumption of the papermaking section, and coat is the energy consumption of the coating section; This represents the predicted total energy consumption at time h ahead. Equipment health correction should consider efficiency degradation and load nonlinearity. ； ； in, H represents the predicted energy consumption for the k-th type of work section after adjusting for actual equipment conditions; t For device health; Age t The equivalent service time or cumulative operating time of the equipment; Load t φ represents the current load factor. k (·) represents the device status correction function; h, a, l correspond to H respectively. t Age t Load t ;α 1k ,α 2k ,α 3k The identification coefficient for the k-th type of energy consumption; a design Reference life; l rated This is the rated load.

6. The intelligent parameter adjustment method for papermaking process oriented towards energy consumption optimization according to claim 1, characterized in that, The quality prediction model employs multi-task learning with shared representations for multiple quality indicators, and constructs task-specific inputs by combining causal paths, as detailed below: Shared representation: ； in, The time-series input feature window is used for quality prediction; Encoder(·) is a shared feature extractor that compresses multivariate time series into a low-dimensional representation; Let be the shared representation vector at time t; Indicator Forecast: ； in, This is a quantitative prediction value leading h; The predicted moisture content is h ahead of time; The intensity prediction value is h ahead of time; g b (·),g m (·),g s (·), representing the regression head for each task, starting from the shared representation z. t Output the point prediction for the corresponding indicator; Quality tolerance weighted loss: ； ； Where j is the quality index; This is the predicted value of indicator j leading h. δ represents the true value of index j; j γ represents the tolerance band for index j; j The weighting coefficient for out-of-band penalty controls the amplification of out-of-tolerance deviations; w q,j L represents the weight of index j in the total mass loss. j L is the segmented weighted loss of index j. quality This represents the total loss for all quality tasks.

7. The intelligent parameter adjustment method for papermaking process oriented towards energy consumption optimization according to claim 1, characterized in that, The multi-objective strategy optimization includes workshop-level global optimization, unit-level predictive control, and equipment-level fast control, resulting in hierarchical optimal setpoint trajectories and operational strategies within feasible regions, as follows: The workshop-level global optimization uses NSGA-II to generate the Pareto front and outputs the target setpoint intervals for each section; the unit-level predictive control uses MPC to consider the uncertainty of quality prediction as a chance constraint; and the equipment-level fast control tracks the MPC setpoint through PID control.

8. A smart parameter adjustment system for papermaking processes aimed at energy consumption optimization, characterized in that, It includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the intelligent parameter adjustment method for papermaking process oriented towards energy consumption optimization as described in any one of claims 1-7.

Images