Intelligent prediction control method and system for whole-process energy consumption of lithium battery separator production

Abstract

The application provides a lithium battery diaphragm production full-process energy consumption intelligent prediction control method and system, and belongs to the lithium battery production technical field. The method comprises the following steps: acquiring multi-source data in a lithium battery diaphragm production process in a prediction target period; inputting the multi-source data into a preset graph neural network model to obtain a prediction energy consumption value in the prediction target period; acquiring an actual energy consumption measurement value in the prediction target period, and calculating an energy consumption deviation value between the prediction energy consumption value and the actual energy consumption measurement value; based on a comparison between an absolute value of the energy consumption deviation value and a preset deviation threshold value, judging whether control optimization is needed; if optimization is needed, a target state variable set is obtained by screening from the multi-source data; the energy consumption deviation value, a preset energy consumption target value and the target state variable set are combined into state information; and the state information is input into a pre-trained reinforcement learning model to obtain a target control strategy. The application can improve the timeliness and accuracy of energy consumption management and control.

Description

Technical Field

[0001] This application relates to the field of lithium battery production technology, and in particular to a method and system for intelligent prediction and control of energy consumption throughout the entire lithium battery separator production process. Background Technology

[0002] In the field of lithium battery separator production, precise control of energy consumption throughout the entire process is a core challenge of green manufacturing. Currently, the production process involves strong coupling between multiple sources of data, including process parameters, equipment status, environmental conditions, and product quality. Single-factor adjustment methods are insufficient to address the complex fluctuations in energy consumption.

[0003] At the current technological level, traditional energy consumption prediction models mostly rely on single-parameter fitting or simple machine learning methods, failing to effectively explore the intrinsic relationships between process parameters, equipment status, and environmental factors. This results in low prediction accuracy and an inability to provide reliable support for real-time control. Meanwhile, existing control methods often fall into the trap of passively intervening after energy consumption exceeds limits, making it difficult to achieve dynamic optimization across the entire process. This leads to serious energy waste and low production efficiency.

[0004] Furthermore, parameter settings relying on human experience are prone to excessive energy consumption or unstable product quality, and existing control methods lack the ability to dynamically integrate multi-source data, making it difficult to achieve the goal of optimizing energy consumption throughout the entire process. With the accelerated green transformation of the manufacturing industry, enterprises are increasingly demanding higher timeliness and accuracy in energy consumption management. Therefore, there is an urgent need for an intelligent predictive control method and system for energy consumption throughout the lithium battery separator production process to meet these improved timeliness and accuracy requirements. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a method and system for intelligent prediction and control of energy consumption throughout the entire lithium battery separator production process, thereby improving the timeliness and accuracy of energy consumption management.

[0006] A first aspect of this application provides a method for intelligent prediction and control of energy consumption throughout the entire lithium battery separator production process, comprising:

[0007] Acquire multi-source data during the lithium battery separator production process for the predicted target period, including process parameters, equipment operating parameters, environmental parameters, and quality inspection parameters;

[0008] The multi-source data is input into a preset graph neural network model to obtain the predicted energy consumption value for the target time period.

[0009] After the predicted target period ends, the actual energy consumption measurement value is obtained, and the energy consumption deviation value between the predicted energy consumption value and the actual energy consumption measurement value is calculated.

[0010] Based on the comparison between the absolute value of the energy consumption deviation and the preset deviation threshold, it is determined whether control optimization is needed.

[0011] If optimization is required, the target state variable set is obtained by filtering from the multi-source data;

[0012] The energy consumption deviation value and the target state variable set are combined into state information;

[0013] The state information is input into a pre-trained reinforcement learning model to obtain a target control strategy; the target control strategy includes: process parameter adjustment amount and / or equipment operating parameter adjustment amount.

[0014] Preferably, a method for intelligent prediction and control of energy consumption throughout the entire lithium battery separator production process further includes:

[0015] During the execution of the target control strategy, the adjustment step size of the process parameter adjustment amount and / or equipment operating parameter adjustment amount is calculated based on the absolute value of the energy consumption deviation value and the historical adjustment record. The historical adjustment record includes the adjustment amount and the corresponding absolute value of energy consumption deviation in the past several control cycles.

[0016] The adjustment amount of the target control strategy is executed within the range of the adjustment step size;

[0017] If the adjustment amount required by the target control strategy is greater than the adjustment step size, then the adjustment amount of the target control strategy is executed based on the adjustment step size, and an alarm signal is generated.

[0018] In this embodiment, during the execution of the target control strategy, the adjustment step size is calculated based on the energy consumption deviation value and historical adjustment records, which can avoid the instability of the production process caused by excessive adjustment. The adjustment amount is executed within the adjustment step size range, and an alarm signal is generated if it is exceeded. This not only ensures the effective execution of the control strategy, but also enhances the stability and reliability of the system, preventing adverse effects on production due to excessive adjustment.

[0019] Preferably, a method for intelligent prediction and control of energy consumption throughout the entire lithium battery separator production process further includes:

[0020] If the absolute value of the energy consumption deviation is greater than the preset deviation threshold for N consecutive control cycles after the alarm signal is generated, the optimization process of lithium battery separator production control will be suspended.

[0021] Based on the correlation analysis between historical multi-source data and energy consumption deviation values, potential faulty equipment or abnormal process links can be located.

[0022] In this embodiment, after generating an alarm signal, if the energy consumption deviation value is still greater than the preset threshold for multiple consecutive control cycles, the optimization process is paused and potential faulty equipment or abnormal process links are located through correlation analysis. This helps to discover deep-seated problems in the production process in a timely manner, avoid blind optimization, improve the efficiency of fault diagnosis, and reduce the increase in energy consumption and production losses caused by equipment failure or process abnormalities.

[0023] Preferably, a method for intelligent prediction and control of energy consumption throughout the entire lithium battery separator production process further includes:

[0024] During the preset evaluation period after the target control strategy is implemented, the actual energy consumption value and the corresponding multi-source evaluation data are obtained;

[0025] The multi-source evaluation data is input into a preset quality prediction model to obtain the predicted quality index;

[0026] Based on the actual energy consumption value, the preset energy consumption target value for the evaluation period, the predicted quality index, the preset quality index target value, and the quality qualification threshold, a reward signal is calculated through a reward function.

[0027] Experience tuples are constructed using the initial state information of the evaluation period, the target control strategy implemented, the reward signal, and the state information at the end of the evaluation period.

[0028] The parameters of the reinforcement learning model are optimized based on the empirical tuples.

[0029] The evaluation multi-source data is input into the graph neural network model to obtain the predicted energy consumption value, and the prediction error between the energy consumption value and the actual energy consumption value is calculated.

[0030] The parameters of the graph neural network model are optimized based on the prediction error.

[0031] In this embodiment, after executing the control strategy, the parameters of the reinforcement learning model are optimized by acquiring actual energy consumption values, evaluating multi-source data, calculating reward signals, and constructing empirical tuples. Simultaneously, the parameters of the graph neural network model are optimized based on the prediction error. This allows the reinforcement learning model and the graph neural network model to continuously learn and adapt to changes in the production process. Over time, the prediction and control become increasingly accurate, further improving the accuracy of energy consumption prediction and the effectiveness of the control strategy, thus achieving continuous optimization of production energy consumption.

[0032] Preferably, a method for intelligent prediction and control of energy consumption throughout the entire lithium battery separator production process further includes:

[0033] When the predicted quality index is less than the quality qualification threshold, the multi-source evaluation data is input into the quality constraint proxy model to generate a safe operating range for process parameters.

[0034] Under the constraints of the safe operating range, the target control strategy is regenerated through a reinforcement learning model.

[0035] In this embodiment, when the predicted quality index is less than the qualified threshold, a safe operating range for process parameters is generated, and the target control strategy is regenerated within this range. This enables energy consumption optimization while ensuring product quality, avoiding product quality degradation due to excessive pursuit of energy reduction. It achieves a balance between energy consumption control and quality control, ensuring that the production process is both energy-saving and meets quality requirements.

[0036] Preferably, the step of calculating the reward signal using a reward function based on actual energy consumption measurements, preset energy consumption target values ​​for the evaluation period, predicted quality indicators, preset quality indicator target values, and quality qualification thresholds includes:

[0037] If the predicted quality index is less than the quality qualification threshold, a negative penalty signal is output as a reward signal.

[0038] If the predicted quality index is greater than or equal to the quality qualification threshold, then the energy consumption bonus component is calculated based on the deviation between the actual energy consumption value and the preset evaluation period energy consumption target value.

[0039] The quality reward component is calculated based on the deviation between the predicted quality indicator and the preset quality indicator target value.

[0040] The energy consumption reward component and the quality reward component are weighted and fused together to form the reward signal.

[0041] In this embodiment, a reward signal is calculated using a reward function based on actual energy consumption, preset energy consumption target values, and predicted quality indicators. For quality issues, a negative penalty signal is output; when quality meets the target, the reward component is calculated and integrated by comprehensively considering energy consumption and quality deviation. This approach combines energy consumption control with quality control, incentivizing the system to reduce energy consumption while ensuring product quality, thus promoting a production process that is both energy-efficient and of high quality.

[0042] Preferably, the target state variable set is obtained by filtering from the multi-source data, including:

[0043] The multi-source data is extracted to obtain a set of state variables;

[0044] The absolute value of the energy consumption deviation is compared with a preset deviation threshold. Based on the comparison result, the set of state variables is extracted to obtain a target set of state variables; wherein...

[0045] If the absolute value of the energy consumption deviation is less than or equal to the preset deviation threshold, then the set of state variables is taken as the target set of state variables.

[0046] If the absolute value of the energy consumption deviation is greater than the preset deviation threshold, then based on the feature importance assessment results, the state variables in the state variable set are extracted as the target state variable set.

[0047] In this embodiment, the target set of state variables is selectively determined based on the comparison between the energy consumption deviation value and a preset threshold. When the deviation is small, the entire set of state variables is used directly to ensure data integrity; when the deviation is large, relevant variables are extracted based on feature importance, which can reduce data redundancy, improve model calculation efficiency, and focus on variables that have a greater impact on energy consumption deviation, making the control strategy more targeted.

[0048] Preferably, the step of extracting state variables from the set of state variables as the target set of state variables based on the feature importance evaluation results includes:

[0049] Based on the pre-trained feature importance evaluation model, all state variables in the state variable set are evaluated to obtain the importance score of each state variable.

[0050] The importance scores corresponding to all state variables in the set of state variables are sorted from high to low.

[0051] The threshold for the cumulative importance percentage of the target is determined based on the absolute value of the energy consumption deviation value;

[0052] State variables are selected sequentially based on the ranking until the sum of the cumulative importance scores of the selected state variables reaches or exceeds the product of the target cumulative importance percentage threshold and the sum of the total importance scores. The selected state variables are then used as the target state variable set.

[0053] In this embodiment, a pre-trained feature importance evaluation model scores and ranks state variables, determines the cumulative importance percentage threshold of the target based on the energy consumption deviation value, and then selects state variables to form a target state variable set based on the cumulative importance percentage threshold. This allows for more accurate screening of variables that have a significant impact on energy consumption. This makes the data input to the reinforcement learning model more effective, helps generate better control strategies, and improves the accuracy and effectiveness of energy consumption control.

[0054] Preferably, determining the target cumulative importance percentage threshold based on the absolute value of the energy consumption deviation includes:

[0055] If the absolute value of the energy consumption deviation is less than the first threshold, then the target cumulative importance ratio threshold is a first predetermined ratio;

[0056] If the absolute value of the energy consumption deviation is greater than or equal to the first threshold and less than or equal to the second threshold, then the target cumulative importance ratio threshold is the second predetermined ratio.

[0057] If the absolute value of the energy consumption deviation is greater than the second threshold, then the target cumulative importance ratio threshold is a third predetermined ratio;

[0058] Wherein, the first predetermined ratio is less than the second predetermined ratio, and the second predetermined ratio is less than the third predetermined ratio.

[0059] In this embodiment, different threshold values ​​for the cumulative importance of targets are determined based on the absolute value of the energy consumption deviation. The larger the deviation, the higher the threshold, and more features with higher importance are selected. This method of adjusting the threshold can adjust the feature selection range according to the degree of energy consumption deviation. When the energy consumption deviation is small, the number of features is reduced to improve computational efficiency; when the energy consumption deviation is large, the number of features is increased to comprehensively consider influencing factors, so as to more accurately find the cause of energy consumption deviation and formulate control strategies.

[0060] A second aspect of this application provides an intelligent predictive control system for energy consumption throughout the entire lithium battery separator production process, comprising:

[0061] The data acquisition module is used to acquire multi-source data during the lithium battery separator production process for the predicted target period. The multi-source data includes process parameters, equipment operating parameters, environmental parameters, and quality inspection parameters.

[0062] The data processing module is used to input the multi-source data into a preset graph neural network model to obtain the predicted energy consumption value for the target time period.

[0063] The data calculation module is used to obtain the actual energy consumption measurement value of the predicted target time period and calculate the energy consumption deviation value between the predicted energy consumption value and the actual energy consumption measurement value.

[0064] The data judgment module is used to determine whether control optimization is needed based on the comparison between the absolute value of the energy consumption deviation value and a preset deviation threshold.

[0065] The control strategy module is used to filter from the multi-source data to obtain a target state variable set if optimization is required; combine the energy consumption deviation value, the preset energy consumption target value, and the target state variable set into state information; input the state information into a pre-trained reinforcement learning model to obtain a target control strategy; the target control strategy includes: process parameter adjustment amount and / or equipment operating parameter adjustment amount.

[0066] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above-described intelligent prediction and control method for energy consumption throughout the lithium battery separator production process.

[0067] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described intelligent prediction and control method for energy consumption throughout the lithium battery separator production process.

[0068] The beneficial effects of the intelligent predictive control method and system for energy consumption throughout the lithium battery separator production process provided in this application are as follows: Firstly, by integrating multi-source data and introducing a graph neural network model, this application can deeply explore the correlations between process, equipment, environment, and quality parameters, significantly improving the accuracy of energy consumption prediction. Secondly, by using a comparison mechanism between energy consumption deviation values ​​and preset thresholds, it can determine in real time whether optimization is needed, avoiding the lag of passive intervention and achieving dynamic control throughout the entire process. Furthermore, by generating target control strategies through reinforcement learning models, it can specifically adjust process and equipment parameters, reducing energy consumption and waste while ensuring product quality; it achieves dynamic response and real-time optimization of energy consumption deviations, solving the lag problem of control methods. In particular, combining energy consumption deviation threshold judgment with target state variable screening makes control optimization more targeted, reduces ineffective adjustments, improves equipment stability, extends the service life of key equipment, and balances the dual goals of energy saving and production quality. Attached Figure Description

[0069] Figure 1 This is a flowchart illustrating an embodiment of the intelligent prediction and control method for energy consumption throughout the lithium battery separator production process provided in this application.

[0070] Figure 2 A structural block diagram of an intelligent predictive control system for energy consumption throughout the lithium battery separator production process provided in an embodiment of this application;

[0071] Figure 3 This is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0072] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0073] To make the purpose, technical solution, and advantages of this application clearer, the following will be described in conjunction with the appendix. Figure 1-3 The following is an explanation using specific examples.

[0074] Please refer to Figure 1 , Figure 1This is a flowchart illustrating an embodiment of the intelligent predictive control method for energy consumption throughout the lithium battery separator production process provided in this application. The method includes:

[0075] S101: Obtain multi-source data during the lithium battery separator production process for the predicted target period. The multi-source data includes process parameters, equipment operating parameters, environmental parameters, and quality inspection parameters.

[0076] In this embodiment, the process parameters include raw material ratio, extrusion temperature, stretching ratio, and heat setting temperature; the equipment operating parameters include: extruder screw speed, heating tube power, fan air volume, vacuum pump vacuum degree, and motor operating current; the environmental parameters include: production workshop temperature, relative humidity, and atmospheric pressure; and the quality testing parameters include lithium battery separator thickness, porosity, tensile strength, and air permeability.

[0077] Process parameters are acquired through sensors installed on the production equipment. For example, temperature sensors on the extruder monitor the extrusion temperature, speed sensors acquire the screw speed, displacement sensors on the stretching machine calculate the stretch ratio, power sensors for the heating elements, airflow sensors for the fans, vacuum sensors for the vacuum pumps, and current sensors for the motors. Environmental parameters are collected by temperature and humidity sensors and atmospheric pressure sensors deployed in different areas of the production workshop. Quality testing parameters are obtained using specialized testing equipment, such as thickness gauges to measure diaphragm thickness, porosity meters to measure porosity, tensile testing machines to test tensile strength, and air permeability testers to test air permeability. Energy consumption data is collected through smart meters installed on the production equipment at the main production line inlet.

[0078] S102: Input multi-source data into a preset graph neural network model to obtain the predicted energy consumption value for the target time period.

[0079] In this embodiment, the construction process of the graph neural network model includes: constructing a graph structure with production processes as nodes and material transfer relationships and energy interaction relationships between processes as edges; using an adjacency matrix to represent the connection strength between nodes, and the node feature vector is composed of multi-source data of the corresponding process; wherein, the graph neural network model includes an input layer, a graph convolutional layer, a pooling layer and an output layer, and the output layer uses a linear activation function to output the predicted energy consumption value; the graph neural network model is trained using historical production data, the training process uses mean squared error as the loss function, the Adam optimizer is used for parameter updates, and L2 regularization is introduced during the training process to prevent overfitting, and iterative training stops when the loss function value is lower than 0.01.

[0080] S103: After the predicted target period ends, obtain the actual energy consumption measurement value and calculate the energy consumption deviation value between the predicted energy consumption value and the actual energy consumption measurement value.

[0081] In this embodiment, the actual energy consumption measurement value is collected by installing smart meters at the main production line and each major equipment. The actual energy consumption measurement value is the cumulative value of the energy consumption of each measurement point within the predicted target time period. The calculation of the energy consumption deviation value also needs to consider the weight of the energy consumption of different equipment. According to the proportion of each equipment in the total energy consumption of production, corresponding weights are assigned. The weights are determined by the analytic hierarchy process.

[0082] S104: Based on the comparison between the absolute value of the energy consumption deviation and the preset deviation threshold, determine whether control optimization is needed.

[0083] In this embodiment, the preset deviation threshold is set according to production process requirements and energy consumption control accuracy. The initial preset deviation threshold is set to 0%-10% of the theoretical energy consumption value for the predicted target period, and can be dynamically adjusted based on historical control effects. Specifically, when the deviation value is less than 80% of the preset deviation threshold after three consecutive control operations, the preset deviation threshold is reduced by 10%; when the deviation value is still greater than the preset deviation threshold after two consecutive control operations, the preset deviation threshold is increased by 5%. In this embodiment, different preset deviation thresholds are used in different production stages. The production stages include the start-up stage, the stable production stage, and the shutdown stage. For example, the preset deviation threshold for the start-up stage can be increased to 15% of the theoretical energy consumption value for the predicted target period, the initial threshold is used in the stable production stage, and the preset deviation threshold for the shutdown stage is reduced to 3%-5% of the theoretical energy consumption value for the predicted target period.

[0084] In this embodiment, when the absolute value of the energy consumption deviation is less than or equal to the preset deviation threshold, it indicates that the current production parameters can control energy consumption well, and the current production parameters can be maintained; when the absolute value of the energy consumption deviation is greater than the preset deviation threshold, it indicates that there is a deviation in energy consumption control, and the control optimization process needs to be started.

[0085] S105: If optimization is required, filter from multi-source data to obtain the target state variable set; combine the energy consumption deviation value and the target state variable set into state information; input the state information into a pre-trained reinforcement learning model to obtain the target control strategy; the target control strategy includes: process parameter adjustment amount and / or equipment operating parameter adjustment amount.

[0086] In this embodiment, if optimization is required, the process of filtering from multi-source data to obtain the target state variable set includes: calculating the Pearson correlation coefficient between each parameter and the energy consumption value, retaining parameters whose absolute correlation coefficient value is greater than a preset absolute threshold; performing principal component analysis on the retained parameters, extracting the original parameters corresponding to the principal components whose cumulative contribution rate is greater than a preset contribution rate threshold, thus forming the target state variable set; the target state variables in the target state variable set have a significant impact on energy consumption and are the key objects for control optimization. Furthermore, for the filtered target state variable set, redundancy checks are also required, and the mutual information value between variables is calculated. When the mutual information value is greater than a preset mutual information value, variables with a smaller impact on energy consumption are removed.

[0087] In this embodiment, the energy consumption deviation value and the target state variable set are combined into state information, which is then input into a pre-trained deep deterministic policy gradient reinforcement learning model to obtain the target control policy. In the reinforcement learning model, the agent uses minimizing energy consumption deviation and quality loss as a dual-objective reward function, and the action space includes process parameter adjustments and / or equipment operating parameter adjustments.

[0088] This embodiment of the reinforcement learning model employs a combination of offline training and online fine-tuning. In the offline phase, historical production data is used for training until convergence. In the online phase, the strategy network parameters are updated based on actual feedback after each batch of production, ensuring the real-time adaptability of the control strategy. The generated target control strategy also needs to be verified through simulation. A digital twin system simulates the energy consumption changes after parameter adjustments. If the simulation results meet the requirements of reduced energy consumption and satisfactory quality indicators, the strategy is sent to the control equipment for execution; otherwise, it is returned to the reinforcement learning model for recalculation. The data from the execution is then incorporated into the model training library as new samples, forming a closed-loop optimization mechanism of prediction-control-feedback, continuously improving the predictive control level of energy consumption in lithium battery separator production.

[0089] As can be seen from the above, this application, by integrating multi-source data and introducing a graph neural network model, can deeply explore the correlations between process, equipment, environment, and quality parameters, significantly improving the accuracy of energy consumption prediction. Secondly, by using a mechanism to compare energy consumption deviation values ​​with preset thresholds, it can determine in real time whether optimization is needed, avoiding the lag of passive intervention and achieving dynamic control throughout the entire process. Furthermore, by generating target control strategies through reinforcement learning models, it can specifically adjust process and equipment parameters, reducing energy consumption and waste while ensuring product quality; it achieves dynamic response and real-time optimization of energy consumption deviations, solving the problem of control method lag. In particular, combining energy consumption deviation threshold judgment with target state variable screening makes control optimization more targeted, reduces ineffective adjustments, improves equipment stability, extends the service life of key equipment, and balances the dual goals of energy saving and production quality.

[0090] In one embodiment of this application, the intelligent prediction and control method for energy consumption throughout the lithium battery separator production process further includes:

[0091] During the execution of the target control strategy, the adjustment step size of the process parameter adjustment amount and / or equipment operating parameter adjustment amount is calculated based on the absolute value of the energy consumption deviation and the historical adjustment record. The historical adjustment record includes the adjustment amount and the corresponding absolute value of the energy consumption deviation in the past several control cycles.

[0092] Adjustments to the target control strategy are made within the range of adjustment steps.

[0093] If the adjustment amount required by the target control strategy is greater than the adjustment step size, then the adjustment amount of the target control strategy is executed based on the adjustment step size, and an alarm signal is generated.

[0094] In this embodiment, during the execution of the target control strategy, the adjustment step size of the process parameter adjustment amount and / or equipment operating parameter adjustment amount is calculated based on the absolute value of the energy consumption deviation and historical adjustment records. The historical adjustment records include the adjustment amounts and corresponding absolute values ​​of energy consumption deviation within a certain number of past control cycles. The formula for calculating the adjustment step size is: Adjustment Step Size = Basic Step Size × (1 + Absolute Value of Energy Consumption Deviation / Historical Average Absolute Value of Energy Consumption Deviation) × (1 - Historical Adjustment Success Rate). The basic step size is set according to the parameter type; for example, the basic step size for process parameters is 1%-3% of its baseline value, and the basic step size for equipment operating parameters is 2%-5% of its baseline value. The historical adjustment success rate is the ratio of the number of times the absolute value of the energy consumption deviation decreased after adjustment within a certain number of past control cycles to the total number of adjustments. The number of past control cycles can be 5-10. The adjustment amount of the target control strategy is executed within the range of the adjustment step size; that is, when the absolute value of the adjustment amount required by the target control strategy is less than or equal to the adjustment step size, the adjustment amount required by the target control strategy is executed. If the adjustment amount required by the target control strategy is greater than the adjustment step size, the adjustment amount of the target control strategy is executed based on the adjustment step size, and an alarm signal is generated. The alarm signal includes information such as the overshoot parameter name, the target adjustment amount, the actual adjustment step size, and the current energy consumption deviation value, and is issued through the alarm module of the industrial control system in the form of audible and visual alarms and system pop-ups.

[0095] In another embodiment, the calculation of the adjustment step size employs a dynamic adaptive algorithm to achieve accuracy and stability in parameter adjustment. Specifically, features are first extracted from historical adjustment records, which include adjustment amounts and corresponding absolute values ​​of energy consumption deviations over multiple past control cycles. A time decay factor is used to assign higher weight to recent data, highlighting the impact of recent data on the current adjustment step size calculation.

[0096] The adjustment step size is calculated using a deviation-step size correlation model. When the absolute value of the energy consumption deviation is less than or equal to the first adjustment deviation threshold, the adjustment step size is taken as the first percentage of the average adjustment amount of the historical deviation interval. For example, if the energy consumption deviation is 1.5% in historical data and the first adjustment deviation threshold is 3%, the average temperature adjustment amount is ±1.2℃, then the current step size is set to ±0.72℃. When the absolute value of the energy consumption deviation is greater than the first adjustment deviation threshold and less than or equal to the second adjustment deviation threshold, the adjustment step size is taken as the second percentage of the average adjustment amount of the historical deviation interval. When the absolute value of the energy consumption deviation is greater than the second adjustment deviation threshold, the adjustment step size is taken as the third percentage of the average adjustment amount of the historical deviation interval. The third percentage is greater than the second percentage, the second percentage is greater than the first percentage, and the second adjustment deviation threshold is greater than the first adjustment deviation threshold. This embodiment also incorporates process constraints to ensure that the adjustment step size is within a reasonable and safe range.

[0097] The adjustment amount of the target control strategy is executed within the adjustment step size. The execution process adopts a step-by-step mechanism. The first step is to execute 50% of the adjustment step size, and energy consumption data is collected at 15-minute intervals. If the energy consumption deviation shows a downward trend, the remaining 50% is executed. If the deviation does not decrease or quality indicators fluctuate, execution is paused and a second evaluation is triggered to avoid adverse effects on production quality due to excessive adjustment. Adjustment commands are transmitted in real time via the industrial bus during execution.

[0098] If the adjustment amount required by the target control strategy is greater than the adjustment step size, the adjustment amount of the target control strategy is executed based on the adjustment step size, and an alarm signal is generated. The alarm signal adopts a three-level warning mechanism. When the adjustment amount is greater than the adjustment step size and the difference is within the first difference range, a yellow warning is triggered, and a text prompt is displayed on the central control system accompanied by a short buzzer sound. When the adjustment amount is greater than the adjustment step size and the difference is within the second difference range, an orange warning is triggered, and in addition to the above prompt, the warning information is automatically pushed to the production supervisor's mobile terminal. When the adjustment amount is greater than the adjustment step size and the difference is within the third difference range, a red warning is triggered, the system suspends automatic adjustment and forcibly switches to manual control mode, and simultaneously activates the audible and visual alarm device. Among these, the value in the third difference range is greater than the value in the second difference range, and the value in the second difference range is greater than the value in the first difference range.

[0099] The alarm signals in this embodiment include: the name of the parameter exceeding the limit, the current adjustment step size, the required adjustment amount of the strategy, and the percentage of deviation. All alarm records are stored in the event log library as feedback data for subsequent model optimization. By analyzing high-frequency alarm parameters, the action space constraints of the reinforcement learning model can be optimized in a targeted manner to reduce the occurrence of exceeding the limit.

[0100] In one embodiment of this application, the intelligent prediction and control method for energy consumption throughout the lithium battery separator production process further includes:

[0101] If the absolute value of the energy consumption deviation is greater than the preset deviation threshold for N consecutive control cycles after an alarm signal is generated, the optimization process of lithium battery separator production control will be suspended.

[0102] Based on the correlation analysis between historical multi-source data and energy consumption deviation values, potential faulty equipment or abnormal process links can be located.

[0103] In this embodiment, after an alarm signal is generated, if the absolute value of the energy consumption deviation exceeds a preset deviation threshold for N consecutive control cycles, the optimization process of lithium battery separator production control is paused. The value of N is dynamically set based on the stability requirements of the production process, typically defaulting to 3 control cycles for faster response to potential problems. The pause command is sent to the control module via the system kernel-level interface.

[0104] In this embodiment, based on the correlation analysis between historical multi-source data and energy consumption deviation values, potential faulty equipment or abnormal process steps are located. The correlation analysis employs a dual verification mechanism. First, frequent itemsets are mined using association rule algorithms. For example, when the co-occurrence frequency of "screw motor current fluctuation > 5%" and "energy consumption deviation > 3%" is greater than 80%, the motor is marked as high-risk equipment. A causal inference model is constructed based on production time-series data. Causal tests are used to analyze the time lag relationship between sudden changes in process parameters and energy consumption deviation. For example, if an energy consumption deviation exceeds the standard within 3 minutes after a sudden 10°C increase in extrusion temperature, it can be determined that the temperature control system is abnormal.

[0105] This embodiment for equipment fault location includes constructing a fault probability assessment matrix based on historical maintenance records of the production equipment (real-time vibration spectrum and abrupt changes in energy consumption curves). For identifying abnormal process steps, this embodiment employs a process energy consumption fingerprint comparison method. Standard energy consumption models for each process are pre-established. When the deviation between the actual energy consumption of a process and the standard model consistently exceeds a preset value, and equipment factors are excluded, the abnormal parameters are located by comparing the process parameter combinations of historically qualified batches with the current parameters. The location results of this embodiment are presented in a visual report format, including 3D location markings of the faulty equipment, trend comparison charts of abnormal process parameters, and correlation confidence scores.

[0106] In one embodiment of this application, based on the correlation analysis between historical multi-source data and energy consumption deviation values, potential faulty equipment or abnormal process steps are located, including:

[0107] Based on multi-source data covering different production batches, raw material status and equipment operation cycles in the historical database, a multivariate statistical benchmark model is constructed to characterize normal production conditions. The benchmark model is used to define the normal joint distribution characteristics of process parameters, equipment operation parameters and environmental parameters.

[0108] Feature extraction and preprocessing are performed on multi-source data for the target time period, and time-series features are fused to generate a feature vector for the target time period; the preprocessing includes data cleaning, missing value handling, and normalization / standardization.

[0109] A multivariate statistical benchmark model is applied to calculate the multivariate statistical control index of the feature vector of the target time period relative to the normal joint distribution.

[0110] The root cause analysis model is trained using historical data. The input of the root cause analysis model includes multivariate statistical control indicators and their decomposition terms, as well as historical fault labels or significant energy consumption deviation labels. The output is the abnormal contribution of each equipment or process link.

[0111] The multivariate statistical control indicators and their decomposition terms calculated for the target forecast period are input into the root cause analysis model to obtain the abnormal contribution score of each equipment or process step; equipment or process steps with an abnormal contribution score greater than the preset contribution threshold are marked as potential abnormal sources.

[0112] When a piece of equipment or a process step is marked as a potential source of anomaly, a fault diagnosis report is generated, which includes the identification of the equipment / process step, the deviation of relevant parameters, and the score of the contribution of the anomaly.

[0113] In this embodiment, the historical database includes multi-source data on different production batches, raw material states, and equipment operating cycles. The multivariate statistical benchmark model adopts a fusion architecture combining principal component analysis and independent component analysis. First, principal component analysis reduces the dimensionality of the high-dimensional data, preserving the main variation information. Then, independent component analysis separates independent latent variables from the principal components, capturing non-Gaussian distribution characteristics. The normal joint distribution of the benchmark model is constructed through kernel density estimation, and confidence intervals are set for the value range of each latent variable, forming a three-dimensional control limit surface (horizontal axis for production batch, vertical axis for equipment operating cycle, and Z-axis for raw material state).

[0114] In this embodiment, feature extraction employs a combination of time and frequency domain methods. Time-domain features include parameter mean, variance, and peak factor, while frequency-domain features extract the dominant frequency component of the equipment vibration signal through Fourier transform. In the preprocessing stage, data cleaning uses the Isolation Forest algorithm to remove jump values ​​caused by sensor drift; missing value handling employs linear interpolation for short-term missing values ​​and uses historical data patterns from the same operating conditions to fill in long-term missing values; normalization / standardization uses the Z-score method to eliminate dimensional differences between different parameters. Time-series feature fusion extracts the parameter change rate and trend persistence through a sliding window, ultimately forming a feature vector for the target time period.

[0115] In this embodiment, the multivariate statistical control indicators include measures of overall deviation and measures of residual deviation. The overall deviation is measured by calculating the weighted distance between the eigenvectors and the principal components of the baseline model; the residual deviation is measured by calculating the orthogonal distance between the eigenvectors and the principal component space.

[0116] In this embodiment, a root cause analysis model is trained using historical data. The input to the root cause analysis model includes multivariate statistical control indicators and their decomposition terms, as well as historical fault labels or significant energy consumption deviation labels. The output is the abnormal contribution of each device or process step. The root cause analysis model adopts a gradient boosting tree architecture. The input layer includes multivariate statistical control indicators and their decomposition terms, as well as historical fault labels or significant energy consumption deviation labels. The training samples include normal records, equipment fault records, and process anomaly records, with each sample labeled with a corresponding fault label and energy consumption deviation label. The root cause analysis model outputs the abnormal contribution of each device or process step by learning the nonlinear mapping relationship between indicators and labels.

[0117] In this embodiment, the preset contribution threshold is dynamically set according to the importance of the equipment. For example, the threshold for critical equipment is set to 40 points, and the threshold for auxiliary equipment is set to 60 points; the threshold for core process steps is set to 35 points, and the threshold for general steps is set to 50 points. When equipment or process steps are marked as potential anomalies, a fault diagnosis report is generated, including the equipment / process step identifier, relevant parameter deviations, and anomaly contribution scores. The report adopts a three-level visualization structure: the first-level page displays a heatmap of anomaly source distribution; the second-level page displays parameter deviation details; and the third-level page displays the contribution traceability chain.

[0118] In one embodiment of this application, the intelligent prediction and control method for energy consumption throughout the lithium battery separator production process further includes:

[0119] During the preset evaluation period after the target control strategy is implemented, the actual energy consumption value and the corresponding multi-source evaluation data are obtained;

[0120] The multi-source data is input into a pre-set quality prediction model to obtain the predicted quality index;

[0121] Based on the actual energy consumption value, the preset energy consumption target value for the assessment period, the predicted quality indicators, the preset quality indicator target value and the quality qualification threshold, the reward signal is calculated through the reward function.

[0122] Experience tuples are constructed using the initial state information of the evaluation period, the target control strategy implemented, the reward signal, and the state information at the end of the evaluation period.

[0123] Parameters of reinforcement learning models are optimized based on empirical tuples;

[0124] The predicted energy consumption value is obtained by inputting multi-source data into the graph neural network model, and the prediction error between the energy consumption value and the actual energy consumption value is calculated.

[0125] Optimize the parameters of the graph neural network model based on prediction error.

[0126] In this embodiment, the preset evaluation period is dynamically set according to the response characteristics of the production process. For example, the evaluation period for the extrusion process is 2 hours, and for the stretching process it is 1.5 hours. The total evaluation time is determined by the principle of the longest process duration plus a 30-minute buffer period to ensure that the impact of parameter adjustments on energy consumption and quality can be fully obtained. The collection range of multi-source evaluation data is consistent with the initial multi-source data, focusing on recording the parameter change trajectory after the control strategy is implemented.

[0127] In this embodiment, the quality prediction model employs a bidirectional LSTM network based on an attention mechanism. The attention layer enhances feature extraction of quality-sensitive parameters, and the output layer uses a sigmoid activation function to output predicted values ​​for five core quality indicators: diaphragm thickness deviation, porosity, tensile strength, thermal shrinkage rate, and puncture strength. This quality prediction model is pre-trained to convergence using production data from the past six months, and its parameters are automatically updated after each batch of production.

[0128] In this embodiment, the reward function is:

[0129]

[0130]

[0131] in, , which is the total reward signal, used to optimize the target of the reinforcement learning model; Energy consumption weighting; The actual energy consumption value represents the measured energy consumption during the evaluation period. The target energy consumption value represents the expected energy consumption level for process optimization. For quality weights; and It is a pre-set positive parameter used to control the decay rate; This is a key characteristic of the quality model output for predicting quality indicators; The ideal quality level set for the process to represent the quality target value; K is the quality acceptance threshold, the minimum acceptable quality standard for the product; K is the quality penalty coefficient, the fixed penalty for non-compliance with quality standards. This is used to ensure a balance among multiple objectives.

[0132] In this embodiment, the experience tuple adopts a four-tuple structure of (St, At, Rt, S{t+1}), where St is the state vector at the beginning of the evaluation period, At is the control policy parameter to be executed, Rt is the calculated reward signal, and S{t+1} is the state vector at the end of the evaluation period. One experience tuple is generated for each evaluation period and stored in the experience replay pool in chronological order. The experience replay pool adopts a circular buffer structure, and when the capacity is full, it automatically overwrites the oldest tuple to ensure the timeliness of the data in the pool.

[0133] In this embodiment, the parameters of the reinforcement learning model are optimized based on empirical tuples. The optimization process employs a priority experience replay mechanism, assigning different sampling weights to empirical tuples according to the absolute value of the reward signal. The larger the absolute value of the reward signal, the higher the weight and the higher the sampling probability. Each training batch consists of 256 tuples randomly selected from the pool. The Actor network and Critic network of the reinforcement learning model are updated synchronously. The Actor network minimizes the loss function using policy gradient descent, while the Critic network is optimized using the mean squared error loss function.

[0134] In this embodiment, the predicted energy consumption value is obtained by evaluating the multi-source data input graph neural network model, and the prediction error between the energy consumption value and the actual energy consumption value is calculated; wherein, the prediction error is calculated using the mean absolute percentage error.

[0135] In one embodiment of this application, the intelligent prediction and control method for energy consumption throughout the lithium battery separator production process further includes:

[0136] When the predicted quality index is less than the quality qualification threshold, the multi-source data will be evaluated and input into the quality constraint proxy model to generate a safe operating range for process parameters.

[0137] Under the constraint of a safe operating range, the target control strategy is regenerated through a reinforcement learning model.

[0138] In this embodiment, the quality constraint surrogate model is a surrogate model built based on Gaussian process regression. Its training data comes from the correspondence between process parameters and quality indicators of all qualified batches in historical production, including process parameters such as extrusion temperature, stretching ratio, and annealing time, as well as the corresponding quality indicator compliance data. The model uses a squared exponential kernel function to capture the nonlinear correlation between parameters and optimizes the kernel function hyperparameters through maximum likelihood estimation.

[0139] In this embodiment, after evaluating the multi-source data input quality constraint surrogate model, the model first extracts the baseline values ​​of the current process parameters. Then, it generates multiple sets of parameter perturbation schemes through Monte Carlo simulation, calculates the predicted quality index values ​​corresponding to each scheme, and selects parameter combinations in which all quality indices meet the qualified threshold to form the initial feasible region. The initial feasible region is then refined using a boundary shrinkage algorithm to eliminate unstable edge regions, ultimately generating a safe operating range for process parameters centered on the baseline values.

[0140] In this embodiment, under the constraint of a safe operating range, the target control strategy is regenerated through a reinforcement learning model. The action space of the reinforcement learning model is restricted to the safe operating range. The reward function of the reinforcement learning model adds a quality constraint penalty term to the original energy consumption optimization objective: when the parameter adjustment scheme touches the boundary of the safe range, the reward value is reduced by 10%; if it exceeds the boundary (due to model calculation error), the reward value is reduced by 50%, thereby guiding the model to find the optimal solution within the safe range.

[0141] When regenerating the strategy, a phased search mechanism is adopted: the first phase conducts a global exploration within the safe operating range, prioritizing the energy consumption optimization effects of different parameter combinations; the second phase focuses on the parameter direction with the largest energy consumption reduction gradient for local optimization. The final generated target control strategy must ensure that all parameter adjustments are within the safe operating range and be verified by the quality prediction model. If a strategy that meets the conditions cannot be generated after multiple iterations, a manual intervention process is triggered, pushing a decision support report including the current quality risk level and the recommended range of parameter adjustments to the process engineer's terminal.

[0142] In one embodiment of this application, a reward signal is calculated using a reward function based on actual energy consumption measurements, a preset energy consumption target value for an evaluation period, a predicted quality index, a preset quality index target value, and a quality qualification threshold. This includes:

[0143] If the predicted quality index is less than the quality qualification threshold, a negative penalty signal is output as a reward signal.

[0144] If the predicted quality index is greater than or equal to the quality qualification threshold, the energy consumption bonus component is calculated based on the deviation between the actual energy consumption value and the preset energy consumption target value for the assessment period.

[0145] The quality reward component is calculated based on the deviation between the predicted quality indicator and the preset quality indicator target value.

[0146] The energy consumption reward component and the quality reward component are weighted and fused together to form the reward signal.

[0147] In this embodiment, if the predicted quality index is less than the quality pass threshold, a negative penalty signal is output as a reward signal. The value of the negative penalty signal is positively correlated with the degree to which the quality index deviates from the pass threshold, and a piecewise linear penalty mechanism is adopted: when the predicted quality index is in the first interval, the penalty value is set as the product of the first percentage coefficient and the preset penalty value; when it is in the second interval, the penalty value is set as the product of the second percentage coefficient and the preset penalty value; when it is in the third interval, the penalty value is set as the preset penalty value; wherein, the value in the third interval is greater than the value in the second interval, the value in the second interval is greater than the value in the first interval, and the second percentage coefficient is greater than the first percentage coefficient.

[0148] In this embodiment, if the predicted quality index is greater than or equal to the quality qualification threshold, an energy consumption bonus component is calculated based on the deviation between the actual energy consumption value and the preset energy consumption target value for the evaluation period. The formula for the energy consumption bonus component is as follows:

[0149]

[0150] in, Energy consumption bonus portion; It is an energy consumption sensitivity coefficient, used to control the rate at which energy consumption rewards decay.

[0151] In this embodiment, a quality reward component is calculated based on the deviation between the predicted quality indicator and the preset quality indicator target value. The formula for the quality reward component is as follows:

[0152]

[0153] in, As a quality bonus; This is the Sigmoid steepness coefficient, which adjusts the gradient of quality reward changes.

[0154] The energy consumption reward component and the quality reward component are weighted and fused together as the reward signal. The weighting and fusion adopts a dynamic weighting mechanism; for example, when production is in "energy consumption priority mode", the weight of the energy consumption reward component is set to 0.6 and the weight of the quality reward component is set to 0.4; when it is in "quality priority mode", the weights are adjusted to energy consumption 0.3 and quality 0.7; in the default mode, the weights are 0.5 for each.

[0155] The fusion formula in this embodiment is: Reward Signal = Energy Consumption Reward Component × Energy Consumption Weight + Quality Reward Component × Quality Weight. After the reward signal is calculated, a detailed score report is generated simultaneously, indicating the scores of each indicator and the basis for weight allocation, providing interpretable feedback for policy optimization of the reinforcement learning model. In one embodiment of this application, a target state variable set is obtained by filtering from multi-source data, including:

[0156] Extract from multi-source data to obtain a set of state variables;

[0157] The absolute value of the energy consumption deviation is compared with a preset deviation threshold. Based on the comparison result, the set of state variables is extracted to obtain the target set of state variables; among which,

[0158] If the absolute value of the energy consumption deviation is less than or equal to the preset deviation threshold, then the set of state variables will be used as the target set of state variables.

[0159] If the absolute value of the energy consumption deviation is greater than the preset deviation threshold, then based on the feature importance assessment results, the state variables in the state variable set are extracted as the target state variable set.

[0160] In this embodiment, multi-source data extraction employs a hierarchical feature selection method. First, the raw data is divided into multiple modules according to production processes. Each module extracts fundamental variables closely related to energy consumption, ultimately summarizing them to form an initial state variable set encompassing three main categories: equipment operation, process parameters, and environmental conditions. During the extraction process, each variable is uniquely identified, and its collection frequency, data type, and unit information are recorded.

[0161] This embodiment extracts the set of state variables based on the comparison between the absolute value of the energy consumption deviation and a preset deviation threshold, thus obtaining the target set of state variables. The preset deviation threshold is set according to the production process requirements to adapt to the energy consumption control accuracy requirements of different processes.

[0162] If the absolute value of the energy consumption deviation is less than or equal to a preset deviation threshold, then the set of state variables is used as the target set of state variables. This indicates that the current production energy consumption is within a controllable range, and there is no need to reduce the dimensionality of variables; all state variables are retained to fully reflect the production status. These variables will be directly used for subsequent state information combination, providing complete input data for the reinforcement learning model and ensuring that the model can make decisions based on comprehensive state information.

[0163] If the absolute value of the energy consumption deviation is greater than a preset deviation threshold, then based on the feature importance assessment results, state variables from the state variable set are extracted as the target state variable set. The feature importance assessment employs a combination of a fusion tree model and mutual information. First, a random forest model is used to calculate the contribution of each state variable to the energy consumption deviation. Then, mutual information is used to calculate the statistical correlation between the variable and the energy consumption deviation value. The two indicators are weighted according to preset weights to obtain a comprehensive importance score. Variables are screened based on the comprehensive importance score, a scoring threshold is set, and state variables with scores higher than the threshold are extracted to form the target state variable set. The dimensions of the screened target state variable set focus on key variables that significantly affect energy consumption deviation, such as stretching ratio, annealing temperature, and fan power. Furthermore, to avoid overlooking potentially important variables, this embodiment manually reviews variables with scores close to the threshold, combining process knowledge to determine whether to include them in the target state variable set. For example, if a certain environmental humidity variable has a score close to the scoring threshold, but process experience indicates that it has a significant impact on energy consumption during high humidity seasons, it is retained.

[0164] Furthermore, the target state variable set employs a dynamic update mechanism, recalculating the feature importance score every 10 control cycles (based on a manually set number of control cycles) and adjusting the variable selection results according to changes in production conditions. This ensures that the target state variable set accurately reflects the key factors affecting energy consumption deviations. After selection, a list of target state variables is generated, marking the importance score and selection criteria for each variable, providing a clear variable basis for subsequent state information combination and control strategy generation.

[0165] In one embodiment of this application, based on the feature importance evaluation results, state variables in the state variable set are extracted as the target state variable set, including:

[0166] Based on a pre-trained feature importance evaluation model, all state variables in the state variable set are evaluated to obtain an importance score for each state variable.

[0167] Sort all state variables in the set of state variables from highest to lowest importance score;

[0168] The threshold for the cumulative importance percentage of targets is determined based on the absolute value of energy consumption deviation.

[0169] State variables are selected sequentially based on the ranking until the sum of the cumulative importance scores of the selected state variables reaches or exceeds the product of the target cumulative importance percentage threshold and the sum of the total importance scores. The selected state variables are then used as the target state variable set.

[0170] In this embodiment, the pre-trained feature importance evaluation model employs an ensemble learning framework, fusing the evaluation results of three tree models: random forest, gradient boosting tree, and extreme gradient boosting. The training data for these feature importance evaluation models is historical production data, which includes multiple normal operating condition samples and multiple samples with excessive energy consumption deviations. The input to each model is a state variable, and the output is an importance score for that state variable. During the training process of each model, the hyperparameters are optimized using 5-fold cross-validation. In this embodiment, when evaluating state variables, the pre-trained feature importance evaluation model first outputs three importance scores for each state variable, and then calculates the final importance score by weighting the three scores based on preset weighting coefficients. To fully leverage the advantages of each model, this embodiment determines the weighting coefficients based on historical experimental data; for example, a weighted fusion of the three importance scores is performed using a 4:3:3 weighting coefficient to obtain the final importance score.

[0171] This embodiment sorts all state variables in the state variable set from highest to lowest importance score. The sorting process employs a stable sorting algorithm; when two variables have the same score, the variable with higher mutual information to the energy consumption deviation value is prioritized. Furthermore, this embodiment generates a variable priority list after sorting, labeling each variable's score and ranking, providing a clear order for subsequent variable selection.

[0172] In this embodiment, the target cumulative importance percentage threshold is determined based on the absolute value of the energy consumption deviation. The target cumulative importance percentage threshold is positively correlated with the absolute value of the energy consumption deviation, employing a segmented increasing mechanism: when the absolute value of the energy consumption deviation is in the first exceeding range (mild exceeding the standard), the threshold is set to the first threshold, meaning variables with a cumulative importance percentage reaching a first predetermined proportion are selected; when the absolute value of the energy consumption deviation is in the second exceeding range (moderate exceeding the standard), the threshold is set to the second threshold, meaning variables with a cumulative importance percentage reaching a second predetermined proportion are selected; when the deviation value is greater than the value in the second exceeding range (severe exceeding the standard), the threshold is set to the second threshold, meaning variables with a cumulative importance percentage reaching a third predetermined proportion are selected. This mechanism in this embodiment achieves precise control with fewer key variables when the energy consumption deviation is small; when the deviation is large, more relevant variables are included to comprehensively investigate the problem. For example, when the absolute value of the energy consumption deviation is moderately exceeding the standard, the target cumulative importance percentage threshold is determined to be 80%.

[0173] This embodiment selects state variables sequentially based on ranking until the sum of the cumulative importance scores of the selected state variables reaches or exceeds the product of the target cumulative importance percentage threshold and the total importance score. The selected state variables are then used as the target state variable set. The total importance score is the sum of the importance scores of all variables in the state variable set. For example, the total score of 85 variables is 5200 points. If the target cumulative importance percentage threshold is 80%, then the target cumulative score is 5200 × 80% = 4160 points. Variables are selected from highest to lowest ranking, and the cumulative score is calculated: the cumulative score is 2800 points for the first 10 variables, 3600 points for the first 15 variables, and 4200 points for the first 20 variables. At this point, 4200 points ≥ 4160 points, so selection stops. Therefore, the first 20 variables constitute the target state variable set.

[0174] In this embodiment, during the selection process, if the importance scores of multiple consecutive variables (greater than or equal to 3) suddenly drop, even if the cumulative score does not reach the target value, the selection will be paused and manually verified to determine whether it is due to model evaluation bias. After the selection is completed, this embodiment outputs a list of variables in the target state variable set, the cumulative importance percentage, and the score percentage of each variable, providing accurate variable basis for subsequent state information combination.

[0175] In one embodiment of this application, determining the target cumulative importance percentage threshold based on the absolute value of the energy consumption deviation includes:

[0176] If the absolute value of the energy consumption deviation is less than the first threshold, then the target cumulative importance percentage threshold is the first predetermined percentage;

[0177] If the absolute value of the energy consumption deviation is greater than or equal to the first threshold and less than or equal to the second threshold, then the target cumulative importance percentage threshold is the second predetermined percentage.

[0178] If the absolute value of the energy consumption deviation is greater than the second threshold, then the target cumulative importance percentage threshold is the third predetermined percentage.

[0179] Among them, the first pre-booking ratio is less than the second pre-booking ratio, and the second pre-booking ratio is less than the third pre-booking ratio.

[0180] In this embodiment, if the absolute value of the energy consumption deviation is less than a first threshold, the target cumulative importance percentage threshold is a first predetermined proportion. The first threshold is an energy consumption control critical value derived from historical production data statistics; when the absolute value of the deviation is less than the first threshold, the production system is typically in a state of slight fluctuation, and the core process parameters have not deviated significantly. The first predetermined proportion is determined by analyzing the variable importance distribution of multiple historical normal production batches.

[0181] If the absolute value of the energy consumption deviation is greater than or equal to the first threshold and less than or equal to the second threshold, then the cumulative importance percentage threshold for the target is a second predetermined proportion. The second threshold is derived from historical production data statistics and corresponds to a critical state of moderate anomaly in the production system. When the absolute value of the deviation energy consumption is greater than or equal to the first threshold and less than or equal to the second threshold, coordinated fluctuations in multiple process stages will occur. The second predetermined proportion is based on the review of moderate deviation cases, which avoids computational redundancy caused by too many variables while ensuring that no key influencing factors are overlooked.

[0182] If the absolute value of the energy consumption deviation exceeds the second threshold, the cumulative importance percentage threshold for the target is set as the third predetermined proportion. Deviations exceeding the second threshold correspond to severe anomalies, potentially involving equipment malfunctions or loss of control over critical process parameters. While such cases account for a relatively small percentage in historical data, energy consumption losses represent a significant portion of total anomaly losses. The third predetermined proportion is determined based on the analysis of multiple severe deviation cases, enabling comprehensive capture of multi-stage interconnected anomalies.

[0183] The system employs a tiered threshold system where the first predetermined proportion is less than the second, and the second is less than the third, forming a step-by-step threshold system positively correlated with the degree of deviation. This system dynamically adjusts the variable selection range, using a simplified variable set to improve control response speed when energy consumption deviation is small, and a comprehensive variable set to ensure anomaly diagnosis accuracy when deviation is large, achieving a balance between efficiency and accuracy. The threshold and proportion settings are validated through Monte Carlo simulations. In multiple random scenario tests, the system's variable selection is accurate, effectively supporting the policy generation efficiency of subsequent reinforcement learning models. Furthermore, the first threshold, second threshold, and each predetermined proportion are periodically calibrated based on the latest production data. If the anomaly handling accuracy decreases more than the decrease threshold in a certain deviation range, the thresholds are optimized by updating the threshold values ​​through linear regression analysis, ensuring the system always adapts to the dynamic changes of the production system.

[0184] Corresponding to the intelligent prediction and control method for energy consumption throughout the lithium battery separator production process described in the above embodiment, Figure 2 This is a structural block diagram of an intelligent predictive control system for energy consumption throughout the lithium battery separator production process, provided in one embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. References Figure 2 The intelligent predictive control system 20 for energy consumption throughout the lithium battery separator production process includes: a data acquisition module 21, a data processing module 22, a data calculation module 23, a data judgment module 24, and a control strategy module 25.

[0185] Among them, the data acquisition module 21 is used to acquire multi-source data in the lithium battery separator production process during the predicted target period. The multi-source data includes process parameters, equipment operating parameters, environmental parameters and quality inspection parameters.

[0186] Data processing module 22 is used to input multi-source data into a preset graph neural network model to obtain the predicted energy consumption value for the target time period;

[0187] Data calculation module 23 is used to obtain the actual energy consumption measurement value of the predicted target period and calculate the energy consumption deviation value between the predicted energy consumption value and the actual energy consumption measurement value.

[0188] The data judgment module 24 is used to determine whether control optimization is needed based on the comparison between the absolute value of the energy consumption deviation value and the preset deviation threshold.

[0189] The control strategy module 25 is used to filter from multi-source data to obtain a target state variable set if optimization is required; combine the energy consumption deviation value, the preset energy consumption target value and the target state variable set into state information; input the state information into a pre-trained reinforcement learning model to obtain the target control strategy; the target control strategy includes: process parameter adjustment amount and / or equipment operating parameter adjustment amount.

[0190] See Figure 3 , Figure 3 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 3 The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the modules in the aforementioned device embodiments, for example... Figure 2 The functions of the data acquisition module 21, data processing module 22, data calculation module 23, data judgment module 24, and control strategy module 25 are shown.

[0191] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0192] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0193] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store device type information.

[0194] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this application can execute the implementation methods described in any embodiment of the intelligent prediction and control method for energy consumption of the entire lithium battery separator production process provided in the embodiments of this application, or they can execute the implementation methods of the electronic devices described in the embodiments of this application, which will not be repeated here.

[0195] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0196] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0197] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0198] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0199] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces or units, or it may be an electrical, mechanical, or other form of connection.

[0200] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0201] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0202] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent prediction and control of energy consumption throughout the entire lithium battery separator production process, characterized in that, include: Acquire multi-source data during the lithium battery separator production process for the predicted target period, including process parameters, equipment operating parameters, environmental parameters, and quality inspection parameters; The multi-source data is input into a preset graph neural network model to obtain the predicted energy consumption value for the target time period. The construction process of the graph neural network model includes: using production processes as nodes and material transfer relationships and energy interaction relationships between processes as edges to construct a graph structure; using an adjacency matrix to represent the connection strength between nodes, and the node feature vector is composed of multi-source data of the corresponding process; the graph neural network model includes an input layer, a graph convolutional layer, a pooling layer and an output layer, and the output layer uses a linear activation function to output the predicted energy consumption value. After the predicted target period ends, the actual energy consumption value is obtained, and the energy consumption deviation value between the predicted energy consumption value and the actual energy consumption value is calculated. Based on the comparison between the absolute value of the energy consumption deviation and the preset deviation threshold, it is determined whether control optimization is needed. If optimization is required, the multi-source data is extracted to obtain a set of state variables; the absolute value of the energy consumption deviation value is compared with a preset deviation threshold, and the set of state variables is extracted based on the comparison result to obtain a target set of state variables; wherein, if the absolute value of the energy consumption deviation value is less than or equal to the preset deviation threshold, the set of state variables is used as the target set of state variables; if the absolute value of the energy consumption deviation value is greater than the preset deviation threshold, the state variables in the set of state variables are extracted as the target set of state variables based on the feature importance evaluation result. The energy consumption deviation value and the target state variable set are combined into state information; The state information is input into a pre-trained reinforcement learning model to obtain a target control strategy; the target control strategy includes: process parameter adjustment amount and / or equipment operating parameter adjustment amount; During a preset evaluation period after the target control strategy is implemented, the actual energy consumption value and the corresponding multi-source evaluation data are obtained; the multi-source evaluation data are then input into a preset quality prediction model to obtain the predicted quality index. Based on the actual energy consumption value, the preset energy consumption target value for the evaluation period, the predicted quality index, the preset quality index target value, and the quality qualification threshold, a reward signal is calculated through a reward function. Using the initial state information of the evaluation period, the executed target control strategy, the reward signal, and the state information at the end of the evaluation period, an experience tuple is constructed; the parameters of the reinforcement learning model are optimized based on the experience tuple. The evaluation multi-source data is input into the graph neural network model to obtain the predicted energy consumption value, and the prediction error between the energy consumption value and the actual energy consumption value is calculated; the parameters of the graph neural network model are optimized based on the prediction error.

2. The intelligent predictive control method for energy consumption throughout the entire lithium battery separator production process according to claim 1, characterized in that, Also includes: During the execution of the target control strategy, the adjustment step size of the process parameter adjustment amount and / or equipment operating parameter adjustment amount is calculated based on the absolute value of the energy consumption deviation value and the historical adjustment record. The historical adjustment record includes the adjustment amount and the corresponding absolute value of energy consumption deviation in the past several control cycles. The adjustment amount of the target control strategy is executed within the range of the adjustment step size; If the adjustment amount required by the target control strategy is greater than the adjustment step size, then the adjustment amount of the target control strategy is executed based on the adjustment step size, and an alarm signal is generated.

3. The intelligent predictive control method for energy consumption throughout the entire lithium battery separator production process according to claim 2, characterized in that, Also includes: If the absolute value of the energy consumption deviation is greater than the preset deviation threshold for N consecutive control cycles after the alarm signal is generated, the optimization process of lithium battery separator production control will be suspended. Based on the correlation analysis between historical multi-source data and energy consumption deviation values, potential faulty equipment or abnormal process links can be located.

4. The intelligent predictive control method for energy consumption throughout the entire lithium battery separator production process according to claim 1, characterized in that, Also includes: When the predicted quality index is less than the quality qualification threshold, the multi-source evaluation data is input into the quality constraint proxy model to generate a safe operating range for process parameters. Under the constraints of the safe operating range, the target control strategy is regenerated through a reinforcement learning model.

5. The intelligent predictive control method for energy consumption throughout the entire lithium battery separator production process according to claim 1, characterized in that, Based on actual energy consumption, preset energy consumption target values ​​for the assessment period, predicted quality indicators, preset quality indicator target values, and quality qualification thresholds, a reward signal is calculated using a reward function, including: If the predicted quality index is less than the quality qualification threshold, a negative penalty signal is output as a reward signal. If the predicted quality index is greater than or equal to the quality qualification threshold, then the energy consumption bonus component is calculated based on the deviation between the actual energy consumption value and the preset evaluation period energy consumption target value. The quality reward component is calculated based on the deviation between the predicted quality indicator and the preset quality indicator target value. The energy consumption reward component and the quality reward component are weighted and fused together to form the reward signal.

6. The intelligent predictive control method for energy consumption throughout the entire lithium battery separator production process according to claim 1, characterized in that, The step of extracting state variables from the set of state variables as the target set of state variables based on the feature importance evaluation results includes: Based on the pre-trained feature importance evaluation model, all state variables in the state variable set are evaluated to obtain the importance score of each state variable. The importance scores corresponding to all state variables in the set of state variables are sorted from high to low. The threshold for the cumulative importance percentage of the target is determined based on the absolute value of the energy consumption deviation value; State variables are selected sequentially based on the ranking until the sum of the cumulative importance scores of the selected state variables reaches or exceeds the product of the target cumulative importance percentage threshold and the sum of the total importance scores. The selected state variables are then used as the target state variable set.

7. The intelligent predictive control method for energy consumption throughout the entire lithium battery separator production process according to claim 6, characterized in that, The determination of the target cumulative importance percentage threshold based on the absolute value of the energy consumption deviation includes: If the absolute value of the energy consumption deviation is less than the first threshold, then the target cumulative importance ratio threshold is a first predetermined ratio; If the absolute value of the energy consumption deviation is greater than or equal to the first threshold and less than or equal to the second threshold, then the target cumulative importance ratio threshold is the second predetermined ratio. If the absolute value of the energy consumption deviation is greater than the second threshold, then the target cumulative importance ratio threshold is a third predetermined ratio; Wherein, the first predetermined ratio is less than the second predetermined ratio, and the second predetermined ratio is less than the third predetermined ratio.

8. A smart predictive control system for energy consumption throughout the entire lithium battery separator production process, characterized in that, include: The data acquisition module is used to acquire multi-source data during the lithium battery separator production process for the predicted target period. The multi-source data includes process parameters, equipment operating parameters, environmental parameters, and quality inspection parameters. The data processing module is used to input the multi-source data into a preset graph neural network model to obtain the predicted energy consumption value for the target time period. The construction process of the graph neural network model includes: using production processes as nodes and material transfer relationships and energy interaction relationships between processes as edges to construct a graph structure; using an adjacency matrix to represent the connection strength between nodes, and the node feature vector is composed of multi-source data of the corresponding process; the graph neural network model includes an input layer, a graph convolutional layer, a pooling layer and an output layer, and the output layer uses a linear activation function to output the predicted energy consumption value. The data calculation module is used to obtain the actual energy consumption value of the predicted target time period and calculate the energy consumption deviation value between the predicted energy consumption value and the actual energy consumption value. The data judgment module is used to determine whether control optimization is needed based on the comparison between the absolute value of the energy consumption deviation value and a preset deviation threshold. The control strategy module is used to extract the multi-source data to obtain a set of state variables if optimization is required; compare the absolute value of the energy consumption deviation value with a preset deviation threshold, and extract the set of state variables based on the comparison result to obtain a target set of state variables; wherein, if the absolute value of the energy consumption deviation value is less than or equal to the preset deviation threshold, the set of state variables is used as the target set of state variables; if the absolute value of the energy consumption deviation value is greater than the preset deviation threshold, the state variables in the set of state variables are extracted as the target set of state variables based on the feature importance evaluation result. The energy consumption deviation value, the preset energy consumption target value, and the target state variable set are combined into state information; the state information is input into a pre-trained reinforcement learning model to obtain a target control strategy; the target control strategy includes: process parameter adjustment amount and / or equipment operating parameter adjustment amount; During a preset evaluation period after the target control strategy is implemented, the actual energy consumption value and the corresponding multi-source evaluation data are obtained; the multi-source evaluation data are then input into a preset quality prediction model to obtain the predicted quality index. Based on the actual energy consumption value, the preset energy consumption target value for the evaluation period, the predicted quality index, the preset quality index target value, and the quality qualification threshold, a reward signal is calculated through a reward function. Using the initial state information of the evaluation period, the executed target control strategy, the reward signal, and the state information at the end of the evaluation period, an experience tuple is constructed; the parameters of the reinforcement learning model are optimized based on the experience tuple. The evaluation multi-source data is input into the graph neural network model to obtain the predicted energy consumption value, and the prediction error between the energy consumption value and the actual energy consumption value is calculated; the parameters of the graph neural network model are optimized based on the prediction error.

Images