Self-adaptive temperature control method for polyphenylene sulfide polymerizer

Abstract

The invention relates to the technical field of process control, in particular to a polyphenylene sulfide polymerizer temperature self-adaptive control method, which comprises the following steps: acquiring melt viscosity, heat release power and residence time deviation of a polymerizer in real time, and constructing a normalized multi-dimensional time sequence input vector by using a sliding window technology; and then, through fusing a weight control model of a long short-term memory network and an attention mechanism, combining static working condition characteristics, and analyzing and outputting adaptive weight coefficients for different deviation dimensions in real time. Meanwhile, based on the shear heat effect and the reaction accumulation effect, specific temperature compensation components for the inlet of the heating medium jacket are calculated, and independent amplitude limiting processing is carried out on all the components. And finally, a target set temperature is synthesized through weighting, and an execution mechanism is driven to act accurately by utilizing cascade PID control and a valve flow characteristic inverse mapping model. According to the invention, through deep fusion and mechanism compensation, accurate steady-state control of the polymerization reaction is realized.

Description

Technical Field

[0001] This invention relates to the field of process control technology, specifically to an adaptive temperature control method for polyphenylene sulfide polymerization reactors. Background Technology

[0002] Continuous polymerization systems, especially those used to produce high-performance polymers such as polyphenylene sulfide, typically employ multiple polymerization reactors operating in series. Because polymerization itself is a typical complex system characterized by high exothermic activity, high viscosity, and large hysteresis, the quality indicators of the reaction products (such as melt viscosity) are extremely sensitive to even minute temperature fluctuations. Therefore, achieving precise, stable, and robust temperature control throughout the continuous production process is crucial for ensuring product quality and operational safety.

[0003] In existing technologies, temperature control in polymerization reactors is mainly based on traditional PID (Proportional-Integral-Derivative) control methods or cascade control strategies extended from these methods. These methods acquire real-time material temperatures within the reactor using sensors and generate control signals based on the deviation between these temperatures and setpoints, thereby adjusting the circulation flow rate of the heat transfer medium (such as heat transfer oil) within the temperature regulating jacket. This control method relies on preset fixed parameters and performs well when the system is in a stable operating state.

[0004] However, in the complex industrial production environment, this fixed-parameter-based control strategy faces numerous challenges. First, the strong coupling within the polymerization reaction system is a core problem. For example, changes in the melt viscosity directly affect the shear heat (i.e., endogenous heat) generated during stirring, thus altering the actual exothermic power of the reaction. Simultaneously, fluctuations in the reactant level affect the actual residence time within the reactor, directly impacting the reaction depth and instantaneous calorific value at any given moment. Traditional control loops struggle to effectively decouple these nonlinear and intense parameter interactions.

[0005] Secondly, existing systems lack adaptability and generalization ability to dynamic operating conditions. During continuous operation, small fluctuations in feed composition, catalytic activity decay, and vessel wall scaling can all cause the optimal operating point of the system to drift slowly. Traditional control algorithms lack an effective mechanism to identify these changes in operating conditions and dynamically adjust their control gain or model parameters according to the new operating background, resulting in a decrease in control accuracy over time.

[0006] Finally, existing control systems have limitations in utilizing multi-dimensional information. Although a large amount of multi-source, high-value data, such as melt viscosity, inlet and outlet temperatures of the heat transfer jacket, and heat transfer circulation flow rate, can be collected in real time during the production process, current technologies mainly use only a few parameters, such as temperature and liquid level, as feedback signals. This makes the system lack the ability to efficiently integrate this heterogeneous multi-dimensional information and transform it into high-precision, adaptive control decisions.

[0007] The core problem facing the temperature control technology of existing continuous polymerization reaction systems is: how to establish a control method that can make full use of multi-source, multi-dimensional real-time operating data, effectively cope with the strong coupling of reactant parameters and the dynamic changes in operating conditions, so as to achieve high-precision adaptive adjustment of the polymerization reactor temperature.

[0008] To address this, an adaptive temperature control method for polyphenylene sulfide polymerization reactor is proposed. Summary of the Invention

[0009] The purpose of this invention is to provide an adaptive temperature control method for polyphenylene sulfide (PPS) polymerization reactors. Through deep integration and mechanistic compensation, precise steady-state control of the polymerization reaction is achieved. This includes real-time acquisition of melt viscosity, exothermic power, and residence time deviations in the polymerization reactor, and the construction of a normalized multi-dimensional time-series input vector using a sliding window technique. Subsequently, a weighted control model integrating a long short-term memory network and an attention mechanism is used, combined with static operating condition characteristics, to analyze and output adaptive weight coefficients for different deviation dimensions in real time. Simultaneously, based on shear heat effect and reaction accumulation effect, a specific temperature compensation component for the heat transfer medium jacket inlet is calculated, and each component is independently limited. Finally, the target set temperature is weighted and synthesized, and the actuator is driven to move precisely using cascade PID control and a valve flow characteristic inverse mapping model.

[0010] To achieve the above objectives, the present invention provides the following technical solution: A method for adaptive temperature control of a polyphenylene sulfide polymerization reactor, comprising: Collect the melt viscosity, liquid level, jacket inlet temperature, jacket outlet temperature, and heat medium circulation flow rate of the polymerization reactor; Calculate the viscosity deviation between the real-time melt viscosity and the target viscosity; calculate the real-time heat release power based on the jacket inlet temperature, jacket outlet temperature and heat medium circulation flow rate and obtain the power deviation relative to the standard heat release power; calculate the actual residence time deviation based on the rate of change of the real-time liquid level. First temperature compensation value, second temperature compensation value, and third temperature compensation value are generated based on viscosity deviation, power deviation, and actual residence time deviation, respectively; viscosity deviation, power deviation, and actual residence time deviation within a preset historical time period are extracted to construct a multi-dimensional time-series sliding window input vector; Input the multidimensional time-series sliding window input vector into the weight control model, and output the first weight coefficient, the second weight coefficient, and the third weight coefficient. The first temperature compensation value, the second temperature compensation value, and the third temperature compensation value are weighted and summed using the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient. The calculation result is superimposed on the reference temperature to obtain the final target set temperature. The opening degree of the heat medium regulating valve is adjusted according to the final target set temperature.

[0011] Preferably, the data collection of melt viscosity, liquid level, jacket inlet temperature, jacket outlet temperature, and heat medium circulation flow rate in the polymerization reactor includes: An online torsional viscometer is used to detect the vibration damping signal of the probe in the reactant. Using a preset viscosity-damping correspondence table, the vibration damping signal is mapped and converted into the melt viscosity. A dual-flange differential pressure transmitter is used to collect the real-time pressure difference between the bottom flange of the polymerization reactor and the gas phase space. Based on the principle of fluid statics, the real-time pressure difference is divided by the product of a preset melt density and gravitational acceleration to calculate the liquid level height. A sheathed platinum resistance temperature sensor is used to collect the resistance value signal that changes with temperature. Based on the resistance-temperature characteristic curve of the platinum resistance, the collected resistance value signal is converted into the jacket inlet temperature and the jacket outlet temperature. A mass flow meter is used to detect the phase difference signal generated by fluid flow in the measuring tube. The phase difference signal is processed through linear proportional conversion to obtain the heat medium circulation flow rate.

[0012] Preferably, the calculation process for the viscosity deviation, power deviation, and actual residence time deviation includes: performing a numerical difference operation between the real-time melt viscosity and the target viscosity to obtain an absolute viscosity difference with positive and negative signs, and dividing the absolute viscosity difference by the target viscosity to obtain the normalized viscosity deviation; reading the pre-stored specific heat capacity parameter of the heat medium, calculating the temperature difference between the jacket outlet temperature and the jacket inlet temperature, and calculating the product of the temperature difference, the heat medium circulation flow rate, and the specific heat capacity of the heat medium to obtain the real-time heat release power at the current moment; comparing the real-time heat release power with the center value of the standard heat release power, and calculating the difference ratio between the two as the power deviation; performing a first-order difference operation on the continuously collected real-time liquid level height, calculating the difference in liquid level height between the current sampling moment and the previous sampling moment, and dividing by the sampling period to obtain the liquid level rise / fall rate per unit time; mapping the liquid level rise / fall rate into the residence time fluctuation value of the reactant based on the effective volume parameter of the polymerization reactor, and using the residence time fluctuation value as the actual residence time deviation.

[0013] Preferably, the process of generating the first temperature compensation value, the second temperature compensation value, and the third temperature compensation value includes: obtaining preset coefficients required for calculation; reading preset viscosity ratio coefficients, power ratio coefficients, time ratio coefficients, and preset shear heat coupling coefficients and reaction cumulative coupling coefficients from the controller storage unit; multiplying the viscosity deviation by the viscosity ratio coefficients to obtain a viscosity basis component that only responds to viscosity changes; multiplying the power deviation by the power ratio coefficients to obtain a power basis component that only responds to exothermic changes; and multiplying the actual residence time deviation by the time ratio coefficients to obtain a time basis component that only responds to time changes. The following steps are performed: Calculate the product of the viscosity deviation and the power deviation, and multiply this product by the shear-thermal coupling coefficient to obtain a shear-thermal correction term for compensating for high-viscosity frictional heat generation; calculate the product of the power deviation and the actual residence time deviation, and multiply this product by the reaction cumulative coupling coefficient to obtain a cumulative correction term for compensating for reaction hysteresis effects; sum the viscosity basis component and the shear-thermal correction term to obtain the first temperature compensation value; subtract the cumulative correction term from the power basis component to obtain the second temperature compensation value; and directly assign the time basis component to the third temperature compensation value.

[0014] Preferably, the process of constructing the multi-dimensional time-series sliding window input vector includes: setting the time depth parameter and sampling step size parameter of the sliding window; opening a first FIFO queue, a second FIFO queue, and a third FIFO queue in memory to store historical sequence data of the viscosity deviation, the power deviation, and the actual residence time deviation, respectively; according to the sampling step size, pushing the data of the current moment and the past N consecutive sampling moments into the corresponding FIFO queues in sequence, and removing the earliest data to maintain a fixed queue length; reading the preset physical limit amplitude of each parameter, and respectively... All data in the first, second, and third FIFO queues undergo max-min normalization to map all values ​​to a dimensionless range of zero to one. The normalized data from the first FIFO queue is extracted as the viscosity feature dimension. The normalized data from the second FIFO queue is extracted as the power feature dimension. The normalized data from the third FIFO queue is extracted as the time feature dimension. The viscosity feature dimension, the power feature dimension, and the time feature dimension are concatenated end-to-end or stacked in parallel according to a preset channel order to generate the multidimensional time-series sliding window input vector.

[0015] Preferably, the weight control model includes: The dynamic trend encoder receives the multi-dimensional time-series sliding window input vector as input, processes it using a long short-term memory network, eliminates short-term random noise through internal forget gate and input gate mechanisms, captures the historical evolution patterns of viscosity deviation, power deviation and residence time deviation, and generates a dynamic trend feature vector containing dynamic time-varying information. The static condition encoder receives preset reference temperature, target viscosity and standard exothermic power as input, and processes them using a fully connected embedded network to map the physical process setting parameters into high-dimensional sparse semantic vectors, generating static condition feature vectors that characterize the current polymerization reaction stage. The multi-source attention fusion unit concatenates the static working condition feature vector with the dynamic trend feature vector, inputs the concatenated composite state vector into the attention generation network, generates attention weight vectors for different response parameters, applies the attention weight vectors to the dynamic trend feature vectors output by the dynamic trend encoder, performs dynamic weighting through a feature recalibration mechanism, and outputs global context-aware features. The adaptive decision output layer receives the global context-aware features as input, uses a multilayer perceptron to reduce the dimensionality of the high-dimensional features and perform linear combination, and maps the operation results to the numerical range of 0 to 1 through a sigmoid nonlinear activation function, and finally outputs the first weight coefficient, the second weight coefficient and the third weight coefficient in parallel.

[0016] Preferably, the process of obtaining the final target set temperature includes: calculating the product of the first weighting coefficient and the first temperature compensation value, the product of the second weighting coefficient and the second temperature compensation value, and the product of the third weighting coefficient and the third temperature compensation value, respectively, and adding the three product results to obtain the total dynamic compensation amount; performing algebraic superposition of the total dynamic compensation amount and the reference temperature to generate the original target set temperature; reading the preset maximum allowable temperature and minimum holding temperature of the polymerization process; determining whether the original target set temperature exceeds the safety range formed by the maximum allowable temperature and the minimum holding temperature; if it exceeds the safety range, forcibly clamping the original target set temperature to the corresponding boundary value; calculating the difference between the original target set temperature of the current calculation cycle and the final target set temperature of the previous cycle; if the absolute value of the difference is greater than the preset maximum heating rate threshold, then performing peak-shaving processing on the original target set temperature and outputting the rate-smoothed value as the final target set temperature.

[0017] Preferably, adjusting the opening of the heat medium regulating valve includes: using the final target set temperature as the set value of the main control loop, using the collected real-time temperature of the polymerization reactor as the feedback value of the main control loop, and calculating the jacket temperature set value through the main PID controller; using the jacket temperature set value as the set value of the secondary control loop, using the collected jacket inlet temperature as the feedback value of the secondary control loop, and calculating the standard control output signal through the secondary PID controller; retrieving a preset heat medium regulating valve flow characteristic curve, the characteristic curve representing the nonlinear relationship between valve opening and heat medium flow; using the inverse function model of the flow characteristic curve to perform reverse mapping on the standard control output signal, and calculating the corrected linearized valve opening command; converting the linearized valve opening command into a current signal to drive the actuator of the heat medium regulating valve to move, so that the valve core position reaches the corresponding opening.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By constructing a multi-factor coupling correction term, the strong coupling problem between viscosity, heat release, and residence time in traditional control is solved. Traditional PID control treats these complex effects as a single disturbance, which can easily lead to misjudgments in the control output. While calculating the basic compensation quantity, shear heat correction terms and cumulative heat correction terms are introduced to perform precise physical decoupling and correction of the dynamic compensation command. This interactive correction mechanism effectively improves the accuracy of the target set temperature.

[0019] 2. A weighted control unit was constructed using a deep learning model based on dual-source input, which significantly enhanced the system's robustness and adaptability to dynamic operating conditions. The model not only receives historical deviation sequences (dynamic information) but also inputs process setpoints such as target viscosity and reference temperature in parallel (static operating condition information). This dual-source feature fusion mechanism enables the model to determine the reaction stage of the current operating condition in real time and dynamically adjust the weight coefficients of the three compensation components. Therefore, in stages of drastic changes in operating conditions, such as the start, transition, or end of the polymerization reaction, the model can adaptively select the most suitable control strategy, effectively avoiding temperature deviations caused by the lag of manual parameter adjustment.

[0020] 3. By inputting three-dimensional time-series data into a model that combines long short-term memory networks and attention mechanisms, it is possible to deeply explore the complex evolution trends in historical data and identify the key information that has the greatest impact on current control. The AI-based decision-making mechanism avoids the complex logic design of multi-parameter parallel connection in traditional control, provides a continuous and smooth weight output, reduces hard jumps in control output, extends the service life of actuators such as valves, and improves the overall system stability. Attached Figure Description

[0021] Figure 1This is a schematic diagram of the process for an adaptive temperature control method for a polyphenylene sulfide polymerization reactor according to the present invention. Figure 2 This is a schematic diagram of the process for generating temperature compensation values ​​according to the present invention; Figure 3 This is a schematic diagram of the weight control model of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figures 1 to 3 This invention provides a method for adaptive temperature control of a polyphenylene sulfide polymerization reactor, the technical solution of which is as follows: Example

[0024] A method for adaptive temperature control of a polyphenylene sulfide polymerization reactor, comprising: Collect the melt viscosity, liquid level, jacket inlet temperature, jacket outlet temperature, and heat medium circulation flow rate of the polymerization reactor; Calculate the viscosity deviation between the real-time melt viscosity and the target viscosity; calculate the real-time heat release power based on the jacket inlet temperature, jacket outlet temperature and heat medium circulation flow rate and obtain the power deviation relative to the standard heat release power; calculate the actual residence time deviation based on the rate of change of the real-time liquid level. First temperature compensation value, second temperature compensation value, and third temperature compensation value are generated based on viscosity deviation, power deviation, and actual residence time deviation, respectively; viscosity deviation, power deviation, and actual residence time deviation within a preset historical time period are extracted to construct a multi-dimensional time-series sliding window input vector; Input the multidimensional time-series sliding window input vector into the weight control model, and output the first weight coefficient, the second weight coefficient, and the third weight coefficient. The first temperature compensation value, the second temperature compensation value, and the third temperature compensation value are weighted and summed using the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient. The calculation result is superimposed on the reference temperature to obtain the final target set temperature. The opening degree of the heat medium regulating valve is adjusted according to the final target set temperature.

[0025] Furthermore, the collection of melt viscosity, liquid level, jacket inlet temperature, jacket outlet temperature, and heat medium circulation flow rate of the polymerization reactor includes: An online torsional viscometer is used to detect the vibration damping signal of the probe in the reactant. Using a preset viscosity-damping correspondence table, the vibration damping signal is mapped and converted into the melt viscosity. A dual-flange differential pressure transmitter is used to collect the real-time pressure difference between the bottom flange of the polymerization reactor and the gas phase space. Based on the principle of fluid statics, the real-time pressure difference is divided by the product of a preset melt density and gravitational acceleration to calculate the liquid level height. A sheathed platinum resistance temperature sensor is used to collect the resistance value signal that changes with temperature. Based on the resistance-temperature characteristic curve of the platinum resistance, the collected resistance value signal is converted into the jacket inlet temperature and the jacket outlet temperature. A mass flow meter is used to detect the phase difference signal generated by fluid flow in the measuring tube. The phase difference signal is processed through linear proportional conversion to obtain the heat medium circulation flow rate.

[0026] Specifically, the melt viscosity is obtained using an online torsional vibration viscometer. This instrument detects the resistance encountered by the probe when vibrating within the material in the polymerization reactor, acquiring this vibration damping signal in real time. The signal intensity has a direct functional relationship with the material viscosity. The control system then consults a viscosity-damping correspondence table pre-calibrated in a laboratory environment and uses interpolation to accurately convert the damping signal into a real-time melt viscosity value. The liquid level is measured using a dual-flange differential pressure transmitter installed on the polymerization reactor. This transmitter acquires the pressure difference between the bottom and top of the reactor in real time. After reading this pressure difference, based on the principles of fluid statics, the specific formula is as follows: ; in, Indicates the liquid level height. Indicates pressure difference, This indicates the preset melt density. It represents the acceleration due to gravity.

[0027] The jacket inlet and outlet temperatures are acquired by high-precision armored platinum resistance temperature sensors. The resistance value output by the sensor is read, and converted into a Celsius temperature value according to the resistance-temperature characteristic curve of the platinum resistance. The heat medium circulation flow rate is measured by a Coriolis mass flow meter. The flow meter detects the phase difference signal generated by the flow, and the phase difference signal is processed by a standard linear proportional conversion to obtain the mass circulation flow rate value of the heat medium.

[0028] By explicitly employing specialized industrial sensors such as torsional viscometers, differential pressure transmitters, and mass flow meters, and supplementing these with rigorous mathematical and physical models such as fluid statics principles and heat balance formulas for calculation, the high precision and reliability of the raw data input to the intelligent control model are ensured. This strict control over data quality is the foundation for the subsequent deep learning model to be stably trained and output high-precision control decisions, effectively avoiding the problems of garbage input and garbage output caused by impure sensor signals or calculation deviations.

[0029] Further, the calculation process for the viscosity deviation, power deviation, and actual residence time deviation includes: performing a numerical difference operation between the real-time melt viscosity and the target viscosity to obtain an absolute viscosity difference with positive and negative signs, and dividing the absolute viscosity difference by the target viscosity to obtain the normalized viscosity deviation; reading the pre-stored specific heat capacity parameter of the heat medium, calculating the temperature difference between the jacket outlet temperature and the jacket inlet temperature, and calculating the product of the temperature difference, the heat medium circulation flow rate, and the specific heat capacity of the heat medium to obtain the real-time heat release power at the current moment; comparing the real-time heat release power with the center value of the standard heat release power, and calculating the difference ratio between the two as the power deviation; performing a first-order difference operation on the continuously collected real-time liquid level height, calculating the difference in liquid level height between the current sampling moment and the previous sampling moment, and dividing by the sampling period to obtain the liquid level rise / fall rate per unit time; mapping the liquid level rise / fall rate into the residence time fluctuation value of the reactant based on the effective volume parameter of the polymerization reactor, and using the residence time fluctuation value as the actual residence time deviation.

[0030] Specifically, the system first reads the current real-time melt viscosity monitoring value and the target viscosity set in the process formula. The calculation unit performs a numerical difference operation, subtracting the target viscosity from the real-time melt viscosity to obtain an absolute viscosity difference value with a positive or negative sign. If the difference is positive, it indicates that the current reaction is too deep or the molecular weight is too large; if it is negative, it indicates that the reaction is insufficient. In order to eliminate the dimensional differences caused by different target viscosity bases for different product grades (e.g., high viscosity grade and low viscosity grade), the system divides the absolute viscosity difference value by the target viscosity to obtain a normalized viscosity deviation. This deviation value is a dimensionless percentage value that directly reflects the degree of deviation of the current product quality from the set target, which is convenient for the neural network model to process uniformly.

[0031] The exothermic reaction state is calculated based on the principle of heat balance. First, the specific heat capacity parameter of the heat transfer oil (which is a constant or a function of temperature) is read from the storage unit. Then, the difference between the jacket outlet temperature and the jacket inlet temperature is calculated to obtain the heat transfer medium temperature rise. Subsequently, the calculation unit multiplies the heat transfer medium temperature rise, the real-time collected heat transfer medium circulation flow rate, and the specific heat capacity of the heat transfer medium. The result is the real-time exothermic power of the polymerization reactor at the current moment (usually in kW). This real-time value is compared with the preset standard exothermic power center value (i.e., the theoretical ideal heat release of this reaction stage), and the ratio of the difference between the two to the standard value is calculated to obtain the power deviation. This step can sensitively detect abnormal changes in the reaction rate caused by catalyst activity fluctuations or impurities.

[0032] By monitoring the rate of change in liquid level to extrapolate fluctuations in material residence time, the computing unit performs a first-order difference operation on the continuously collected real-time liquid level height. This involves calculating the difference between the liquid level height in the current sampling period and the liquid level height in the previous sampling period, and dividing this difference by the duration of the sampling period to obtain the rate of liquid level rise / fall per unit time. Subsequently, the geometric parameters of the polymerization reactor (including effective cross-sectional area and effective volume) are invoked, and a preset volume-time mapping model is used to convert the rate of liquid level rise / fall into a fluctuation value of the residence time of the reactants. Finally, this fluctuation value is used as the actual residence time deviation. This calculation process transforms static liquid level data into a dynamic time-dimensional deviation, accurately reflecting the effect of prolonged or shortened reaction time caused by imbalances in the infeed and outfeed.

[0033] By normalizing the viscosity calculation, the interference of differences in the target viscosity base of different product grades on the algorithm weights is eliminated, enhancing the versatility of the control strategy. The power deviation is calculated based on the thermodynamic balance principle, which can more intuitively and essentially reflect the real-time dynamics of the polymerization reaction rate compared to simple temperature monitoring. Furthermore, by using the first-order difference of the liquid level to deduce the residence time deviation, the static liquid level signal is successfully transformed into a dynamic time signal reflecting the material flow trend, thus providing data support with clear physical meaning and high signal-to-noise ratio for subsequent adaptive weighted control.

[0034] Further, the process of generating the first temperature compensation value, the second temperature compensation value, and the third temperature compensation value includes: obtaining the preset coefficients required for calculation; reading preset viscosity ratio coefficients, power ratio coefficients, time ratio coefficients, as well as preset shear heat coupling coefficients and reaction cumulative coupling coefficients from the controller storage unit; multiplying the viscosity deviation by the viscosity ratio coefficients to obtain a viscosity basis component that only responds to viscosity changes; multiplying the power deviation by the power ratio coefficients to obtain a power basis component that only responds to exothermic changes; and multiplying the actual residence time deviation by the time ratio coefficients to obtain a time basis component that only responds to time changes. The following steps are taken: Calculate the product of the viscosity deviation and the power deviation, and multiply this product by the shear-thermal coupling coefficient to obtain a shear-thermal correction term for compensating for high-viscosity frictional heat generation; calculate the product of the power deviation and the actual residence time deviation, and multiply this product by the reaction cumulative coupling coefficient to obtain a cumulative correction term for compensating for reaction hysteresis effects; sum the viscosity base component and the shear-thermal correction term to obtain the first temperature compensation value; subtract the cumulative correction term from the power base component to obtain the second temperature compensation value; and directly assign the time base component to the third temperature compensation value.

[0035] Specifically, before the control loop begins, the controller first accesses its internal non-volatile memory unit to read five pre-calibrated sets of key control coefficients. These coefficients include viscosity proportionality coefficient, power proportionality coefficient, and time proportionality coefficient, which are used to set the system's response sensitivity to a single variable; and shear heat coupling coefficient and reaction cumulative coupling coefficient, which are used to quantify the intensity of the interaction between different variables. These coefficients are usually set based on regression analysis or expert experience using historical operating data of the polymerization reactor, forming the basic model parameters of the algorithm. The controller first assumes that the physical quantities do not interfere with each other and calculates the linear contribution value of each single factor. The processor performs a multiplication operation between the real-time input viscosity deviation and the viscosity proportionality coefficient, and the result is used as the viscosity base component. This component represents the temperature range that the system should adjust when considering only viscosity changes. Similarly, the processor multiplies the power deviation by the power proportionality coefficient to obtain the power base component, and multiplies the actual residence time deviation by the time proportionality coefficient to obtain the time base component. These three base components constitute the reference value for temperature regulation.

[0036] The physical basis of the shear heat correction term lies in the frictional heat generated by the relative motion between the stirring blades and the high-viscosity material during the polymerization reactor stirring process. This frictional heat is mainly influenced by both viscosity and stirring power. Although rigorous thermodynamic derivation requires complex parameters such as stirrer geometry and rotational speed, empirical statistics based on historical operating data show that this interaction can be described by the following approximate relationship: Additional frictional heat ≈ Base frictional heat × f (viscosity deviation, power deviation), where f is the interaction function. Considering the continuity and smoothness of the system, a linear form is adopted: f ≈ Viscosity deviation × Power deviation × Shear coefficient. This correction term is used as a higher-order correction to the base temperature compensation, compensating for the additional heat generated by the simultaneous increase in viscosity and power.

[0037] To address the strong coupling issue in the polymerization reaction, the controller then calculates two key nonlinear correction terms. First, it calculates the product of viscosity deviation and power deviation, and multiplies this product by the shear heat coupling coefficient to obtain the shear heat correction term. This term quantifies the additional mechanical frictional heat generated when high viscosity and high reaction rate occur simultaneously. Second, it calculates the product of power deviation and actual residence time deviation, and multiplies this product by the reaction accumulation coupling coefficient to obtain the accumulation correction term. This term compensates for the cumulative effect of reaction depth caused by changes in the material residence time distribution (RTD). In a continuous polymerization reactor, when the liquid level change rate indicates an extension of the actual residence time of the material, it means that the average heating and reaction time of the reactants in the reactor have increased, which may lead to the reaction degree exceeding the preset steady-state equilibrium point, thereby generating an unexpected cumulative reaction heat. The reaction accumulation coupling coefficient is used to quantify this time-heat conversion sensitivity. By calculating the product of power deviation and time deviation, the model can proactively identify local overheating trends caused by slow flow and output cooling signals in advance to suppress them. Finally, the controller generates the final output command through algebraic addition and subtraction operations, summing the viscosity basis component and the shear heat correction term to obtain the first temperature compensation value. This adds compensation against shear heat to the basic adjustment. Simultaneously, the cumulative correction term is subtracted from the power basis component to obtain the second temperature compensation value, thereby mitigating over-adjustment that may be caused by hysteresis and preventing temperature overshoot. The time basis component is directly assigned the third temperature compensation value. Through this basic + correction synthesis logic, precise decoupled control of complex polymerization conditions is achieved.

[0038] By constructing a composite computational model consisting of a basic linear component and a nonlinear coupling correction term, not only is the basic adjustment amount of a single deviation calculated, but also a shear heat correction term and a cumulative correction term are innovatively introduced. The former accurately quantifies the superimposed effect of mechanical friction on exothermic reactions under high viscosity conditions, while the latter proactively suppresses heat accumulation caused by reaction lag. This decoupling strategy, which integrates physical mechanisms into the algorithm, achieves dynamic compensation for complex operating conditions. While ensuring response speed, it significantly reduces the risk of temperature overshoot and ensures the steady-state accuracy of the polymerization reaction at different process stages.

[0039] Further, the process of constructing the multi-dimensional time-series sliding window input vector includes: setting the time depth parameter and sampling step size parameter of the sliding window; opening a first FIFO queue, a second FIFO queue, and a third FIFO queue in memory to store historical sequence data of the viscosity deviation, the power deviation, and the actual residence time deviation, respectively; according to the sampling step size, pushing the data of the current moment and the past N consecutive sampling moments into the corresponding FIFO queues in sequence, and removing the earliest data to maintain a fixed queue length; reading the preset physical limit amplitude values ​​of each parameter, respectively... Perform max-min normalization on all data in the first, second, and third FIFO queues to map all values ​​to a dimensionless range of 0 to 1; extract the normalized data from the first FIFO queue as the viscosity feature dimension; extract the normalized data from the second FIFO queue as the power feature dimension; extract the normalized data from the third FIFO queue as the time feature dimension; concatenate or stack the viscosity feature dimension, power feature dimension, and time feature dimension in a preset channel order to generate the multidimensional time-series sliding window input vector.

[0040] First, the controller reads the preset time depth parameter (set to N=60) and sampling step size parameter (set to 1 second). Based on these two parameters, it allocates three independent First-In-First-Out (FIFO) queues in the random access memory (RAM), defined as queue one, queue two, and queue three, respectively. These three queues are initialized as floating-point arrays of fixed length 60, and are used to store historical sequence data of viscosity deviation, power deviation, and actual residence time deviation, respectively. During the system cold start phase, if the queue data is not full, a zero-padding strategy is used to fill the empty spaces until the queue reaches full capacity. Every [period]... The latest real-time calculation data is acquired every 1 second (i.e., the sampling step size). The controller performs a strict enqueue-dequeue synchronization operation: the viscosity deviation value calculated at the current moment is pushed to the tail of the first queue, the power deviation value is pushed to the tail of the second queue, and the actual residence time deviation value is pushed to the tail of the third queue. At the same time, the oldest data at the head of each queue is forcibly removed. This operation ensures that the queue always retains only the complete data segment within 60 seconds from the current moment, forming a sliding window that moves forward over time. This window locks the dynamic evolution trajectory of the reaction parameters in the near future.

[0041] To eliminate numerical differences between different physical dimensions, a max-min normalization operation is performed on the data in the queues. The controller reads the preset physical limit amplitude, which defines the upper and lower boundaries of the allowable fluctuations of each parameter in the process (for example, setting the allowable range of viscosity deviation to -50% to +50%). It iterates through each data point in the three queues, subtracts the corresponding physical lower limit value from its value, and divides the result by the difference between the physical upper and lower limits. Through this deterministic linear transformation, all data in the three queues are uniformly mapped to the dimensionless interval of 0 to 1, ensuring the equality of numerical weights for all input features. Finally, feature fusion is performed to generate the final input command. The controller extracts the normalized data from the first queue as the viscosity feature channel, the data from the second queue as the power feature channel, and the data from the third queue as the time feature channel. Using a parallel stacking method, these three one-dimensional arrays of length 60 are combined into a two-dimensional matrix structure with a shape of 60 rows × 3 columns. This matrix structure is the multidimensional time-series sliding window input vector, which is directly transmitted to the subsequent neural network model input layer. Each row represents a time step, and each column represents a physical feature, thus fully preserving the temporal and feature correlations of the multi-source data.

[0042] By employing a sliding window mechanism and preserving continuous historical sampling sequences, static instantaneous detection values ​​are transformed into time-series data containing evolutionary patterns. This enables the model to handle the delayed response characteristics of systems with large time lags. The normalization operation based on physical limits unifies the numerical magnitudes of different physical parameters such as viscosity, power, and time, effectively avoiding model weight allocation biases caused by differences in dimensions. Furthermore, the matrix vectors generated using a parallel stacking method achieve strict alignment of multi-source heterogeneous data in the time dimension, providing highly consistent input features for subsequent deep learning algorithms and improving the computational stability of the control strategy.

[0043] Furthermore, the weight control model includes: The dynamic trend encoder receives the multi-dimensional time-series sliding window input vector as input, processes it using a long short-term memory network, eliminates short-term random noise through internal forget gate and input gate mechanisms, captures the historical evolution patterns of viscosity deviation, power deviation and residence time deviation, and generates a dynamic trend feature vector containing dynamic time-varying information. The static condition encoder receives preset reference temperature, target viscosity and standard exothermic power as input, and processes them using a fully connected embedded network to map the physical process setting parameters into high-dimensional sparse semantic vectors, generating static condition feature vectors that characterize the current polymerization reaction stage. The multi-source attention fusion unit concatenates the static working condition feature vector with the dynamic trend feature vector, inputs the concatenated composite state vector into the attention generation network, generates attention weight vectors for different response parameters, applies the attention weight vectors to the dynamic trend feature vectors output by the dynamic trend encoder, performs dynamic weighting through a feature recalibration mechanism, and outputs global context-aware features. The adaptive decision output layer receives the global context-aware features as input, uses a multilayer perceptron to reduce the dimensionality of the high-dimensional features and perform linear combination, and maps the operation results to the numerical range of 0 to 1 through a sigmoid nonlinear activation function, and finally outputs the first weight coefficient, the second weight coefficient and the third weight coefficient in parallel.

[0044] The dynamic trend encoder is connected to the multi-dimensional temporal sliding window input vector generated in the previous step. The input tensor has a shape of (60, 3), which means 60 time steps and 3 feature channels. The encoder is internally constructed as a two-layer stacked long short-term memory network. The first layer LSTM contains 64 hidden neurons, which are responsible for receiving the original input sequence and filtering out high-frequency random noise through the internal forget gate mechanism. The output of the whole sequence hidden state is directly used as the input of the second layer LSTM. The second layer LSTM contains 32 hidden neurons, which are used to further refine the deep temporal patterns. The model only extracts the hidden state output of the second layer LSTM at the last time step (t=60) to generate a dynamic trend feature vector with a dimension of 32.

[0045] The static condition encoder receives three scalar data points at its input: a preset reference temperature, a target viscosity, and a standard exothermic power. The encoder is constructed using a single-layer fully connected network with 16 neurons. The input layer and the fully connected layer are fully connected via a weight matrix. Instead of using a nonlinear activation function, the 3D coordinates of the physical space are directly mapped and projected onto the 16D coordinates of the semantic space to generate a static condition feature vector. This vector provides the model with contextual information about the current reaction stage (such as the initial, steady, or final stage of the reaction).

[0046] The multi-source attention fusion unit performs a feature concatenation operation, concatenating the 32-dimensional dynamic trend feature vector with the 16-dimensional static condition feature vector along the channel dimension to form a 48-dimensional composite state vector. This composite state vector is then input into an attention generation sub-network, which contains a 32-node hidden layer (using ReLU activation) and a 32-node output layer (using Softmax activation). The output of the output layer is the 32-dimensional attention weight vector. Finally, feature recalibration is performed by element-wise multiplying the attention weight vector with the original 32-dimensional dynamic trend feature vector. Through this operation, the model automatically enhances or suppresses the signal strength of specific dimensions in the dynamic features based on the static condition, outputting globally context-aware features.

[0047] The decision output layer is responsible for mapping high-dimensional features to the final control coefficients. This layer receives 32-dimensional global context-aware features as input and is connected to a multilayer perceptron (MLP). The MLP contains a hidden layer with 16 neurons (using the ReLU activation function) and an output layer with 3 neurons. The output layer uses the sigmoid activation function, which strictly restricts the values ​​of the three output nodes to a closed interval between 0.0 and 1.0. These three output nodes correspond to the first weight coefficient, the second weight coefficient, and the third weight coefficient, respectively.

[0048] To ensure the model possesses accurate decision-making capabilities, offline supervised training is required before the system is deployed online. The training dataset consists of high-quality historical operating data from the polymerization reactor over the past year. Stable periods with temperature control deviations within ±0.5℃ are selected, and the corresponding viscosity, power, and liquid level deviations are extracted as input samples X. Using reverse engineering or PID parameter optimization algorithms, the optimal weight combination that maintains system stability during these periods is calculated and used as the label value Y. The loss function is set as the mean squared error loss function to measure the difference between the model's predicted output Y and the true optimal label Y. The optimization process uses the Adam optimizer (adaptive moment estimator) with an initial learning rate of 0.001. The constructed dataset is divided into training and validation sets in an 8:2 ratio. The gradient of the loss function with respect to the weights and biases of each layer in the network is calculated using the backpropagation algorithm, and all network parameters are iteratively updated until the loss value on the validation set converges and no longer decreases, thus completing the model training.

[0049] By constructing a deep weighted control model that integrates a long short-term memory network and an attention mechanism, the dynamic trend encoder uses an LSTM structure to remove short-term random noise from the data and accurately capture the historical evolution of the deviation signal. Combined with the process background information introduced by the static condition encoder, the attention mechanism can adaptively adjust the attention weights of each feature dimension according to the current reaction stage (such as the initiation period or the decay period), realizing the dynamic optimization of the control strategy based on real-time operating conditions and enhancing the robustness and stability of the temperature regulation process in nonlinear environments.

[0050] Further, the process of obtaining the final target set temperature includes: calculating the product of the first weighting coefficient and the first temperature compensation value, the product of the second weighting coefficient and the second temperature compensation value, and the product of the third weighting coefficient and the third temperature compensation value, respectively, and adding the three product results to obtain the total dynamic compensation amount; performing algebraic superposition of the total dynamic compensation amount and the reference temperature to generate the original target set temperature; reading the preset maximum allowable temperature and minimum holding temperature of the polymerization process; determining whether the original target set temperature exceeds the safety range formed by the maximum allowable temperature and the minimum holding temperature; if it exceeds the safety range, forcibly clamping the original target set temperature to the corresponding boundary value; calculating the difference between the original target set temperature of the current calculation cycle and the final target set temperature of the previous cycle; if the absolute value of the difference is greater than the preset maximum heating rate threshold, then performing peak-shaving processing on the original target set temperature and outputting the rate-smoothed value as the final target set temperature.

[0051] Specifically, the controller first reads in parallel the first, second, and third temperature compensation values ​​generated by the preceding steps, as well as the first, second, and third weight coefficients output by the weighted control model. The arithmetic unit performs a weighted summation operation: calculating the product of the first weight coefficient and the first temperature compensation value, the product of the second weight coefficient and the second temperature compensation value, and the product of the third weight coefficient and the third temperature compensation value, respectively. Then, the three product results are summed to obtain the total dynamic compensation amount. This value represents the optimal temperature adjustment range (e.g., +1.5℃ or -0.8℃) determined by the system based on the current complex coupled operating conditions (viscosity, power, time). Next, the total dynamic compensation amount is algebraically superimposed with the preset reference temperature in the process formula to generate the original target set temperature.

[0052] To prevent temperature setpoints from exceeding the limits of equipment or materials due to abnormal algorithm output, a safety range check is performed. The controller reads the pre-stored maximum allowable temperature (the upper limit to prevent material degradation) and minimum sustaining temperature (the lower limit to prevent reaction stagnation) of the polymerization process, and determines whether the original target set temperature falls within this safety range. If the original value is higher than the maximum allowable temperature, it is forcibly modified to the maximum allowable temperature; if the original value is lower than the minimum sustaining temperature, it is forcibly modified to the minimum sustaining temperature. This forced clamping mechanism ensures that no matter how the algorithm calculates, the final command will never violate the absolute safety boundaries of the chemical process.

[0053] To avoid thermal shock to the heating system or frequent oscillation of the regulating valve caused by drastic fluctuations in the temperature setpoint, a rate limiting logic is implemented. The controller reads the final target setpoint output from the previous control cycle and calculates the difference between it and the original target setpoint of the current cycle. The absolute value of this difference is compared with a preset maximum heating rate threshold (e.g., a change of no more than 2°C per minute). If the difference exceeds the threshold, the original target setpoint is smoothed out: that is, the previous cycle value + maximum rate threshold (when heating) or the previous cycle value - maximum rate threshold (when cooling) is used as the output. Finally, the value after this rate smoothing is output as the final target setpoint and sent to the PID controller or valve positioner to execute the specific heating / cooling operation.

[0054] By implementing multiple security verification and smoothing mechanisms, this application ensures the safety and stability of intelligent control commands during execution. The weighted synthesis step transforms the AI's decision-making logic into specific physical adjustment quantities, eliminating over-temperature accidents caused by calculation anomalies. Meanwhile, the rate smoothing process effectively filters out high-frequency jitter in control commands, avoiding thermal shock to the heat medium circulation system and extending the service life of actuators such as regulating valves.

[0055] Further, adjusting the opening of the heat medium regulating valve includes: using the final target set temperature as the set value of the main control loop, using the collected real-time temperature of the polymerization reactor as the feedback value of the main control loop, and calculating the jacket temperature set value through the main PID controller; using the jacket temperature set value as the set value of the secondary control loop, using the collected jacket inlet temperature as the feedback value of the secondary control loop, and calculating the standard control output signal through the secondary PID controller; retrieving a preset heat medium regulating valve flow characteristic curve, the characteristic curve representing the nonlinear relationship between valve opening and heat medium flow; using the inverse function model of the flow characteristic curve to perform reverse mapping on the standard control output signal, calculating the corrected linearized valve opening command; converting the linearized valve opening command into a current signal to drive the actuator of the heat medium regulating valve to actuate, so that the valve core position reaches the corresponding opening.

[0056] To address the temperature lag issue caused by the large volume of the polymerization reactor, a main control loop (outer loop) is first constructed. The controller uses the final target setpoint temperature calculated in the previous steps as the setpoint (SV) of the main loop and the real-time collected material temperature of the polymerization reactor as the feedback value (PV) of the main loop. Based on the deviation between the setpoint and the feedback value, the main PID controller outputs an intermediate variable through proportional, integral, and derivative operations. This intermediate variable does not directly drive the valve but represents the ideal jacket temperature required under the current operating conditions and is defined as the jacket temperature setpoint. Following this, a secondary control loop is constructed (… The inner loop controller uses the jacket temperature setpoint output from the main loop as the setpoint (SV) of the secondary loop and the real-time collected jacket inlet temperature as the feedback value (PV) of the secondary loop. The secondary PID controller performs rapid calculations based on the deviation between these two values ​​and outputs a standard control output signal. Since the response speed of the secondary loop is much faster than that of the main loop, it can quickly sense and reduce the interference caused by the pressure fluctuation of the heat medium main pipe or the temperature fluctuation of the heat medium source, ensuring that the energy of the heat medium entering the jacket is stable. This standard control output signal is usually expressed as a percentage from 0% to 100%, representing the linear demand of the system on the heat medium flow rate.

[0057] Given that heat transfer medium control valves in industrial settings typically exhibit nonlinear flow characteristics (such as equal percentage characteristics), meaning that changes in the valve's physical opening are not proportional to changes in flow rate, directly using PID output to drive the valve can lead to inconsistent control gain. Therefore, the flow characteristic curve of the control valve is pre-programmed into the storage unit. The controller then calls the inverse function model of this curve to perform a reverse mapping operation on the standard control output signal of the secondary PID controller. For example, when the PID controller requests 50% flow, the corresponding physical opening command calculated through the inverse function model might be 65% (the specific value depends on the valve characteristics). This step achieves theoretical linearization compensation at the software level. While ignoring drastic changes in the pipe friction coefficient, the inverse function model improves the inherent nonlinear flow characteristics of the control valve. This approach aims to ensure that the control system achieves relatively consistent regulation sensitivity across most of its operating range (especially in the originally highly nonlinear low-opening or high-opening regions), thereby reducing the difficulty of controller parameter tuning.

[0058] Ideally, the calculated, corrected, linearized valve opening command is converted into a standard 4-20mA analog current signal via a digital-to-analog converter (D / A). This current signal is transmitted to the electric valve positioner of the heat medium regulating valve. After receiving the signal, the positioner drives the pneumatic diaphragm or electric actuator to move, pushing the valve core to the corresponding physical opening position. In this way, the actual flow cross-sectional area of ​​the valve is precisely adjusted, thereby changing the mass flow rate of the heat medium flowing through the jacket, and ultimately achieving precise control of the polymerization reactor temperature.

[0059] By adopting a master-slave cascade control architecture, the source disturbance of the heat medium system is effectively isolated by utilizing the fast response characteristics of the slave loop, thus solving the control problem of objects with large time lag. The introduced linearization compensation mechanism based on the inverse function model eliminates the influence of the inherent nonlinear characteristics of the regulating valve on the control accuracy at the algorithm level. This strategy avoids valve oscillation at small openings and sluggish response at large openings, ensuring that the system can maintain a uniform and linear control response throughout the entire operating range, significantly improving the steady-state accuracy and dynamic tracking capability of temperature regulation.

[0060] By fusing heterogeneous data from multiple sources, such as viscosity, exothermic power, and residence time, an adaptive weighted control system based on deep learning was constructed. This system effectively solved the control challenges caused by strong coupling and large hysteresis of multiple physics fields in continuous polymerization reactions. By extracting time-series features using a sliding window and normalization processing, and dynamically adjusting the control weights of each deviation variable in conjunction with an attention mechanism, precise perception of the reaction stage was achieved. At the same time, by combining decoupling compensation of physical mechanisms and cascade execution strategies, shear heat and cumulative hysteresis effects were suppressed at the algorithm level, and valve nonlinear interference was eliminated at the execution level. This significantly improved the steady-state accuracy and dynamic response capability of polymerization temperature control, ensuring the safety of the production process and the consistency of product quality.

[0061] Example 2: To improve the temperature control effect in the continuous polymerization production line of polyphenylene sulfide, an adaptive temperature control method for the polyphenylene sulfide polymerization reactor is introduced in this application. The specific operation process of this control method when the continuous polymerization production line of polyphenylene sulfide enters the peak exothermic stage is as follows: the working condition characteristics in the reactor are: the melt viscosity increases sharply, and the heat released by the chemical reaction reaches its peak. The system is in a critical state with strong nonlinearity and high coupling.

[0062] During the peak of the exothermic period, significant changes in operating conditions were captured by the sensor array. The online torsional viscometer detected an exponential increase in the probe damping signal, and the converted real-time melt viscosity rapidly climbed to 5500 mPa·s (higher than the target value of 5000 mPa·s). At the same time, the armored platinum resistance temperature sensor detected that the jacket outlet temperature was significantly higher than the inlet temperature, and the temperature difference widened in a short period of time. The mass flow meter continuously uploaded the heat medium circulation flow data. At this time, although the liquid level height was basically stable as monitored by the differential pressure transmitter, due to the slight fluctuation of the feed flow rate, the residence time deviation calculated by differential analysis showed that the actual residence time of the material in the reactor was slightly prolonged.

[0063] The controller receives the raw data and immediately performs a standardization operation, calculating a viscosity deviation of +10% (i.e., (5500-5000) / 5000), indicating that the material consistency exceeds expectations. Based on the heat balance logic, the calculated real-time heat release power reaches 65kW, while the standard heat release power at this stage is only 50kW, resulting in a power deviation of +30%, clearly indicating that the reaction is too violent. At the same time, the actual residence time deviation obtained based on the liquid level change rate mapping is normalized to +2%, indicating a slight reaction lag. These normalized data not only eliminate dimensional differences but also transform complex physical phenomena into out-of-range signals that the model can read.

[0064] The three deviation indicators mentioned above are pushed into a sliding window queue of length 60 with a sampling step of 1 second. At this time, the data in the queue shows a clear upward trend (viscosity continues to rise) and high-level oscillation (power remains high). After physical limit normalization of the queue data, a multi-dimensional time-series sliding window input vector is generated. Subsequently, the weight control model intervenes: the dynamic trend encoder keenly captures the synchronous upward trend of power and viscosity, and identifies this as a typical superposition mode of shear heat and reaction heat; the static condition encoder generates the corresponding condition features based on the input reaction mid-term flag; the multi-source attention fusion unit automatically increases the attention weight of power deviation and viscosity deviation, because these two parameters have the greatest impact on safety during the peak of exothermic period; the decision output layer finally outputs a set of specific weight coefficients: greatly increasing the second weight coefficient (for power) and the first weight coefficient (for viscosity) to strengthen the suppression of heat.

[0065] The calculated shear heat correction term is positive and large, indicating that high-viscosity stirring generates a large amount of frictional heat, requiring additional cooling compensation. Simultaneously, the cumulative correction term also shows heat accumulation; when generating the second temperature compensation value, the controller subtracts a large cumulative correction term from the base power component, resulting in a significantly negative compensation value (e.g., -3.5℃), intending to remove excess heat by drastically reducing the jacket temperature. Finally, the reference temperature (set to 260℃) is algebraically superimposed with the calculated total dynamic compensation (-4.2℃), yielding the original target set temperature of 255.8℃. Verification shows that 255.8℃ is within the safe range allowed by the process (2...). Within the temperature range of 50℃-280℃, the rate limiting logic detected that the set value decreased too quickly compared to the previous cycle (exceeding the cooling rate limit of 2℃ / min). To prevent thermal shock, the system automatically performed peak shaving, smoothly adjusting the final target set temperature to 258.0℃. Based on the set value of 258.0℃, the main circuit commanded the secondary circuit to reduce the jacket temperature. After calculating the standard output signal, the secondary circuit, through inverse mapping of the valve flow characteristics, output a precise linearized valve opening command, driving the cold oil valve opening to increase to the precise position. As the flow rate of the low-temperature heat transfer medium increased, the excess heat in the polymerization reactor was rapidly removed, the reaction temperature steadily decreased, and the viscosity increase rate returned to normal, successfully avoiding a runaway temperature accident.

[0066] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for adaptive temperature control of a polyphenylene sulfide polymerization reactor, characterized in that, include: Collect the melt viscosity, liquid level, jacket inlet temperature, jacket outlet temperature, and heat medium circulation flow rate of the polymerization reactor; Calculate the viscosity deviation between the real-time melt viscosity and the target viscosity; calculate the real-time heat release power based on the jacket inlet temperature, jacket outlet temperature and heat medium circulation flow rate and obtain the power deviation relative to the standard heat release power; calculate the actual residence time deviation based on the rate of change of the real-time liquid level. First temperature compensation value, second temperature compensation value, and third temperature compensation value are generated based on viscosity deviation, power deviation, and actual residence time deviation, respectively; viscosity deviation, power deviation, and actual residence time deviation within a preset historical time period are extracted to construct a multi-dimensional time-series sliding window input vector; Input the multidimensional time-series sliding window input vector into the weight control model, and output the first weight coefficient, the second weight coefficient, and the third weight coefficient. The first temperature compensation value, the second temperature compensation value, and the third temperature compensation value are weighted and summed using the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient. The calculation result is superimposed on the reference temperature to obtain the final target set temperature. The opening degree of the heat medium regulating valve is adjusted according to the final target set temperature.

2. The method for adaptive temperature control of a polyphenylene sulfide polymerization reactor according to claim 1, characterized in that, The data collected from the polymerization reactor include melt viscosity, liquid level, jacket inlet temperature, jacket outlet temperature, and heat medium circulation flow rate. An online torsional viscometer is used to detect the vibration damping signal of the probe in the reactant. Using a preset viscosity-damping correspondence table, the vibration damping signal is mapped and converted into the melt viscosity. A dual-flange differential pressure transmitter is used to collect the real-time pressure difference between the bottom flange of the polymerization reactor and the gas phase space. Based on the principle of fluid statics, the real-time pressure difference is divided by the product of a preset melt density and gravitational acceleration to calculate the liquid level height. A sheathed platinum resistance temperature sensor is used to collect the resistance value signal that changes with temperature. Based on the resistance-temperature characteristic curve of the platinum resistance, the collected resistance value signal is converted into the jacket inlet temperature and the jacket outlet temperature. A mass flow meter is used to detect the phase difference signal generated by fluid flow in the measuring tube. The phase difference signal is processed through linear proportional conversion to obtain the heat medium circulation flow rate.

3. The method for adaptive temperature control of a polyphenylene sulfide polymerization reactor according to claim 1, characterized in that, The calculation process for viscosity deviation, power deviation, and actual residence time deviation includes: performing a numerical difference operation between the real-time melt viscosity and the target viscosity to obtain an absolute viscosity difference with positive and negative signs, and dividing the absolute viscosity difference by the target viscosity to obtain the normalized viscosity deviation; reading the pre-stored specific heat capacity parameter of the heat medium, calculating the temperature difference between the jacket outlet temperature and the jacket inlet temperature, and calculating the product of the temperature difference, the heat medium circulation flow rate, and the specific heat capacity of the heat medium to obtain the real-time heat release power at the current moment; comparing the real-time heat release power with the center value of the standard heat release power, and calculating the difference ratio between the two as the power deviation; performing a first-order difference operation on the continuously collected real-time liquid level height, calculating the difference in liquid level height between the current sampling moment and the previous sampling moment, and dividing by the sampling period to obtain the liquid level rise / fall rate per unit time; mapping the liquid level rise / fall rate to the residence time fluctuation value of the reactant based on the effective volume parameter of the polymerization reactor, and using the residence time fluctuation value as the actual residence time deviation.

4. The method for adaptive temperature control of a polyphenylene sulfide polymerization reactor according to claim 1, characterized in that, The process of generating the first temperature compensation value, the second temperature compensation value, and the third temperature compensation value includes: obtaining the preset coefficients required for calculation; reading preset viscosity ratio coefficient, power ratio coefficient, time ratio coefficient, as well as preset shear heat coupling coefficient and reaction cumulative coupling coefficient from the controller storage unit; multiplying the viscosity deviation by the viscosity ratio coefficient to obtain a viscosity basis component that only responds to viscosity changes; multiplying the power deviation by the power ratio coefficient to obtain a power basis component that only responds to exothermic changes; and multiplying the actual residence time deviation by the time ratio coefficient to obtain a time basis component that only responds to time changes. Calculate the product of the viscosity deviation and the power deviation, and multiply this product by the shear-thermal coupling coefficient to obtain a shear-thermal correction term for compensating for high-viscosity frictional heat generation; calculate the product of the power deviation and the actual residence time deviation, and multiply this product by the reaction cumulative coupling coefficient to obtain a cumulative correction term for compensating for reaction hysteresis effects; sum the viscosity basis component and the shear-thermal correction term to obtain the first temperature compensation value; subtract the cumulative correction term from the power basis component to obtain the second temperature compensation value; and directly assign the time basis component as the third temperature compensation value.

5. The method for adaptive temperature control of a polyphenylene sulfide polymerization reactor according to claim 1, characterized in that, The process of constructing the multidimensional temporal sliding window input vector includes: setting the temporal depth parameter and sampling step size parameter of the sliding window; First, second, and third FIFO queues are created in memory to store historical sequence data of the viscosity deviation, power deviation, and actual residence time deviation, respectively. According to the sampling step size, data from the current moment and the past N consecutive sampling moments are sequentially pushed into the corresponding FIFO queues, and the earliest data is removed, maintaining a fixed queue length. Preset physical limit amplitudes for each parameter are read, and max-min normalization operations are performed on all data in the first, second, and third FIFO queues to map all values ​​to a dimensionless range of zero to one. The normalized data from the first FIFO queue is extracted as the viscosity feature dimension; the normalized data from the second FIFO queue is extracted as the power feature dimension; and the normalized data from the third FIFO queue is extracted as the time feature dimension. The viscosity feature dimension, power feature dimension, and time feature dimension are concatenated end-to-end or stacked in parallel according to a preset channel order to generate the multidimensional time-series sliding window input vector.

6. The method for adaptive temperature control of a polyphenylene sulfide polymerization reactor according to claim 1, characterized in that, The weight control model includes: The dynamic trend encoder receives the multi-dimensional time-series sliding window input vector as input, processes it using a long short-term memory network, eliminates short-term random noise through internal forget gate and input gate mechanisms, captures the historical evolution patterns of viscosity deviation, power deviation and residence time deviation, and generates a dynamic trend feature vector containing dynamic time-varying information. The static condition encoder receives preset reference temperature, target viscosity and standard exothermic power as input, and processes them using a fully connected embedded network to map the physical process setting parameters into high-dimensional sparse semantic vectors, generating static condition feature vectors that characterize the current polymerization reaction stage. The multi-source attention fusion unit concatenates the static working condition feature vector with the dynamic trend feature vector, inputs the concatenated composite state vector into the attention generation network, generates attention weight vectors for different response parameters, applies the attention weight vectors to the dynamic trend feature vectors output by the dynamic trend encoder, performs dynamic weighting through a feature recalibration mechanism, and outputs global context-aware features. The adaptive decision output layer receives the global context-aware features as input, uses a multilayer perceptron to reduce the dimensionality of the high-dimensional features and perform linear combination, and maps the operation results to the numerical range of 0 to 1 through a sigmoid nonlinear activation function, and finally outputs the first weight coefficient, the second weight coefficient and the third weight coefficient in parallel.

7. The method for adaptive temperature control of a polyphenylene sulfide polymerization reactor according to claim 1, characterized in that, The process of obtaining the final target set temperature includes: calculating the product of the first weighting coefficient and the first temperature compensation value, the product of the second weighting coefficient and the second temperature compensation value, and the product of the third weighting coefficient and the third temperature compensation value, respectively, and adding the three product results to obtain the total dynamic compensation amount; algebraically superimposing the total dynamic compensation amount with the reference temperature to generate the original target set temperature; reading the preset maximum allowable temperature and minimum holding temperature of the polymerization process; determining whether the original target set temperature exceeds the safety range formed by the maximum allowable temperature and the minimum holding temperature; if it exceeds the safety range, forcibly clamping the original target set temperature to the corresponding boundary value; calculating the difference between the original target set temperature of the current calculation cycle and the final target set temperature of the previous cycle; if the absolute value of the difference is greater than the preset maximum heating rate threshold, performing peak-shaving processing on the original target set temperature and outputting the rate-smoothed value as the final target set temperature.

8. The method for adaptive temperature control of a polyphenylene sulfide polymerization reactor according to claim 1, characterized in that, The process of adjusting the opening of the heat medium regulating valve includes: using the final target set temperature as the set value of the main control loop, using the collected real-time temperature of the polymerization reactor as the feedback value of the main control loop, and calculating the jacket temperature set value through the main PID controller; using the jacket temperature set value as the set value of the secondary control loop, using the collected jacket inlet temperature as the feedback value of the secondary control loop, and calculating the standard control output signal through the secondary PID controller; retrieving a preset heat medium regulating valve flow characteristic curve, which represents the nonlinear relationship between valve opening and heat medium flow; using the inverse function model of the flow characteristic curve to perform reverse mapping on the standard control output signal, and calculating the corrected linearized valve opening command; converting the linearized valve opening command into a current signal to drive the actuator of the heat medium regulating valve to actuate, so that the valve core position reaches the corresponding opening.

Images