A control method and system for a carrier gas evaporator and condenser

Abstract

This invention belongs to the field of control technology, specifically relating to a control method and system for a carrier gas evaporative condenser. The method includes the following steps: acquiring historical and real-time operating parameters of the carrier gas evaporative condenser; constructing a process model consisting of a mechanistic sub-model and a neural network sub-model connected in parallel; calculating the difference between the actual equipment output and the process model output to obtain a model error signal; when the integral square value of the model error signal exceeds a threshold within a preset time window, updating the weight parameters of the neural network sub-model online using historical operating parameters based on a sliding time window regularized parameter estimation method, and recording the norm change of the weight parameters. This invention improves the control accuracy and operational stability of the carrier gas evaporative condenser while suppressing frequent actuator movements caused by high-frequency noise and reducing equipment wear.

Description

Technical Field

[0001] This invention belongs to the field of control technology, specifically relating to a control method and system for a carrier gas evaporator condenser. Background Technology

[0002] Carrier gas evaporator-condenser, as an important heat exchange device, is widely used in chemical, power, refrigeration, and new energy industries. Its operating efficiency and stability directly affect the energy consumption level and safe production of the process system. In practical industrial applications, traditional methods often struggle to establish high-precision control models. While relying solely on pure mechanistic models can characterize the steady-state operation of the equipment, the modeling process typically requires the introduction of idealized assumptions, making it difficult to accurately describe nonlinear behavior under complex operating conditions and prone to significant prediction biases. Furthermore, when operating conditions deviate from historical data distributions, significant model mismatch may occur. Therefore, constructing a high-fidelity process model that conforms to the physical laws of the equipment and can identify and compensate for nonlinear deviations is a crucial problem that urgently needs to be solved to achieve high-quality control.

[0003] Patent application CN113606817A discloses an intelligent control system for a deep-cooling evaporative condenser, comprising: a controller, which includes an input terminal, an output terminal, and a processing unit, the processing unit serving as the data processing hub and having built-in arithmetic logic; a temperature and humidity sensor, electrically connected to the controller input terminal, for acquiring external temperature and humidity information; a pressure sensor, electrically connected to the controller input terminal, for acquiring internal condensing pressure information; at least one condenser, serving as the controller output load, including a heat exchange device and an air inlet device and an air outlet device respectively connected to the heat exchange device; and a frequency converter, located between the controller and the condenser, for receiving controller signals and adjusting the condenser power according to the controller signals.

[0004] Conventional control strategies typically process feedback error signals holistically, lacking the ability to separate and specifically suppress interference across different frequency bands. This can easily make the controller overly sensitive to high-frequency measurement noise, leading to frequent fluctuations or violent movements in critical components such as cooling medium actuators and fans. This not only accelerates mechanical wear of the actuators but may also disrupt the operational stability of the heat transfer system. With equipment aging, scaling, and continuous changes in environmental conditions, the operating characteristics of carrier gas evaporative condensers exhibit significant time-varying characteristics. Traditional controllers with fixed control parameters and fixed prediction step sizes lack sufficient flexibility and struggle to balance response speed, anti-interference capability, and actuator stability. Therefore, a control method capable of separating multi-source interference and possessing adaptive adjustment capabilities is urgently needed to improve the operational stability and control accuracy of carrier gas evaporative condensers. Summary of the Invention

[0005] This invention provides a control method and system for a carrier gas evaporator condenser to solve the technical problems in the prior art, such as poor model accuracy and easy mismatch, frequent actuator operation caused by noise, and poor adaptability of fixed control parameters.

[0006] In a first aspect, the present invention provides a control method for a carrier gas evaporator-condenser, comprising the following steps: S1. Obtain historical and real-time operating parameters of the carrier gas evaporator condenser, construct a process model consisting of a mechanism sub-model and a neural network sub-model connected in parallel, calculate the difference between the actual equipment output and the process model output to obtain the model error signal, and when the integral square value of the model error signal exceeds the threshold within the preset time window, use the regularized parameter estimation method based on the sliding time window to update the weight parameters of the neural network sub-model online using historical operating parameters, and record the norm change of the weight parameters. S2, using multi-resolution wavelet transform to decompose the model error signal into low-frequency components representing quasi-steady-state disturbances and high-frequency components representing measurement noise, and estimating the time-varying variance of the high-frequency components in real time to construct an internal model controller, using the difference between the set target value and the low-frequency components of the model error signal as the reference input; S3, each control cycle adjusts the future prediction time domain length of the internal model controller according to the norm change of the weight parameters, and calculates the control output increment sequence online by minimizing the cost function. The weight coefficient used to penalize the rate of change of the control output in the cost function is determined according to the time-varying variance of the high-frequency components. The first term of the control output increment sequence is taken as the control command reference to generate the actual control commands for the cooling medium actuator and the fan actuator in the carrier gas evaporator condenser.

[0007] By relying on the parallel modeling of mechanism and neural network, the model can be adaptively corrected. Wavelet splitting is used to distinguish between steady-state disturbances and random noise. Then, based on the changes in network parameters and noise variance, the prediction time domain and control penalty coefficient are adaptively optimized. The model accuracy, interference suppression and control output are managed in layers. This can not only reduce the problem of frequent start-stop wear of actuators induced by high-frequency noise, but also dynamically adapt to the time-varying operating conditions of condenser, and comprehensively improve the accuracy of condensing temperature control and the stability of the whole machine operation.

[0008] Furthermore, the mechanistic sub-model is used to represent steady-state heat and mass transfer characteristics, while the neural network sub-model is used to compensate for the prediction bias of the mechanistic sub-model and identify nonlinear characteristics.

[0009] Furthermore, historical and real-time operating parameters of the carrier gas evaporator condenser are obtained to construct a process model, including: collecting inlet air temperature, inlet air flow rate, cooling medium flow rate, fan speed or air-side heat exchange air volume, and ambient temperature as input variables, and condensate outlet temperature as output variable; calculating the steady-state predicted condensate outlet temperature using the energy conservation equation and mass conservation equation as the output of the mechanism sub-model; collecting the actual condensate outlet temperature; calculating the first difference between the actual condensate outlet temperature and the output of the mechanism sub-model; using inlet air temperature, inlet air flow rate, cooling medium flow rate, fan speed or air-side heat exchange air volume, and ambient temperature as input variables, and the first difference as the target output; training a neural network using the gradient descent algorithm; using the trained neural network as the neural network sub-model; and adding the output of the mechanism sub-model to the output of the neural network sub-model to obtain the output of the process model.

[0010] By clearly defining multiple process parameters as model inputs and condensate outlet temperature as output, and constructing a mechanism model based on conservation equations, the measured residuals are used to train a neural network to compensate for errors. The network training is completed and implemented based on measured industrial data, which fits the actual heat exchange process parameters of the condenser. This effectively eliminates the system prediction bias caused by the simplification assumptions of the fixed mechanism model, and ensures that the parallel model fits the actual on-site operating rules.

[0011] Furthermore, the regularized parameter estimation method based on a sliding time window updates the weight parameters of the neural network sub-model online using historical operating parameters, including: [the method then adjusts the parameters based on the most recent data]. The input and output variables of each control cycle are combined to construct a data time window. A sliding time window regularized parameter estimation objective function is constructed. The data fitting term of the sliding time window regularized parameter estimation objective function is the sum of the squared differences between the actual condensate outlet temperature at each sampling time within the data time window and the sum of the outputs of the mechanistic sub-model and the neural network sub-model, plus the product of the preset penalty weight coefficient and the sum of the squared changes of the weight parameters to be updated within the total number of weight parameters.

[0012] A parameter update method using a sliding time window combined with regularization constraints is adopted. The network weights are updated using real-time historical data within the window, while a penalty term for parameter changes is added. This method not only relies on the latest working condition data to complete the online adaptive iteration of the model and improve the model mismatch problem caused by working condition drift, but also avoids excessive weight mutations that could lead to model overfitting through regularization constraints, thereby improving the stability of parameter updates.

[0013] Furthermore, the norm change of the weight parameters is recorded, including: obtaining the updated network parameter set by using an optimization algorithm to find the optimal solution of the objective function, calculating the set of differences between the updated network parameter set and the network parameter set before the update, expanding the set of differences into parameter vectors, calculating the Euclidean norm, and recording it as the norm change of the weights and bias parameters. The norm change is equal to the square root of the sum of the squares of the differences between each weight parameter and the differences between each bias parameter.

[0014] By calculating the magnitude of the Euclidean quantification model change of network parameters before and after the update, the parameter changes are quantified into monitorable numerical indicators, providing a precise quantitative basis for the subsequent dynamic adjustment of the controller in the time domain, realizing the linkage between the degree of model change and control parameters, and making the controller adjustment logic based on evidence.

[0015] Furthermore, multi-resolution wavelet transform is used to decompose the model error signal into low-frequency components representing quasi-steady-state disturbances and high-frequency components representing measurement noise, and the time-varying variance of the high-frequency components is estimated in real time, including: using Daubechies wavelets as basis functions to estimate the variance of the preceding components. The model error signal for each sampling period is subjected to a 3-level discrete wavelet decomposition, where The approximate coefficients of the output of the third layer decomposition are extracted to reconstruct the gradually changing signal feature components as low-frequency components. The detail coefficients generated by the first layer decomposition are extracted to form a high-frequency feature sequence. The median of the absolute value of the first layer detail coefficient sequence is calculated, and the median of the absolute value is divided by 0.6745 to obtain the noise standard deviation. The square of the noise standard deviation is used as the time-varying variance of the high-frequency component at this time.

[0016] The db wavelet is selected to complete the three-level discrete decomposition, and the low-frequency process disturbance and high-frequency measurement noise are separated in layers. The time-varying variance of noise is accurately estimated by using the absolute median difference method, so as to achieve accurate separation of disturbance types. On the one hand, the noise interference on the closed-loop control is eliminated, and on the other hand, the noise intensity is quantified in real time, providing accurate noise data support for the dynamic tuning of control cost weights.

[0017] Furthermore, in each control cycle, the future prediction time domain length of the internal model controller is adjusted according to the norm change of the weight parameters. This includes: calculating the ratio of the norm change of the weight parameters to a preset benchmark norm constant; multiplying the ratio by the proportional coefficient constant to obtain a product; rounding the product down to obtain a rounded value; subtracting the rounded value from the preset benchmark prediction time domain length to obtain a first result; comparing the first result with the prediction time domain lower limit constraint threshold; taking the maximum value of the two as the adjusted future prediction time domain length; and using the calculated adjusted future prediction time domain length as the prediction time window span value of the internal model controller in the current control cycle.

[0018] Furthermore, the control output increment sequence is calculated online by minimizing the cost function, including: constructing a cost function, which is the sum of the squares of the differences between the predicted output set of the process model at future times and the future target reference trajectory signal sequence over the future prediction time domain length, plus the product of the weight coefficient and the sum of the squares of the control output increment sequence over the control time domain length, wherein the control time domain length is not greater than the adjusted future prediction time domain length corresponding to the current control cycle, and the product of the preset system scaling coefficient and the time-varying variance of the high-frequency component is used as the weight coefficient in the cost function, solving for the control output increment sequence that satisfies the optimal condition of quadratic programming constraints, and extracting the first term in the control output increment sequence as the control command benchmark at the current time.

[0019] Furthermore, the first term of the control output increment sequence is taken as the control command reference to generate actual control commands for the cooling medium actuator and the fan actuator in the carrier gas evaporator condenser. This includes: adding the control command reference to the total cooling demand control quantity at the previous moment to obtain the total cooling demand control quantity at the current moment; using a preset 2×1 allocation column vector to decompose the total cooling demand control quantity into the cooling water regulation target quantity and the fan speed regulation target quantity; converting the cooling water regulation target quantity into a standard current signal of 4 to 20 mA within the range through linear mapping and outputting it to the cooling medium actuator to adjust the valve opening; and converting the fan speed regulation target quantity into a standard DC voltage signal of 0 to 10 V within the range through proportional conversion and outputting it to the frequency converter to control the speed of the fan actuator.

[0020] Secondly, the present invention provides a control system for a carrier gas evaporator condenser, including a memory and a processor. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned control method for the carrier gas evaporator condenser is implemented.

[0021] The beneficial effects are as follows: This invention constructs a process model composed of a mechanistic sub-model and a neural network sub-model in parallel. It fully combines the ability of the mechanistic model to characterize steady-state heat and mass transfer laws with the advantages of the neural network in compensating for prediction errors and identifying nonlinear operating characteristics. Furthermore, it utilizes a sliding time window regularized parameter estimation method to update model parameters online, thereby improving the model's prediction accuracy and long-term adaptability under complex operating conditions. Multi-resolution wavelet transform is used to decompose model errors, separating high-frequency measurement noise from low-frequency quasi-steady-state disturbances. The low-frequency quasi-steady-state disturbance component is used to correct the reference input of the internal model controller, enhancing the control system's anti-interference capability. The future prediction time domain length is adjusted online based on the change in network weight parameters, and the penalty weight for the control output change rate is determined based on the time-varying variance of the high-frequency components, achieving adaptive optimization of control commands. While suppressing frequent actuator actions caused by high-frequency noise and reducing equipment wear, this invention improves the control accuracy and operational stability of the carrier gas evaporator condenser. Attached Figure Description

[0022] Figure 1 This is a flowchart of the control method for a carrier gas evaporator-condenser.

[0023] Figure 2 A schematic diagram illustrating the extraction of model error signal feature components using multi-resolution wavelet transform.

[0024] Figure 3 This is a schematic diagram comparing the condensate outlet temperature response under a step change in flow rate. Detailed Implementation

[0025] 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, not all, of the embodiments of the present invention. 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.

[0026] An embodiment of the control method for a carrier gas evaporator condenser provided by the present invention: like Figure 1 As shown, the control method for the carrier gas evaporator-condenser includes the following steps: S1, dual-model parallel modeling, online update of network weights when exceeding threshold.

[0027] Historical and real-time operating parameters of the carrier gas evaporator-condenser are obtained, and a process model consisting of a mechanistic sub-model and a neural network sub-model in parallel is constructed. The mechanistic sub-model is used to represent the steady-state heat and mass transfer characteristics, while the neural network sub-model is used to compensate for the prediction bias of the mechanistic sub-model and identify nonlinear characteristics. The difference between the actual equipment output and the process model output is calculated to obtain the model error signal. When the integral square value of the model error signal exceeds the threshold within a preset time window, the weight parameters of the neural network sub-model are updated online using the historical operating parameters based on the regularized parameter estimation method of the sliding time window, and the norm change of the weight parameters is recorded.

[0028] Historical and real-time operating parameters of the carrier gas evaporator-condenser, such as temperature, flow rate, and actuator status, are acquired. Historical CSV data files are read, and real-time sensor data from the programmable logic controller (PLC) is acquired using the PyModbus library. A process model is constructed by combining a mechanistic sub-model and a neural network sub-model. The mechanistic sub-model establishes partial differential equations for heat and mass transfer based on the first law of thermodynamics and the law of conservation of mass, and uses the fsolve function to solve the spatially discretized steady-state algebraic equations to represent the steady-state heat and mass transfer characteristics. The neural network sub-model is built using a feedforward backpropagation neural network. Based on the steady-state condensate outlet predicted temperature output by the steady-state algebraic equations, first-order inertial elements and pure time-delay elements are introduced, or the condensate outlet temperature, cooling medium flow rate, and fan speed of the previous control cycle are introduced as dynamic state variables to construct a dynamic process model capable of rolling prediction for multiple future control cycles.

[0029] A feedforward neural network consisting of an input layer, hidden layers, and an output layer is constructed using the Linear module to compensate for prediction biases in the mechanistic sub-model and identify nonlinear characteristics. The arithmetic logic unit of the microprocessor calculates the difference between the real-time acquired actual equipment output temperature and the predicted temperature output in parallel by the process model, obtaining the model error signal. The time window length is set to ten sampling periods. The trapezoidal numerical integration algorithm is used to solve for the definite integral of the squared error signal within this time window. When the squared integral exceeds a set empirical constant threshold, a sliding time window regularization parameter update mechanism is triggered. An objective function containing the sum of squared prediction errors within the historical window and a penalty term for weight changes is constructed. The CasADi library, in conjunction with the IPOPT nonlinear optimization solver, is used to solve the optimization problem online, thereby updating the weight matrix parameters of each layer of the neural network sub-model. The updated network parameter set is subtracted from the original network parameter set to obtain the parameter difference set. This parameter difference set is then expanded into a parameter vector, and the Euclidean norm is calculated. This scalar is recorded as the norm change of the weights and bias parameters in a memory data list.

[0030] As an optional implementation, historical and real-time operating parameters of the carrier gas evaporator-condenser are obtained, and a process model consisting of a mechanistic sub-model and a neural network sub-model connected in parallel is constructed. The mechanistic sub-model is used to represent the steady-state heat and mass transfer characteristics, and the neural network sub-model is used to compensate for the prediction bias of the mechanistic sub-model and identify nonlinear characteristics, including: The system collects intake air temperature, intake air flow rate, cooling medium flow rate, fan speed or air-side heat exchange air volume, and ambient temperature as input variables, and condensate outlet temperature as output variable. It calculates the predicted condensate outlet temperature under steady-state conditions using the energy conservation equation and the mass conservation equation, and uses this as the output of the mechanistic sub-model. It also collects the actual condensate outlet temperature and calculates the first difference between the actual condensate outlet temperature and the output of the mechanistic sub-model. Using intake air temperature, intake air flow rate, cooling medium flow rate, fan speed or air-side heat exchange air volume, and ambient temperature as input variables, and the first difference as the target output, it trains a neural network using the gradient descent algorithm. The trained neural network is then used as the neural network sub-model. Finally, the output of the process model is obtained by adding the output of the mechanistic sub-model to the output of the neural network sub-model.

[0031] During model construction, input variables are collected in real time through sensors of the distributed control system. The operating range and examples are as follows: inlet air temperature 150℃ to 250℃, example input value 200℃; inlet air flow rate 500 to 1500 m³ / h, example 1000 m³ / h; cooling medium flow rate 10 to 50 t / h, example 30 t / h; fan speed 300 to 1500 r / min, example 900 r / min; or air-side heat exchange air volume 5000 to 20000 m³ / h, example 12000 m³ / h; and ambient temperature -10℃ to 40℃, example 25℃. The condensate outlet temperature is used as the target control variable, with a range of 40℃ to 80℃, example 50℃. Based on the steady-state energy and mass conservation equations of heat transfer, a mechanism sub-model is constructed using formulas such as the logarithmic mean temperature difference method. The predicted outlet temperature under steady-state conditions is calculated, and the actual measured outlet temperature is subtracted from the predicted value to obtain the first difference error sequence representing the unmodeled condition.

[0032] In the error compensation stage, a feedforward backpropagation neural network is used as the neural network sub-model. The input layer of this network structure contains five variable nodes, receiving the intake air temperature. Intake flow rate Cooling medium flow rate Fan speed or air-side heat exchange air volume and ambient temperature The input vector X consists of the five variables mentioned above. The number of hidden layer nodes is set to 15, and the ReLU activation function is used. The input vector X is left-multiplied by the weight matrix from the input layer to the hidden layer. And add the bias vector of the hidden layer Then, the hidden layer output vector H is obtained by processing with the ReLU activation function. The output layer is a single node, namely the predicted temperature difference Y. The hidden layer output vector H is left-multiplied by the weight vector from the hidden layer to the output layer. And add the output layer bias The output Y of the output layer is obtained. The Adam gradient descent optimization algorithm is used for training, with an initial learning rate of 0.005, a batch size of 64, and a maximum number of training iterations of 1000. The output values ​​of the mechanistic sub-model and the neural network sub-model are added together to obtain the output value of the parallel process model.

[0033] As an optional implementation, the sliding time window-based regularized parameter estimation method updates the weight parameters of the neural network sub-model online using historical operating parameters and records the norm change of the weight parameters, including: Recently The input and output variables of each control cycle are combined to construct a data time window. A sliding time window regularized parameter estimation objective function is constructed. The data fitting term of the sliding time window regularized parameter estimation objective function is the sum of the squared differences between the actual condensate outlet temperature at each sampling time within the data time window and the sum of the outputs of the mechanistic sub-model and the neural network sub-model. This sum is added to the product of the preset penalty weight coefficient and the sum of the squared changes of the weight parameters to be updated within the total number of weight parameters. An optimization algorithm is used to find the optimal solution of the objective function to obtain the updated network parameter set. The difference set between the updated network parameter set and the network parameter set before the update is calculated. The difference set is expanded into a parameter vector and the Euclidean norm is calculated and recorded as the norm change of the weights and bias parameters. The norm change is equal to the square root of the sum of the squared differences of the weight parameters and the differences of the bias parameters.

[0034] The control system operates with a fixed control cycle, such as 2 seconds, storing historical operating parameters in a first-in-first-out (FIFO) circular queue, and retrieving the most recent parameters. A historical sliding data time window spanning 200 seconds is constructed using data from 100 sampling periods. A moving-time domain estimation objective function, including a data fitting term and a parameter regularization term, is constructed for this time window. The fitting term is the sum of squares of the differences between the actual condensate outlet temperature and the temperature predicted by the parallel process model at 100 time points; the regularization term consists of the sum of squares of the changes in network weights multiplied by a preset penalty weight coefficient of 0.05. In the objective function solution and state recording stages, a sequential quadratic programming algorithm or the Levenberg-Marquardt algorithm is used as the optimizer to obtain the updated network weight parameter matrix that minimizes the objective function within a set convergence accuracy (e.g., 1e-4) and maximum number of iterations (e.g., 50). After completing the online parameter replacement, the difference matrix is ​​obtained by subtracting the old weight matrix from the previous control period using the new weight matrix. Following a feedforward neural network structure with 5 nodes in the input layer, 15 nodes in the hidden layer, and 1 node in the output layer, the sum of squares of the weight and bias differences at each layer is calculated. The square root of this sum is then taken to determine the norm change of the weight and bias parameters. This norm value is extracted as a numerical indicator representing the current network model's adaptability to unknown conditions; for example, a result of 0.12 represents the norm change of the weight parameters.

[0035] S2, wavelet decomposition error component, is the input of the low-frequency tuned inner model.

[0036] The model error signal is decomposed into a low-frequency component representing quasi-steady-state disturbance and a high-frequency component representing measurement noise by using multi-resolution wavelet transform. The time-varying variance of the high-frequency component is estimated in real time, and an internal model controller is constructed. The difference between the set target value and the low-frequency component of the model error signal is used as the reference input.

[0037] One-dimensional discrete multi-resolution wavelet decomposition was performed on the continuously acquired model error signal sequence. The db4 wavelet from the Daubechies wavelet family was selected as the mother wavelet, and the decomposition level was set to three levels. The inverse wavelet transform function was called to reconstruct the low-frequency components representing quasi-steady-state disturbances from the single branch of the third-level approximation coefficients. The detail coefficients generated by the first-level decomposition were extracted to form a high-frequency feature sequence representing measurement noise. For the separated high-frequency feature sequence, the time-varying variance was estimated in real time using the absolute median difference estimation method. The absolute median of the first-level detail coefficient sequence was calculated within each sliding window, and the absolute median was divided by 0.6745 to obtain the noise standard deviation. The noise standard deviation was then squared to obtain the time-varying variance of the high-frequency component corresponding to the current window. An internal model controller was constructed in the industrial control computer. The desired condensation temperature required by the system process was used as the target value. The difference between the target value and the low-frequency component was calculated by executing a subtraction instruction. After eliminating the steady-state disturbance, this difference was used as the reference input of the internal model controller and passed to the downstream optimization calculation module.

[0038] As an optional implementation, multi-resolution wavelet transform is used to decompose the model error signal into a low-frequency component representing quasi-steady-state disturbance and a high-frequency component representing measurement noise, and the time-varying variance of the high-frequency component is estimated in real time, including: Using Daubechies wavelets as basis functions for the preceding The model error signal for each sampling period is subjected to a 3-level discrete wavelet decomposition, where The approximate coefficients of the output of the third layer decomposition are extracted to reconstruct the gradually changing signal feature components as low-frequency components. The detail coefficients generated by the first layer decomposition are extracted to form a high-frequency feature sequence. The median of the absolute value of the first layer detail coefficient sequence is calculated, and the median of the absolute value is divided by 0.6745 to obtain the noise standard deviation. The square of the noise standard deviation is used as the time-varying variance of the high-frequency component at this time.

[0039] To separate frequency band features from the coupled error signal, pre-acquisition and buffering are performed. The model prediction error signal with 256 sampling periods is used to form a signal vector of length 256. The Daubechies4 wavelet is selected as the basis function, and the vector is subjected to a three-level discrete wavelet decomposition using the standard Mallat algorithm. With a control period of 2 seconds, a sampling frequency of 0.5Hz, and a Nyquist frequency of 0.25Hz, after the three-level decomposition, the third-level approximation coefficients mainly retain low-frequency gradual transformation components below approximately 0.03125Hz, with the specific frequency band boundaries affected by the wavelet filter transition band. These approximation coefficients are extracted and reconstructed into a time-domain waveform using inverse wavelet transform, yielding a quasi-steady-state disturbance signal representing extremely low-frequency, smooth-behavior disturbances such as surface scaling and slow ambient temperature drift—the low-frequency component.

[0040] The detail coefficients generated from the first-level decomposition are extracted as a high-frequency feature sequence. This sequence corresponds to the higher frequency band within the resolvable frequency band under the current sampling conditions, mainly located in the range of approximately 0.125Hz to 0.25Hz, rather than the frequency band above 0.5Hz. It includes measurement noise such as sensor electrical noise and random fluctuations caused by local fluid turbulence. Variance estimation is performed using the absolute median method: After the original signal of length 256 undergoes the first-level binary downsampling decomposition, the number of first-level detail coefficients is 128. The median of the absolute values ​​of these 128 first-level detail coefficients is calculated. Assuming the measurement result is 0.05℃, this median is divided by the Gaussian white noise correction factor of 0.6745, yielding a noise standard deviation of approximately 0.0741℃. Squaring this standard deviation yields a time-varying variance of 0.0055℃². This variance parameter represents the degree of high-frequency interference of the sensor under the current operating conditions. Figure 2As shown, the original error signal is split into a low-frequency slowly varying disturbance component and a high-frequency random noise component, thus realizing the separation of signals in different frequency bands. Based on the separated high-frequency component, the time-varying variance of noise can be calculated.

[0041] S3, based on weighted variance, adaptively tunes the prediction time domain and control quantity.

[0042] In each control cycle, the future prediction time domain length of the internal model controller is adjusted according to the norm change of the weight parameters, and the control output increment sequence is calculated online by minimizing the cost function. The weight coefficients used to penalize the rate of change of the control output in the cost function are determined according to the time-varying variance of the high-frequency components. The first term of the control output increment sequence is taken as the control command reference to generate the actual control commands for the cooling medium actuator and the fan actuator in the carrier gas evaporator condenser.

[0043] Within each control cycle, the norm change of the weight parameters recorded in the previous time step is read from memory. A negative correlation linear mapping algorithm is used to calculate the adjustment step size in the prediction time domain. When the norm change increases, the prediction time domain length is shortened to improve the control response speed under rapidly changing model conditions and reduce the accumulation of long-term prediction errors. If the current control cycle does not trigger online updates of the weight parameters of the neural network sub-model, the norm change of the weight parameters in the current control cycle is set to 0, or the norm change of the previous control cycle is exponentially decayed according to a preset decay coefficient, so that the future prediction time domain length gradually recovers to the baseline prediction time domain length after the process model re-stabilizes. The linear mapping result is rounded down to update the integer parameter of the future prediction time domain length in the internal model controller's predictive control algorithm module. A quadratic programming type cost function is constructed internally by the controller. This cost function includes the sum of squares of the deviations between the reference input and the predicted output in the future prediction time domain, as well as a penalty term for the square of the control output increment. The time-varying variance of the high-frequency components updated in real time is extracted, and the weighting coefficient used to penalize the rate of change of the control output in the cost function is calculated by multiplying it by a fixed scaling factor. This makes the weighting coefficient larger when the noise variance increases, so as to suppress the high-frequency violent oscillation of the control action.

[0044] A quadratic programming problem with actuator amplitude constraints and control increment constraints is established using the CVXPY library. An online minimization solution is then performed using a quadratic programming solver to obtain a sequence of control output increments for several future cycles. The first element of this increment sequence is extracted as the control command reference. This reference is then added to the total cooling demand control value from the previous time step to obtain the current total cooling demand control value. Before issuing the obtained control command, a hardware-level limiting program restricts it to within the actuator's allowable range. The asyncua asynchronous communication library function, based on the OPC standard protocol, is called to decompose the current total cooling demand control value into a preset 2×1 allocation column vector into a cooling water regulation target value and a fan speed regulation target value. These are then converted into communication messages and written to the opening command register of the electric regulating valve and the frequency control register of the inverter, thereby driving the cooling medium actuator and fan actuator in the carrier gas evaporator condenser.

[0045] As an optional implementation, the future prediction time domain length of the internal model controller is adjusted each control cycle based on the norm change of the weight parameters, including: The ratio of the norm change of the weight parameters to the preset benchmark norm constant is calculated. The ratio is multiplied by the proportional coefficient constant to obtain the product. The product is rounded down to obtain the rounded value. The preset benchmark prediction time domain length is subtracted from the rounded value to obtain the first result. The first result is compared with the prediction time domain lower limit constraint threshold. The maximum value of the two is taken as the adjusted future prediction time domain length. The calculated adjusted future prediction time domain length is used as the prediction time window span value of the internal model controller in the current control cycle.

[0046] The step size of the control algorithm is adjusted inversely by the fluctuation of the model's own parameters. Key scheduling parameters are set in the controller: the baseline norm constant (representing the steady-state neural network parameter update amplitude) is set to 0.05, the proportional gain constant is set to 20, the baseline prediction time domain length is set to 40 control cycles (80s), and the prediction time domain lower limit constraint threshold is set to 15 cycles (30s) to ensure the optimization space of the internal model controller. When the norm change of the neural network weight parameters at the current moment is obtained, a scaling transformation calculation is performed. For example, when encountering a change in operating conditions or a sudden change in feed composition leading to model mismatch, the network updates to adapt to the new operating conditions, and the obtained norm change is 0.12. Dividing the norm change of 0.12 by the baseline norm constant 0.05 yields a ratio of 2.4. Multiplying 2.4 by the proportional gain 20 yields a product of 48, which is rounded down to 48. Subtracting 48 from the baseline time domain length of 40 yields a first result of -8. The value of -8 is compared with the lower limit constraint threshold of 15, and the larger of the two, 15, is taken. Therefore, the prediction time window span is shortened from the baseline of 40 and limited to 15 periods. This adjustment logic allows for a shorter prediction time domain to improve system response speed and avoid control oscillations caused by accumulated prediction errors when the process model is in a rapid update state; when the model tends to converge and stabilize, such as when the norm change is only 0.01, it can automatically recover to a longer time domain to achieve high-precision stable control.

[0047] As an optional implementation, the control output increment sequence is computed online by minimizing the cost function, including: A cost function is constructed, which is the sum of the squares of the differences between the predicted output set of the process model at future times and the future target reference trajectory signal sequence over the future prediction time domain length, plus the product of the weight coefficients and the sum of the squares of the control output increment sequence over the control time domain length. The control time domain length is no greater than the adjusted future prediction time domain length corresponding to the current control cycle. The product of the preset system scaling coefficient and the time-varying variance of the high-frequency components is used as the weight coefficients in the cost function. The optimal control output increment sequence that satisfies the quadratic programming constraint is solved, and the first term in the control output increment sequence is extracted as the control command benchmark at the current time.

[0048] The cost function is constructed based on the standard model predictive control architecture, combining trajectory tracking deviation and control energy penalty. The prediction time domain length is an adjusted value, such as 15 to 40 cycles, while the control time domain length is configured to 5 control cycles, satisfying the constraint that it is no greater than the adjusted future prediction time domain length corresponding to the current control cycle, thus reducing the computational dimensionality of online optimization. To allocate control weights according to the degree of change in the measurement environment, the time-varying variance of the high-frequency components estimated by the front-end module (e.g., 0.0055℃²) is multiplied by a system scaling factor for dimension unification and amplitude amplification, typically set between 500 and 1000 (e.g., 500), resulting in a product of 2.75. This product, 2.75, is assigned as the weighting coefficient of the control output increment penalty term. This mechanism ensures that as measurement noise increases and variance grows, the penalty weight increases synchronously, reducing the amplitude of control quantity changes and suppressing frequent wear of the actuator caused by tracking high-frequency noise. At the mathematical solution execution level, the predicted output equation is substituted into the cost function, which is then transformed into a standard quadratic programming problem. By calling the quadratic programming solver, under constraints on control output increment, actuator amplitude, and upper and lower limits of total cooling demand control, a control output increment sequence satisfying the optimal conditions of the quadratic programming constraints is solved. Within one control cycle of 2 seconds, a control output increment sequence containing five future increment elements—+2.5%, +1.0%, -0.5%, 0%, and 0%—is obtained. After solving the sequence, following the moving-time window principle, subsequent predicted actions are discarded, and only the first term of the sequence is extracted as the baseline increment of the control command applied at the current moment.

[0049] As an optional implementation, the first term of the control output increment sequence is taken as the control command reference to generate actual control commands for the cooling medium actuator and the fan actuator in the carrier gas evaporator condenser, including: The total cooling demand control quantity at the current moment is obtained by adding the control command reference with the total cooling demand control quantity at the previous moment. The total cooling demand control quantity is decomposed into the cooling water regulation target quantity and the fan speed regulation target quantity using a preset 2×1 allocation column vector. The cooling water regulation target quantity is linearly mapped into a standard current signal of 4 to 20 mA within the range and output to the cooling medium actuator to adjust the valve opening. The fan speed regulation target quantity is proportionally converted into a standard DC voltage signal of 0 to 10 V within the range and output to the frequency converter to control the speed of the fan actuator.

[0050] The dimensionless control increment calculated by the algorithm is converted into an analog electrical signal recognizable by the industrial field actuators. Assuming the extracted current-moment control command baseline increment is +2.5%, and the total cooling demand control quantity recorded in the buffer at the previous moment is 60.0%, the arithmetic sum of the two yields a current-moment total percentage cooling demand control quantity of 62.5%. To coordinate the water-cooling and air-cooling systems, the distributed control system pre-configures a fixed-ratio 2×1 allocation column vector, for example, a transpose matrix configured with 0.7 and 0.3. Through vector multiplication, the 62.5% total demand is decoupled into two parallel targets: 70% is allocated to the water-cooling end, resulting in a cooling water regulation target of 43.75%; the remaining 30% is allocated to the air-cooling end, resulting in a fan speed regulation target of 18.75%. In the electrical signal conversion and output drive stage, the controller's digital-to-analog conversion module performs signal range matching. For the cooling water regulating loop, using the industry-standard 4-20 mA current loop as a reference, 43.75% of the target quantity is interpolated and mapped, resulting in an output current value of approximately 11.0 mA. This standard current signal is transmitted via an isolation safety barrier to the positioner of the electric or pneumatic cooling water regulating valve, driving the valve to open to the corresponding stroke. For the fan control loop, based on a standard DC voltage protocol of 0-10 V, 18.75% of the target quantity is proportionally converted to obtain a standard DC voltage of approximately 1.875 V. This control voltage is directly connected to the analog input of the fan frequency converter, instructing the frequency converter to output AC power of the matching frequency to drive the condenser fan at a controlled speed. Throughout the process, hardware-level limiting programs restrict out-of-range commands, achieving multi-variable low-level execution control.

[0051] This comparative experiment was conducted on a semi-simulation platform for a carrier gas evaporator-condenser. The inlet gas temperature was set to fluctuate between 150 and 250°C, the inlet gas flow rate was set to vary from 500 to 1500 cubic meters per hour, the ambient temperature was set to a randomly fluctuating baseline of 25°C, and the target condensate outlet temperature was set to 50°C. The experiment was divided into a control group and an experimental group. The control group used a traditional model predictive control algorithm with a fixed time domain length and fixed penalty weights. The experimental group used the complete control mechanism of this invention, which includes a parallel model of mechanism and neural network, adjusts the prediction time domain based on weight norm, and updates the control weights based on time-varying variance. Both algorithms ran continuously for 500 control cycles under the same hardware and disturbance conditions, with the sampling control cycle uniformly set to 2 seconds.

[0052] Under conditions of a sudden change in intake airflow, data recordings showed that the control group experienced a maximum overshoot of 6.5℃ in the condensate outlet temperature, a temperature stabilization time of 125s, and a steady-state error oscillating between ±1.2℃. The actuator variance of the cooling water regulating valve was 45.2. In contrast, under the same sudden disturbance, the experimental group reduced the maximum overshoot to 2.1℃, and the stabilization time to the target temperature of 50℃ was only 48s, suppressing the steady-state error to within ±0.3℃. Regarding actuator losses, the experimental group showed a cooling water regulating valve variance of only 12.6, smoother fan speed changes, and a decrease in the amplitude of the first term fluctuation in the control increment sequence for both types of actuators. Figure 3 As shown in the figure, the control group exhibited overshoot and prolonged oscillation after the disturbance, while the experimental group showed reduced overshoot and faster recovery speed in its temperature curve, verifying the advantages of this invention in suppressing overshoot and shortening the settling time.

[0053] A parallel model based on mechanistic principles and neural networks compensates for unmodeled errors, improving the accuracy of temperature prediction. A mechanism that adjusts the prediction time-domain length based on changes in the norm of the weight parameters rapidly shortens the step size during sudden changes in operating conditions, accelerating the system response and reducing control oscillations caused by accumulated prediction errors. Using wavelet transform to separate high-frequency noise and real-time estimation of time-varying variance to adjust the control cost weight coefficients allows the experimental group to increase control energy penalties when encountering measurement noise interference, reducing actuator action variance by 72.1%. This achieves both high-precision temperature trajectory tracking and reduced mechanical wear on the equipment.

[0054] An embodiment of the control system for the carrier gas evaporator condenser provided by the present invention: The control system of the carrier gas evaporator condenser includes a processor and a memory. The memory stores computer program instructions, which are executed by the processor to implement the above-described control method for the carrier gas evaporator condenser.

[0055] The control system of the carrier gas evaporator condenser also includes other components well known to those skilled in the art, such as communication interfaces. Their settings and functions are known in the art and will not be described in detail here.

[0056] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A control method for a carrier gas evaporator-condenser, characterized in that, Includes the following steps: S1. Obtain historical and real-time operating parameters of the carrier gas evaporator condenser, construct a process model consisting of a mechanism sub-model and a neural network sub-model connected in parallel, calculate the difference between the actual equipment output and the process model output to obtain the model error signal, and when the integral square value of the model error signal exceeds the threshold within the preset time window, use the regularized parameter estimation method based on the sliding time window to update the weight parameters of the neural network sub-model online using historical operating parameters, and record the norm change of the weight parameters. S2, using multi-resolution wavelet transform to decompose the model error signal into low-frequency components representing quasi-steady-state disturbances and high-frequency components representing measurement noise, and estimating the time-varying variance of the high-frequency components in real time to construct an internal model controller, using the difference between the set target value and the low-frequency components of the model error signal as the reference input; S3, each control cycle adjusts the future prediction time domain length of the internal model controller according to the norm change of the weight parameters, including: calculating the ratio of the norm change of the weight parameters to the preset benchmark norm constant, multiplying the ratio by the proportional coefficient constant to obtain the product, rounding the product down to obtain the rounded value, subtracting the rounded value from the preset benchmark prediction time domain length to obtain the first result, comparing the first result with the prediction time domain lower limit constraint threshold, taking the maximum value of the two as the adjusted future prediction time domain length, and taking the calculated adjusted future prediction time domain length as the prediction time window span value of the internal model controller in the current control cycle; and calculating the control output increment sequence online by minimizing the cost function. The weight coefficient used to penalize the rate of change of control output in the cost function is determined according to the time-varying variance of the high-frequency components. The first term of the control output increment sequence is taken as the control command benchmark, and the actual control commands for the cooling medium actuator and the fan actuator in the carrier gas evaporator condenser are generated.

2. The control method for the carrier gas evaporator condenser according to claim 1, characterized in that, The mechanistic sub-model is used to represent steady-state heat and mass transfer characteristics, while the neural network sub-model is used to compensate for the prediction bias of the mechanistic sub-model and identify nonlinear characteristics.

3. The control method for the carrier gas evaporator condenser according to claim 2, characterized in that, Historical and real-time operating parameters of the carrier gas evaporator condenser are obtained to construct a process model, including: collecting inlet air temperature, inlet air flow rate, cooling medium flow rate, fan speed or air-side heat exchange air volume, and ambient temperature as input variables, and condensate outlet temperature as output variable; calculating the steady-state predicted condensate outlet temperature using the energy conservation equation and mass conservation equation as the output of the mechanism sub-model; collecting the actual condensate outlet temperature; calculating the first difference between the actual condensate outlet temperature and the output of the mechanism sub-model; using inlet air temperature, inlet air flow rate, cooling medium flow rate, fan speed or air-side heat exchange air volume, and ambient temperature as input variables, and the first difference as the target output; training a neural network using the gradient descent algorithm; using the trained neural network as the neural network sub-model; and adding the output of the mechanism sub-model to the output of the neural network sub-model to obtain the output of the process model.

4. The control method for the carrier gas evaporator condenser according to claim 3, characterized in that, The regularized parameter estimation method based on sliding time windows updates the weight parameters of the neural network sub-model online using historical operating parameters, including: […]. The input and output variables of each control cycle are combined to construct a data time window. A sliding time window regularized parameter estimation objective function is constructed. The data fitting term of the sliding time window regularized parameter estimation objective function is the sum of the squared differences between the actual condensate outlet temperature at each sampling time within the data time window and the sum of the outputs of the mechanistic sub-model and the neural network sub-model, plus the product of the preset penalty weight coefficient and the sum of the squared changes of the weight parameters to be updated within the total number of weight parameters.

5. The control method for the carrier gas evaporator condenser according to claim 4, characterized in that, Record the norm change of the weight parameters, including: using an optimization algorithm to find the optimal solution of the objective function to obtain the updated network parameter set, calculating the set of differences between the updated network parameter set and the network parameter set before the update, expanding the set of differences into parameter vectors and calculating the Euclidean norm, and recording it as the norm change of the weight and bias parameters. The norm change is equal to the square root of the sum of the squares of the differences between each weight parameter and the differences between each bias parameter.

6. The control method for a carrier gas evaporator-condenser according to claim 1, characterized in that, Multi-resolution wavelet transform is used to decompose the model error signal into low-frequency components representing quasi-steady-state disturbances and high-frequency components representing measurement noise. The time-varying variance of the high-frequency components is estimated in real time, including: using Daubechies wavelets as basis functions to estimate the variance of the preceding components. The model error signal for each sampling period is subjected to a 3-level discrete wavelet decomposition, where The approximate coefficients of the output of the third layer decomposition are extracted to reconstruct the gradually changing signal feature components as low-frequency components. The detail coefficients generated by the first layer decomposition are extracted to form a high-frequency feature sequence. The median of the absolute value of the first layer detail coefficient sequence is calculated, and the median of the absolute value is divided by 0.6745 to obtain the noise standard deviation. The square of the noise standard deviation is used as the time-varying variance of the high-frequency component at this time.

7. The control method for a carrier gas evaporator-condenser according to claim 1, characterized in that, The control output increment sequence is calculated online by minimizing the cost function, including: constructing a cost function, which is the sum of the squares of the differences between the predicted output set of the process model at future times and the future target reference trajectory signal sequence over the future prediction time domain length, plus the product of the weight coefficient and the sum of the squares of the control output increment sequence over the control time domain length, wherein the control time domain length is not greater than the adjusted future prediction time domain length corresponding to the current control cycle; using the product of the preset system scaling coefficient and the time-varying variance of the high-frequency components as the weight coefficient in the cost function; solving for the control output increment sequence that satisfies the optimal condition of quadratic programming constraints; and extracting the first term in the control output increment sequence as the control command reference at the current time.

8. The control method for the carrier gas evaporator condenser according to claim 1, characterized in that, The first term of the control output increment sequence is taken as the control command reference to generate actual control commands for the cooling medium actuator and the fan actuator in the carrier gas evaporator condenser. This includes: adding the control command reference to the total cooling demand control quantity at the previous moment to obtain the total cooling demand control quantity at the current moment; using a preset 2×1 allocation column vector to decompose the total cooling demand control quantity into the cooling water regulation target quantity and the fan speed regulation target quantity; converting the cooling water regulation target quantity into a standard current signal of 4 to 20 mA within the range through linear mapping and outputting it to the cooling medium actuator to adjust the valve opening; and converting the fan speed regulation target quantity into a standard DC voltage signal of 0 to 10 V within the range through proportional conversion and outputting it to the frequency converter to control the speed of the fan actuator.

9. A control system for a carrier gas evaporator-condenser, characterized in that, It includes a memory and a processor, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the control method of the carrier gas evaporator condenser according to any one of claims 1-8 is implemented.

Images