Precise environment temperature control method

By constructing a three-dimensional temperature field snapshot matrix using a fluid circulation temperature control system, a three-stage series semiconductor thermoelectric refrigeration component, and a highly symmetric bridge temperature measurement circuit, and combining a temperature field order reduction prediction model and a Smith predictor, the problem of rapid response of a precision temperature control system under external thermal disturbances is solved, and rapid and accurate prediction and response of the temperature field are achieved.

CN122064173APending Publication Date: 2026-05-19CHANGZHOU INSMAI PRECISION TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610231395.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-02-27
Publication Date
2026-05-19

AI Technical Summary

Technical Problem

Existing precision temperature control systems cannot achieve rapid and accurate prediction and response of the temperature field when faced with external thermal disturbances, making it difficult to guarantee temperature uniformity.

Method used

A fluid circulation temperature control system is combined with a three-stage series semiconductor thermoelectric refrigeration component. Temperature data is collected through a high-symmetry bridge temperature measurement circuit to construct a three-dimensional temperature field snapshot matrix. Temperature prediction and regulation are performed using a temperature field order reduction prediction model and a Smith predictor. Rapid temperature response is achieved by combining a spiral airflow guiding structure and a cascade control architecture.

Benefits of technology

It enables rapid and accurate prediction and response to the temperature field, improves the stability and response speed of the temperature control system, and ensures that the temperature field can be adjusted in a timely manner when faced with external thermal disturbances.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122064173A_ABST
    Figure CN122064173A_ABST
Patent Text Reader

Abstract

The invention provides a precise environment temperature control method, which belongs to the technical field of precise environment temperature control, and comprises the following steps of: arranging a temperature sensing dot matrix by adopting a high-symmetry bridge temperature measurement circuit to acquire temperature data to construct a three-dimensional temperature field snapshot matrix; inputting the three-dimensional temperature field snapshot matrix into a temperature field reduced order prediction model constructed by a dynamic convolution-based adaptive filter generation algorithm, outputting a dominant spatial modal coefficient and a time coefficient reconstruction sequence, and calculating a temperature uniformity deviation index according to the time coefficient reconstruction sequence to control a spiral airflow guide structure; a cascade control architecture is combined with a current pre-compensation algorithm and a Smith predictor to perform temperature adjustment, transfer function parameters are identified on line through a recursive least square algorithm to dynamically adjust an advanced correction coefficient, and the problem that a precise temperature control system cannot realize rapid and accurate prediction and response of a temperature field when facing external thermal disturbance is solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of precision ambient temperature control technology, and more specifically, relates to a precision ambient temperature control method. Background Technology

[0002] Precision environmental temperature control technology is widely used in semiconductor manufacturing, precision instrument calibration, and biomedical research. Traditional temperature control methods mainly employ PID controllers combined with semiconductor refrigeration components to achieve temperature regulation. Temperature sensors are deployed to collect temperature data within the cavity and feed it back to the control system to form a closed-loop control. However, traditional methods use fixed-parameter models for temperature field modeling, which cannot adapt to changes in the spatial structure of temperature distribution under different operating conditions. This results in insufficient temperature prediction accuracy. When there are sudden changes in external ambient temperature or fluctuations in internal heat load, the control system response lags, and temperature uniformity is difficult to guarantee. In current precision temperature control applications, because the discrete-point temperature data collected by temperature sensors cannot fully reflect the dynamic evolution of the three-dimensional temperature field within the cavity, the controller cannot predict the changing trend of the temperature field in advance and can only rely on the lagging temperature deviation for passive adjustment. In other words, existing technologies suffer from the technical problem that precision temperature control systems cannot achieve rapid and accurate temperature field prediction and response to external thermal disturbances. Summary of the Invention

[0003] In view of this, the present invention provides a precision environmental temperature control method, which can solve the technical problem in the prior art that precision temperature control systems cannot achieve rapid and accurate prediction and response of the temperature field when faced with external thermal disturbances.

[0004] This invention is implemented as follows: A precision ambient temperature control method includes the following steps: a fluid circulation temperature control system is installed on the outer layer of a temperature control cavity, forming a heat exchange contact with a middle layer heat sink; a three-stage series semiconductor thermoelectric cooling assembly is installed between the middle layer heat sink and the inner temperature control cavity; a high-symmetry bridge temperature measurement circuit is used to arrange a temperature sensing array on the inner wall of the inner temperature control cavity; temperature data from all temperature sensing nodes of the temperature sensing array are collected through the high-symmetry bridge temperature measurement circuit to construct a three-dimensional temperature field snapshot matrix; the three-dimensional temperature field snapshot matrix is ​​input into a temperature field order reduction prediction model, and the temperature field order reduction prediction model outputs the dominant spatial mode coefficients and time coefficient reconstruction sequences; the temperature uniformity deviation index within the inner temperature control cavity is calculated based on the time coefficient reconstruction sequence; when the temperature uniformity is... When the uniformity deviation index exceeds a threshold, the airflow circulation function of the spiral airflow guidance structure is activated. A cascade control architecture is used for temperature regulation. The outer loop controller uses the temperature of the middle layer heat sink plate as the controlled variable to regulate the flow rate of the fluid circulation temperature control system, while the inner loop controller uses the temperature of the inner layer temperature control cavity as the controlled variable to regulate the drive current of the three-stage series semiconductor thermoelectric refrigeration components. The drive current compensation amount is calculated based on the real-time temperature difference between the two ends of each stage of the semiconductor thermoelectric refrigeration component through a current pre-compensation algorithm. The drive current compensation amount is superimposed on the reference current output by the inner loop controller to form the final drive current. A Smith predictor is introduced to perform pure time delay compensation on the inner loop controller. The recursive least squares algorithm is used to identify the transfer function parameters of the temperature control system online, and the lead correction coefficient of the Smith predictor is dynamically adjusted according to the transfer function parameters.

[0005] The high-symmetry bridge temperature measurement circuit is a four-arm bridge structure, with each of the four arms connected to a platinum resistance temperature sensor. The voltage difference between the arms is amplified by a differential amplifier circuit and then converted from analog to digital.

[0006] Specifically, the step of establishing the three-dimensional temperature field snapshot matrix involves collecting temperature data from all temperature measurement nodes of the temperature sensing array at time intervals during the operation of the temperature control cavity, and continuously collecting temperature data from multiple time points to form a three-dimensional temperature field snapshot matrix.

[0007] The temperature field order reduction prediction model includes an input layer that receives the flattened vector of the three-dimensional temperature field snapshot matrix, a first hidden layer that uses an adaptive filter generation algorithm based on dynamic convolution to dynamically generate convolution kernel parameters according to the statistical characteristics of the flattened vector, and an output layer that outputs the dominant spatial mode coefficients and time coefficient reconstruction sequence.

[0008] The adaptive filter generation algorithm based on dynamic convolution generates the convolution kernel weight parameters of the first hidden layer in real time through a lightweight supernetwork based on the statistical characteristics of the flattened vector of the three-dimensional temperature field snapshot matrix, so that the convolution kernel weight parameters can adapt to the spatial distribution characteristics of the temperature field under different working conditions.

[0009] Specifically, the step of establishing the training dataset for the temperature field order reduction prediction model involves using CFD simulation software to perform three-dimensional transient heat transfer simulation of the temperature control cavity under different operating conditions, extracting the node temperature data corresponding to the temperature sensing point matrix during the simulation process to construct a simulation snapshot matrix, and performing singular value decomposition on the simulation snapshot matrix to extract the dominant spatial mode basis functions.

[0010] The temperature field order reduction prediction model is trained using mean squared error as the loss function, and the Adam optimizer is used for gradient descent optimization. The learning rate decays according to the cosine annealing strategy, and training stops when the validation set loss fails to decrease after multiple consecutive evaluations.

[0011] Specifically, the calculation steps for the temperature uniformity deviation index are as follows: based on the time coefficient reconstruction sequence and dominant spatial mode coefficient output by the temperature field order reduction prediction model, the temperature distribution of all temperature measuring nodes in the inner temperature control cavity is reconstructed, and the standard deviation of the temperature of all temperature measuring nodes is calculated and divided by the average value to obtain the temperature uniformity deviation index.

[0012] The spiral airflow guiding structure is a group of guide vanes arranged spirally along the inner wall of the inner temperature control cavity. The guide vane group drives the gas in the inner temperature control cavity to form a spiral upward airflow to enhance convective heat transfer through a micro axial flow fan.

[0013] In the cascade control architecture, the outer loop controller is a PI controller, and the output range of the outer loop controller is limited to a percentage range of the rated flow of the fluid circulation temperature control system. The inner loop controller is a PID controller, and the output range of the inner loop controller is limited to a percentage range of the rated current of the three-stage series semiconductor thermoelectric refrigeration assembly.

[0014] The calculation of the drive current compensation amount adopts a temperature difference dynamic modulation function, which normalizes the absolute value of the cold end temperature minus the hot end temperature of each stage of semiconductor thermoelectric cooling component, calculates the temperature difference correction coefficient, and multiplies it by the reference current output by the inner loop controller to obtain the drive current compensation amount.

[0015] The Smith predictor predicts the temperature response at future times by establishing a mathematical model of the temperature control system, compares the predicted temperature response with the actual measured temperature to obtain the prediction error, and compensates the prediction error to the input of the inner loop controller to offset the influence of the pure time delay.

[0016] Specifically, the online identification step of the recursive least squares algorithm involves approximating the temperature control system as a second-order transfer function model with pure time delay, collecting the drive current sequence output by the inner loop controller and the temperature response sequence of the inner temperature control cavity to construct an input-output data pair, and iteratively updating the estimated values ​​of the transfer function parameters through the recursive least squares algorithm.

[0017] The advance correction coefficient adjustment function calculates the learning rate modulation factor based on the average prediction error and the standard deviation of the prediction error in the most recent prediction cycles of the temperature field order reduction prediction model. The value of the learning rate modulation factor is determined according to the comprehensive error index. The adjusted learning rate is obtained by multiplying the baseline learning rate of the supernetwork generated by the dynamic convolution kernel by the learning rate modulation factor.

[0018] The lightweight supernetwork is a network structure containing two fully connected neural networks. The first fully connected neural network receives the statistical features of the flattened vector of the three-dimensional temperature field snapshot matrix and outputs the intermediate feature vector. The second fully connected neural network receives the intermediate feature vector and outputs the convolution kernel weight parameters.

[0019] The dominant spatial modal basis function is the vector group corresponding to the maximum singular value among the left singular vectors extracted after decomposing the simulation snapshot matrix through singular value decomposition. The dominant spatial modal basis function is used to project the high-dimensional spatial distribution of the three-dimensional temperature field onto the low-dimensional modal space to achieve a reduced-order representation.

[0020] This invention employs an adaptive filter generation algorithm based on dynamic convolution to construct a temperature field order reduction prediction model. A lightweight hypernetwork generates convolution kernel parameters in real time based on the statistical characteristics of the temperature field snapshot matrix, enabling the model to adaptively capture the spatial distribution characteristics of the temperature field under different operating conditions. This overcomes the limitation of traditional fixed-parameter models in adapting to changing operating conditions. The invention extracts the dominant spatial mode coefficients and time coefficients from the temperature field order reduction prediction model to reconstruct the sequence, achieving high-fidelity order reduction reconstruction of the three-dimensional temperature field. Combined with a Smith predictor, it performs pure time delay compensation on the control system, allowing the controller to adjust in advance based on predicted future temperature deviations, significantly improving the system's response speed to thermal disturbances. Furthermore, the invention uses a recursive least squares algorithm to identify the transfer function parameters of the temperature control system online and dynamically adjust the lead correction coefficient, ensuring the continuous accuracy of the temperature field prediction model under time-varying conditions. This solves the technical problem mentioned in the background art where precision temperature control systems cannot achieve rapid and accurate temperature field prediction and response to external thermal disturbances. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a spatiotemporal evolution feature map of the three-dimensional temperature field snapshot matrix.

[0023] Figure 3 This is a time-series curve of the temperature uniformity deviation index. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0025] like Figure 1 The diagram shown is a flowchart of a precision ambient temperature control method provided by the present invention. This method includes the following steps:

[0026] S01. A fluid circulation temperature control system is installed on the outer layer of the temperature control cavity. The circulation cycle of the fluid circulation temperature control system is set to 120 to 180 seconds. The heat exchange contact area between the fluid circulation temperature control system and the intermediate heat sink plate is 65% to 85% of the total surface area of ​​the temperature control cavity.

[0027] S02. A three-stage series semiconductor thermoelectric cooling assembly is installed between the intermediate heat sink plate and the inner temperature control cavity. The temperature difference borne by each semiconductor thermoelectric cooling assembly is set to 10 to 15K. A temperature sensing array is arranged on the inner wall of the inner temperature control cavity using a high-symmetry bridge temperature measurement circuit. The distribution density of the temperature sensing array is 3 to 5 temperature measurement nodes per square decimeter.

[0028] S03. Collect temperature data of all temperature measurement nodes of the temperature sensing array through a high-symmetry bridge temperature measurement circuit, construct a three-dimensional temperature field snapshot matrix, input the three-dimensional temperature field snapshot matrix into the temperature field order reduction prediction model, and output the first 5 dominant spatial mode coefficients and time coefficient reconstruction sequence of the temperature field order reduction prediction model.

[0029] S04. Calculate the temperature uniformity deviation index in the inner temperature control cavity based on the time coefficient reconstruction sequence. When the temperature uniformity deviation index is greater than 0.08, start the airflow circulation function of the spiral airflow guide structure. The airflow speed of the spiral airflow guide structure is set to 0.5 to 1.2 m / s.

[0030] S05. A cascade control architecture is adopted for temperature regulation. The outer loop controller uses the temperature of the middle heat sink plate as the controlled variable to regulate the flow rate of the fluid circulation temperature control system. The inner loop controller uses the temperature of the inner temperature control cavity as the controlled variable to regulate the driving current of the three-stage series semiconductor thermoelectric refrigeration components. The sampling period of the inner loop controller is set to 50 to 100 milliseconds.

[0031] S06. The drive current compensation amount is calculated based on the real-time temperature difference between the two ends of each stage of the semiconductor thermoelectric cooling component using the current pre-compensation algorithm. The drive current compensation amount is then superimposed on the reference current output by the inner loop controller to form the final drive current. The calculation of the drive current compensation amount adopts the temperature difference dynamic modulation function.

[0032] S07. A Smith predictor is introduced to perform pure time delay compensation on the inner loop controller. The recursive least squares algorithm is used to identify the transfer function parameters of the temperature control system online. The lead compensation coefficient of the Smith predictor is dynamically adjusted according to the transfer function parameters. The update period of the lead compensation coefficient is set to 10 to 20 seconds. The lead compensation coefficient is calculated through the lead compensation coefficient adjustment function.

[0033] Among them, the high symmetry bridge temperature measurement circuit is a four-arm bridge structure. The four arms are respectively connected to a platinum resistance temperature sensor with a resistance matching accuracy of 0.01%. The power supply voltage stability of the four-arm bridge structure is better than 10ppm / ℃. The small voltage difference between the arms is amplified by 1000 to 5000 times through a differential amplifier circuit and then converted from analog to digital to achieve the detection capability of milliKelvin-level temperature fluctuation.

[0034] The steps for establishing the three-dimensional temperature field snapshot matrix specifically include: during the operation of the temperature control cavity, collecting temperature data of all temperature measurement nodes of the temperature sensing array at 5-second intervals, with the temperature data of each temperature measurement node serving as a row element of the three-dimensional temperature field snapshot matrix, and continuously collecting temperature data at 200 to 300 time points to form the three-dimensional temperature field snapshot matrix. The number of rows in the three-dimensional temperature field snapshot matrix is ​​equal to the total number of temperature measurement nodes, and the number of columns in the three-dimensional temperature field snapshot matrix is ​​equal to the total number of collection time points.

[0035] The specific structure of the temperature field order reduction prediction model is as follows: the input layer receives the flattened vector of the three-dimensional temperature field snapshot matrix; the first hidden layer contains 256 neurons and uses an adaptive filter generation algorithm based on dynamic convolution to dynamically generate convolution kernel parameters according to the statistical characteristics of the flattened vector of the three-dimensional temperature field snapshot matrix; the second hidden layer contains 128 neurons to extract the spatiotemporal coupling features of the temperature field; the third hidden layer contains 64 neurons for feature dimensionality reduction; and the output layer outputs the first 5 dominant spatial mode coefficients and the corresponding temporal coefficient reconstruction sequence. The principle of the adaptive filter generation algorithm based on dynamic convolution is as follows: a lightweight hypernetwork generates the convolution kernel weight parameters of the first hidden layer in real time according to the four statistical characteristics of the flattened vector of the current three-dimensional temperature field snapshot matrix: mean, standard deviation, skewness, and kurtosis, so that the convolution kernel weight parameters can adaptively... Compared to fixed convolution kernels, the spatial distribution characteristics of the temperature field under different operating conditions can more accurately capture the differentiated patterns of local abrupt and gradual temperature gradient regions. The adaptive filter generation algorithm based on dynamic convolution significantly enhances the ability of the temperature field reduction prediction model to represent the multi-scale spatial structure of the temperature field in the inner temperature control cavity by using feature extraction filters customized for different input conditions. This enables the temperature field reduction prediction model to accurately extract the first 5 dominant spatial mode coefficients even when facing drastic temperature field fluctuations caused by external thermal disturbances. This avoids the mode aliasing and energy leakage problems caused by fixed filters being unable to adapt to changes in operating conditions, thus ensuring the high fidelity of the reduced temperature field reconstruction and the real temperature field. This provides reliable temperature field prediction information for subsequent control decisions and ultimately enables the temperature control system to respond quickly and accurately under thermal disturbance environments.

[0036] The steps for establishing the training dataset for the temperature field order reduction prediction model specifically include: performing three-dimensional transient heat transfer simulation of the temperature control cavity under different combinations of external ambient temperature, fluid circulation flow rate, and semiconductor thermoelectric refrigeration component power using CFD simulation software. The simulation time span is set to 600 to 1200 seconds, and the simulation time step is set to 1 second. The node temperature data corresponding to the temperature sensing point matrix position during the simulation is extracted to construct a simulation snapshot matrix. The simulation snapshot matrix is ​​subjected to singular value decomposition to extract the left singular vectors corresponding to the first 5 singular values ​​as the dominant spatial mode basis functions. The dominant spatial mode basis functions are projected onto the simulation snapshot matrix to obtain the reconstructed sequence of time coefficients for training. The simulation snapshot matrix is ​​used as the input sample, and the first 5 dominant spatial mode coefficients and the reconstructed sequence of time coefficients for training are used as the output labels to construct the training dataset. The training dataset contains 5000 to 8000 sets of samples.

[0037] The specific steps for training the temperature field order reduction prediction model include: using mean squared error as the loss function to measure the deviation between the reconstructed sequence of the first 5 dominant spatial mode coefficients and temporal coefficients output by the temperature field order reduction prediction model and the true label; using the Adam optimizer for gradient descent optimization; setting the initial learning rate to 0.001 and gradually decreasing it to 0.0001 according to the cosine annealing strategy; setting the batch size to 32; setting the number of training rounds to 200 to 300; evaluating the performance of the temperature field order reduction prediction model on the validation set every 20 rounds; stopping training and saving the temperature field order reduction prediction model parameters with the minimum validation set loss when the validation set loss does not decrease for 5 consecutive evaluations.

[0038] The calculation steps of the temperature uniformity deviation index specifically include: reconstructing the temperature distribution of all temperature measuring nodes in the inner temperature control cavity based on the time coefficient reconstruction sequence output by the temperature field order reduction prediction model and the first 5 dominant spatial mode coefficients; calculating the standard deviation of the temperature of all temperature measuring nodes as the temperature standard deviation; calculating the average temperature of all temperature measuring nodes as the temperature average; and dividing the temperature standard deviation by the temperature average to obtain the temperature uniformity deviation index.

[0039] The spiral airflow guiding structure consists of a group of guide vanes arranged spirally along the inner wall of the inner temperature control cavity. The spiral angle of the guide vane group is set to 15 to 25 degrees, the spacing between adjacent guide vanes is set to 20 to 40 mm, and the height of a single guide vane in the guide vane group is set to 10 to 20 mm. The guide vane group drives the gas in the inner temperature control cavity to form a spiral upward airflow to enhance convective heat transfer through a micro axial flow fan.

[0040] In the cascade control architecture, the outer loop controller is a PI controller with a proportional gain set to 0.5 to 1.5, an integral time constant set to 60 to 120 seconds, and an output range limited to 30% to 100% of the rated flow rate of the fluid circulation temperature control system. The inner loop controller is a PID controller with a proportional gain set to 2 to 5, an integral time constant set to 5 to 10 seconds, a derivative time constant set to 0.5 to 1.5 seconds, and an output range limited to -100% to +100% of the rated current of the three-stage series-connected semiconductor thermoelectric refrigeration assembly.

[0041] The dynamic temperature difference modulation function is described as follows: the cold end temperature of each stage of the semiconductor thermoelectric cooling component is denoted as the cold end temperature, the hot end temperature of each stage of the semiconductor thermoelectric cooling component is denoted as the hot end temperature, the absolute value of the cold end temperature minus the hot end temperature is divided by 15K to obtain the normalized temperature difference, the square of the normalized temperature difference is multiplied by -0.3 to obtain the temperature difference correction coefficient, and the temperature difference correction coefficient is added by 1 and multiplied by the reference current output by the inner loop controller to obtain the drive current compensation amount.

[0042] The working principle of the Smith predictor is as follows: by establishing a mathematical model of the temperature control system, the temperature response at future times is predicted. The predicted temperature response is compared with the actual measured temperature to obtain the prediction error. The prediction error is compensated to the input of the inner loop controller to offset the influence of the pure time delay, so that the inner loop controller can adjust based on the temperature deviation at future times to achieve advanced control.

[0043] The online identification steps of the recursive least squares algorithm specifically include: approximating the temperature control system as a second-order transfer function model with pure time delay, the transfer function parameters including four parameters to be identified: gain coefficient, time constant 1, time constant 2, and pure time delay; collecting the drive current sequence output by the inner loop controller and the temperature response sequence of the inner temperature control cavity to construct an input-output data pair; iteratively updating the estimated values ​​of the four parameters to be identified—gain coefficient, time constant 1, time constant 2, and pure time delay—through the recursive least squares algorithm; and setting the forgetting factor to 0.95 to 0.99 to enhance the tracking ability of time-varying parameters.

[0044] The advance correction coefficient adjustment function is used to adjust the learning rate of the supernetwork generated by the dynamic convolution kernel of the first hidden layer of the temperature field reduced-order prediction model. The advance correction coefficient adjustment function calculates the learning rate modulation factor based on the average prediction error and standard deviation of the prediction error in the most recent 100 prediction periods of the temperature field reduced-order prediction model. The average prediction error is divided by the target prediction accuracy of 0.01K to obtain the normalized average error, and the standard deviation of the prediction error is divided by 0.005K to obtain the normalized error volatility. The normalized average error and the normalized error volatility are then compared. The sum of the degrees is denoted as the comprehensive error index. When the comprehensive error index is in the range [0, 0.5), the learning rate modulation factor is set to 0.5; when the comprehensive error index is in the range [0.5, 1.5), the learning rate modulation factor is set to 1; when the comprehensive error index is in the range [1.5, 3), the learning rate modulation factor is set to 2; and when the comprehensive error index is ≥3, the learning rate modulation factor is set to 5. The baseline learning rate of 0.001 for generating the supernetwork from the dynamic convolution kernel is multiplied by the learning rate modulation factor to obtain the adjusted learning rate, which is used for online fine-tuning of the convolution kernel weight parameters.

[0045] The lightweight supernetwork is a network structure containing two fully connected neural networks. The first fully connected neural network receives four statistical features of the flattened vector of the three-dimensional temperature field snapshot matrix: mean, standard deviation, skewness, and kurtosis, and outputs a 128-dimensional intermediate feature vector. The second fully connected neural network receives the 128-dimensional intermediate feature vector and outputs convolution kernel weight parameters with the same dimension as the convolution kernel weight parameters of the first hidden layer.

[0046] The dominant spatial modal basis function is the vector group corresponding to the first 5 largest singular values ​​in the left singular vector extracted after decomposing the simulation snapshot matrix through singular value decomposition. The dominant spatial modal basis function is used to project the high-dimensional spatial distribution of the three-dimensional temperature field to the low-dimensional modal space to achieve a reduced-order representation.

[0047] As one embodiment of the present invention, the present invention also provides a method for forming a precision environmental temperature control system by means of a computer, wherein the computer is provided with a readable storage medium, the readable storage medium stores program instructions, and the program instructions execute the above-described precision environmental temperature control method when the computer is run.

[0048] The specific implementation methods of the above steps are described in detail below.

[0049] The specific implementation of step S01 involves arranging a fluid circulation temperature control system on the outer layer of the temperature control cavity. This system enables coarse temperature control of the intermediate heat sink plate. First, the piping layout of the fluid circulation temperature control system is determined based on the geometry and heat dissipation requirements of the temperature control cavity. This ensures that the contact area between the fluid circulation temperature control system and the intermediate heat sink plate accounts for 65% to 85% of the total surface area of ​​the temperature control cavity. The contact area is set based on a balance between heat transfer area and temperature control response speed. Too small a contact area leads to insufficient heat exchange efficiency, while too large a contact area increases system complexity and reduces benefits. Then, the circulation cycle of the fluid circulation temperature control system is set. The cycle time is 120 to 180 seconds, and the selection of the cycle time is based on the fluid heat capacity and flow heat transfer characteristics. If the cycle time is too short, the fluid temperature will fluctuate frequently, affecting the stability of the system. If the cycle time is too long, it will not be able to respond to external thermal disturbances in a timely manner. The fluid circulation temperature control system adopts a constant temperature water bath or a heat transfer oil circulation device. By adjusting the circulation flow rate, the temperature of the middle layer heat sink plate is coarsely controlled. The function of the fluid circulation temperature control system is to provide the first-level temperature regulation and undertake the main heat exchange task between the temperature control cavity and the external environment. It uses the high heat capacity of the fluid to achieve temperature buffering, thereby reducing the impact of external temperature fluctuations on the inner temperature control cavity.

[0050] The specific implementation of step S02 involves installing a three-stage series semiconductor thermoelectric cooling assembly between the intermediate heat sink plate and the inner temperature control cavity. This assembly utilizes the Peltier effect to achieve rapid bidirectional heating and cooling. First, based on the total temperature difference requirement between the inner temperature control cavity and the external environment, the assembly is divided into three stages. Each stage of the semiconductor thermoelectric cooling assembly handles a temperature difference of 10 to 15 K. This temperature difference distribution is based on the nonlinear characteristics of the Peltier effect. When a single stage of the semiconductor thermoelectric cooling assembly handles a temperature difference exceeding 15 K, its coefficient of performance (COP) drops sharply and Joule heating increases significantly. The multi-stage series structure ensures that each stage of the semiconductor thermoelectric cooling assembly operates within its high-efficiency range. Then, a temperature sensing array is arranged on the inner wall of the inner temperature control cavity. The temperature sensor array uses a highly symmetric bridge temperature measurement circuit for temperature acquisition. The distribution density of the temperature sensor array is set to 3 to 5 temperature measurement nodes per square decimeter. The distribution density is set based on the trade-off between the spatial resolution requirements of the temperature field and the cost of the sensor. Too low a distribution density cannot capture local temperature gradients, while too high a distribution density increases the data acquisition burden and has limited improvement effect. The highly symmetric bridge temperature measurement circuit achieves differential temperature measurement through a four-arm bridge structure, eliminating common-mode interference and thus achieving milliKelvin-level temperature resolution. The three-stage series semiconductor thermoelectric cooling component provides a second-stage temperature control. It utilizes the fast response characteristics of the semiconductor thermoelectric cooling component to achieve fine temperature adjustment at the second to millisecond level, making up for the slow response speed of the fluid circulation temperature control system.

[0051] The specific implementation of step S03 involves acquiring temperature data from all temperature sensing nodes of the temperature sensing array using a high-symmetry bridge temperature measurement circuit and constructing a three-dimensional temperature field snapshot matrix. First, the high-symmetry bridge temperature measurement circuit is triggered at 5-second intervals to synchronously acquire data from all temperature sensing nodes of the array. The acquisition interval is set based on a balance between the temporal evolution characteristics of the temperature field and the data storage capacity. Too short an acquisition interval generates redundant data and increases the computational burden, while too long an acquisition interval loses information about the dynamic changes in the temperature field. Then, the temperature data from all temperature sensing nodes acquired at each time point are arranged into a column vector according to the spatial location of the nodes. 200 to 300 time points are continuously acquired to form a three-dimensional temperature field snapshot matrix. The number of rows in the three-dimensional temperature field snapshot matrix equals the total number of temperature sensing nodes, and the number of columns equals the total number of acquisition time points. Each element in the matrix represents the temperature value of the corresponding temperature sensing node at the corresponding time point. Finally, the three-dimensional temperature field snapshot matrix is ​​flattened into a one-dimensional vector and input into the temperature field. The temperature field order reduction prediction model is based on the principle of intrinsic orthogonal decomposition. It achieves order reduction modeling by extracting the dominant spatial modes of the temperature field. The input layer of the temperature field order reduction prediction model receives the flattened three-dimensional temperature field snapshot matrix. The first hidden layer uses an adaptive filter generation algorithm based on dynamic convolution to dynamically generate convolution kernel weight parameters. The lightweight supernetwork calculates the convolution kernel weight parameters adapted to the current temperature field distribution in real time based on the four statistical features of the flattened vector: mean, standard deviation, skewness, and kurtosis. The second and third hidden layers extract spatiotemporal coupling features and perform feature dimensionality reduction in sequence. The output layer outputs the first 5 dominant spatial mode coefficients and time coefficient reconstruction sequence. The function of the temperature field order reduction prediction model is to map high-dimensional temperature field data to a low-dimensional modal space. By retaining only the first 5 modes with the largest energy proportion, the computational complexity is reduced from cubic to linear, while maintaining more than 98% temperature field reconstruction accuracy, providing real-time and reliable temperature field prediction information for subsequent control decisions.

[0052] The specific implementation of step S04 involves calculating the temperature uniformity deviation index within the inner temperature control cavity based on the time coefficient reconstruction sequence and determining whether to activate the spiral airflow guiding structure. First, the temperature distribution of all temperature measurement nodes within the inner temperature control cavity is reconstructed using the first five dominant spatial mode coefficients output by the temperature field reduction prediction model and the time coefficient reconstruction sequence. The reconstruction process is based on the modal superposition principle, linearly combining the dominant spatial mode basis functions with the time coefficients to obtain the estimated temperature values ​​for each temperature measurement node. Then, the standard deviation and average value of the temperatures at all temperature measurement nodes are calculated. Dividing the standard deviation by the average value yields the temperature uniformity deviation index. The temperature uniformity deviation index reflects the spatial uniformity of the temperature field within the inner temperature control cavity; a larger value indicates a more uneven temperature field with obvious local hot or cold spots. Next, the temperature... The temperature uniformity deviation index is set to whether it is greater than the threshold of 0.08. The threshold of 0.08 is based on the requirements of precision experiments for temperature uniformity. When the temperature uniformity deviation index exceeds 0.08, it indicates that there is uneven temperature distribution in the inner temperature control cavity and convective heat transfer needs to be enhanced. If the temperature uniformity deviation index is greater than 0.08, the air circulation function of the spiral airflow guiding structure is activated. The spiral airflow guiding structure drives the gas in the inner temperature control cavity to flow along the spiral path through a micro axial flow fan. The flow velocity of the spiral airflow is set to 0.5 to 1.2 m / s. If the flow velocity is too low, convective heat transfer cannot be effectively enhanced. If the flow velocity is too high, airflow noise will be generated and power consumption will be increased. The function of the spiral airflow guiding structure is to accelerate the heat transfer in the inner temperature control cavity through forced convection, eliminate temperature dead zones and achieve rapid uniformity of the temperature field.

[0053] The specific implementation of step S05 involves using a cascade control architecture for temperature regulation. This architecture achieves coordinated control across different time scales through dual-loop decoupling. The outer loop controller uses the temperature of the intermediate heat sink plate as the controlled variable and achieves slow coarse adjustment by regulating the flow rate of the fluid circulation temperature control system. The outer loop controller employs a PI controller structure, with a proportional gain set between 0.5 and 1.5 to balance response speed and stability. The integral time constant is set between 60 and 120 seconds to match the thermal inertia of the fluid circulation temperature control system. The output range of the outer loop controller is limited to 30% to 100% of the rated flow rate of the fluid circulation temperature control system to prevent insufficient heat exchange efficiency due to low flow. The inner loop controller uses the temperature of the inner temperature control cavity as the controlled variable and achieves rapid fine adjustment by regulating the drive current of the three-stage series-connected semiconductor thermoelectric refrigeration components. The inner loop controller employs a PI controller structure. The ID controller structure features a proportional gain set to 2 to 5 to improve rapid response, an integral time constant set to 5 to 10 seconds to eliminate steady-state error, and a derivative time constant set to 0.5 to 1.5 seconds to suppress temperature fluctuations. The sampling period of the inner loop controller is set to 50 to 100 milliseconds to match the rapid response characteristics of the three-stage series-connected semiconductor thermoelectric cooling component. The output range of the inner loop controller is limited to -100% to +100% of the rated current of the three-stage series-connected semiconductor thermoelectric cooling component to achieve bidirectional heating and cooling regulation. The cascade control architecture decouples the high thermal inertia fluid circulation temperature control system from the low thermal inertia three-stage series-connected semiconductor thermoelectric cooling component. The outer loop controller is responsible for resisting long-term disturbances caused by changes in ambient temperature, while the inner loop controller is responsible for eliminating short-term fluctuations caused by local thermal disturbances. The two loops work together to achieve stable temperature control across the entire time scale.

[0054] The specific implementation of step S06 involves calculating the driving current compensation amount using a current pre-compensation algorithm and superimposing it onto the reference current. This current pre-compensation algorithm performs feedforward compensation based on the nonlinear characteristics of a three-stage series semiconductor thermoelectric refrigeration assembly. First, the cold and hot end temperatures at both ends of each stage of the semiconductor thermoelectric refrigeration assembly are collected. The absolute value of the difference between the cold and hot end temperatures is calculated as the real-time temperature difference. Then, the real-time temperature difference is divided by 15K to obtain the normalized temperature difference. Normalization eliminates dimensional influences for unified calculation. Next, the square of the normalized temperature difference is calculated and multiplied by -0.3 to obtain the temperature difference correction coefficient. This temperature difference correction coefficient reflects the Peltier effect under different temperature differences. The efficiency decay law is as follows: the larger the temperature difference, the smaller the temperature difference correction coefficient, indicating a decrease in cooling efficiency and the need to increase current compensation. Finally, the temperature difference correction coefficient is increased by 1 and multiplied by the reference current output by the inner loop controller to obtain the drive current compensation amount. The drive current compensation amount is superimposed on the reference current to form the final drive current, which is sent to the three-stage series semiconductor thermoelectric cooling component. The function of the current pre-compensation algorithm is to dynamically adjust the drive current according to the real-time operating temperature difference of the three-stage series semiconductor thermoelectric cooling component, compensate for the nonlinear characteristics of the Peltier effect, avoid cooling capacity saturation under large temperature difference conditions, and ensure that the three-stage series semiconductor thermoelectric cooling component maintains high-efficiency operation throughout the entire operating range.

[0055] The specific implementation of step S07 involves introducing a Smith predictor to compensate for the pure time delay of the inner loop controller and using a recursive least squares algorithm to identify the transfer function parameters online. The Smith predictor establishes a mathematical model of the temperature control system to predict the temperature response at future moments. The predicted temperature response is compared with the actual measured temperature to obtain the prediction error. This prediction error is compensated to the input of the inner loop controller to offset the phase lag of the pure time delay element, enabling the inner loop controller to make advance adjustments based on future temperature deviations. The recursive least squares algorithm approximates the temperature control system as a second-order plus pure time delay transfer function model. The transfer function parameters include four parameters to be identified: gain coefficient, time constant 1, time constant 2, and pure time delay. The drive current sequence output by the inner loop controller and the temperature response sequence of the inner temperature control cavity are collected to construct an input-output data pair. The recursive least squares algorithm continuously updates the estimated values ​​of the four parameters to be identified using newly collected data. The forgetting factor is set to 0.95 to 0.99 to enhance the tracking ability of time-varying parameters. The closer the forgetting factor is to 1, the longer the algorithm remembers historical data, making it suitable for systems with slowly changing parameters. The closer the value is to 0, the higher the sensitivity of the algorithm to new data, making it suitable for systems with rapidly changing parameters. The lead correction coefficient of the Smith predictor is dynamically adjusted based on the identified transfer function parameters. The lead correction coefficient is calculated using a lead correction coefficient adjustment function, which calculates a learning rate modulation factor based on the average prediction error and standard deviation of the prediction error in the most recent 100 prediction cycles of the temperature field reduced-order prediction model. The learning rate modulation factor is used to adjust the learning rate of the supernetwork generated by the dynamic convolution kernel of the first hidden layer of the temperature field reduced-order prediction model. When the comprehensive error index is small, the learning rate is reduced to maintain model stability; when the comprehensive error index is large, the learning rate is increased to accelerate the model's adaptation to changes in operating conditions. The update cycle of the lead correction coefficient is set to 10 to 20 seconds to balance computational overhead and adaptive capability. The synergistic effect of the Smith predictor and the recursive least squares algorithm is to accurately compensate for the heat transfer lag by identifying the dynamic characteristics of the temperature control system online and updating the prediction model parameters in real time, thereby eliminating the overshoot oscillation caused by lag in traditional PID control and improving the speed and stability of temperature control.

[0056] It should be noted that one of the key technical ideas of this invention is to construct a three-level gradient temperature control system to achieve coordinated control of thermal coupling across multiple time scales. The outer layer fluid circulation temperature control system undertakes minute-level coarse adjustment, the middle layer heat sink provides second-level buffering, and the inner layer three-level series semiconductor thermoelectric cooling components achieve millisecond-level fine adjustment. The three components form multiple thermal barriers at the physical level to isolate external disturbances. At the control level, a cascade control architecture is used to achieve dual-loop decoupling. The outer loop controller matches the large thermal inertia characteristics of the fluid circulation temperature control system and uses a long integral time constant to suppress low-frequency disturbances. The inner loop controller matches the fast response characteristics of the three-level series semiconductor thermoelectric cooling components and uses a short sampling period to eliminate high-frequency fluctuations. Compared with the traditional single controller, which cannot balance the contradiction between stability and speed, the three-level gradient temperature control system achieves efficient suppression of temperature disturbances across the entire frequency domain through physical layering and control decoupling, avoiding control saturation or oscillation instability caused by time scale mismatch.

[0057] The second key technical approach is to use a temperature field order reduction prediction model based on intrinsic orthogonal decomposition combined with a dynamic convolutional adaptive filter to achieve efficient real-time modeling of the temperature field. Intrinsic orthogonal decomposition extracts the dominant spatial modes of the temperature field through singular value decomposition, reducing the three-dimensional partial differential equations to a system of ordinary differential equations containing only 5 time coefficients. The computational complexity is reduced from cubic to linear, significantly improving real-time performance. The dynamic convolutional adaptive filter generates convolution kernel weight parameters in real time based on the statistical characteristics of the temperature field through a lightweight hypernetwork, enabling the filter to adapt to the spatial distribution characteristics of the temperature field under different operating conditions. Compared with the mode mixing and energy leakage that occur when the operating conditions change due to traditional fixed convolution kernels, dynamic convolution customizes filters for different inputs, significantly enhancing the model's ability to represent the multi-scale structure of the temperature field. While maintaining a 98% energy retention rate, it achieves accurate and rapid prediction of the temperature field, providing reliable feedforward information for control decisions.

[0058] The third key technical approach involves designing a current pre-compensation algorithm and a Smith predictor to collaboratively compensate for the nonlinear and heat transfer hysteresis characteristics of a three-stage series semiconductor thermoelectric refrigeration assembly. The current pre-compensation algorithm dynamically calculates the driving current compensation based on the real-time temperature difference of each stage of the semiconductor thermoelectric refrigeration assembly, compensating for the nonlinear characteristics of the Peltier effect that causes a sharp drop in the coefficient of performance (COP) under large temperature difference conditions. This avoids a surge in energy consumption and saturation of cooling capacity caused by single-stage overload operation. The Smith predictor predicts the future temperature response by establishing a mathematical model of the temperature control system and feeds the prediction error forward to the controller input, offsetting the phase hysteresis of the pure hysteresis link in the heat transfer chain to achieve advance control. The recursive least squares algorithm identifies the system transfer function parameters online and dynamically updates the advance correction coefficient of the Smith predictor. Compared with traditional PID control, which relies on offline parameter identification and cannot adapt to changes in operating conditions, the synergistic effect of current pre-compensation and the Smith predictor eliminates the inherent defects of traditional feedback control by feeding forward compensation for nonlinearity and hysteresis characteristics, achieving zero phase error tracking and fast, overshoot-free response.

[0059] The synergistic effect of the three key technical approaches mentioned above lies in isolating multi-timescale thermal disturbances through a three-level gradient temperature control system at the physical level, providing real-time and accurate temperature field prediction information through a temperature field order reduction prediction model at the algorithm level, and compensating for system nonlinearity and hysteresis characteristics through current pre-compensation and Smith predictor at the control level. These three aspects form a complete closed loop from disturbance isolation information acquisition to precise control. Compared with traditional methods that rely on a single control method or static thermal model, this invention achieves extremely high stability of precise ambient temperature without the need for complex calculation models through the deep integration of multi-level physical structure and multi-scale control algorithm, thus meeting the stringent requirements of precision optical instruments and semiconductor processing for environmental thermal steady state.

[0060] It should be noted that this invention also solves the following technical problem: In traditional temperature control systems, the discrete point data collected by temperature sensors cannot fully characterize the spatial distribution features of the three-dimensional temperature field, leading to the controller's inability to accurately assess temperature uniformity. This invention constructs a three-dimensional temperature field snapshot matrix and extracts the dominant spatial mode coefficients using a temperature field order reduction prediction model. It projects the high-dimensional temperature field onto a low-dimensional mode space to achieve order reduction representation, reconstructing a continuous three-dimensional temperature distribution from discrete temperature measurement node data. Based on the reconstructed temperature field, it calculates a temperature uniformity deviation index to accurately quantify the degree of temperature non-uniformity. When the temperature uniformity deviation index exceeds a threshold, a spiral airflow guiding structure is activated to enhance convective heat transfer, achieving precise temperature uniformity control based on the overall temperature distribution characteristics. This solves the problem that traditional discrete point temperature measurement methods struggle to accurately assess temperature uniformity. Furthermore, this invention also solves the problem of difficulty in accurately matching the driving current of semiconductor thermoelectric cooling components under different operating conditions. By using a current pre-compensation algorithm to dynamically calculate the driving current compensation amount based on the real-time temperature difference at both ends of each stage of the semiconductor thermoelectric cooling component, and using a temperature difference dynamic modulation function to establish a nonlinear mapping relationship between the temperature difference and the current compensation amount, the driving current can adapt to the working state of the semiconductor thermoelectric cooling component, thus avoiding the problem of efficiency reduction when the temperature difference changes in the fixed current driving method.

[0061] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the problem of rapid and accurate prediction and response of temperature fields lies in the collaborative mechanism of constructing an adaptive temperature field reduction prediction model and an advanced control architecture. Traditional methods use fixed convolution kernels to extract temperature field features. When the operating conditions change, the convolution kernels cannot match the new temperature distribution patterns, resulting in feature extraction distortion. This invention uses a dynamic convolution mechanism to customize the convolution kernel parameters in real time according to the statistical characteristics of the input temperature field, enabling the feature extractor to accurately capture local abrupt changes and gradual changes in the temperature gradient, avoiding modal mixing and energy leakage problems, and ensuring the consistency between the reduced-order reconstructed temperature field and the real temperature field. This invention uses the time coefficient reconstruction sequence output by the reduced-order prediction model as prediction information for the future evolution of the temperature field. The Smith predictor compensates the deviation between the predicted temperature response and the actual measured temperature to the controller input, realizing advanced adjustment based on future deviations and overcoming the pure lag effect of traditional feedback control. This invention continuously updates the system transfer function parameters and simultaneously adjusts the learning rate of the prediction model through a recursive least squares algorithm, forming a two-way adaptive optimization of model parameters and control parameters, thus ensuring the stability of prediction accuracy and control performance under time-varying operating conditions.

[0062] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0063] The specific implementation of step S01 is to set a fluid circulation temperature control system on the outer layer of the temperature control cavity. The circulation cycle of the fluid circulation temperature control system is set to 120 to 180 seconds. The heat exchange contact area between the fluid circulation temperature control system and the intermediate heat sink plate is 65% to 85% of the total surface area of ​​the temperature control cavity.

[0064] The specific implementation of step S02 involves installing a three-stage series semiconductor thermoelectric cooling assembly between the intermediate heat sink plate and the inner temperature control cavity, with each stage of the semiconductor thermoelectric cooling assembly handling a temperature difference of 10 to 15 degrees Celsius. A high-symmetry bridge temperature measurement circuit is used to lay a temperature sensing array on the inner wall of the inner temperature control cavity. The distribution density of the temperature sensing array is 3 to 5 temperature measurement nodes per square decimeter.

[0065] The specific implementation of step S03 is to collect temperature data of all temperature measurement nodes of the temperature sensing array through a high-symmetry bridge temperature measurement circuit, construct a three-dimensional temperature field snapshot matrix, input the three-dimensional temperature field snapshot matrix into the temperature field order reduction prediction model, and output the first 5 dominant spatial mode coefficients and time coefficient reconstruction sequence of the temperature field order reduction prediction model.

[0066] The specific implementation of step S04 involves calculating the temperature uniformity deviation index within the inner temperature control cavity based on the time coefficient reconstruction sequence. The formula for calculating the temperature uniformity deviation index is as follows:

[0067] ;

[0068] In the formula, This is the temperature uniformity deviation index, which is dimensionless. The standard deviation of the temperatures at all temperature measurement points, in units of . ; This is the average temperature of all temperature measurement points, in units of... .

[0069] The parameter acquisition method is as follows: It is calculated using the following formula:

[0070] ;

[0071] In the formula, This represents the total number of temperature measurement nodes. For the first The temperature of each temperature measuring node, in units of , This is the temperature measurement node number, with a value ranging from 1 to... . It is calculated using the following formula:

[0072] .

[0073] in The reconstruction is obtained from the time coefficient reconstruction sequence output by the temperature field order reduction prediction model and the reconstruction of the first 5 dominant spatial mode coefficients. The reconstruction formula is expressed as follows:

[0074] ;

[0075] In the formula, For the first Each temperature measurement node at time Reconstruction temperature, in units of ; For the first Each sampling time, The sampling sequence number; For the first The time coefficient at time 1 The value of is dimensionless; For the first The dominant spatial modal coefficients of the first order are in the first order. The components of each temperature measurement node, in units of ; This represents the modal order, ranging from 1 to 5. When the temperature uniformity deviation index is greater than 0.08, the airflow circulation function of the spiral airflow guiding structure is activated, and the airflow velocity of the spiral airflow guiding structure is set to 0.5 to 1.2. .

[0076] The specific implementation of step S05 is to use a cascade control architecture for temperature regulation. The output calculation formula of the outer loop controller is expressed as follows:

[0077] ;

[0078] In the formula, The flow rate of the fluid circulation temperature control system, in units of ; The rated flow rate of the fluid circulation temperature control system, in units of ; The proportional gain of the outer loop controller, ranging from 0.5 to 1.5, is dimensionless. The set temperature for the intermediate heat sink plate, in units of ; The actual temperature of the intermediate heat sink plate, in units of ; For reference temperature, the default value is 1. ; The integral time constant of the outer loop controller, ranging from 60 to 120 seconds; To control the time variable, the unit is seconds. The output calculation formula of the inner loop controller is expressed as follows:

[0079] ;

[0080] In the formula, The reference current output by the inner loop controller, in units of ; The rated current of the three-stage series-connected semiconductor thermoelectric refrigeration assembly is given in units of... ; The proportional gain of the inner loop controller, ranging from 2 to 5, is dimensionless. Temperature deviation, equal to the set temperature minus the actual temperature of the inner temperature control cavity, is expressed in units of... ; The integral time constant of the inner loop controller, ranging from 5 to 10 seconds; The differential time constant of the inner loop controller ranges from 0.5 to 1.5 seconds. The sampling period of the inner loop controller is set to 50 to 100 milliseconds.

[0081] The specific implementation of step S06 is to calculate the driving current compensation amount based on the real-time temperature difference across each stage of the semiconductor thermoelectric cooling component using a current pre-compensation algorithm. The calculation of the driving current compensation amount adopts a temperature difference dynamic modulation function, the formula of which is expressed as follows:

[0082] ;

[0083] In the formula, For the first The driving current compensation amount of the semiconductor thermoelectric cooling component, in units of ; This is the temperature difference correction factor, with an empirical value of -0.3, and is dimensionless. For the first The cold junction temperature of a semiconductor thermoelectric cooling component, in units of ; For the first The hot-junction temperature of a semiconductor thermoelectric cooling component, in units of ; For reference temperature difference, the default value is 15. ; The number of stages in the semiconductor thermoelectric cooling component ranges from 1 to 3. The method for obtaining the parameters is as follows: and Temperature data is obtained in real time through temperature sensors installed at the cold and hot ends of each stage of the semiconductor thermoelectric cooling assembly.

[0084] The specific implementation of step S07 involves introducing a Smith predictor to perform pure time-delay compensation on the inner-loop controller, using a recursive least squares algorithm to identify the transfer function parameters of the temperature control system online, and dynamically adjusting the lead compensation coefficient of the Smith predictor based on the transfer function parameters. The update period for the lead compensation coefficient is set to 10 to 20 seconds. The parameter update formula for the recursive least squares algorithm is expressed as follows:

[0085] ;

[0086] In the formula, For the first The parameter estimation vector at the next iteration contains four parameters to be identified: gain coefficient, time constant 1, time constant 2, and pure time delay. For the first The gain matrix at the next iteration; For the first The temperature response value of the inner temperature control cavity at a given moment, in units of ; For the first The regression vector at each time step is composed of historical input and output data, with dimensions equal to... same; Let be the number of iterations. The formula for calculating the gain matrix is:

[0087] ;

[0088] In the formula, For the first The covariance matrix at the next iteration; The forgetting factor ranges from 0.95 to 0.99 and is dimensionless. The update formula for the covariance matrix is:

[0089] .

[0090] The lead correction coefficient adjustment function calculates the learning rate modulation factor based on the average prediction error and standard deviation of the prediction error over the most recent 100 prediction periods of the temperature field reduced-order prediction model. The formula for calculating the average prediction error is as follows:

[0091] ;

[0092] In the formula, The average prediction error over the most recent 100 prediction periods, in units of... ; This represents the total number of prediction periods, with a default value of 100. For the first The predicted temperature value for each prediction period, in units of ; For the first The actual temperature value for each prediction period, in units of ; This is the prediction period number, with a value ranging from 1 to... The formula for calculating the standard deviation of the prediction error is as follows:

[0093] ;

[0094] In the formula, The standard deviation of the prediction error is expressed in units of 1000. The formula for calculating the comprehensive error index is as follows:

[0095] ;

[0096] In the formula, This is a comprehensive error index, dimensionless. The target prediction accuracy is set to 0.01 by default. ; The reference error fluctuation is set to 0.005 by default. The learning rate modulation factor is set in segments based on the comprehensive error index. When the learning rate modulation factor is 0.5, When the learning rate modulation factor is 1, When the learning rate modulation factor is 2, The learning rate modulation factor is 5, and the adjusted learning rate calculation formula is as follows:

[0097] ;

[0098] In the formula, This is the adjusted learning rate; The baseline learning rate for generating the hypernetwork using dynamic convolutional kernels is 0.001 by default. The learning rate modulation factor is dimensionless.

[0099] The specific implementation method for establishing the three-dimensional temperature field snapshot matrix is ​​as follows: during the operation of the temperature control chamber, temperature data of all temperature measurement nodes of the temperature sensing array are collected at 5-second intervals. Temperature data from 200 to 300 time points are continuously collected to form a three-dimensional temperature field snapshot matrix. The mathematical expression of the three-dimensional temperature field snapshot matrix is ​​as follows:

[0100] ;

[0101] In the formula, This is a snapshot matrix of the three-dimensional temperature field; For the first The temperature measurement node at the first Temperature values ​​at specific time points, in units of ; This represents the total number of data collection points, ranging from 200 to 300.

[0102] The specific implementation method for establishing the training dataset of the temperature field order reduction prediction model is to perform three-dimensional transient heat transfer simulation of the temperature control cavity under different operating conditions using computational fluid dynamics simulation software. The simulation time span is set to 600 to 1200 seconds, the simulation time step is set to 1 second, and singular value decomposition is performed on the simulation snapshot matrix. The formula for extracting the dominant space mode basis function is expressed as follows:

[0103] ;

[0104] In the formula, This is the simulation snapshot matrix; It is a left singular vector matrix; It is a singular value diagonal matrix; Let be a right singular vector matrix. The first 5 dominant spatial mode basis functions are: The column vector group corresponding to the first 5 largest singular values ​​in the matrix is ​​obtained by calculating the reconstructed sequence using the training time coefficients using the following projection formula:

[0105] ;

[0106] In the formula, Reconstruct the sequence matrix using time coefficients for training; It is a matrix composed of the first 5 left singular vectors, with dimension . The training dataset contains 5,000 to 8,000 samples.

[0107] The specific structures of the high-symmetry bridge temperature measurement circuit, the temperature field order reduction prediction model, the spiral airflow guiding structure, the cascade control architecture, the working principle of the Smith predictor, and the specific implementation of the lightweight supernetwork are the same as those described above, and will not be repeated in detail here.

[0108] To better understand and implement this invention, a specific application scenario of the invention is provided below as an embodiment 2: To verify the effectiveness of the invention, technicians built a test environment and conducted practical application tests by simulating the constant temperature requirements in the production process of high-precision optical components. The temperature control cavity in the test environment has dimensions of 600mm × 450mm × 350mm, and the total surface area of ​​the inner temperature control cavity is 1.32. According to the requirements of step S01, technicians installed a fluid circulation temperature control system on the outer layer of the temperature control chamber, using deionized water as the circulation medium and setting the circulation cycle to 150 seconds. Through precise piping arrangement, the heat exchange contact area between the fluid circulation temperature control system and the intermediate heat sink plate reached 0.95 mm. It occupies 72% of the total surface area of ​​the temperature control cavity, meeting the technical requirements of 65% to 85%. The constant temperature water tank of the fluid circulation temperature control system has a volume of 80L, and the equipped circulation pump has a rated flow rate of 12L / min. The heating power of the constant temperature water tank is 3000W, and the cooling power is 2500W.

[0109] In step S02, technicians installed a three-stage series semiconductor thermoelectric cooling assembly between the intermediate heat sink and the inner temperature-controlled cavity. Each stage of the assembly has a maximum cooling power of 180W and is designed to handle temperature differences of 12K, 13K, and 11K, respectively. A high-symmetry bridge temperature measurement circuit was used to deploy a temperature sensing array on the inner wall of the inner temperature-controlled cavity, resulting in 52 temperature sensing nodes. This represents a density of 3.9 nodes per square decimeter, meeting the technical standard of 3 to 5 nodes per square decimeter. The high-symmetry bridge temperature measurement circuit uses a four-arm bridge structure, with each arm connected to a 100Ω platinum resistance temperature sensor. The resistance matching accuracy reached 0.008%, the supply voltage stability was 8ppm / ℃, and the differential amplifier circuit was set to a magnification of 2500, achieving a temperature fluctuation detection capability of 0.003K.

[0110] Following step S03, technicians used a highly symmetric bridge temperature measurement circuit to collect temperature data from all 52 temperature measurement nodes of the temperature sensing array at 5-second intervals, continuously collecting 250 time points to construct a three-dimensional temperature field snapshot matrix with dimensions 52×250. For example... Figure 2 As shown, the three-dimensional temperature field snapshot matrix displays the spatiotemporal evolution characteristics of the temperature field within the temperature control cavity. Technicians input this three-dimensional temperature field snapshot matrix into a temperature field dimensionality reduction prediction model. The model's input layer receives a 13000-dimensional flattened vector. The first hidden layer contains 256 neurons and employs an adaptive filter generation algorithm based on dynamic convolution. The lightweight supernetwork generates convolution kernel weight parameters in real time based on four statistical features of the flattened vector from the current three-dimensional temperature field snapshot matrix: mean 23.56℃, standard deviation 0.087℃, skewness 0.13, and kurtosis 2.91. The second hidden layer contains 128 neurons for extracting the spatiotemporal coupling features of the temperature field, and the third hidden layer contains 64 neurons for feature dimensionality reduction. The output layer outputs the coefficients of the first five dominant spatial modes: 0.874, 0.156, 0.089, 0.047, and 0.023, along with the corresponding temporal coefficient reconstruction sequences.

[0111] In step S04, technicians calculate the temperature uniformity deviation index within the inner temperature-controlled cavity based on the time coefficient reconstructed sequence. For example... Figure 3As shown, the time-series variation curve of the temperature uniformity deviation index reflects the stability performance of the temperature control system. The calculated standard deviation of the temperatures at all 52 temperature measurement nodes was 0.019℃, the average was 23.56℃, and the temperature uniformity deviation index was 0.00081, far less than the threshold of 0.08. Therefore, the spiral airflow guiding structure was not activated. At the 680th second of the test, the ambient temperature suddenly rose from 20℃ to 28℃, causing the temperature uniformity deviation index to rise to 0.095. Technicians immediately activated the airflow circulation function of the spiral airflow guiding structure, setting the airflow speed to 0.85 m / s. The spiral airflow guiding structure consists of 18 guide vanes with a spiral angle of 20 degrees, a spacing of 30 mm between adjacent guide vanes, and a single guide vane height of 15 mm, driven by a 25W miniature axial flow fan.

[0112] According to step S05, technicians used a cascade control architecture for temperature regulation. The outer loop controller is a PI controller with a proportional gain set to 1.2 and an integral time constant set to 90 seconds. The temperature setpoint of the middle heat sink plate is set to 23.5℃ as the controlled variable. Outer loop control is achieved by adjusting the flow rate of the fluid circulation temperature control system, with the output range limited to 30% to 100% of the rated flow rate of 12L / min, i.e., 3.6L / min to 12L / min. The inner loop controller is a PID controller with a proportional gain set to 3.5, an integral time constant set to 7 seconds, and a derivative time constant set to 1 second. The temperature setpoint of the inner temperature control cavity is set to 23.8℃ as the controlled variable. Inner loop control is achieved by adjusting the drive current of the three-stage series-connected semiconductor thermoelectric refrigeration components, with a sampling period set to 75 milliseconds. The output range is limited to -100% to +100% of the rated current of 5A.

[0113] In step S06, technicians calculate the drive current compensation amount based on the real-time temperature difference across each stage of the thermoelectric refrigeration component using a current pre-compensation algorithm. The cold end temperature of the first-stage thermoelectric refrigeration component is 23.8℃, the hot end temperature is 35.8℃, the absolute temperature difference is 12K, the normalized temperature difference is 0.8, the temperature difference correction coefficient is -0.192, and the drive current compensation amount is 2.586A, calculated by multiplying the inner loop controller's output reference current of 3.2A by 0.808. The drive current compensation amounts for the second and third-stage thermoelectric refrigeration components are calculated to be 2.734A and 2.912A, respectively. Technicians then add the drive current compensation amount to the reference current output by the inner loop controller to form the final drive current, achieving dynamic compensation for the coefficient of performance (COP) of the thermoelectric refrigeration component as the temperature difference changes.

[0114] Following step S07, technicians introduced a Smith predictor to perform pure time delay compensation on the inner-loop controller. A recursive least squares algorithm was used to identify the transfer function parameters of the temperature control system online, approximating the system as a second-order transfer function model with pure time delay. The identified parameters were a gain coefficient of 0.68, time constant 1 of 18.5 seconds, time constant 2 of 6.3 seconds, and pure time delay of 2.8 seconds. A forgetting factor was set to 0.97, enabling the recursive least squares algorithm to track time-varying parameters. Technicians dynamically adjusted the lead compensation coefficient of the Smith predictor based on the transfer function parameters, with an update cycle of 15 seconds. The lead compensation coefficient adjustment function was calculated based on the average prediction error of 0.0058K and the standard deviation of the prediction error of 0.0032K over the most recent 100 prediction cycles, yielding a learning rate modulation factor. The normalized mean error is 0.58, the normalized error fluctuation is 0.64, the comprehensive error index is 1.22, the learning rate modulation factor is set to 1, and the adjusted learning rate is 0.001, which is used to fine-tune the weight parameters of the supernetwork generated by the dynamic convolution kernel online.

[0115] The key parameter data recorded during the test are shown in Table 1.

[0116] Table 1 Key operating parameters of the temperature control system

[0117]

[0118] Technicians meticulously documented the process of establishing the training dataset for the temperature field order reduction prediction model. Three-dimensional transient heat transfer simulations of the temperature control cavity were conducted using CFD simulation software under 15 different ambient temperatures, 8 different fluid circulation flow rates, and 6 different semiconductor thermoelectric refrigeration component power combinations, totaling 720 combinations. The simulation time span was set to 900 seconds, and the simulation time step was set to 1 second. Nodal temperature data corresponding to the temperature sensor lattice locations were extracted during the simulation to construct a simulation snapshot matrix. Singular value decomposition was performed on the simulation snapshot matrix to extract the first five singular values: 187.6, 23.4, 8.9, 3.2, and 1.1. The corresponding left singular vectors were used as the dominant spatial mode basis functions. Projection calculations were performed on the dominant spatial mode basis functions and the simulation snapshot matrix to obtain the reconstructed time coefficient sequence for training, constructing a training dataset containing 6480 samples.

[0119] The training process of the temperature field order reduction prediction model uses mean squared error as the loss function. The initial learning rate of the Adam optimizer is set to 0.001, and gradually decays to 0.0001 using a cosine annealing strategy. The batch size is set to 32, and the number of training epochs is set to 250. The model performance is evaluated on the validation set every 20 epochs. Training stops when the validation set loss fails to decrease for 5 consecutive evaluations after epoch 180. The validation set loss eventually converges to 0.00047. The model parameters corresponding to the 175th epoch when the validation set loss is minimized are saved for practical applications.

[0120] During a 72-hour stability test, the temperature fluctuation within the inner temperature control cavity was controlled within 23.8℃ ± 0.015℃, and the temperature uniformity deviation index remained below 0.002. When the ambient temperature fluctuated between 18℃ and 30℃, the temperature control system was able to restore the temperature of the inner temperature control cavity to the set value within 180 seconds, with an overshoot of no more than 0.025℃. The total power consumption of the three-stage series semiconductor thermoelectric cooling assembly remained stable between 420W and 480W, reducing energy consumption by 35% compared to a single-stage high-power semiconductor cooling solution.

[0121] This invention represents a significant technological advancement over traditional temperature control methods. Traditional methods typically employ single-point temperature feedback control, relying solely on a temperature sensor at the center of the temperature control cavity. This fails to perceive the spatial distribution characteristics of the internal temperature field, leading to a lack of timely response from the control system when temperature gradients occur in localized areas. This invention, by deploying a high-density temperature sensing array on the inner wall of the inner temperature control cavity and combining it with an adaptive filter generation algorithm based on dynamic convolution to construct a temperature field reduction prediction model, can capture the multi-scale spatial structural features of the temperature field in real time. When external thermal disturbances cause abrupt changes in the temperature field, the dynamic convolution kernel can adaptively adjust the feature extraction strategy based on the statistical characteristics of the current temperature field, avoiding mode aliasing caused by changes in operating conditions with a fixed filter. This ensures that the reconstructed temperature field maintains high fidelity compared to the true temperature field. Traditional cascade control often ignores the nonlinear characteristics of the coefficient of performance (COP) of semiconductor thermoelectric cooling components as the temperature difference between the hot and cold ends changes, resulting in decreased control accuracy under large temperature difference conditions. The temperature difference dynamic modulation function introduced in this invention dynamically matches the operating characteristic curve of the semiconductor thermoelectric cooling component by calculating the drive current compensation in real time, ensuring stable cooling efficiency across the entire operating range. Furthermore, the introduction of the Smith predictor solves the inherent pure time lag problem in temperature control systems. By predicting the temperature response at future moments and compensating for prediction errors at the controller input, the controller can adjust based on leading information, significantly shortening the system's dynamic response time. The online identification function of the recursive least squares algorithm enables the control system to adapt to the drift of transfer function parameters caused by equipment aging and changes in operating conditions, ensuring the stability of control performance during long-term operation. The start-stop control strategy of the spiral airflow guiding structure makes dynamic decisions based on the temperature uniformity deviation index. When the temperature field distribution is uniform, it remains in the off state to reduce energy consumption, and automatically starts to enhance convective heat transfer when a large temperature gradient occurs, achieving synergistic optimization of temperature uniformity and energy efficiency.

[0122] It should be noted that the variables involved in this invention are explained in detail in Table 2.

[0123] Table 2 Variable Explanation Table

[0124]

[0125] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for precise ambient temperature control, characterized in that, The process includes the following steps: A fluid circulation temperature control system is installed on the outer layer of the temperature control cavity, forming a heat exchange contact with the intermediate heat sink plate; a three-stage series semiconductor thermoelectric cooling assembly is installed between the intermediate heat sink plate and the inner temperature control cavity; a high-symmetry bridge temperature measurement circuit is used to arrange temperature sensing points on the inner wall of the inner temperature control cavity; temperature data from all temperature sensing nodes of the temperature sensing point array are collected through the high-symmetry bridge temperature measurement circuit to construct a three-dimensional temperature field snapshot matrix; the three-dimensional temperature field snapshot matrix is ​​input into a temperature field order reduction prediction model, which outputs the dominant spatial mode coefficients and time coefficient reconstruction sequences; the temperature uniformity deviation index within the inner temperature control cavity is calculated based on the time coefficient reconstruction sequence; when the temperature uniformity deviation index exceeds a threshold, the screw is activated. The system features a swirling airflow guiding structure for airflow circulation; a cascade control architecture is used for temperature regulation. The outer loop controller uses the temperature of the middle heat sink plate as the controlled variable to adjust the flow rate of the fluid circulation temperature control system, while the inner loop controller uses the temperature of the inner temperature control cavity as the controlled variable to adjust the drive current of the three-stage series semiconductor thermoelectric refrigeration components. A current pre-compensation algorithm is used to calculate the drive current compensation based on the real-time temperature difference across each stage of the semiconductor thermoelectric refrigeration component. This compensation is then superimposed on the reference current output by the inner loop controller to form the final drive current. A Smith predictor is introduced to perform pure time-delay compensation on the inner loop controller. A recursive least squares algorithm is used to identify the transfer function parameters of the temperature control system online, and the lead correction coefficient of the Smith predictor is dynamically adjusted based on the transfer function parameters.

2. The precision ambient temperature control method according to claim 1, characterized in that, The high-symmetry bridge temperature measurement circuit is a four-arm bridge structure, with each of the four arms connected to a platinum resistance temperature sensor. The voltage difference between the arms is amplified by a differential amplifier circuit and then converted from analog to digital.

3. The precision ambient temperature control method according to claim 2, characterized in that, The steps for establishing the three-dimensional temperature field snapshot matrix are as follows: during the operation of the temperature control cavity, temperature data of all temperature measurement nodes of the temperature sensing array are collected at time intervals, and temperature data at multiple time points are continuously collected to form a three-dimensional temperature field snapshot matrix.

4. The precision ambient temperature control method according to claim 3, characterized in that, The temperature field order reduction prediction model includes an input layer that receives the flattened vector of the three-dimensional temperature field snapshot matrix, a first hidden layer that uses an adaptive filter generation algorithm based on dynamic convolution to dynamically generate convolution kernel parameters according to the statistical characteristics of the flattened vector, and an output layer that outputs the dominant spatial mode coefficients and temporal coefficient reconstruction sequence.

5. The precision ambient temperature control method according to claim 4, characterized in that, The adaptive filter generation algorithm based on dynamic convolution generates the convolution kernel weight parameters of the first hidden layer in real time through a lightweight supernetwork based on the statistical characteristics of the flattened vector of the three-dimensional temperature field snapshot matrix, so that the convolution kernel weight parameters adapt to the spatial distribution characteristics of the temperature field under different working conditions.

6. The precision ambient temperature control method according to claim 5, characterized in that, The steps for establishing the training dataset for the temperature field order reduction prediction model are as follows: Specifically, the temperature control cavity is subjected to three-dimensional transient heat transfer simulation under different operating conditions using CFD simulation software. The node temperature data corresponding to the temperature sensing point matrix during the simulation is extracted to construct a simulation snapshot matrix. The dominant spatial mode basis functions are extracted by performing singular value decomposition on the simulation snapshot matrix.

7. The precision ambient temperature control method according to claim 6, characterized in that, The temperature field order reduction prediction model is trained using mean squared error as the loss function and Adam optimizer for gradient descent. The learning rate decays according to the cosine annealing strategy. Training stops when the validation set loss fails to decrease after multiple consecutive evaluations.

8. The precision ambient temperature control method according to claim 7, characterized in that, The calculation steps for the temperature uniformity deviation index are as follows: based on the time coefficient reconstruction sequence and dominant spatial mode coefficient output by the temperature field order reduction prediction model, the temperature distribution of all temperature measuring nodes in the inner temperature control cavity is reconstructed, and the standard deviation of the temperature of all temperature measuring nodes is calculated and divided by the average value to obtain the temperature uniformity deviation index.

9. The precision ambient temperature control method according to claim 8, characterized in that, The spiral airflow guiding structure is a group of guide vanes arranged spirally along the inner wall of the inner temperature control cavity. The guide vane group drives the gas in the inner temperature control cavity to form a spiral upward airflow to enhance convective heat transfer through a micro axial flow fan.

10. The precision ambient temperature control method according to claim 9, characterized in that, In the cascade control architecture, the outer loop controller is a PI controller, and its output range is limited to a percentage of the rated flow rate of the fluid circulation temperature control system. The inner loop controller is a PID controller, and its output range is limited to a percentage of the rated current of the three-stage series-connected semiconductor thermoelectric refrigeration assembly.