Efficient and self-adjustable air cooling control system and method
By improving the Kalman filter algorithm and the KAN-MLP neural network model combined with the self-disturbance rejection control algorithm, the precise adjustment of the cooling pipe temperature of the air cooler is achieved, which solves the condensation problem of the turbine unit cooler and improves the safety and stability of the equipment.
Patent Information
- Application Number
- CN202510830436.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-20
- Publication Date
- 2025-09-05
AI Technical Summary
Existing technologies cannot effectively prevent condensation on cooling pipes in turbine air coolers, leading to metal corrosion, electrical equipment short circuits, and reduced heat exchange efficiency. Traditional methods also have problems such as high energy consumption, poor results, or poor adaptability.
The data acquisition module, data processing module, dew point prediction module, control module and execution module are used, combined with the improved Kalman filter algorithm and KAN-MLP neural network model, and the automatic anti-disturbance control algorithm is used to achieve precise adjustment of the cooling pipe temperature to prevent condensation.
It effectively prevents condensation on cooling pipes, improves the safety and stability of equipment operation, has strong environmental adaptability, accurately predicts dew point temperature, improves temperature regulation accuracy and dynamic response capabilities, and reduces the risk of equipment damage.
Smart Images
Figure CN120593352A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of dehumidification and anti-condensation, and in particular to a highly efficient and self-adjustable air cooling control system and method. Background Art
[0002] With the continuous development of hydropower systems, the cooling system of turbine units plays a vital role in ensuring stable equipment operation. Among them, air coolers, as a critical component of turbine units, face condensation problems on their cooling pipes, which directly impact system efficiency and equipment safety. Condensation on the cooling pipes can cause metal corrosion, short circuits in electrical equipment, and reduced heat exchange efficiency. In severe cases, it can even affect the reliability of the entire turbine unit. Therefore, effectively preventing condensation on the cooling pipes of air coolers has become an unresolved engineering challenge.
[0003] Currently, a variety of anti-condensation technologies are being applied in various fields, including traditional heating and insulation, surface hydrophobic coatings, and environmental parameter control. However, these methods have numerous limitations in practical application: heating and insulation methods increase energy consumption and affect cooling efficiency; surface hydrophobic coatings are prone to aging and have poor long-term effectiveness; and environmental parameter control often relies on simple humidity threshold settings, which are difficult to adapt to complex operating conditions. Therefore, existing technologies still have shortcomings in the application of anti-condensation in turbine air coolers, and a more intelligent, efficient, and reliable solution is urgently needed. Summary of the Invention
[0004] To solve the technical problem of condensation on cooling pipes of conventional air coolers caused by environmental fluctuations and model mismatch, the present invention provides an efficient and self-adjustable air cooling control system, comprising:
[0005] A data acquisition module is used to collect corresponding data of the condensation temperature of the cooling pipe in the air cooler and the ambient temperature and humidity;
[0006] a data processing module for optimizing the smoothness of the corresponding data of the condensation temperature of the cooling pipe in the air cooler and the ambient temperature and humidity by storing the mean of the historical data of the previous d moments using an improved Kalman filter algorithm;
[0007] A dew point prediction module is configured to input the corresponding data of the ambient temperature and humidity and output a predicted dew point temperature value based on a KAN-MLP neural network model; the KAN-MLP model comprises a fusion structure of a KAN network and an MLP network, wherein the KAN network uses an activation function that is a superposition of a spline function and a basis function; during the model training phase, the predicted value is compared with the condensation temperature of the cooling pipe in the air cooler as the true value, and the error is calculated and back-propagated to optimize the network parameters;
[0008] Dual computing unit architecture module, used to exchange data through serial communication protocol, realizing the separation deployment of dew point prediction and cooling pipe temperature control;
[0009] A control module is configured to adopt an active disturbance rejection control algorithm, use the dew point temperature prediction value as a set value, use the output of the improved Kalman filter as the output of the controlled system, and output a control signal through nonlinear feedback;
[0010] The execution module adjusts the temperature of the cooling pipe according to the control signal and uses PWM to drive the motor to achieve closed-loop control.
[0011] Furthermore, the basic function is: ;
[0012] Obtain, among which, As the basic function, is the characteristic value of the ambient temperature and humidity or cooling pipe temperature data, is the Sigmoid activation function;
[0013] The spline function is obtained by: ;
[0014] Obtain, among which, is the spline function, is the adjustable parameter of the model, is the kth (k+1th order) basis function.
[0015] Furthermore, the active disturbance rejection control algorithm processes the setpoint dynamics through a tracking differentiator, and an extended observer actively estimates the disturbance; wherein the tracking differentiator is implemented by:
[0016] Implementation, where is the sampling time, is the convergence speed, is the input set value curve, To track the output of the differentiator, To track the intermediate variables in the differentiator;
[0017] The expanded observer is constructed by:
[0018] get, is the error between the estimated and actual measurement of the observed variable by the augmented observer, where is the estimated value of the system state at step k, is the actual output of the system, 、 、 is the system variable observed by the extended observer, 、 、 、 is the adjustable parameter of the extended observer, is the sampling time of the current system, For control quantity The coefficient of influence on the system output.
[0019] Furthermore, the nonlinear feedback is achieved by:
[0020] Implementation, where is the output of the nonlinear feedback output, and are the parameters of the nonlinear function, is the feedback gain, is the tracking error.
[0021] Furthermore, the improved Kalman filter algorithm is implemented by: ;
[0022] Implementation, where To predict the value, The system measures data. is the Kalman gain in the Kalman filter, is the observation matrix.
[0023] Furthermore, the dual-computing unit architecture module includes a dew point calculation unit and a controller unit;
[0024] The dew point calculation unit performs dew point prediction based on the IMX6ULL processor and the Linux system, and the controller unit completes temperature closed-loop control based on the STM32 microcontroller and the FreeRTOS real-time system. The two achieve highly reliable data interaction through a customized serial port protocol.
[0025] A highly efficient and self-adjustable air cooling control method is also provided, comprising:
[0026] S1, a data collection step for collecting corresponding data of the condensation temperature of the cooling pipe in the air cooler and the ambient temperature and humidity;
[0027] S2, a data processing step for optimizing the smoothness of the corresponding data of the condensation temperature of the cooling pipe in the air cooler and the ambient temperature and humidity by storing the mean of the historical data of the previous d moments using an improved Kalman filter algorithm;
[0028] S3, a dew point prediction step, for inputting the corresponding data of the ambient temperature and humidity based on a KAN-MLP neural network model, and outputting a dew point temperature prediction value; the KAN-MLP model includes a fusion structure of a KAN network and an MLP network, wherein the KAN network uses an activation function that superimposes a spline function and a basis function; during the model training phase, the predicted value and the condensation temperature of the cooling pipe in the air cooler are used as the true value to perform error calculation and back propagate, thereby optimizing the network parameters;
[0029] S4, dual computing unit architecture step, for exchanging data through a serial communication protocol to achieve separate deployment of dew point prediction and cooling pipe temperature control;
[0030] S5, a control step for adopting an active disturbance rejection control algorithm, taking the dew point temperature prediction value as a set value, using the output of the improved Kalman filter as the output of the controlled system, and outputting a control signal through nonlinear feedback;
[0031] S6, executing step, adjusting the temperature of the cooling pipe according to the control signal, and using PWM to drive the motor to achieve closed-loop control.
[0032] The beneficial effects of the present invention are:
[0033] Through dew point prediction, data optimization, and active disturbance rejection control (ADRC), the cooling pipe temperature is precisely adjusted, effectively preventing condensation and improving equipment operation safety and stability. First, this invention effectively prevents condensation on the cooling pipe surface, reducing the risk of equipment damage. Using the KAN-MLP deep learning model to predict the ambient dew point, it ensures that the cooling pipe temperature remains above the dew point, fundamentally suppressing condensation. Second, this invention exhibits strong environmental adaptability, accurately predicting the dew point temperature and adapting to varying temperature and humidity conditions. Compared to traditional control methods based on fixed thresholds, this solution utilizes a deep learning model to adjust the dew point prediction in real time, maintaining stable anti-condensation performance even in high-temperature and high-humidity environments. In terms of control strategy, this invention replaces traditional PID control with active disturbance rejection control (ADRC), improving temperature regulation accuracy and dynamic response. ADRC offers stronger anti-interference capabilities and faster response speed, enabling precise adjustment of cooling pipe temperature despite fluctuations in ambient temperature and humidity, ensuring stable system operation. Furthermore, to enhance the reliability of sensor data, this invention utilizes an improved Kalman filter algorithm to reduce sudden and jumpy changes in temperature data, improve measurement accuracy, and ensure the accuracy of dew point prediction data, providing a more precise decision-making basis for the temperature control system. This invention also employs a dual computing unit architecture to improve system computational efficiency and real-time performance. The IMX6ULL (Linux) runs the deep learning model responsible for dew point temperature prediction, while the STM32F407 (FreeRTOS) processes sensor data and performs temperature control, ensuring high real-time performance. Communication between the two is achieved via a high-speed UART, incorporating header and footer checksums to ensure stable data transmission. DMA is also used to reduce CPU load and improve overall system efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 It is a system structure diagram of the present invention;
[0035] Figure 2 It is the structural diagram of the active disturbance rejection control algorithm;
[0036] Figure 3 This is the hardware structure diagram of the dual computing unit architecture module;
[0037] Figure 4 This is the program flow chart of the dew point calculation unit;
[0038] Figure 5 is a program flow chart of the controller unit; DETAILED DESCRIPTION
[0039] To make the technical solutions and advantages of the embodiments of the present invention more clearly understood, exemplary embodiments of the present invention are further described in detail below with reference to the accompanying drawings. It should be noted that the embodiments described are only a portion of the embodiments of the present invention, and are not an exhaustive list of all embodiments. It should be noted that the embodiments of the present invention and the features thereof may be combined with each other unless they conflict.
[0040] Example 1, combined Figure 1 This embodiment describes an efficient and self-regulating air cooling control system, comprising:
[0041] A data acquisition module is used to collect corresponding data of the condensation temperature of the cooling pipe in the air cooler and the ambient temperature and humidity;
[0042] a data processing module for optimizing the smoothness of the corresponding data of the condensation temperature of the cooling pipe in the air cooler and the ambient temperature and humidity by storing the mean of the historical data of the previous d moments using an improved Kalman filter algorithm;
[0043] A dew point prediction module is configured to input the corresponding data of the ambient temperature and humidity and output a predicted dew point temperature value based on a KAN-MLP neural network model; the KAN-MLP model comprises a fusion structure of a KAN network and an MLP network, wherein the KAN network uses an activation function that is a superposition of a spline function and a basis function; during the model training phase, the predicted value is compared with the condensation temperature of the cooling pipe in the air cooler as the true value, and the error is calculated and back-propagated to optimize the network parameters;
[0044] Dual computing unit architecture module, used to exchange data through serial communication protocol, realizing the separation deployment of dew point prediction and cooling pipe temperature control;
[0045] A control module is configured to adopt an active disturbance rejection control algorithm, use the dew point temperature prediction value as a set value, use the output of the improved Kalman filter as the output of the controlled system, and output a control signal through nonlinear feedback;
[0046] The execution module adjusts the temperature of the cooling pipe according to the control signal and uses PWM to drive the motor to achieve closed-loop control.
[0047] In the data collection module, the same temperature and humidity sensors as those used in the air coolers must be used to collect data. This ensures consistent ambient temperature and humidity data during model simulation. The dataset should contain approximately 2,000 to 3,000 locations. The data should cover the entire air cooler's operating environment as closely as possible, and the temperature and humidity data for each segment should be distributed as evenly as possible.
[0048] In the data processing module, the improved Kalman filter differs from the standard Kalman filter in that it introduces a memory block to store data from the previous d moments during the state update process. It then replaces the traditional measured value with the average of this historical data to enhance the prediction capability for the next moment. This fully utilizes historical information, resulting in smoother prediction results and improved prediction accuracy and stability.
[0049] In the dew point prediction module, the KAN portion of the KAN-MLP network model differs from traditional MLP networks in that it directly modifies the parameters of nonlinear functions to achieve function fitting. The KAN network structure removes the linear portion of the MLP network. Nonlinear functions with variable parameters are directly summed. The KAN-MLP network model consists of a four-layer KAN component, with an input layer containing two nodes and hidden layers containing 8, 16, and 8 nodes, respectively. The MLP component consists of two layers, each with 16 neurons. The input to each neuron is the output of the KAN component. The output layer consists of one neuron.
[0050] The basic function is achieved by: ;
[0051] Obtain, among which, As the basic function, is the characteristic value of the ambient temperature and humidity or cooling pipe temperature data, is the Sigmoid activation function;
[0052] The spline function is obtained by: ;
[0053] Obtain, among which, is the spline function, is the adjustable parameter of the model, is the kth (k+1th order) basis function.
[0054] Furthermore, the active disturbance rejection control algorithm processes the setpoint dynamics through a tracking differentiator, and an extended observer actively estimates the disturbance; wherein the tracking differentiator is implemented by:
[0055] Implementation, where is the sampling time, is the convergence speed, is the input set value curve, To track the output of the differentiator, To track the intermediate variables in the differentiator;
[0056] The expanded observer is constructed by:
[0057] get, is the error between the estimated and actual measurement of the observed variable by the augmented observer, where is the estimated value of the system state at step k, is the actual output of the system, 、 、 is the system variable observed by the extended observer, 、 、 、 is the adjustable parameter of the extended observer, is the sampling time of the current system, For control quantity The coefficient of influence on the system output.
[0058] Specifically, when the parameters of the extended observer are selected appropriately, The original system state variables The estimated The formula is: ; in In the function The function is ;
[0059] The nonlinear feedback is achieved by: ;
[0060] Implementation, where is the output of the nonlinear feedback output, and are the parameters of the nonlinear function, is the feedback gain, is the tracking error.
[0061] Specifically, through Figure 2 It can be seen that the tracking differentiator (TD) smoothes the input signal v0 into the reference trajectory v1 and its differential v2, and the nonlinear state error feedback (NL SEF) generates the control quantity u based on the errors e1 and e2 ∗ After the controlled object outputs y, the extended observer (ESO) estimates the system states z1, z2 and disturbance z3 in real time, and completes dynamic compensation through the proportional links 1 / b and b to form an anti-disturbance closed loop.
[0062] The improved Kalman filter algorithm is achieved by: ;
[0063] Implementation, where To predict the value, The system measures data. is the Kalman gain in the Kalman filter, is the observation matrix.
[0064] Furthermore, the dual-computing unit architecture module includes a dew point calculation unit and a controller unit;
[0065] The dew point calculation unit performs dew point prediction based on the IMX6ULL processor and the Linux system, and the controller unit completes temperature closed-loop control based on the STM32 microcontroller and the FreeRTOS real-time system. The two achieve highly reliable data interaction through a customized serial port protocol.
[0066] Specifically, through Figure 3 As can be seen, the temperature and humidity sensors are connected to the IMX6ULL via the I2C bus, monitoring the environmental parameters around the cooling pipe in real time. The PC host communicates with the IMX6ULL via the UDP protocol, supporting remote data exchange. The IMX6ULL sends processed environmental data to the STM32 via the UART serial port. The STM32 drives the water pump based on the received data and monitors the actual temperature of the cooling pipe using a thermocouple temperature sensor. It also provides real-time feedback on the water flow rate and pipe wall condition, forming a closed-loop control circuit.
[0067] In terms of data communication, the data frame sent from the dew point calculation unit to the cooling pipe controller consists of a char data type and a float data type, with a header of 0xA0 and a trailer of 0xD0 to ensure data integrity and correct parsing. The reverse data transmission format consists of two float data types, a header, and a trailer.
[0068] During system operation, the sensor data acquisition thread and the model inference thread work together to ensure that the dew point temperature calculation is updated in real time and accurately transmitted to the controller unit, enabling more precise cooling pipe temperature regulation. DMA technology is used for data transmission, effectively reducing CPU load and data transmission latency, thereby improving system response speed and stability.
[0069] To optimize serial port data transmission efficiency, the system organizes data using a union nested within a structure. Structures store the data to be sent, and unions contain both structures and byte arrays of equal size. Because all members of a union share the same storage space, the structure data can be converted into an 8-bit data stream by accessing the byte array members and sent byte by byte to the serial port. The receiving end uses the same data organization to ensure correct data parsing and restoration, thereby improving transmission stability.
[0070] During data transmission, the system uses DMA (Direct Memory Access) technology to further reduce CPU load and improve transmission efficiency. The DMA controller (DMAC) triggers an interrupt when the transmission is complete or an exception occurs, and sends a notification to the CPU. When the complete data frame is transmitted, the DMA controller triggers an interrupt. The system clears the flag bit in the interrupt service routine and performs frame header and frame footer checks to ensure data integrity. The DMA data transmission process is as follows: Figure 4 shown.
[0071] On the Linux side, the system uses multi-threading technology to optimize CPU resource utilization and ensure the real-time performance of dew point calculation. When the system is running, two threads are started: one for sensor data reading and the other for model inference.
[0072] The sensor data reading thread is responsible for collecting ambient temperature and humidity data. During operation, this thread enters kernel mode and obtains temperature and humidity data through the IIC driver. The collected data undergoes RCR verification to ensure its accuracy. The data transmitted by the IIC is then parsed through bit operations and passed to the application layer.
[0073] The model inference thread's primary task is to execute the NCNN inference model to calculate the dew point temperature. This thread is initially blocked and awakens when the sensor thread provides temperature and humidity data to perform inference. Upon receiving the data, the NCNN model performs forward calculations to derive the dew point temperature. The result is transferred to kernel state via the write function and then sent to the cooling pipe temperature controller via the serial port driver. After the data is sent, the model inference thread re-enters the blocked state, awaiting new data input.
[0074] On the controller side, the system primarily monitors and regulates the temperature of the air cooler's cooling pipes. An SPI driver is used to read data from the sensor module, ensuring reliable and efficient data transmission. This data is processed using an improved Kalman filter to reduce noise and mutations, improving data stability and accuracy.
[0075] The filtered data is passed to another task via a global variable, which uses it as input for the Active Disturbance Rejection Control (ADRC) algorithm. Simultaneously, the system retrieves setpoints via DMA interrupt processing and stores and shares them in global variables to ensure real-time and stable operation. This data transfer method reduces CPU load and improves system responsiveness.
[0076] After the ADRC algorithm completes its calculations, the system generates a control output signal that drives the actuator. Specifically, this control value adjusts the duty cycle of the PWM signal to control the motor speed and precisely regulate the cooling pipe temperature. This closed-loop control system ensures that the cooling pipe temperature remains above the dew point, effectively preventing condensation and improving system stability and dynamic response.
[0077] An efficient and self-adjustable air cooling control method, comprising:
[0078] S1, a data collection step for collecting corresponding data of the condensation temperature of the cooling pipe in the air cooler and the ambient temperature and humidity;
[0079] S2, a data processing step for optimizing the smoothness of the corresponding data of the condensation temperature of the cooling pipe in the air cooler and the ambient temperature and humidity by storing the mean of the historical data of the previous d moments using an improved Kalman filter algorithm;
[0080] S3, a dew point prediction step, for inputting the corresponding data of the ambient temperature and humidity based on a KAN-MLP neural network model, and outputting a dew point temperature prediction value; the KAN-MLP model includes a fusion structure of a KAN network and an MLP network, wherein the KAN network uses an activation function that superimposes a spline function and a basis function; during the model training phase, the predicted value and the condensation temperature of the cooling pipe in the air cooler are used as the true value to perform error calculation and back propagate, thereby optimizing the network parameters;
[0081] S4, dual computing unit architecture step, for exchanging data through a serial communication protocol to achieve separate deployment of dew point prediction and cooling pipe temperature control;
[0082] S5, a control step for adopting an active disturbance rejection control algorithm, taking the dew point temperature prediction value as a set value, using the output of the improved Kalman filter as the output of the controlled system, and outputting a control signal through nonlinear feedback;
[0083] S6. Execute the step of adjusting the temperature of the cooling pipe according to the control signal and using PWM to drive the motor to achieve closed-loop control.
[0084] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0085] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.
Claims
1. An efficient and self-adjustable air cooling control system, characterized in that: include: A data acquisition module is used to collect corresponding data of the condensation temperature of the cooling pipe in the air cooler and the ambient temperature and humidity; a data processing module for optimizing the smoothness of the corresponding data of the condensation temperature of the cooling pipe in the air cooler and the ambient temperature and humidity by storing the mean of the historical data of the previous d moments using an improved Kalman filter algorithm; A dew point prediction module is configured to input the corresponding data of the ambient temperature and humidity and output a predicted dew point temperature value based on a KAN-MLP neural network model; the KAN-MLP model comprises a fusion structure of a KAN network and an MLP network, wherein the KAN network uses an activation function that is a superposition of a spline function and a basis function; during the model training phase, the predicted value is compared with the condensation temperature of the cooling pipe in the air cooler as the true value, and the error is calculated and back-propagated to optimize the network parameters; Dual computing unit architecture module, used to exchange data through serial communication protocol, realizing the separation deployment of dew point prediction and cooling pipe temperature control; A control module is configured to adopt an active disturbance rejection control algorithm, use the dew point temperature prediction value as a set value, use the output of the improved Kalman filter as the output of the controlled system, and output a control signal through nonlinear feedback; The execution module adjusts the temperature of the cooling pipe according to the control signal and uses PWM to drive the motor to achieve closed-loop control.
2. The efficient and self-adjustable air cooling control system according to claim 1, characterized in that: The basic function is achieved by: ； Obtain, among which, As the basic function, is the characteristic value of the ambient temperature and humidity or cooling pipe temperature data, is the Sigmoid activation function; The spline function is obtained by: ； Obtain, among which, is the spline function, is the adjustable parameter of the model, is the kth (k+1th order) basis function.
3. The efficient and self-adjustable air cooling control system according to claim 2, characterized in that: The active disturbance rejection control algorithm processes the setpoint dynamics through a tracking differentiator, and the extended observer actively estimates the disturbance; wherein the tracking differentiator is implemented by: ； Implementation, where is the sampling time, is the convergence speed, is the input set value curve, To track the output of the differentiator, To track the intermediate variables in the differentiator; The expanded observer is constructed by: ； get, is the error between the estimated and actual measurement of the observed variable by the augmented observer, where is the estimated value of the system state at step k, is the actual output of the system, 、 、 is the system variable observed by the extended observer, 、 、 、 is the adjustable parameter of the extended observer, is the sampling time of the current system, For control quantity The coefficient of influence on the system output.
4. The efficient and self-adjustable air cooling control system according to claim 3, characterized in that: The nonlinear feedback is achieved by: ； Implementation, where is the output of the nonlinear feedback output, and are the parameters of the nonlinear function, is the feedback gain, is the tracking error.
5. The efficient and self-adjustable air cooling control system according to claim 4, characterized in that: The improved Kalman filter algorithm is achieved by: ； Implementation, where To predict the value, The system measures data. is the Kalman gain in the Kalman filter, is the observation matrix.
6. The efficient and self-adjustable air cooling control system according to claim 5, characterized in that: The dual calculation unit architecture module includes a dew point calculation unit and a controller unit; The dew point calculation unit performs dew point prediction based on the IMX6ULL processor and the Linux system, and the controller unit completes temperature closed-loop control based on the STM32 microcontroller and the FreeRTOS real-time system. The two achieve highly reliable data interaction through a customized serial port protocol.
7. An efficient and self-adjustable air cooling control method, characterized in that: include: S1, a data collection step for collecting corresponding data of the condensation temperature of the cooling pipe in the air cooler and the ambient temperature and humidity; S2, a data processing step for optimizing the smoothness of the corresponding data of the condensation temperature of the cooling pipe in the air cooler and the ambient temperature and humidity by storing the mean of the historical data of the previous d moments using an improved Kalman filter algorithm; S3, a dew point prediction step, for inputting the corresponding data of the ambient temperature and humidity based on a KAN-MLP neural network model, and outputting a dew point temperature prediction value; the KAN-MLP model includes a fusion structure of a KAN network and an MLP network, wherein the KAN network uses an activation function that superimposes a spline function and a basis function; during the model training phase, the predicted value and the condensation temperature of the cooling pipe in the air cooler are used as the true value to perform error calculation and back propagate, thereby optimizing the network parameters; S4, dual computing unit architecture step, for exchanging data through a serial communication protocol to achieve separate deployment of dew point prediction and cooling pipe temperature control; S5, a control step for adopting an active disturbance rejection control algorithm, taking the dew point temperature prediction value as a set value, using the output of the improved Kalman filter as the output of the controlled system, and outputting a control signal through nonlinear feedback; S6, executing step, adjusting the temperature of the cooling pipe according to the control signal, and using PWM to drive the motor to achieve closed-loop control.