Cold plate surface temperature prediction method and system based on RNN-KAN network

Abstract

The invention provides a cold plate surface temperature prediction method and system based on an RNN-KAN network, and the method comprises the steps: 1, building a cold plate surface temperature dynamic test board under multiple working conditions, and collecting the change conditions of the cold plate surface temperature under different working conditions; 2, noise data in the collected data are screened out, and vacancy values are complemented in a quadratic spline interpolation mode; 3, building a prediction model; 4, using a mean square error (MSE) as a loss function, using the training set to train the model, updating the weight and bias of the model through a gradient descent algorithm according to a gradient obtained by back propagation in time, and finally forming a prediction model; and 5, verifying the model by using the test set. The surface temperature of the cold plate can be controlled by effectively predicting the surface temperature of the cold plate, the risk of condensation of the electronic equipment is remarkably reduced, the phenomena of short circuit and corrosion are reduced, and the safety and reliability of the equipment are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of temperature prediction, and in particular to a method and system for predicting the surface temperature of a cold plate based on an RNN-KAN network. Background Art

[0002] With the increasing development of power electronics technology and the rapid improvement of the integration level of electrified equipment, the heat dissipation demand of high-power electronic devices has further increased. In recent years, liquid cooling systems have gradually replaced air cooling systems as the mainstream cooling method due to their high cooling efficiency, small footprint, and low noise. Among them, the indirect liquid cooling solution - cold plate liquid cooling is widely used in the field of industrial production due to its advantages such as simple maintenance and low initial investment. The rapid increase in the heat flux density of electronic devices has led to a significant increase in the heat exchange area required for the indirect liquid cooling system, and the water supply temperature has been continuously reduced, especially in high humidity environments, which may cause serious problems: condensation. In a high humidity environment, if the surface temperature of an object is lower than the dew point temperature of the surrounding environment, condensation is likely to occur. This local low temperature condition causes moisture to condense on the surface of the object to form water droplets. These water droplets not only cause the growth of bacteria and mold, causing hygiene and health problems, but also have a negative impact on the indoor living environment. In the field of industrial production, the condensation problem should not be ignored. It not only accelerates the corrosion of metal structures, but also may cause short circuits or damage to the circuits of electronic devices if the electronic equipment is not adequately insulated, thereby endangering the safe operation and production efficiency of the equipment. Therefore, solving the condensation problem is crucial to ensuring equipment safety and extending its service life.

[0003] Currently, there is limited research on moisture and dew resistance of electronic devices, mainly focusing on the application of waterproof polymer coatings in electronic device thermal management systems under high humidity environments. Although these studies have been proven to effectively reduce the impact of humid environments, the economic feasibility, maintenance requirements and durability of the coatings used in indirectly exposed electronic device thermal management systems have high requirements, making it difficult to be widely used.

[0004] In addition to solving the impact of humid environment on the thermal management system of electronic equipment, keeping the surface temperature of the liquid cold plate always above the safe operating temperature can also effectively prevent the occurrence of condensation. Because the method of controlling the surface temperature of the liquid cold plate does not require additional environmental control equipment, this method is a simpler and more effective anti-condensation measure. Therefore, accurately predicting the surface temperature of the cold plate is crucial to the implementation of effective anti-condensation measures. Since the heat exchange has a certain delay when the liquid cooling system is running, the surface temperature of the cold plate is not only affected by the current environmental conditions, but also related to past environmental changes and cooling water parameters. It is necessary to establish a dynamic prediction model that can reflect time dependence to accurately predict the real-time surface temperature of the cold plate. This multi-dimensional parameter coupled dynamic prediction model can provide accurate data support for the operation monitoring of equipment such as cold plates, thereby helping to prevent condensation, and can also ensure the reliable and stable operation of equipment such as cold plates, thereby improving the efficiency and safety of the entire system. Summary of the invention

[0005] In view of the defects in the prior art, the object of the present invention is to provide a cold plate surface temperature prediction method and system based on RNN-KAN network.

[0006] The cold plate surface temperature prediction method based on the RNN-KAN network provided by the present invention includes:

[0007] Step 1: Establish a dynamic test bench for the cold plate surface temperature under multiple working conditions, and collect the changes in the cold plate surface temperature under different working conditions;

[0008] Step 2: Screen out the noise data in the collected data and fill in the missing values ​​by quadratic spline interpolation;

[0009] Step 3: Build a prediction model, which is a mixture of a recurrent neural network (RNN) and a Kolmogorov-Arnold network (KAN). After the time series data is learned by n neurons in the RNN layer, the output of each neuron is used as the input of the KAN layer, and the final output is obtained by learning the KAN layer again.

[0010] Step 4: Take the mean square error (MSE) as the loss function and use the training set to train the RNN-KAN network. The model uses the gradient descent algorithm to update the weights and biases according to the gradient obtained by back propagation in time, and finally forms a prediction model.

[0011] Step 5: Use the test set to verify RNN-KAN. Take the cooling water inlet temperature, cooling water flow, outdoor temperature and outdoor humidity in the test set as input, predict the cold plate surface temperature as the final output result, and use the mean square error (MSE) and determination coefficient between the predicted value and the actual value as the evaluation indicators of the prediction effect.

[0012] Preferably, step 2 comprises:

[0013] Quadratic spline interpolation is to determine a parabola through three points P1, P2, and P3 that are not on a straight line, where P1 and P3 are the endpoints of the parabola, and P2 is the midpoint. Its parameterized expression is:

[0014] P(t)=a2+a2t+a3t 2 ,0≤t≤1# (1)

[0015] Among them, a1, a2, a3 are parameter coefficients; t is the independent variable of the parabola expression;

[0016] The parabola satisfies: when t=0, the curve segment passes through point P1; when t=1, the curve segment passes through point P3; when t=0.5, the curve passes through point P2 and the tangent vector is P3-P1;

[0017] When the positions of the three points are expressed in coordinate form, the coefficients of the curve segment are derived and substituted back into equation (1) to obtain the final curve expression:

[0018]

[0019] Among them, x1, x2, x3 represent the horizontal coordinates of points P1, P2 and P3 respectively; y1, y2, y3 represent the vertical coordinates of points P1, P2 and P3 respectively;

[0020] According to formula (2), missing values ​​can be filled based on existing data points;

[0021] The processed data are grouped according to the sliding window grouping method, where the window length is K, to form a data set, which is randomly divided into a training set and a test set in a ratio of 8:2.

[0022] Preferably, the step 3 comprises:

[0023] In the RNN layer, the system state calculation formula for the i-th neuron is as follows:

[0024] h t =f(U i X t +W i h t-1 +b i )#(3)

[0025] Among them, h t and h t-1 is the system state of the current and previous time step; X t is the input of the current time step; U i , W i , b iis the weight and bias of layer i, and the weight and bias of a single neuron for inputs at different time steps are the same; f() is the activation function or encapsulated feedforward neural network;

[0026] The KAN layer expresses the activation function as the sum of the SiLU function and the B-spline function, making it parameterizable, and finds the most suitable activation function through learning and training the activation function. Its expression is as follows:

[0027]

[0028] Among them, V is the weight, φ p,i () is h i The corresponding p-th activation function, Φ p () is φ p,i () corresponds to the activation function of the next layer, y is the final prediction output of the KAN network; h i represents the system state of the i-th neuron in the RNN layer at the last moment, which is the input of the KAN layer; B i is a B-spline curve, c i For B i The linear combination coefficient of ; n represents the number of neurons in the RNN layer.

[0029] Preferably, the loss function is expressed as:

[0030]

[0031] Among them, y m is the predicted value, is the actual value, M is the number of samples, m∈M.

[0032] Preferably, the determination coefficient R 2 The expression is:

[0033]

[0034] in, is the mean of the actual values.

[0035] The cold plate surface temperature prediction system based on the RNN-KAN network provided by the present invention includes:

[0036] Module M1: Establish a dynamic test bench for the surface temperature of the cold plate under multiple working conditions to collect the changes in the surface temperature of the cold plate under different working conditions;

[0037] Module M2: Screen out the noise data in the collected data and fill in the missing values ​​by quadratic spline interpolation;

[0038] Module M3: Build a prediction model, which is a mixture of a recurrent neural network (RNN) and a Kolmogorov-Arnold network (KAN). After the time series data is learned by n neurons in the RNN layer, the output of each neuron is used as the input of the KAN layer, and the final output is obtained by learning the KAN layer again.

[0039] Module M4: The mean square error (MSE) is used as the loss function and the training set is used to train the RNN-KAN network. The model uses the gradient descent algorithm to update the weights and biases according to the gradient obtained by back propagation in time, and finally forms a prediction model.

[0040] Module M5: Use the test set to verify RNN-KAN. Take the cooling water inlet temperature, cooling water flow, outdoor temperature and outdoor humidity in the test set as input, predict the cold plate surface temperature as the final output result, and use the mean square error (MSE) and determination coefficient between the predicted value and the actual value as evaluation indicators of the prediction effect.

[0041] Preferably, the module M2 comprises:

[0042] Quadratic spline interpolation is to determine a parabola through three points P1, P2, and P3 that are not on a straight line, where P1 and P3 are the endpoints of the parabola, and P2 is the midpoint. Its parameterized expression is:

[0043] P(t)=a2+a2t+a3t 2 ,0≤t≤1# (1)

[0044] Among them, a1, a2, a3 are parameter coefficients; t is the independent variable of the parabola expression;

[0045] The parabola satisfies: when t=0, the curve segment passes through point P1; when t=1, the curve segment passes through point P3; when t=0.5, the curve passes through point P2 and the tangent vector is P3-P1;

[0046] When the positions of the three points are expressed in coordinate form, the coefficients of the curve segment are derived and substituted back into equation (1) to obtain the final curve expression:

[0047]

[0048] Among them, x1, x2, x3 represent the horizontal coordinates of points P1, P2 and P3 respectively; y1, y2, y3 represent the vertical coordinates of points P1, P2 and P3 respectively;

[0049] According to formula (2), missing values ​​can be filled based on existing data points;

[0050] The processed data are grouped according to the sliding window grouping method, where the window length is K, to form a data set, which is randomly divided into a training set and a test set in a ratio of 8:2.

[0051] Preferably, the module M3 comprises:

[0052] In the RNN layer, the system state calculation formula for the i-th neuron is as follows:

[0053] h t =f(U i X t +W i h t-1 +b i )#(3)

[0054] Among them, h t and h t-1 is the system state of the current and previous time step; X t is the input of the current time step; U i , W i , b i is the weight and bias of layer i, and the weight and bias of a single neuron for inputs at different time steps are the same; f() is the activation function or encapsulated feedforward neural network;

[0055] The KAN layer expresses the activation function as the sum of the SiLU function and the B-spline function, making it parameterizable, and finds the most suitable activation function through learning and training the activation function. Its expression is as follows:

[0056]

[0057] Among them, V is the weight, φ p,i () is h i The corresponding p-th activation function, Φ p () is φ p,i () corresponds to the activation function of the next layer, y is the final prediction output of the KAN network; h i represents the system state of the i-th neuron in the RNN layer at the last moment, which is the input of the KAN layer; B i is a B-spline curve, c i For B i The linear combination coefficient of ; n represents the number of neurons in the RNN layer.

[0058] Preferably, the loss function is expressed as:

[0059]

[0060] Among them, y m is the predicted value, is the actual value, M is the number of samples, m∈M.

[0061] Preferably, the determination coefficient R 2 The expression is:

[0062]

[0063] in, is the mean of the actual values.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] (1) Improved safety: By effectively predicting the surface temperature of the cold plate, the surface temperature of the cold plate can be controlled, significantly reducing the risk of condensation on electronic equipment, reducing short circuits and corrosion, and improving the safety and reliability of the equipment;

[0066] (2) Reduce maintenance costs: By actively controlling the temperature, equipment failures and maintenance requirements caused by condensation are reduced, significantly reducing long-term maintenance costs.

[0067] (3) Extend equipment life: Protecting electronic components from condensation helps extend the service life of equipment and improve the return on investment. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0069] Figure 1 This is a flow chart of the cold plate surface temperature prediction method based on the RNN-KAN network;

[0070] Figure 2 It is a curve chart of actual value and predicted value of cold plate surface temperature changing with time;

[0071] Figure 3 Schematic diagram of KAN network. DETAILED DESCRIPTION

[0072] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0073] Example

[0074] The present invention proposes a dynamic temperature prediction model for the surface of the cold plate for heat dissipation of electronic components. The input variables of the model are ambient temperature, ambient humidity, cold plate cooling water temperature and flow, and liquid cooling system operation time, and the output variable is the surface temperature of the cold plate at this moment. By coupling the ambient temperature, humidity and cooling water status with the operation time of the liquid cooling system, a dynamic temperature prediction model is constructed, which can identify potential condensation risks in advance. The model takes into account the time delay of heat exchange, and can adjust the prediction of the cold plate surface temperature in real time according to previous environmental changes and system operation status, providing more accurate temperature control information. Through this dynamic prediction model with multi-dimensional parameter coupling, accurate data support can be provided for the operation monitoring of cold plates and other equipment.

[0075] The present invention conducts multi-condition experiments on the four input variables of the model respectively, and obtains data under multiple conditions including extreme conditions through scientific and reasonable CFD simulation, and records the temperature changes on the surface of the cold plate within a certain period of time (for example, 10 minutes) under each condition to form a multi-dimensional dynamic condition data set. By conducting multi-condition experiments on the four input variables and combining scientific and reasonable computational fluid dynamics (CFD) simulation, a dynamic data set covering a wide range of working conditions is generated. This method not only improves the accuracy of the model, but also enhances its applicability. By analyzing the temperature changes under different working conditions, the model can generate prediction results of the cold plate surface temperature in real time. Since the heat exchange of the liquid cooling system has a time delay, the model can integrate historical data and real-time data to provide more accurate temperature predictions.

[0076] The model combines the RNN network and the KAN network. The RNN captures the data dependency in the time series through the cyclic structure, while the KAN processes the relationship between parameters, and together they build a prediction model with time dependency. The output is used as the input of the next time step, thereby capturing the time dependency in the sequence. The KAN network uses the spline function as the activation function, which can adaptively adjust the activation mode according to the dynamic characteristics of the time series data. Compared with the traditional fixed activation function, it can more accurately capture the complex patterns and changes in the time series. By combining RNN and KAN, the model can not only capture the time dependency of the time series data, but also handle the complex nonlinear relationship between parameters, thereby more accurately predicting the temperature of the cold plate surface.

[0077] like Figure 1 , specifically including the following steps:

[0078] 1. Establish a test bench and collect data: Establish a dynamic test bench for the cold plate surface temperature under multiple working conditions, where the cooling water inlet temperature varies in the range of 20°C, 25°C and 30°C, the cooling water flow rate is 0.7L / min, 1.0L / min and 1.3L / min, the outdoor temperature is 25°C, 30°C and 35°C, and the outdoor humidity is 60%, 70%, 80% and 90%. A total of 108 groups of experimental conditions were set up to obtain the changes in the cold plate surface temperature within 600s under different working conditions.

[0079] 2. Data processing and grouping: filter out the noise data in the collected data, and fill in the missing values ​​by quadratic spline interpolation. Quadratic spline interpolation is to determine a parabola through three points P1, P2, and P3 that are not on a straight line, where P1 and P3 are the endpoints of the parabola, and P2 is the midpoint. Its parameterized expression is:

[0080] P(t)=a2+a2t+a3t 2 ,0≤t≤1# (1)

[0081] Among them, a1, a2, a3 are parameter coefficients; t is the independent variable of the parabola expression;

[0082] The parabola should satisfy the following conditions: when t=0, the curve segment passes through point P1; when t=1, the curve segment passes through point P3; when t=0.5, the curve passes through point P2 and the tangent vector is P3-P1.

[0083] When the positions of the three points are expressed in coordinate form, the coefficients of the curve segment can be deduced and substituted back into equation (1) to obtain the final curve expression:

[0084]

[0085] Among them, x1, x2, x3 represent the horizontal coordinates of points P1, P2 and P3 respectively; y1, y2, y3 represent the vertical coordinates of points P1, P2 and P3 respectively;

[0086] According to formula (2), missing values ​​can be filled based on existing data points.

[0087] The processed data are grouped according to the "sliding window" grouping method, where the window length is K, to form a data set. The data set is randomly divided into a training set and a test set in a ratio of 8:2.

[0088] 3. Model building: The prediction model is a mixture of recurrent neural network (RNN) and Kolmogorov-Arnold network (KAN). After the time series data is learned by n neurons in the RNN layer, the output of each neuron is used as the input of the KAN layer, and the final output is obtained by learning the KAN layer again. In the RNN layer, the system state calculation formula of the i-th neuron is as follows:

[0089] h t =f(I i X t +W i h t-1 +b i )#(3)

[0090] Among them, h t and h t-1 is the system state of the current and previous time step; X t is the input of the current time step; U i , W i ,

[0091] b i are the weights and biases of layer i. The weights and biases of a single neuron for inputs at different time steps are the same; f() is the activation function or encapsulated feedforward neural network.

[0092] The KAN layer expresses the activation function as the sum of the SiLU function and the B-spline function, making it parameterizable, and finds the most suitable activation function through learning and training the activation function. Its expression is as follows:

[0093]

[0094] Among them, V is the weight, φ p,i () is h i The corresponding p-th activation function, Φ p () is φ p,i () corresponds to the activation function of the next layer, y is the final prediction output of the KAN network; h i Represents the system state of the i-th neuron in the RNN layer at the last moment, which is the input of the KAN layer (the KAN network diagram is shown in Figure 3 );B i is a B-spline curve, c i For B i The linear combination coefficient of ; n represents the number of neurons in the RNN layer.

[0095] 4. Model training: Using mean square error (MSE) as the loss function, the RNN-KAN network is trained using the training set. The model uses the gradient descent algorithm to update weights and biases according to the gradient obtained by back propagation in time, and finally forms a prediction model.

[0096]

[0097] Among them, y m is the predicted value, is the actual value, M is the number of samples, m∈M.

[0098] 5. Model prediction: The RNN-KAN is verified using the above test set. The cooling water inlet temperature, cooling water flow, outdoor temperature and outdoor humidity in the test set are used as input to predict the surface temperature of the cold plate as the final output result. The mean square error (MSE) and the determination coefficient (R) between the predicted value and the actual value are used to calculate the surface temperature of the cold plate. 2 It is an evaluation index of the prediction effect.

[0099]

[0100] in, is the mean of the actual values.

[0101] Prediction results: 108 sets of working condition data were integrated into [41040*30*6] sets of data, 32832 sets of data were randomly selected as training sets, and the rest were test sets. The test results show that the MSE of the test set is 0.30℃ 2 , R 2 is 0.99, indicating that this method can effectively predict the surface temperature of the cold plate. In order to more intuitively demonstrate the prediction effect, an arbitrary set of working conditions (cooling water inlet temperature of 30°C, cooling water flow rate of 1.3L / min, outdoor temperature of 35°C, outdoor humidity of 90%) is selected, and the actual value and predicted value of the cold plate surface temperature over time are shown in Figure 2. Figure 2 shown.

[0102] Those skilled in the art know that, in addition to implementing the system, device and its various modules provided by the present invention in a purely computer-readable program code, it is entirely possible to implement the same program in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers and embedded microcontrollers by logically programming the method steps. Therefore, the system, device and its various modules provided by the present invention can be considered as a hardware component, and the modules included therein for implementing various programs can also be considered as structures within the hardware component; the modules for implementing various functions can also be considered as both software programs for implementing the method and structures within the hardware component.

[0103] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A cold plate surface temperature prediction method based on RNN-KAN network, characterized in that: include: Step 1: Establish a dynamic test bench for the cold plate surface temperature under multiple working conditions, and collect the changes in the cold plate surface temperature under different working conditions; Step 2: Screen out the noise data in the collected data and fill in the missing values ​​by quadratic spline interpolation; Step 3: Build a prediction model, which is a mixture of a recurrent neural network (RNN) and a Kolmogorov-Arnold network (KAN). After the time series data is learned by n neurons in the RNN layer, the output of each neuron is used as the input of the KAN layer, and the final output is obtained by learning the KAN layer again. Step 4: Take the mean square error (MSE) as the loss function and use the training set to train the RNN-KAN network. The model uses the gradient descent algorithm to update the weights and biases according to the gradient obtained by back propagation in time, and finally forms a prediction model. Step 5: Use the test set to verify RNN-KAN. Take the cooling water inlet temperature, cooling water flow, outdoor temperature and outdoor humidity in the test set as input, predict the cold plate surface temperature as the final output result, and use the mean square error (MSE) and determination coefficient between the predicted value and the actual value as the evaluation indicators of the prediction effect.

2. The cold plate surface temperature prediction method based on RNN-KAN network according to claim 1 is characterized in that: The step 2 comprises: Quadratic spline interpolation is to determine a parabola through three points P1, P2, and P3 that are not on a straight line, where P1 and P3 are the endpoints of the parabola, and P2 is the midpoint. Its parameterized expression is: <h2 style=";text-align:left;direction:ltr">P(t) = a2 + a2t + a3t<h2 style=";text-align:left;direction:ltr"> 2 <h2 style=";text-align:left;direction:ltr"> ,0≤t≤1# (1) Among them, a1, a2, a3 are parameter coefficients; t is the independent variable of the parabola expression; The parabola satisfies: when t=0, the curve segment passes through point P1; when t=1, the curve segment passes through point P3; when t=0.5, the curve passes through point P2 and the tangent vector is P3-P1; When the positions of the three points are expressed in coordinate form, the coefficients of the curve segment are derived and substituted back into equation (1) to obtain the final curve expression: Among them, x1, x2, x3 represent the horizontal coordinates of points P1, P2 and P3 respectively; y1, y2, y3 represent the vertical coordinates of points P1, P2 and P3 respectively; According to formula (2), missing values ​​can be filled based on existing data points; The processed data are grouped according to the sliding window grouping method, where the window length is K, to form a data set, which is randomly divided into a training set and a test set in a ratio of 8:

2.

3. The cold plate surface temperature prediction method based on RNN-KAN network according to claim 2 is characterized in that: The step 3 comprises: In the RNN layer, the system state calculation formula for the i-th neuron is as follows: h t =f(U i X t +W i h t-1 +b i )#(3) Among them, h t and h t-1 is the system state of the current and previous time step; X t is the input of the current time step; U i , W i 、b i is the weight and bias of layer i, and the weight and bias of a single neuron for inputs at different time steps are the same; f() is the activation function or encapsulated feedforward neural network; The KAN layer expresses the activation function as the sum of the SiLU function and the B-spline function, making it parameterizable, and finds the most suitable activation function through learning and training the activation function. Its expression is as follows: φ(x)=V(SiLU(x)+Σ i c i B i (x))#(4) Among them, V is the weight, φ p，i () is h i The corresponding p-th activation function, Φ p () is φ p，i () corresponds to the activation function of the next layer, y is the final prediction output of the KAN network; h i represents the system state of the i-th neuron in the RNN layer at the last moment, which is the input of the KAN layer; B i is a B-spline curve, c i For B i The linear combination coefficient of ; n represents the number of neurons in the RNN layer.

4. The cold plate surface temperature prediction method based on RNN-KAN network according to claim 3 is characterized in that: The expression of the loss function is: Among them, y m is the predicted value, is the actual value, M is the number of samples, m∈M.

5. The cold plate surface temperature prediction method based on RNN-KAN network according to claim 4 is characterized in that: The coefficient of determination R 2 The expression is: in, is the mean of the actual values.

6. A cold plate surface temperature prediction system based on RNN-KAN network, characterized in that: include: Module M1: Establish a dynamic test bench for the surface temperature of the cold plate under multiple working conditions to collect the changes in the surface temperature of the cold plate under different working conditions; Module M2: Screen out the noise data in the collected data and fill in the missing values ​​by quadratic spline interpolation; Module M3: Build a prediction model, which is a mixture of a recurrent neural network (RNN) and a Kolmogorov-Arnold network (KAN). After the time series data is learned by n neurons in the RNN layer, the output of each neuron is used as the input of the KAN layer, and the final output is obtained by learning the KAN layer again. Module M4: The mean square error (MSE) is used as the loss function and the training set is used to train the RNN-KAN network. The model uses the gradient descent algorithm to update the weights and biases according to the gradient obtained by back propagation in time, and finally forms a prediction model. Module M5: Use the test set to verify RNN-KAN. Take the cooling water inlet temperature, cooling water flow, outdoor temperature and outdoor humidity in the test set as input, predict the cold plate surface temperature as the final output result, and use the mean square error (MSE) and determination coefficient between the predicted value and the actual value as evaluation indicators of the prediction effect.

7. The cold plate surface temperature prediction system based on RNN-KAN network according to claim 6 is characterized in that: The module M2 comprises: Quadratic spline interpolation is to determine a parabola through three points P1, P2, and P3 that are not on a straight line, where P1 and P3 are the endpoints of the parabola, and P2 is the midpoint. Its parameterized expression is: <h2 style=";text-align:left;direction:ltr">P(t) = a2 + a2t + a3t<h2 style=";text-align:left;direction:ltr"> 2 <h2 style=";text-align:left;direction:ltr"> ,0≤t≤1# (1) Among them, a1, a2, a3 are parameter coefficients; t is the independent variable of the parabola expression; The parabola satisfies: when t=0, the curve segment passes through point P1; when t=1, the curve segment passes through point P3; when t=0.5, the curve passes through point P2 and the tangent vector is P3-P1; When the positions of the three points are expressed in coordinate form, the coefficients of the curve segment are derived and substituted back into equation (1) to obtain the final curve expression: Among them, x1, x2, x3 represent the horizontal coordinates of points P1, P2 and P3 respectively; y1, y2, y3 represent the vertical coordinates of points P1, P2 and P3 respectively; According to formula (2), missing values ​​can be filled based on existing data points; The processed data are grouped according to the sliding window grouping method, where the window length is K, to form a data set, which is randomly divided into a training set and a test set in a ratio of 8:

2.

8. The cold plate surface temperature prediction system based on RNN-KAN network according to claim 7 is characterized in that: The module M3 comprises: In the RNN layer, the system state calculation formula for the i-th neuron is as follows: h t =f(U i X t +W i h t-1 +b i )#(3) Among them, h t and h t-1 is the system state of the current and previous time step; X t is the input of the current time step; U i , W i 、b i is the weight and bias of layer i, and the weight and bias of a single neuron for inputs at different time steps are the same; f() is the activation function or encapsulated feedforward neural network; The KAN layer expresses the activation function as the sum of the SiLU function and the B-spline function, making it parameterizable, and finds the most suitable activation function through learning and training the activation function. Its expression is as follows: φ(x)=V(SiLU(x)+Σ i c i B i (x))#(4) Among them, V is the weight, φ p，i () is h i The corresponding p-th activation function, Φ p () is φ p，i () corresponds to the activation function of the next layer, y is the final prediction output of the KAN network; h i represents the system state of the i-th neuron in the RNN layer at the last moment, which is the input of the KAN layer; B i is a B-spline curve, c i For B i The linear combination coefficient of ; n represents the number of neurons in the RNN layer.

9. The cold plate surface temperature prediction system based on RNN-KAN network according to claim 8, characterized in that: The expression of the loss function is: Among them, y m is the predicted value, is the actual value, M is the number of samples, m∈M.

10. The cold plate surface temperature prediction system based on RNN-KAN network according to claim 9, characterized in that: The coefficient of determination R 2 The expression is: in, is the mean of the actual values.

Images