A Topological Optimization Design Method for the Flow Channel of a Phase Change Heat Storage Device Based on FCN

Abstract

The present invention proposes a method for optimizing the flow channel topology design of a phase change heat storage device based on FCN. Taking the hot water flow channel of a simplified phase change heat storage device as a model, the input sample data information is optimized and iterated by the traditional orthotropic penalty material density method. The multi-model attributes of the device are changed to perform finite element analysis to obtain multiple sets of phase change heat storage material distribution maps and corresponding tensor matrices. The tensor matrices are cut into small block matrices, sampled and processed by the signed distance function. The obtained SDF values and PCM densities are respectively encoded, and the U-net network is trained with the SDF values. Then, the SIMP method is used as an optimizer for the phase change heat storage device model to be optimized, and the model attributes to be optimized are input to indirectly obtain the flow channel topology configuration. By combining the deep learning algorithm with SDF processing, the present invention provides a new method for optimizing the flow channel structure topology configuration of a phase change heat storage device to replace the traditional numerical calculation method.

Description

Technical Field

[0001] The present invention belongs to the technical field of phase change heat storage, and relates to the design of a flow channel structure by combining a fully convolutional neural network, in particular to a method for topologically optimizing the flow channel of a phase change heat storage device based on FCN. Background Art

[0002] In order to make full use of industrial waste heat and waste heat with volatility and intermittency, it is very necessary to develop efficient energy storage technologies. Among energy storage technologies, the phase change heat storage technology has important application prospects and values due to its high phase change latent heat value and small working temperature fluctuation range. In order to efficiently absorb heat, using a fluid loop to exchange heat with a phase change device is a relatively common solution. To enhance the heat transfer performance between the fluid working medium and the phase change heat storage material, the optimization of the flow channel structure of the phase change heat storage device is extremely crucial. Most of the existing flow channel optimization methods follow the ideas of size optimization and shape optimization. Considering the shell radius and channel length, different eccentricities and diameters of the channels, etc., have led to limitations in the optimal design of the flow channel structure.

[0003] The basic principle of topology optimization is to optimize the target performance under given constraints by finding the optimal configuration of the position and quantity of materials inside the structure, with a high degree of freedom. Structural mechanics is the most mature application field of topology optimization theory. In the field of conjugate heat transfer, this method has been applied to the optimal design of flow channel configurations with good results. However, all existing material structure topology optimization design methods require a large number of numerical iterative calculations to obtain the best layout of materials within the design domain. As the number of parameters and the number of iterations of the topology optimization design problem increase, the time cost and memory consumption will increase significantly, resulting in low computational efficiency. Summary of the Invention

[0004] A method for topologically optimizing the flow channel structure of a phase change heat storage device based on a fully convolutional neural network (FCN) introduces deep learning into the topological optimization design of the flow channel structure of the phase change heat storage device. Through deep learning and training of the flow channel topology configuration and the phase change heat storage material density distribution initially optimized by the traditional Solid Isotropic Material with Penalization (SIMP) method, the final heat conduction channel topology configuration is obtained, and a dimensionality reduction model is constructed. The technical solution to achieve the object of the present invention while ensuring the prediction accuracy and improving the speed of the topological optimization design of the flow channel structure of the phase change heat storage device is as follows:

[0005] A method for topologically optimizing the flow channel of a phase change heat storage device based on FCN includes the following steps:

[0006] Step 1: Select a phase change heat storage device, establish a numerical analysis model, apply model properties, and input the SIMP optimizer for iterative calculation. After finite element mesh division analysis, multiple sets of flow channel topological configurations, phase change heat storage material distribution maps, and corresponding multi-channel tensor matrices are obtained. Select the phase change heat storage material density value as the label information characterizing the topological structure.

[0007] Step 2: Regard the background mesh in the finite element analysis in Step 1 as image pixels, preprocess the numerical points obtained from the mesh division, crop and sample the obtained phase change heat storage material distribution map into a numerical matrix of the same shape, and calculate the SDF value of the phase change heat storage material distribution matrix to effectively express geometric information through signed distance function processing.

[0008] Step 3: Divide the SDF matrix values obtained in Step 2 into a training set, a validation set, and a test set. The training set is used to train the U-Net convolutional neural network, the validation set is used to validate the U-Net convolutional neural network, and the test set is used to test the U-Net convolutional neural network.

[0009] Step 4: Construct and train a deep learning network. Based on the U-net network, use the training set obtained in Step 3 to train the model properties in the fully convolutional neural network, and train it with the backpropagation algorithm until the new model converges to obtain the required U-net network.

[0010] Step 5: Input the model properties to be optimized to establish a numerical analysis model of the phase change heat storage device to be optimized. Input the boundary conditions of the flow channel to be optimized into the trained U-Net convolutional neural network. After iterative convergence by the SIMP optimizer, obtain the phase change heat storage material density distribution. If the phase change heat storage material density is 0, it is the topological optimization flow channel region, and the rest are all parts filled with the phase change heat storage material.

[0011] Compared with the prior art, the remarkable advantages of the present invention are:

[0012] (1) The present invention proposes a topological optimization design method for the flow channel structure of a phase change heat storage device based on a fully convolutional neural network (FCN). The topological optimization flow channel configuration is indirectly obtained by predicting the position of the phase change heat storage material through the neural network (if the phase change heat storage material density is 0, it is the topological optimization flow channel region, and the rest are all parts filled with the phase change heat storage material). The traditional SIMP method of orthotropic penalty material density is used for complete optimization iteration. The phase change heat storage material distribution map after iteration is preprocessed, and then the SDF value of the sampled tensor matrix is calculated as the input and training set of the deep learning algorithm to construct a fully convolutional neural network. Finally, the trained network model is used to predict the phase change heat storage material density of the phase change heat storage device model to be optimized, and the SIMP optimizer is used for iteration. Multiple designed model properties are input to obtain the finally optimized topological configuration.

[0013] (2) In the process of processing the data set, the present invention uses the SDF (Signed Distance Function) as the geometric representation at the network input; using the SDF, the geometric shape of the object is represented as a level set function defined on the simulated space in which the object is embedded; compared with ordinary binary images, the SDF provides more physical and mathematical information for the FCN; in the SDF representation method, a zero-level set is created to represent the position of the object boundary; compared with traditional binary processing, the data set processed by the SDF can be better captured by the network for geometric features, improving the prediction accuracy of the FCN network for the optimal topological configuration.

[0014] (3) The deep learning algorithm of the present invention adopts the Adam optimization algorithm, and the elu non-linear activation function is introduced into both the convolutional layer and the deconvolutional layer of the deep learning algorithm, increasing the non-linear ability of the FCN in each convolutional or deconvolutional layer.

[0015] (4) The present invention uses a fully convolutional neural network (FCN) to replace the fully connected layer of the general CNN classification network with a convolutional layer, changing the category probability information from 1D to 2D, that is, realizing the category of each pixel point. Description of the Drawings

[0016] Figure 1 It is a flow chart for the topological optimization design of the flow channel of the phase change heat storage device.

[0017] Figure 2 It is a schematic diagram of the boundary conditions and constraint conditions of the design domain.

[0018] Figure 3 It is a fully convolutional neural network model. Detailed Embodiments

[0019] The following further introduces the present invention in conjunction with the drawings and specific embodiments.

[0020] A method for topological optimization design of the flow channel of a phase change heat storage device based on FCN of the present invention, and the optimization design process of the flow channel of the phase change heat storage device is as Figure 1 shown, and the specific steps are as follows:

[0021] Step 1, establish a numerical model of the phase change heat storage device, obtain various model attributes, and input them into the SIMP optimizer respectively. After finite element analysis, multiple groups of flow channel topological configurations, phase change heat storage material distribution maps, and corresponding multi-channel tensor matrices are obtained. When any one of the model attributes of the channel structure changes, at least one channel of the multi-channel matrix tensor matrix changes accordingly. Select the phase change heat storage material density value as the label information characterizing the topological structure;

[0022] Multiple model properties include boundary conditions (hot water inlet position; hot water outlet position; hot water temperature; adiabatic for other boundaries except hot water inlets and outlets) and constraint conditions (number of finite element meshes; volume fraction of phase change heat storage material; penalty factor of SIMP algorithm; the optimization goal is to minimize the average temperature gradient in the design domain).

[0023] Step 2: Consider the background grid in the finite element analysis in Step 1 as image pixels. To retain the detailed information of topology optimization, perform data preprocessing (mainly denoising and cropping). Sample the obtained phase change heat storage material distribution map (multi-channel tensor matrix) into a numerical matrix of the same shape to achieve the preset size matching of the input and output channels of the corresponding neural network. It also includes unifying the naming of the boundary conditions, constraint conditions, and phase change heat storage material density values, realizing the correspondence between the boundary conditions, constraint conditions, and the corresponding unit density values, and calculating the SDF value (signed distance function processing) of the phase change heat storage material distribution matrix. The geometric information can be effectively expressed through the signed distance function processing.

[0024] The calculation method of the SDF function is to create a zero level set to represent the position of the design domain boundary. That is, the boundary Σ of the object is represented as the zero level set of the continuous level set function φ defined in the domain R 2 represents a two-dimensional bounded domain.

[0025] ∑={X∈R 2 :φ(X)=0}

[0026] The level set function φ(X) is defined everywhere in the domain Ω. X represents a pixel point, and X(x i ,y i ) represents the position of the i-th point in the two-dimensional bounded domain. φ(X) = 0 means X(x i ,y i ) is on the boundary of the object. φ(X) < 0 means X(x i ,y i ) is in the flow channel area; φ(X) > 0 means X(x i ,y i ) is outside the flow channel area. The SDF related to the edge of the flow channel and the level set function φ(X) is defined as:

[0027]

[0028] D(X) is a signed distance function, X2 is a point on the object boundary, and X2(x′,y′) represents the position of point X2 on the boundary. The sign function "sign" is defined as:

[0029]

[0030] Step 3: Obtain the density distribution dataset of the phase change heat storage material. Use the sufficient number of small images obtained after the processing in Step 2 as the dataset and divide it into a training set, a validation set, and a test set according to the ratio of 6:2:2. The training set is the data samples used to train the model; the validation set is the sample set set aside separately during the model training process. It can be used to adjust the hyperparameters of the model and to conduct a preliminary evaluation of the model's capabilities. It is usually used to verify the generalization ability (accuracy, recall rate, etc.) of the current model during iterative training of the model to decide whether to stop further training; the test set is used to evaluate the generalization ability of the final model, but cannot be used as the basis for algorithm-related selections such as tuning parameters and feature selection.

[0031] Step 4: Build and train a deep learning network based on the U-net network, and use the training set to train various parameters in the fully convolutional neural network. Train it with the backpropagation algorithm until the new model converges to obtain the required fully convolutional neural network. In each round of training, all training set samples will participate in updating the network weights, and all validation set samples are only used to evaluate the quality of the network model.

[0032] 4.1 Design the structure of the fully convolutional neural network. The network structure of the fully convolutional neural network mainly consists of an encoding part (Recoding) and a decoding part (Decoding). The encoding part aims to reduce the dimension of the input data and extract the SDF matrix and the latent spatial features between adjacent points in each convolutional layer. In the encoding part of the model, multiple convolutional layers are applied to extract a highly encoded geometric representation from the SDF (the input matrix of the network). The decoding part aims to analyze the highly encoded features and decode them into recognizable density values. In the encoding part of the model, transposed convolutions are used to construct multiple stacked decoding layers and process the encoding results of the encoding layer, so as to decode and generate the distribution of the phase change heat storage material. Design the network structure so that the resolution at the input end of the network is the same as the resolution of the output prediction result of the network.

[0033] 4.2 For the topology optimization structure, each pixel value is different. Therefore, the present invention sets the mean squared error loss function (MSE) error to solve the sum of the squared distances between the predicted value and the true value. The loss function is used to represent the gap between the prediction and the actual data and measure the quality of the model, as shown in the following formula:

[0034]

[0035] In the formula, y i represents the density value of the phase change heat storage material at position i in the density matrix of the phase change heat storage material, represents the predicted value of the density of the phase change heat storage material at position i in the network output numerical matrix.

[0036] 4.3 Introduce the elu non-linear activation function to increase the non-linear ability of the FCN in each convolutional or transposed convolutional layer. In a neural network, the artificially added functional relationship between the output of the upper-layer nodes and the input of the lower-layer nodes is called the activation function. If the activation function is not used in the network, no matter how many layers the neural network has, the convolutional layers are all linear combinations, and then the output and input parts are also linear combinations. This situation will lead to very limited fitting or approximation ability of the designed network. The mathematical expression of the elu non-linear activation function is:

[0037] σ(x) = max(0, x) + min(0, α × (exp(x) - 1))

[0038] In the formula, x refers to the input matrix, and α refers to the set learning rate.

[0039] In the present invention, the phase change material density prediction network reduces the loss function value through the efficient optimization algorithm Adam (an extension of the stochastic gradient descent method); set the maximum number of training epochs, training batch size (Batch size), and learning rate.

[0040] Among them, the maximum number of training epochs represents the number of times a complete data set passes through the neural network. At the same time, during the training process of a complete data set, the data set needs to be divided into several batches. The size of a single batch of data set is the training batch size (batch size). After each batch of network training is completed, the network weights are adjusted according to the final loss function and the set learning rate.

[0041] 4.4 Use the training set to train various parameters in the fully convolutional neural network, and train until the new model converges by the backpropagation algorithm. Input the prepared training set data set into the fully convolutional neural network designed in step 4.1 for training. After inputting the matrix tensor according to the set training batch size and passing it through the forward propagation of the network, a matrix tensor of the same size is output, and then calculate the corresponding gradients of each layer of parameter matrices, and update the parameters from back to front, that is, the gradient iterative update of the backpropagation algorithm. By calculating the loss function of the last layer and continuously updating back layer by layer, the weights of each layer of the network are updated according to the finally output loss function.

[0042] 4.5 In each round of training, all training set samples will participate in updating the network weights. The update of a certain value of the weight is equal to the original value minus the learning rate multiplied by the value of the derivative, as shown in the following formula:

[0043]

[0044] In the formula, ω i represents the matrix obtained by cutting in step two, and the subscript k represents which layer of the network. ω ki ′ is the updated network weight of the i-th matrix in the k-th layer of the network, ωki $W_{i,k}$ represents the original network weights of the $i$-th matrix in the $k$-th layer of the network, and loss represents the loss function, which is a scalar cost function with network parameters as independent variables.

[0045] All validation set samples are only used to output through the network and evaluate the quality of the network model. The model is evaluated based on the difference between the prediction results and the true annotation results of all validation set data; after several rounds of training, the learned network weight parameters are obtained.

[0046] Step 5: Establish a numerical analysis model of the phase change heat storage device to be optimized, input various model attributes to be optimized, use the SIMP optimizer to predict the density of the phase change heat storage material through the U-net network trained in Step 4 to obtain the distribution area of the phase change heat storage material (if the density of the phase change heat storage material is 0, it is the topology optimization flow channel area, and the rest are the parts filled with the phase change heat storage material), and then indirectly obtain the topology optimization flow channel configuration (the density of the phase change heat storage material is zero).

[0047] Embodiment

[0048] A method for topology optimization design of the flow channel of a phase change heat storage device based on a fully convolutional neural network provided in this embodiment is specifically as follows:

[0049] Step 1: Select a phase change heat storage device, such as Figure 2 Establish a numerical analysis model with a design domain size of 100mm×100mm, apply various model attributes (including boundary conditions and constraint conditions), and input the SIMP optimizer for iterative calculation. After finite element mesh division, sample data information, namely multiple groups of flow channel topology configurations and phase change heat storage material distribution diagrams, as well as the corresponding multi-channel tensor matrices, is obtained. When any one of the model attributes of the channel structure changes, at least one channel of the multi-channel matrix tensor matrix changes accordingly. Select the density value of the phase change heat storage material as the label information characterizing the topological structure;

[0050] 1.1 The calculation model of the phase change heat storage device includes the following contents:

[0051] Control equation

[0052] Taking the incompressible N-S fluid equation as the control equation, a one-way coupling solution method is used to solve the conjugate heat transfer problem. The control equation is:

[0053] Continuity equation

[0054]

[0055] In the formula, $u$ is the fluid velocity in the $x$-direction in the two-dimensional coordinate system, and $v$ is the fluid velocity in the $y$-direction in the two-dimensional coordinate system.

[0056] Momentum equation:

[0057]

[0058]

[0059] Without considering other body forces,

[0060] f x = -α(X)u, f y = -α(X)v,

[0061] where ρ is the fluid density, p is the fluid pressure, f x and f y are the body force terms in the x - direction and y - direction respectively. To distinguish the fluid material from the PCM (phase change material), a Brinkman penalty term is introduced, that is, a damping term related to the velocity is added to the momentum equation to simulate the frictional force when the fluid passes through the ideal porous medium. Among them, α(X) represents the osmotic - reverse rate at position X. When α(X)=0, it corresponds to the free - flowing fluid region. When α(X) is infinite, it corresponds to the solid region of the phase - change material.

[0062] Energy equation:

[0063]

[0064] where C p is the specific heat capacity at constant pressure of the fluid, T is the temperature, L is the latent heat of the PCM, k is the thermal conductivity, f is the mass fraction, and its expression is:

[0065]

[0066] where ρ phase1 is the solid - phase density of the PCM, ρ phase2 is the liquid - phase density of the PCM, θ is the phase function related to the temperature, which is used to describe the solid component of the phase - change material. It is defined as follows:

[0067]

[0068] Its function is to take the phase - change temperature Tm as the center and regard the phase - change process as a continuous change within a certain temperature range.

[0069] 1.2 The applied boundary conditions specifically include the following:

[0070] (1) The position of the hot - water inlet;

[0071] (2) The hot - water temperature range is 60 - 80 °C;

[0072] (3) The upper and lower boundaries are adiabatic;

[0073] 1.3 The constraint conditions specifically include the following:

[0074] 1) The number of meshes divided by the finite element method. In the numerical examples, the mesh size is divided into 10 mm, and the number of meshes obtained is 10,000;

[0075] 2) The volume fraction of the phase change heat storage material (the proportion in the entire design domain) is 0.2;

[0076] 3) The penalty factor of the SIMP algorithm is 0.3;

[0077] 4) The optimization goal is to minimize the average temperature gradient within the design domain.

[0078] Step 2: Regard the background mesh in the finite element analysis in Step 1 as image pixels, preprocess (crop) the numerical points obtained by the mesh division, and also include unifying the naming of the boundary conditions, constraint conditions, and the density value of the phase change heat storage material, so as to realize the correspondence between the boundary conditions, constraint conditions, and the corresponding unit density values. For the phase change heat storage material distribution map obtained from the numerical example as the sample data, use Python to sample and crop it into a numerical matrix of n×n, where n is taken as 250, to obtain a small image of 250×250, and calculate its SDF value.

[0079] Step 3: Use a part of the SDF matrix values obtained in Step 2 to train the U-Net convolutional neural network, a part for validating the U-Net convolutional neural network, and the remaining part for testing the U-Net convolutional neural network. For example, the training set, validation set, and test set can be divided in a ratio of 6:2:2.

[0080] Step 4: Build and train a deep learning network based on the U-net network. Use the training set to train various parameters in the fully convolutional neural network, and train them by the backpropagation algorithm until the new model converges to obtain the required fully convolutional neural network U-net model.

[0081] 4.1 Design the structure of the fully convolutional neural network. The structure of the fully convolutional neural network is as Figure 3 shown.

[0082] 4.2 For the topology optimization structure, the value of each pixel point is different. Therefore, the present invention sets the mean square error loss function (MSE) error to solve the sum of the squares of the distances between the predicted value and the true value. As follows:

[0083]

[0084] In the formula, y i represents the density value of the phase change heat storage material at position i in the phase change heat storage material density matrix, represents the predicted value of the density of the phase change heat storage material at position i in the network output numerical matrix.

[0085] During the training process, the data set is continuously input into the U-Net convolutional neural network, and the learning rate of the U-Net convolutional neural network is manually adjusted until the loss function value calculated by the MSE loss function is less than the preset value or converges, then the training of the U-Net convolutional neural network is completed.

[0086] 4.3 Introduce the elu non-linear activation function to increase the non-linear ability of the FCN in each convolutional or deconvolutional layer. The mathematical expression of the Elu non-linear activation function is:

[0087] σ(x) = max(0, x) + min(0, α × (exp(x) - 1))

[0088] In the formula, x refers to the input SDF matrix, and α refers to the set learning rate.

[0089] In this embodiment, Adam is used as the optimizer for updating the weights of the neural network; considering the size of the GPU memory and the required training accuracy, the batch size (i.e., the number of data samples grabbed in one training) is set to 64, the initial learning rate is 0.0001, and the regularization coefficient is 0.00001.

[0090] 4.4 Use the training set to train various parameters in the fully convolutional neural network, and train until the new model converges by the backpropagation algorithm. Input the prepared training set images with a size of 250×250×1 into the fully convolutional neural network designed in step 4.1 for training. Conv is the convolutional layer, Deconv is the deconvolutional layer. The convolutional encoder at the backend of the deep learning algorithm described in the present invention includes 7 convolutional layers, each convolutional layer includes several convolutional kernels, the convolutional decoder includes 5 deconvolutional layers. After each input matrix passes through the forward propagation of the network, an output of 250×250×1 is obtained, and then the loss function value is calculated, as shown in the following table, w, h are the sizes of the convolutional kernels, n F is the number of convolutional kernels, Sride is the convolutional stride, N F is the number of channels, that is, the thickness of the convolutional layer, N w is the width of the convolutional kernel, N l is the length of the convolutional kernel. Starting from an image of 250×250×1, after calculation by the Conv1 convolutional layer, it becomes 83×83×32, and so on. When entering the seventh convolutional layer, deconvolution starts to convert the low-resolution numerical matrix into a high-resolution picture.

[0091]

[0092]

[0093] Table 1 Training process of the fully convolutional network

[0094] 4.5 To prevent overfitting, in each round of training, all training set samples will participate in updating the network weights. The update of a certain value of the weight is equal to the original value minus the learning rate multiplied by the value of the derivative, as shown in the following formula:

[0095]

[0096] In the formula, ω i represents the matrix obtained by cutting in step 2. The subscript k represents which layer of the network. ω ki ′ is the updated network weight of the i-th matrix in the k-th layer of the network. ω ki represents the original network weight of the i-th matrix in the k-th layer of the network. Loss represents the loss function, which is a scalar cost function with network parameters as independent variables.

[0097] All validation set samples are only used to output through the network and evaluate the quality of the network model. The model is evaluated based on the difference between the prediction results and the true annotation results of all validation set data; after several rounds of training, the learned network weight parameters are obtained.

[0098] Step 5: Establish a numerical analysis model of the phase change heat storage device to be optimized. Input various model attributes to be optimized, input the boundary conditions of the flow channel to be optimized into the trained U-Net convolutional neural network, and use the U-net network trained in step 4 by the SIMP optimizer to predict the density of the phase change heat storage material to obtain the distribution area of the phase change heat storage material (if the density of the phase change heat storage material is 0, it is the topologically optimized flow channel area, and the rest are the parts filled with the phase change heat storage material), and then indirectly obtain the topologically optimized flow channel configuration (the density of the phase change heat storage material is zero).

Claims

1. A method for topological optimization design of the flow channel of a phase change heat storage device based on FCN, characterized in that, It includes the following steps: Step 1: Select a phase change heat storage device, establish a numerical analysis model, apply model attributes, input the SIMP optimizer for iterative calculation, and obtain multiple groups of flow channel topological configurations, phase change heat storage material distribution maps, and corresponding multi-channel tensor matrices after finite element mesh division analysis. Select the phase change heat storage material density value as the label information characterizing the topological structure; Step 2: Regard the background grid in the finite element analysis in Step 1 as image pixels, preprocess the numerical points obtained from the grid division, crop and sample the phase change heat storage material distribution map obtained into a numerical matrix of the same shape, and calculate the SDF value of the phase change heat storage material distribution matrix, and effectively express geometric information through signed distance function processing; Step 3: Divide the SDF matrix values obtained in Step 2 into a training set, a validation set, and a test set. The training set is used to train the U-Net convolutional neural network, the validation set is used to verify the U-Net convolutional neural network, and the test set is used to test the U-Net convolutional neural network; Step 4: Construct and train a deep learning network. Based on the U-net convolutional neural network, use the training set obtained in Step 3 to train the model attributes in the fully convolutional neural network, and train it to convergence of a new model by the backpropagation algorithm to obtain the required U-net network; Step 5: Input the model attributes to be optimized to establish a numerical analysis model of the phase change heat storage device to be optimized, input the boundary conditions of the flow channel to be optimized into the trained U-Net convolutional neural network, and obtain the phase change heat storage material density distribution after iterative convergence by the SIMP optimizer. If the phase change heat storage material density is 0, it is the topological optimization flow channel area, and the rest are all parts filled with phase change heat storage materials.

2. The method for optimizing the flow channel topology design of the phase change heat storage device based on FCN according to claim 1, characterized in that The numerical analysis model in Step 1 includes: Continuity equation: In the formula, u is the velocity of the fluid in the x direction in the two-dimensional coordinate system, and v is the velocity of the fluid in the y direction in the two-dimensional coordinate system; Momentum equation Without considering other body forces, f x = -α(X)u, f y = -α(X)v, where γ is the fluid density, p is the fluid pressure, f x and f y are the body force terms in the x and y directions, respectively. Here, α(X) represents the reverse osmosis rate at position X. When α(X) = 0, it corresponds to the free-flowing fluid region, and when α(X) is infinite, it corresponds to the phase change material solid region; Energy equation where C p is the specific heat capacity at constant pressure of the fluid, T is the temperature, L is the latent heat of the PCM, k is the thermal conductivity, and f is the mass fraction: where ρ phase1 is the solid-phase density of the PCM, ρ phase2 is the liquid-phase density of the PCM, and θ is a temperature-dependent phase function.

3. The method for optimizing the flow channel topology design of the phase change heat storage device based on FCN according to claim 2, wherein The phase function θ related to temperature is defined as follows: where T m is the phase transition temperature and ΔT is the temperature increment.

4. The method for optimizing the flow channel topology design of the phase change heat storage device based on FCN according to claim 1, wherein The model attributes in Step 1 include: Boundary conditions: (1) Hot water inlet position; (2) Hot water outlet position; (3) Hot water temperature; (4) Other boundaries except the hot water inlet and outlet are adiabatic; Constraint conditions: (1) Number of finite element division grids; (2) Volume fraction of phase change heat storage materials; (3) Penalty factor of the SIMP algorithm; (4) The optimization goal is to minimize the average temperature gradient in the design domain.

5. The method for optimizing the flow channel topology design of the phase change heat storage device based on FCN according to claim 1, wherein Step 2 calculates the SDF value of the phase change heat storage material distribution matrix, specifically: The sign function "sign" is defined as: where D(X) is a directed distance function, X2 is a point on the object boundary, X2(x′, y′) represents the position of point X2 on the boundary, φ(X) is the level set function, X represents a pixel point, and X(x i , y i ) represents the position of the i-th point in the two-dimensional bounded domain.

6. The method for optimizing the flow channel topology design of the phase change heat storage device based on FCN according to claim 1, wherein Step 4 to construct and train a deep learning network includes the following steps: 4.1 Design the structure of the fully convolutional neural network; 4.2 Set the mean square error loss function; 4.3 Introduce the elu non-linear activation function to increase the non-linear ability of the FCN in each convolutional or deconvolutional layer; 4.4 Use the training set to train various parameters in the fully convolutional neural network, and train it to convergence of a new model by the backpropagation algorithm; 4.5 In each round of training, all training set samples will participate in updating the network weights, and the update of a certain value of the weight is equal to the original value minus the learning rate multiplied by the value of the derivative.

Images