Evaporator modeling method based on dynamic load model and improved variational auto-encoder

Abstract

The invention relates to the related technical field of evaporators, in particular to an evaporator modeling method based on a dynamic load model and an improved variational auto-encoder, and the method comprises the steps: collecting refrigerator operation data, and carrying out the preprocessing; constructing a dynamic load model according to the internal load change of the refrigerator, and obtaining dynamic load information of the refrigerator; constructing an improved variational auto-encoder model by combining a bidirectional long-short term memory network BiLSTM and a variational auto-encoder, and then inputting dynamic load information to generate reconstruction data; and training and verifying the improved variational auto-encoder model. According to the method, the time sequence dynamic change of the household refrigerator evaporator can be captured, meanwhile, BiLSTM is embedded into a VAE auto-encoder, a traditional variational auto-encoder is improved, a mixed distribution sampling method is designed, the heat exchange characteristics of the evaporator under different loads and environments are accurately modeled, and the modeling accuracy is improved. And the fitting capability of the model to complex load fluctuation and a nonlinear relationship is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field related to evaporators, and in particular to an evaporator modeling method based on a dynamic load model and an improved variational autoencoder. Background Art

[0002] Household refrigerators are essential appliances in daily life. The evaporator, a core component of the refrigerator, plays a crucial role in the cooling process. Therefore, accurately modeling and optimizing this component can effectively improve the refrigerator's energy efficiency, thereby reducing operating energy consumption.

[0003] Existing evaporator modeling methods mostly rely on traditional physical models or fitting predictions based on static data, such as variational autoencoders. These methods struggle to accurately capture the impact of changes in the evaporator's operating state on the model. This is especially true in home use scenarios, where frequent door openings and a wide variety of refrigerated foods lead to significant fluctuations in the evaporator's load and operating state, placing higher demands on evaporator modeling.

[0004] During the use of household refrigerators, the rational use of dynamic changes in data can further improve the accuracy of the evaporator data model, but it also significantly increases the difficulty of modeling.

[0005] Traditional modeling methods based on physical models often struggle to handle the nonlinear effects of dynamic heat load changes, resulting in an inadequate representation of heat exchange characteristics after data fitting. Traditional fitting networks, such as variational autoencoders (VAEs), are less adaptable to modeling refrigerators in household settings due to their limited ability to capture time-series data. This suggests that both methods struggle to adapt to the complex time-series data generated in household settings, reducing modeling accuracy and impacting subsequent device improvements and control strategy optimization. Summary of the Invention

[0006] In order to solve the above technical problems existing in the prior art, the present invention proposes an evaporator modeling method based on a dynamic load model and an improved variational autoencoder. The specific technical solution is as follows:

[0007] An evaporator modeling method based on a dynamic load model and an improved variational autoencoder includes:

[0008] Collect refrigerator operation data and pre-process it;

[0009] According to the load changes inside the refrigerator, a dynamic load model is constructed to obtain the dynamic load information of the refrigerator;

[0010] Combining the bidirectional long short-term memory network BiLSTM with the variational autoencoder, an improved variational autoencoder model is obtained, which is then fed with dynamic load information to generate reconstructed data.

[0011] Train and validate the improved variational autoencoder model.

[0012] Furthermore, the collected operating data include: the internal temperature of the refrigerator T in , external ambient temperature T out , evaporator inlet and outlet pressure P in and P out , wind speed v, refrigerant flow Q cool , number of door openings and door opening duration O door ;

[0013] Preprocessing is to normalize the collected data. The normalization formula is as follows:

[0014]

[0015] Among them, x o is the original data, μ is the mean of the data, and σ is the standard deviation.

[0016] Furthermore, the specific expression of the dynamic load model is as follows:

[0017] L(t)=f1(M(t),T in (t),T out (t),Q cool (t),O door (t)),

[0018] Among them, L(t) represents the load state at the moment, M(t) represents the type and quantity of food stored in the refrigerator, T in (t),T out (t) are the internal and external temperatures of the refrigerator, Q cool (t) represents the refrigerant flow rate, O door (t) indicates the door status, including the number of times the door is opened and the duration of the door opening;

[0019] The dynamic load model captures the dynamic load fluctuations of the household refrigerator evaporator through the multivariable function f1, providing more input features for the subsequent improved variational autoencoder model modeling.

[0020] Furthermore, the bidirectional long short-term memory network BiLSTM encodes the time series data in the dynamic load information to generate a bidirectional hidden state h t , h t The hidden state of the forward LSTM and the backward LSTM at time t and Together, they represent time information from the past to the future, and the expression is as follows:

[0021]

[0022] and The calculation method is:

[0023]

[0024] Among them, x t is the input data at time t, W x and W h is the weight matrix of input and hidden state, and b is the bias term.

[0025] Furthermore, the variational autoencoder has an encoder, a decoder, and a latent space;

[0026] The encoder maps the input data to the latent space and outputs the mean μ(x t ) and standard deviation σ(x t ), both based on the hidden state h of BiLSTM t Calculation yields:

[0027] μ(x t )=f μ (h t )

[0028] σ(x t )=f σ (h t ),

[0029] Among them, f μ (h t ) and f σ (h t ) are neural network functions that calculate the mean and standard deviation of the latent variables through the BiLSTM hidden state;

[0030] The distribution of the latent space is represented by a mixture of Gaussian distributions:

[0031]

[0032] Where K is the number of mixed distributions, π k is the weight of the kth Gaussian distribution, and μ k (x) and are the mean and variance of the kth Gaussian distribution respectively;

[0033] For each latent variable z, we first select a distribution p(k|x)=π from K Gaussian distributions. k , sample from this distribution to generate the latent variable z k :

[0034] z k =μk (x)+σ k (x)e∈ k ,

[0035] in, is noise sampled from a standard normal distribution;

[0036] The final latent variable sampling is:

[0037]

[0038] in is an indicator function that is 1 when the kth Gaussian distribution is selected and 0 otherwise;

[0039] The decoder reconstructs the data from the final latent variable z, generating reconstructed data x' that approximates x.

[0040] Furthermore, the loss function expression of the variational autoencoder is:

[0041]

[0042] in, is the reconstruction error, i.e., the difference between the reconstructed data and the true data;

[0043]

[0044] is the KL divergence of the Gaussian mixture distribution.

[0045] Furthermore, the collected freezer operation data are divided into a training set and a validation set. The training set is used to train the improved variational autoencoder model, and the validation set is used for model validation and performance evaluation through the cross-validation method.

[0046] Furthermore, the model validation adopts a cross-validation method.

[0047] The beneficial effects of this invention are as follows: The evaporator modeling method described herein accurately models the heat exchange process of a household refrigerator evaporator under dynamic loads, improving modeling accuracy and enhancing the refrigerator's energy efficiency. Specifically, this method captures the temporal dynamics of a household refrigerator evaporator and accurately models the evaporator's heat exchange characteristics under varying loads and environments. Compared to traditional methods, this method improves the model's ability to fit complex load fluctuations and nonlinear relationships. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a schematic diagram of the overall process of an evaporator modeling method based on a dynamic load model and an improved variational autoencoder according to this embodiment;

[0049] Figure 2is a data processing flow chart of the improved variational autoencoder of this embodiment;

[0050] Figure 3 This is a comparison diagram of indicators of traditional physical modeling methods;

[0051] Figure 4 This is a comparison diagram of indicators of traditional variational autoencoder network modeling methods;

[0052] Figure 5 3 is a schematic diagram showing the comparison of indicators of the improved variational autoencoder network modeling method of this embodiment. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and technical effect of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0054] like Figure 1 As shown, an evaporator modeling method based on a dynamic load model and an improved variational autoencoder in this embodiment includes the following contents:

[0055] Step 1: Data collection and preprocessing.

[0056] Temperature sensors, humidity sensors, pressure sensors, wind speed sensors, and door opening sensors are installed inside and outside the household refrigerator to collect the following data in real time: the internal temperature of the refrigerator T in , external ambient temperature T out , evaporator inlet and outlet pressure P in and P out , wind speed v, refrigerant flow Q cool , number of door openings and door opening duration O door .

[0057] The collected data is normalized to have a unified dimension range. The normalization formula is as follows:

[0058]

[0059] Among them, x o is the original data, μ is the mean of the data, and σ is the standard deviation.

[0060] Step 2: Dynamic load model building.

[0061] Considering the load changes inside the household refrigerator, including the number of door openings, temperature fluctuations, and changes in the types of refrigerated food, a dynamic load model is constructed, expressed as follows:

[0062] L(t)=f1(M(t),T in (t),T out (t),Qcool (t),O door (t)),

[0063] Among them, L(t) represents the load state at the moment, M(t) represents the type and quantity of food stored in the refrigerator, T in (t),T out (t) are the internal and external temperatures of the refrigerator, Q cool (t) represents the refrigerant flow rate, O door (t) indicates the door status, including the number of times the door is opened and the duration of the door opening.

[0064] The dynamic load model captures the dynamic load fluctuations of the household refrigerator evaporator through the multivariable function f1, providing more input features for subsequent VAE modeling.

[0065] Step 3: Improved variational autoencoder (VAE) modeling.

[0066] Standard variational autoencoder (VAE) models are typically applied to disordered static data and can handle low-dimensional representations of data well. However, for time series data, such as the dynamic load data generated by the household refrigerator evaporator during operation, the VAE model's timing characteristics modeling capability is relatively weak.

[0067] To solve this problem, the present invention combines the bidirectional long short-term memory network (BiLSTM) with VAE, such as Figure 2 As shown in Figure 2, BiLSTM can capture both forward and reverse information flows in time series data and is particularly suitable for processing time-dependent dynamic data, such as load changes and temperature fluctuations of household refrigerator evaporators in different usage scenarios.

[0068] First, when the dynamic load model information in step 2 is input, BiLSTM will encode the time series data and generate a bidirectional hidden state h t .h t The hidden state of the forward LSTM and the backward LSTM at time t and Together, they represent time information from the past to the future, namely:

[0069]

[0070] and The calculation method is:

[0071]

[0072] Among them, x t is the input data at time t, W x and W his the weight matrix of input and hidden state, and b is the bias term.

[0073] Then, the encoder, decoder, and latent space of the variational autoencoder are constructed, and improvements are made to the calculation of the mean and standard deviation. This paper proposes a computational method for embedding BiLSTM into the VAE encoder. In the latent variable sampling process, a mixed distribution sampling method is proposed, and the loss function is improved. Together, these methods are used to process time series data of household refrigerator evaporators, thereby enhancing the VAE's ability to model time dependencies.

[0074] Specifically, the VAE encoder can map the input data to the latent space and output the mean μ(x t ) and standard deviation σ(x t ), both based on the hidden state h of BiLSTM t The calculation result is as follows:

[0075] μ(x t )=f μ (h t )

[0076] σ(x t )=f σ (h t ),

[0077] Among them, f μ (h t ) and f σ (h t ) are the neural network functions that calculate the mean and standard deviation of the latent variables through the BiLSTM hidden state.

[0078] The distribution of the latent space can be represented by a mixture of Gaussian distributions:

[0079]

[0080] Where K is the number of mixed distributions. k is the weight of the kth Gaussian distribution, and μ k (x) and are the mean and variance of the k-th Gaussian distribution respectively.

[0081] For each latent variable z, we first need to select a distribution p(k|x)=π from K Gaussian distributions. k , sample from this distribution to generate the latent variable z k :

[0082] z k =μ k (x)+σ k(x)e∈ k ,

[0083] in, is noise sampled from a standard normal distribution.

[0084] The final latent variable sampling is

[0085]

[0086] in is an indicator function that is 1 if the kth Gaussian distribution is selected and 0 otherwise.

[0087] Next, we construct the decoder part of the VAE. The decoder reconstructs the data from the latent variables z and regenerates the reconstructed data x′ that is approximately x.

[0088] Finally, we design the loss function of VAE. The goal of VAE is to minimize the variational lower bound, that is, to maximize the log-likelihood of the data. However, for the mixed Gaussian sampling method model, the KL divergence needs to be expanded to adapt to the mixed Gaussian distribution. The KL divergence of the mixed Gaussian distribution can be written as:

[0089]

[0090] Therefore, the innovative loss function is:

[0091]

[0092] in, is the reconstruction error, i.e., the difference between the data generated by the model and the true data.

[0093] Step 4: Model training and validation.

[0094] The collected freezer operating data was divided into a training set and a validation set. The training set was used to train the improved VAE model, while the validation set was used for model verification and performance evaluation. Cross-validation was used to verify the model's accuracy and generalization ability, ensuring its applicability under different operating conditions.

[0095] refer to Figures 3 to 5 , and the comparison of the indicators shown in Table 1 below, it can be seen that the indicators of the present invention are more prominent.

[0096] Table 1

[0097] Model MSE RMSE <![CDATA[R 2 ]]> Physical Model 0.049028095099641164 0.2214228874792332 0.634097468358847 Traditional VAE 0.03524766278609412 0.1877436091751038 0.7369424812109437 Improved VAE 0.014113851093556203 0.11880173017913587 0.8946666429555762

[0098] In summary, the method of the present invention proposes a dynamic load model to capture complex dynamic information for this application scenario. At the same time, it improves the traditional variational autoencoder, embeds BiLSTM into the VAE autoencoder, and designs a mixed distribution sampling method, which effectively improves the model's ability to fit complex load fluctuations and nonlinear relationships.

[0099] The above description is only a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the implementation process of the present invention is described in detail above, it is still possible for those familiar with the art to modify the technical solutions described in the above examples or to replace some of the technical features therein with equivalents. Any modifications, equivalent replacements, etc. made within the spirit and principles of the present invention shall be included in the scope of protection of the present invention.

Claims

1. An evaporator modeling method based on a dynamic load model and an improved variational autoencoder, characterized in that: include: Collect refrigerator operation data and pre-process it; According to the load changes inside the refrigerator, a dynamic load model is constructed to obtain the dynamic load information of the refrigerator; Combining the bidirectional long short-term memory network BiLSTM with the variational autoencoder, an improved variational autoencoder model is constructed, and then dynamic load information is input to generate reconstructed data; Train and validate the improved variational autoencoder model.

2. The evaporator modeling method according to claim 1, wherein: The collected operating data include: the internal temperature of the refrigerator T in , external ambient temperature T out , evaporator inlet and outlet pressure P in and P out , wind speed v, refrigerant flow Q cool , number of door openings and door opening duration O door ; Preprocessing is to normalize the collected data. The normalization formula is as follows: Among them, x o is the original data, μ is the mean of the data, and σ is the standard deviation.

3. The evaporator modeling method according to claim 1, wherein: The specific expression of the dynamic load model is as follows: L(t)=f1(M(t),T in (t),T out (t),Q cool (t),O door (t)) Where L(t) represents the load state at time t, M(t) represents the type and quantity of food stored in the refrigerator, and T in (t),T out (t) are the internal and external temperatures of the refrigerator, Q cool (t) represents the refrigerant flow rate, O door (t) indicates the door status, including the number of times the door is opened and the duration of the door opening; The dynamic load model captures the dynamic load fluctuations of the household refrigerator evaporator through the multivariable function f1, providing more input features for the improved variational autoencoder model modeling.

4. The evaporator modeling method according to claim 1, wherein: The bidirectional long short-term memory network BiLSTM encodes the time series data in the dynamic load information to generate a bidirectional hidden state h t , h t The hidden state of the forward LSTM and the backward LSTM at time t and Together, they represent time information from the past to the future, and the expression is as follows: and The calculation method is: Among them, x t is the input data at time t, W x and W h is the weight matrix of input and hidden state, and b is the bias term.

5. The evaporator modeling method according to claim 4, characterized in that: The variational autoencoder has an encoder, a decoder, and a latent space; The encoder maps the input data x to the latent space and outputs the mean μ(x t ) and standard deviation σ(x t ), both based on the hidden state h of BiLSTM t Calculation yields: μ(x t )=f μ (h t ) σ(x t )=f σ (h t ) Among them, f μ (h t ) and f σ (h t ) are neural network functions that calculate the mean and standard deviation of the latent variables through the BiLSTM hidden state; The distribution of the latent space is represented by a mixture of Gaussian distributions: Where K is the number of mixed distributions, π k is the weight of the kth Gaussian distribution, and μ k (x) and are the mean and variance of the kth Gaussian distribution respectively; For each latent variable z, we first select a distribution p(k|x)=π from K Gaussian distributions. k , sample from this distribution to generate the latent variable z k : z k =μ k (x)+σ k (x)e∈ k in, is noise sampled from a standard normal distribution; The final latent variable sampling is: in is an indicator function that is 1 when the kth Gaussian distribution is selected and 0 otherwise; The decoder reconstructs the data from the final latent variable z, generating reconstructed data x' that approximates the input data x.

6. The evaporator modeling method according to claim 5, characterized in that: The loss function expression of the variational autoencoder is: in, is the reconstruction error, i.e., the difference between the reconstructed data and the true data; is the KL divergence of the Gaussian mixture distribution.

7. The evaporator modeling method according to claim 1, wherein: The collected freezer operation data are divided into training set and validation set. The training set is used to train the improved variational autoencoder model, and the validation set is used for model validation and performance evaluation through the cross-validation method.

8. The evaporator modeling method according to claim 7, characterized in that: The model validation adopts the cross validation method.

Images