Freezer compressor modeling method based on state space construction

Abstract

The invention belongs to the technical field of compressor modeling, and discloses a refrigerator compressor modeling method based on state space construction, which comprises the following steps of: preparing and preprocessing data, acquiring time sequence numbers of input variables and output variables, and performing standardization processing on the time sequence numbers; defining a model structure, constructing a continuous time nonlinear state space model, modeling a state derivative function by adopting a multi-layer sensor (MLP), modeling output mapping by adopting a shallow neural network, and introducing a device embedded vector to improve generalization ability; a state track is solved, and the state track is solved through a four-order Runge-Kutta method; introducing a physical regular term, and adding an energy conservation constraint into a loss function; and training and testing the model, training by adopting an Adam optimizer, and verifying the performance of the model. According to the method, the physical constraint and the neural network fitting capability are combined, the dynamic performance, precision and generalization of the model are remarkably improved, and the method is suitable for optimization control and energy efficiency improvement of the refrigerator compressor.

Description

Technical Field

[0001] The present invention belongs to the technical field of compressor modeling, and in particular relates to a refrigerator compressor modeling method based on state space construction. Background Art

[0002] The compressor is a core component in the refrigeration system, and its performance directly affects the stability and temperature control accuracy of the entire freezer. Therefore, mathematical modeling of the compressor and optimizing its control scheme are crucial for improving freezer energy efficiency. Existing freezer compressor modeling methods primarily include empirical and data-driven models. Empirical models typically include performance curve modeling and simplified curve modeling, while data-driven modeling is primarily achieved through the design of neural networks.

[0003] Common approaches to modeling refrigerator compressors include empirical models derived from thermodynamic or kinetic equations or fitted using experimental test data from compressor manufacturers. These models ignore the compressor's dynamic behavior over time and make it difficult to predict continuous responses. Data-driven models, on the other hand, are fitted using large amounts of data but often ignore physical constraints such as thermodynamics and energy conservation. This can lead to misjudgments at boundary conditions or outside the data distribution.

[0004] At the same time, the two commonly used modeling methods have poor generalization on different models of refrigerator compressors and the cost of model reconstruction is high. Summary of the Invention

[0005] The purpose of the present invention is to provide a refrigerator compressor modeling method based on state space construction to solve the technical problems of poor dynamic performance, insufficient model accuracy and weak generalization performance in existing compressor modeling technology.

[0006] To solve the above technical problems, the specific technical solution of the refrigerator compressor modeling method based on state space construction of the present invention is as follows:

[0007] A refrigerator compressor modeling method based on state space construction includes the following steps:

[0008] Step 1: Data preparation and preprocessing;

[0009] Step 2: Model structure definition;

[0010] Step 3: Solve the state trajectory;

[0011] Step 4: Introduce physical regularization terms;

[0012] Step 5: Model training and testing.

[0013] Furthermore, the step 1 includes the following steps:

[0014] Time series data for the following variables were collected from the freezer system:

[0015] Input variables u(t): including suction temperature T in , condensation temperature T cond , voltage frequency f, ambient temperature T env ;

[0016] Output variable y(t): exhaust temperature T out , compression power P, cooling capacity Q;

[0017] Sampling interval Δt: fixed sampling time step;

[0018] Remove outliers and noise, standardize all variables, and finally divide the data set.

[0019] Furthermore, the step 2 includes the following steps:

[0020] A data model for the continuous-time nonlinear state space structure of a refrigerator is proposed, which can be expressed as:

[0021]

[0022] in:

[0023] Indicates the hidden status inside the compressor;

[0024] represents the control input;

[0025] Indicates the measurement output;

[0026] f θ Represents the state derivative function, modeled by the multilayer perceptron MLP;

[0027] g φ Represents the state-output mapping, modeled by a shallow neural network;

[0028] f θ and g φ The structure is defined as follows:

[0029] f θ Represents a state evolution network composed of a 3-layer MLP network, whose input consists of x(t) and u(t), and the output is

[0030] g φ Represents the output mapping network, the input is x(t), and the output is the predicted y(t).

[0031] Furthermore, the step 2 proposes a parameter sharing mechanism, introducing the device embedding vector ed , the final state evolution function becomes:

[0032]

[0033] Furthermore, the step three includes the following steps:

[0034] First, after obtaining the state evolution function, the device embedding is concatenated with the original input to obtain the new input z(t):

[0035] z(t)=concat(x(t),u(t),e d )

[0036] Second, the MLP neural network uses this expanded input to model the state derivative, expressed as:

[0037]

[0038] Finally, the updated state trajectory is numerically solved by using the fourth-order Runge-Kutta method. The state trajectory derivation formula becomes:

[0039]

[0040] Each of the k i All contain device embedded e d .

[0041] Furthermore, the step 4 includes the following steps:

[0042] The energy conservation constraint is introduced into the loss function, which is defined as follows:

[0043]

[0044] Among them, the prediction error term is expressed as:

[0045]

[0046] The physical constraint term is expressed as:

[0047]

[0048] Where η is the efficiency coefficient, Q in / out is the thermal power, λ phy Controls the intensity of the influence of physical items in training.

[0049] Furthermore, the step five includes the following steps:

[0050] The model is trained using the Adam optimizer and the variable learning rate strategy. After initializing x0, the ODE deduction is performed on each time series to solve x t, through g φ Decode it into y(t), compare it with the true value, and finally calculate the loss function and perform backpropagation until the model converges;

[0051] Finally, the generalization performance and model accuracy are verified on simulated datasets.

[0052] The refrigerator compressor modeling method based on state space construction of the present invention has the following advantages:

[0053] The present invention enables the model to have both physical constraint capabilities and neural network fitting capabilities by constructing a continuous-time nonlinear state space structure, designing a parameter sharing mechanism, and introducing physical regularization terms to improve the loss function. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a flow chart of the refrigerator compressor modeling method based on state space construction of the present invention. DETAILED DESCRIPTION

[0055] In order to better understand the purpose, structure and function of the present invention, the following is a further detailed description of a refrigerator compressor modeling method based on state space construction of the present invention in conjunction with the accompanying drawings.

[0056] like Figure 1 As shown, a refrigerator compressor modeling method based on state space construction of the present invention includes the following steps:

[0057] Step 1: Data preparation and preprocessing

[0058] Time series data for the following variables were collected from the freezer system:

[0059] Input variables u(t): including suction temperature T in , condensation temperature T cond , voltage frequency f, ambient temperature T env .

[0060] Output variable y(t): exhaust temperature T out , compression power P, cooling capacity Q.

[0061] Sampling interval Δt: fixed sampling time step.

[0062] Remove outliers and noise, standardize all variables, and finally divide the data set.

[0063] Step 2: Model structure definition

[0064] The present invention proposes a data model for the continuous-time nonlinear state space structure of a refrigerator, which can be specifically expressed as:

[0065]

[0066] in:

[0067] Indicates the hidden status inside the compressor;

[0068] represents the control input;

[0069] Indicates the measurement output;

[0070] f θ represents the state derivative function, modeled by a multilayer perceptron (MLP);

[0071] gφ represents the state-output mapping, which is modeled by a shallow neural network.

[0072] Furthermore, f θ and g φ The structure is defined as follows:

[0073] f θ Represents a state evolution network composed of a 3-layer MLP network, whose input consists of x(t) and u(t), and the output is

[0074] g φ Represents the output mapping network, the input is x(t), and the output is the predicted y(t).

[0075] Aiming at the problem of insufficient generalization ability of traditional modeling methods, this paper proposes a parameter sharing mechanism by introducing the device embedding vector e d , so that when modeling a new device, only the embedding vector and model parameters need to be fine-tuned to complete the adaptation without the need for overall retraining. The final state evolution function becomes:

[0076]

[0077] Step 3: Solving the state trajectory

[0078] First, after obtaining the state evolution function, the device embedding is concatenated with the original input to obtain the new input z(t):

[0079] z(t)=concat(x(t),u(t),e d )

[0080] Second, the MLP neural network uses this expanded input to model the state derivative, which can be expressed as:

[0081]

[0082] Finally, the updated state trajectory is numerically solved by using the fourth-order Runge-Kutta method. The state trajectory derivation formula becomes:

[0083]

[0084] Each of the k i All contain device embedded e d .

[0085] Step 4: Introducing physical regularization terms

[0086] To enhance the physical consistency of the model, this paper improves the traditional MLP loss function by introducing an energy conservation constraint into the loss function, which can be defined as follows:

[0087]

[0088] Among them, the prediction error term can be expressed as:

[0089]

[0090] The physical constraints can be expressed as:

[0091]

[0092] Where η is the efficiency coefficient, Q in / out is the thermal power. phy Controls the intensity of the influence of physical items in training.

[0093] Step 5: Model training and testing

[0094] The model is trained by using Adam optimizer and variable learning rate strategy. After initializing x0, ODE deduction is performed on each time series to solve x t , through g φ Decoded into y(t), compared with the true value, and finally calculated the loss function and backpropagated until the model converges.

[0095] Finally, the generalization performance and model accuracy are verified on simulated datasets.

[0096] By comparing the physical information model, the traditional neural network model and the method of the present invention on a unified data set, it can be seen that the mean square error (MSE), R 2 and mean absolute percentage error (MAPE), the method proposed in the present invention has certain advantages.

[0097] Table 1 Indicators of different methods on the unified dataset

[0098] method MSE <![CDATA[R 2 ]]> MAPE Traditional neural networks 0.0488 0.9323 5.61％ Physical Information Model 0.0424 0.9420 4.98％ Method of the present invention 0.00247 0.9714 3.28％

[0099] It will be understood that the present invention is described by way of some embodiments, and it will be appreciated by those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are intended to be protected by the present invention.

Claims

1. A refrigerator compressor modeling method based on state space construction, characterized in that: The steps include: Step 1: Data preparation and preprocessing; Step 2: Model structure definition; Step 3: Solve the state trajectory; Step 4: Introduce physical regularization terms; Step 5: Model training and testing.

2. The refrigerator compressor modeling method based on state space construction according to claim 1 is characterized in that: The step 1 includes the following steps: Time series data for the following variables were collected from the freezer system: Input variables u(t): including suction temperature T in , condensation temperature T cond , voltage frequency f, ambient temperature T env ; Output variable y(t): exhaust temperature T out , compression power P, cooling capacity Q; Sampling interval Δt: fixed sampling time step; Remove outliers and noise, standardize all variables, and finally divide the data set.

3. The refrigerator compressor modeling method based on state space construction according to claim 1 is characterized in that: The second step includes the following steps: A data model for the continuous-time nonlinear state space structure of a refrigerator is proposed, which can be expressed as: in: Indicates the hidden status inside the compressor; represents the control input; Indicates the measurement output; f θ Represents the state derivative function, modeled by the multilayer perceptron MLP; gφ represents the state-output mapping, modeled by a shallow neural network; f θ and g φ The structure is defined as follows: f θ Represents a state evolution network composed of a 3-layer MLP network, whose input consists of x(t) and u(t), and the output is g φ Represents the output mapping network, the input is x(t), and the output is the predicted y(t).

4. The refrigerator compressor modeling method based on state space construction according to claim 1 is characterized in that: The second step proposes a parameter sharing mechanism, introducing the device embedding vector e d , the final state evolution function becomes:

5. The refrigerator compressor modeling method based on state space construction according to claim 1 is characterized in that: The step three includes the following steps: First, after obtaining the state evolution function, the device embedding is concatenated with the original input to obtain the new input z(t): z(t)=concat(x(t),u(t),e d ) Second, the MLP neural network uses this expanded input to model the state derivative, expressed as: Finally, the updated state trajectory is numerically solved by using the fourth-order Runge-Kutta method. The state trajectory derivation formula becomes: Each of the k i All contain device embedded e d .

6. The refrigerator compressor modeling method based on state space construction according to claim 1 is characterized in that: The step 4 includes the following steps: The energy conservation constraint is introduced into the loss function, which is defined as follows: Among them, the prediction error term is expressed as: The physical constraint term is expressed as: Where η is the efficiency coefficient, Q in / out is the thermal power, λ phy Controls the intensity of the influence of physical items in training.

7. The refrigerator compressor modeling method based on state space construction according to claim 1 is characterized in that: The step five includes the following steps: The model is trained using the Adam optimizer and the variable learning rate strategy. After initializing x0, the ODE deduction is performed on each time series to solve x t , through g φ Decode it into y(t), compare it with the true value, and finally calculate the loss function and perform backpropagation until the model converges; Finally, the generalization performance and model accuracy are verified on simulated datasets.

Images