Central air-conditioning system digital twinborn model construction method based on automatic machine learning

Abstract

The invention relates to a central air-conditioning system digital twin model construction method based on automatic machine learning, and the method comprises the steps: selecting at least one alternative scheme of a step from an alternative scheme library corresponding to each model construction step; if yes, at least one hyper-parameter value is selected from the corresponding hyper-parameter range, and a modeling scheme alternative library is formed; a modeling scheme set is generated based on the modeling scheme alternative library, the modeling scheme set has a plurality of modeling scheme samples, and an alternative scheme corresponding to each model construction step is configured in each modeling scheme sample; and respectively constructing digital twin models of the central air-conditioning system based on each modeling scheme sample, calculating predicted output parameters of each digital twin model by combining actual input parameters of each device in the central air-conditioning system, and constructing an objective function based on the predicted output parameters and the actual output parameters to perform optimization so as to determine an optimal modeling scheme. The method is suitable for the field of central air-conditioning system digital twinning modeling.

Description

Technical Field

[0001] The present invention relates to a method for constructing a digital twin model of a central air-conditioning system based on automatic machine learning, which is applicable to the field of digital twin modeling of a central air-conditioning system. Background Art

[0002] The central air-conditioning system centrally produces low-temperature chilled water through chillers, which is delivered to each air-conditioning terminal through chilled water pumps. At the same time, high-temperature cooling water is delivered to the cooling tower through cooling water pumps for heat dissipation, thereby meeting the cooling load requirements of each air-conditioning terminal user.

[0003] Actual central air-conditioning systems usually have characteristics such as numerous equipment, complex topology, and strong thermal coupling, which brings great difficulties to the design and operation of central air-conditioning systems.

[0004] Digital twin technology can accurately reflect the thermal characteristics of the system by constructing a high-fidelity virtual image of the central air-conditioning system in digital space. It is the technical foundation for realizing the planning, design, monitoring, management and optimized operation of the central air-conditioning system, and is becoming a research hotspot in the field of central air-conditioning.

[0005] The data-driven method based on machine learning is most widely used in the digital twin modeling tasks of central air-conditioning systems. Compared with the traditional mechanism-driven modeling method, this method has the advantages of wide applicability and strong learning ability.

[0006] In practical applications, the accuracy of data-driven methods is heavily dependent on the selection of modeling algorithms and corresponding hyperparameters. Existing model selection methods usually use two methods: empirical setting and grid search. The former is based on expert experience for manual selection and setting, while the latter calculates a set of optimal modeling methods by traversing all possible modeling algorithms and corresponding hyperparameters. However, the empirical setting-based method relies too much on the modeling experts' own understanding of data and algorithms, has high knowledge barriers, and the modeling accuracy is usually insufficient; although the grid search method has high accuracy, the computational cost of the traversal process is too high, and it is not operational for the actual more complex modeling process. Summary of the invention

[0007] The technical problem to be solved by the present invention is: in response to the above-mentioned problems, a method for constructing a digital twin model of a central air-conditioning system based on automatic machine learning is provided.

[0008] The technical solution adopted by the present invention is: a method for constructing a digital twin model of a central air-conditioning system based on automatic machine learning, comprising: Select at least one alternative solution for each model building step from the alternative solution library corresponding to the step. If the selected alternative solution has a hyperparameter, select at least one hyperparameter value from the corresponding hyperparameter range to form a modeling solution library. Generate a modeling scheme set based on the modeling scheme candidate library, wherein the modeling scheme set has multiple modeling scheme samples, and each modeling scheme sample has an alternative scheme configured for each model building step. If the alternative scheme has hyperparameters, a hyperparameter value is configured for each hyperparameter; Based on the samples of each modeling scheme, a digital twin model of the central air-conditioning system is constructed respectively. Combined with the actual input parameters of each equipment in the central air-conditioning system, the predicted output parameters of each digital twin model are calculated. Based on the predicted output parameters and the actual output parameters, an objective function is constructed to optimize and determine the optimal modeling scheme.

[0009] The digital twin model modeling process is decomposed into multiple model building steps, including missing value processing step, data standardization step, algorithm selection step and algorithm training step.

[0010] The alternative solutions corresponding to the missing value processing step include: Delete, directly delete the sample sequence with missing values; Forward filling, using the data of the previous time series to fill the missing values; Linear interpolation: Use the data of the previous and next time series to fill in the missing values ​​through linear interpolation.

[0011] Alternative solutions corresponding to the data standardization step include: No standardization is performed, no data standardization operation is performed, and the original data is directly used for training; minimum and maximum standardization is performed, and the data is uniformly converted into numbers in the [0, 1] interval. The calculation method is shown in the following formula: , in, is the original sample data, including input data and output data, and are the maximum and minimum values ​​of this type of data, respectively. The data is after standardization operation; Z-score standardization, the calculation method is as follows: , in, is the mean of this type of data, is the standard deviation of this type of data.

[0012] The alternative solutions corresponding to the algorithm selection step include: For artificial neural networks, hyperparameters include the number of hidden layers, the number of hidden layer nodes, and the activation function.

[0013] The hyperparameters corresponding to the algorithm training step include batch size, learning rate and loss function.

[0014] The particle swarm algorithm based on hierarchical coding is used to optimize the alternative solutions for each model construction step and the corresponding hyperparameters.

[0015] A central air-conditioning system digital twin model construction device based on automatic machine learning, comprising: An alternative library construction module is used to select at least one alternative for each model building step from the alternative library corresponding to the step, and if the selected alternative has hyperparameters, at least one hyperparameter value is selected from the corresponding hyperparameter range to form a modeling solution alternative library; A scheme set building module is used to generate a modeling scheme set based on a modeling scheme candidate library, wherein the modeling scheme set has multiple modeling scheme samples, and each modeling scheme sample is configured with an alternative scheme corresponding to each model building step. If the alternative scheme has hyperparameters, a hyperparameter value is configured for each hyperparameter; The modeling scheme optimization module is used to construct the digital twin model of the central air-conditioning system based on each modeling scheme sample, calculate the predicted output parameters of each digital twin model based on the actual input parameters of each device in the central air-conditioning system, and construct the objective function based on the predicted output parameters and the actual output parameters to optimize and determine the optimal modeling scheme.

[0016] A storage medium stores a computer program that can be executed by a processor, and when the computer program is executed, the steps of the method for constructing a digital twin model of a central air-conditioning system based on automatic machine learning are implemented.

[0017] A central air-conditioning system digital twin model construction device comprises a memory and a processor. The memory stores a computer program that can be executed by the processor. When the computer program is executed, the steps of the central air-conditioning system digital twin model construction method based on automatic machine learning are implemented.

[0018] The beneficial effect of the present invention is: the present invention generates a modeling scheme set based on a modeling scheme alternative library, and constructs a digital twin model based on the modeling scheme set to obtain the predicted output parameters of each model, and then optimizes based on the predicted output parameters and the actual output parameters, searches for the optimal alternative scheme and the corresponding optimal hyperparameters for each step of the central air-conditioning digital twin modeling, and automatically calculates the optimal central air-conditioning system digital twin modeling scheme, which does not rely on the user's understanding of data and algorithms, has a low computational cost, and achieves the optimal balance between computational efficiency and computational accuracy in the central air-conditioning digital twin modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of a method for building a digital twin model of a central air-conditioning system according to an embodiment of the present invention.

[0020] Figure 2 Schematic diagram of input and output parameters of the digital twin model of a chiller according to an embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to better understand the technical solution of the present application, the embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0022] It should be clear that the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0023] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. The singular forms "a", "said" and "the" used in the embodiments of the present application and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.

[0024] Example 1: This example uses a water chiller in a central air conditioning system for illustration. The water chiller is the main cooling source equipment in the central air conditioning system, and its model accuracy is of great significance for the operation and regulation of the central air conditioning system. Figure 1 As shown, this embodiment is a method for constructing a digital twin model of a central air-conditioning system based on automatic machine learning, which specifically includes the following steps: S100. Select at least one alternative for each model building step from the alternative library corresponding to the step. If the selected alternative has hyperparameters, select at least one hyperparameter value from the corresponding hyperparameter range to form a modeling solution library.

[0025] In this embodiment, the digital twin model modeling process is decomposed into multiple model building steps, including missing value processing step, data standardization step, algorithm selection step and algorithm training step.

[0026] The modeling solution library in this example is shown in Table 1 below: Table 1

[0027] The alternatives for the missing value processing step in this embodiment include deletion, forward filling and linear interpolation, where deletion refers to directly deleting the sample sequence with missing values; forward filling refers to filling the missing values ​​with the data of the previous time series; linear interpolation refers to filling the missing values ​​with linear interpolation using the data of the previous and next time series.

[0028] The alternatives for the data normalization step in this example include no normalization, minimum and maximum normalization, and Z-score normalization. No normalization means not using any data normalization operation and directly using the original data for training; minimum and maximum normalization means converting the data uniformly into numbers in the [0, 1] interval, and the calculation method is shown in the following formula: , in, is the original sample data, including input data and output data, and are the maximum and minimum values ​​of this type of data, respectively. is the data after standardization. The calculation method of Z-score standardization is as follows: , in, is the mean of this type of data, is the standard deviation of this type of data.

[0029] In the algorithm selection step of this embodiment, the alternative scheme of the algorithm selection step is an artificial neural network, and the hyperparameters include the number of hidden layers, the number of hidden layer nodes and the activation function.

[0030] The model prediction formula of artificial neural network is as follows:

[0031] in, represents the number of hidden layers, and Represent the weight and bias of the neural network respectively, and their total dimension is equal to the number of nodes in the hidden layer of this layer, and are the input vector and output vector of the model respectively, is the activation function. The calculation formulas of the three types of activation functions are shown as follows: , , .

[0032] The alternative scheme for the algorithm training step in this embodiment is the current common method of training using the input and output parameter training set. The hyperparameters include batch size, learning rate and loss function. The update formula is as follows: , in, is the learning rate, is the batch size, For sample For parameters The gradient of the loss function. In this embodiment, the loss function uses the root mean square error .

[0033] S200. Generate a modeling scheme set based on a library of modeling scheme alternatives. The modeling scheme set includes multiple modeling scheme samples. Each modeling scheme sample is configured with an alternative scheme corresponding to each model building step. If the alternative scheme has hyperparameters, a hyperparameter value is configured for each hyperparameter.

[0034] The following table is a sample modeling solution in this embodiment: Table 2

[0035] S300. Based on each modeling scheme sample, a digital twin model of the central air-conditioning system is constructed respectively. Combined with the actual input parameters of each device in the central air-conditioning system, the predicted output parameters of each digital twin model are calculated. Based on the predicted output parameters and the actual output parameters, an objective function is constructed to perform optimization and determine the optimal modeling scheme.

[0036] In this example, the actual input parameters and actual output parameters of the digital twin model of the chiller are as follows: Figure 2 As shown, the input parameters include chilled water inlet temperature, chilled water flow, cooling water inlet temperature, cooling water flow and chiller chilled water outlet temperature set value, and the output parameters include chilled water outlet temperature, cooling water outlet temperature and chiller power.

[0037] In this embodiment, a particle swarm algorithm based on layered coding is used to optimize the optional methods and corresponding hyperparameters of each modeling step. The specific steps are as follows: S310, set particle swarm size , inertia weight , acceleration factor , , maximum number of iterations and iteration counter ; S320, randomly dividing the modeling solution set into a training set and a validation set in a ratio of 7:3, wherein the training set is used for training and establishing the model, and the validation set is used for evaluating the accuracy of the model; S330, randomly initialize the position vector of each particle in the search space , as shown below: , in, and method selection variables representing missing value handling and data standardization steps, respectively, , and They represent the number of hidden layers, number of hidden layer nodes and activation function variables of the artificial neural network respectively. and They represent the batch size variable and learning rate variable for network training respectively.

[0038] The value range of is {1, 2, 3}, which means deletion, forward filling and linear interpolation respectively; The value range of is {1, 2, 3}, which means no standardization, minimum and maximum standardization, and Z-score standardization respectively; The value range of is {1, 2, 3, 4}; The value range of is {4, 5,..., 32}; The value range of is {1, 2, 3}, which respectively indicates the use of ReLU, sigmoid and tanh activation functions; The value range is {8, 16, 32, 64, 128, 256}; The value range of is [0.0001, 0.5], which is a continuous variable. When the position vector is initialized, the initial value of each variable is randomly generated from the corresponding value range.

[0039] Furthermore, the velocity vector of each particle is randomly initialized , as shown below: , in, is the number of variables to be optimized, which is 7 in this embodiment.

[0040] S340, each particle uses the modeling solution set to perform model training based on the alternative solutions represented by the current position vector and the corresponding hyperparameters, and tests it in the validation set.

[0041] For example, the position variable of a particle in a certain iteration is [3, 2, 2, 16, 3, 64, 0.01]. In this case, linear interpolation is used to fill missing values, minimum and maximum standardization is used for data standardization, the number of hidden layers of the artificial neural network is 2, the number of nodes in each layer is 16, the activation function is tanh, the batch size is 64 during network training, the learning rate is 0.01, and the loss function is .

[0042] After training, the root mean square error of the model under this scheme is calculated in the validation set. , as shown below: , in, is the sample size, is the number of model output parameters, and The samples are No. The actual and predicted values ​​of the output parameters; S350, update the velocity vector of each particle , as shown below: , in, and is a random number in the interval [0, 1], For particles In the iterative process The position vector corresponding to the minimum value, For all particles in the iteration process The position vector corresponding to the minimum value; S360, update the position vector of each particle , the updates for continuous variables and discrete variables are as follows: , , , in, is the hyperbolic tangent function; S370, if the number of iterations reaches Or within a certain number of consecutive iterations If both remain unchanged, the algorithm stops and The characterized alternative solutions and corresponding hyperparameters are output. Otherwise, return to step S330.

[0043] S380. Establish a digital twin model of the chiller based on the optimal modeling solution output in step S370.

[0044] In this embodiment, the objective function uses the average relative deviation to measure the prediction accuracy of the model, as shown in the following formula: , This embodiment uses the traditional method based on experience setting to compare and verify the performance of the method of the present invention. The artificial neural network algorithm is uniformly written in the pytorch library. The optimal modeling parameters automatically optimized by the method of this embodiment are compared with the traditional modeling parameters manually set by the modeling engineer. In the test set, the average relative deviations of the digital twin models of the chillers established by the traditional method and the method of this embodiment are 9.2% and 4.8%, respectively.

[0045] The results show that the method in this embodiment uses the particle swarm algorithm to optimize the modeling method and hyperparameters, and can establish a digital twin model tailored to local conditions according to the characteristics of the actual training data set, effectively improving the prediction accuracy of the model.

[0046] Embodiment 2: This embodiment is a central air-conditioning system digital twin model construction device based on automatic machine learning, including: an alternative library construction module, a solution set construction module and a modeling solution optimization module, etc.

[0047] In this example, the alternative library construction module is used to select at least one alternative for each model construction step from the alternative library corresponding to the step. If the selected alternative has hyperparameters, at least one hyperparameter value is selected from the corresponding hyperparameter range to form a modeling solution alternative library.

[0048] In this embodiment, the solution set construction module is used to generate a modeling solution set based on a modeling solution alternative library. The modeling solution set has multiple modeling solution samples. Each modeling solution sample is configured with an alternative solution corresponding to each model construction step. If the alternative solution has hyperparameters, a hyperparameter value is configured for each hyperparameter.

[0049] In this embodiment, the modeling scheme optimization module is used to construct a digital twin model of the central air-conditioning system based on each modeling scheme sample, calculate the loss function of each digital twin model based on the actual input and output parameters of each device in the central air-conditioning system, and construct an objective function based on the loss function to optimize and determine the optimal modeling scheme.

[0050] Embodiment 3: This embodiment is a storage medium on which a computer program that can be executed by a processor is stored. When the computer program is executed, the steps of the method for constructing a digital twin model of a central air-conditioning system based on automatic machine learning in Embodiment 1 are implemented.

[0051] Example 4: This example is a device for building a digital twin model of a central air-conditioning system, comprising a memory and a processor. The memory stores a computer program that can be executed by the processor. When the computer program is executed, the steps of the method for building a digital twin model of a central air-conditioning system based on automatic machine learning in Example 1 are implemented.

Claims

1. A method for constructing a digital twin model of a central air-conditioning system based on automatic machine learning, characterized in that: include: Select at least one alternative solution for each model building step from the alternative solution library corresponding to the step. If the selected alternative solution has a hyperparameter, select at least one hyperparameter value from the corresponding hyperparameter range to form a modeling solution library. Generate a modeling scheme set based on the modeling scheme alternative library, wherein the modeling scheme set has multiple modeling scheme samples, and each modeling scheme sample has an alternative scheme configured for each model building step. If the alternative scheme has hyperparameters, a hyperparameter value is configured for each hyperparameter; Based on the samples of each modeling scheme, a digital twin model of the central air-conditioning system is constructed respectively. Combined with the actual input parameters of each equipment in the central air-conditioning system, the predicted output parameters of each digital twin model are calculated. Based on the predicted output parameters and the actual output parameters, an objective function is constructed to optimize and determine the optimal modeling scheme.

2. The method for constructing a digital twin model of a central air-conditioning system based on automatic machine learning according to claim 1, characterized in that: The digital twin model modeling process is decomposed into multiple model building steps, including missing value processing step, data standardization step, algorithm selection step and algorithm training step.

3. The method for constructing a digital twin model of a central air-conditioning system based on automatic machine learning according to claim 2, characterized in that: The alternative solutions corresponding to the missing value processing step include: Delete, directly delete the sample sequence with missing values; Forward filling, using the data of the previous time series to fill the missing values; Linear interpolation: Use the data of the previous and next time series to fill in the missing values ​​through linear interpolation.

4. The method for constructing a digital twin model of a central air-conditioning system based on automatic machine learning according to claim 2, characterized in that: Alternative solutions corresponding to the data standardization step include: No standardization is performed, no data standardization operation is performed, and the original data is directly used for training; minimum and maximum standardization is performed, and the data is uniformly converted into numbers in the [0, 1] interval. The calculation method is shown in the following formula: , in, is the original sample data, including input data and output data, and are the maximum and minimum values ​​of this type of data, respectively. The data is after standardization operation; Z-score standardization, the calculation method is as follows: , in, is the mean of this type of data, is the standard deviation of this type of data.

5. The method for constructing a digital twin model of a central air-conditioning system based on automatic machine learning according to claim 2, characterized in that: The alternative solutions corresponding to the algorithm selection step include: For artificial neural networks, hyperparameters include the number of hidden layers, the number of hidden layer nodes, and the activation function.

6. The method for constructing a digital twin model of a central air-conditioning system based on automatic machine learning according to claim 2, characterized in that: The hyperparameters corresponding to the algorithm training step include batch size, learning rate and loss function.

7. The method for constructing a digital twin model of a central air-conditioning system based on automatic machine learning according to claim 1, characterized in that: The particle swarm algorithm based on hierarchical coding is used to optimize the alternative solutions for each model construction step and the corresponding hyperparameters.

8. A device for constructing a digital twin model of a central air-conditioning system based on automatic machine learning, characterized in that: include: An alternative library construction module is used to select at least one alternative solution for each model construction step from the alternative solution library corresponding to the step. If the selected alternative solution has hyperparameters, at least one hyperparameter value is selected from the corresponding hyperparameter range to form a modeling solution alternative library; A scheme set building module is used to generate a modeling scheme set based on a modeling scheme candidate library, wherein the modeling scheme set has multiple modeling scheme samples, and each modeling scheme sample is configured with an alternative scheme corresponding to each model building step. If the alternative scheme has hyperparameters, a hyperparameter value is configured for each hyperparameter; The modeling scheme optimization module is used to construct the digital twin model of the central air-conditioning system based on each modeling scheme sample, calculate the loss function of each digital twin model based on the actual input and output parameters of each device in the central air-conditioning system, and construct the objective function based on the loss function to optimize and determine the optimal modeling scheme.

9. A storage medium having stored thereon a computer program executable by a processor, characterized in that: When the computer program is executed, the steps of the method for constructing a digital twin model of a central air-conditioning system based on automatic machine learning as described in any one of claims 1 to 7 are implemented.

10. A central air conditioning system digital twin model construction device, comprising a memory and a processor, wherein the memory stores a computer program executable by the processor, characterized in that: When the computer program is executed, the steps of the method for constructing a digital twin model of a central air-conditioning system based on automatic machine learning as described in any one of claims 1 to 7 are implemented.

Images