Air conditioner load generation method based on fusion of air conditioner physical model and generative adversarial network

Abstract

The invention provides an air conditioner load generation method integrating an air conditioner physical model and a generative adversarial network, which can realize the dynamic generation capability of the air conditioner load, and the generation capability can change according to the sizes, models, environments and other characteristics of different air conditioners to dynamically adapt to the generation of the air conditioner load. The method supports the precision correction capability of air conditioner load generation, and by utilizing the generative adversarial network and model-driven generation method, on one hand, the constructed physical model can be utilized to perform load prediction, and on the other hand, the model correction can be performed by utilizing a data-driven mode, so that the precision of air conditioner load generation is guaranteed; according to the method, the physical model and the data driving model of the generative adversarial network are combined, physical interpretability is provided, unknown data can be generated, and therefore the requirement for air conditioner load simulation is better met, and the requirement for analysis of air conditioner load characteristics can be better met.

Description

Technical Field

[0001] The present invention belongs to the technical field of air-conditioning load characteristic analysis of commercial and other public buildings, and specifically designs an air-conditioning load generation method that integrates an air-conditioning physical model with a generative adversarial network. Background Art

[0002] The growing air conditioning cooling load has become the main factor for the continuous increase in the peak load of the power grid in summer. In order to guide the air conditioning load with huge resources and flexible regulation to participate in demand response to achieve "source-load interaction" and ensure the safe and economical operation of the power grid, major research units have carried out simulation and practical research on real-time control of building air conditioning demand response. However, how to accurately estimate the load of air conditioning and predict the air conditioning load is an important challenge currently faced. The current mainstream methods include two method systems: model-driven and data-driven. Among them, the model-driven method relies on the modeling of the air conditioning load, and it is difficult to simulate the complex changes of the air conditioning load. Data-driven relies on a large amount of data for model training, but it is difficult to cover the various characteristics of the air conditioning load. To this end, this patent intends to perform intelligent fitting of the air conditioning load from the perspective of combining model-driven and data-driven to improve the accuracy and adaptability of air conditioning load generation prediction. Summary of the invention

[0003] The purpose of the present invention is to provide a new method for generating air conditioning load.

[0004] The present invention adopts the following technical means to achieve: an air conditioning load generation method integrating an air conditioning physical model with a generative adversarial network, comprising the following steps:

[0005] 1) Constructing a basic air conditioning model, which includes a compressor thermal model, an evaporator thermal model and an environmental model of the air conditioning;

[0006] 2) Convert the basic air conditioning model into the polynomial form required for nonlinear regression to obtain the air conditioning load model;

[0007] 3) Using the data set to perform nonlinear optimization on the air conditioning load model to determine the polynomial parameters in the air conditioning load model and obtain the air conditioning load polynomial model;

[0008] 4) Generate forecast load data using the air conditioning load polynomial model;

[0009] 5) The data set is updated according to the predicted load data, and when the data in the updated data set meets the preset conditions, the updated data set is used as the air conditioning load generation data set.

[0010] The data in the updated data set satisfies the preset condition means that when the proportion of real data in the updated data set is greater than or equal to a first preset value, the data in the updated data set satisfies the preset condition, otherwise the data in the updated data set does not satisfy the preset condition.

[0011] In step 5), updating the data set according to the predicted load data refers to using the discriminator network of the training process to evaluate the authenticity of the predicted load data. When the authenticity evaluation result is true, the predicted load data is used to update the data set; otherwise, the air-conditioning load model is re-solved by nonlinear optimization using the data set, and the predicted load data is generated according to the air-conditioning load polynomial model obtained after the re-solution.

[0012] The method further comprises:

[0013] 1) Constructing a generative adversarial network for air conditioning load, wherein the generative adversarial model includes an initial discriminator network and an initial generator network;

[0014] 2) When the data in the updated data set does not meet the preset conditions, the updated data set is used to train the initial generator network and the initial discriminator network to obtain the trained discriminator network and the generator network, and the trained discriminator network is used to perform authenticity evaluation on the load data generated during the training process of the initial generator network. When the authenticity evaluation result is true, the load data generated during the training process of the initial generator network is used to update the data set.

[0015] The method further comprises:

[0016] When the data in the updated data set does not meet the preset conditions, the air conditioning load model is re-solved by nonlinear optimization using the updated data set, and the predicted load data is generated according to the air conditioning load polynomial model obtained after the re-solution.

[0017] Update the dataset with forecasted load data, including:

[0018] The existing data in the data set is merged with the predicted load data to form a new data set.

[0019] Update the dataset using the load data generated during the initial generator network training process, including:

[0020] The existing data in the dataset is merged with the load data generated during the initial generator network training process to form a new dataset.

[0021] The initial generator consists of several deconvolution layers, batch normalization layers, a ReLU activation function and a Tanh activation function;

[0022] The initial discriminator is composed of several convolutional layers, batch normalization layers, a LeakyReLU activation function and a Sigmoid function.

[0023] The basic model of air conditioning is: air =f(P comp ,Q evap ,W comp ),

[0024] Where W comp is the compressor power, W comp =Q evap +Q loss , Q evap is the heat absorbed by the evaporator, Q loss is the heat lost by the system;

[0025] Q evap is the heat transfer coefficient of the evaporator, Q evap =U evap ·A evap ΔT evap , A evap is the heat transfer area of ​​the evaporator, ΔT evap is the temperature difference across the evaporator;

[0026] Where P comp is the environmental impact factor, P comp =α·T ambient +β·RH+γ·v wind , where T ambient is the ambient temperature, RH is the relative humidity, v wind is the wind speed, α, β, γ are model parameters, which quantitatively describe the influence of various environmental factors on the load of the air-conditioning system.

[0027] The air conditioning load model is composed of the P air The nonlinear regression transformation is as follows:

[0028] y=f(x)+∈

[0029] y is the output load value, x is the input feature vector, ∈ is the error term, and f(x) is a nonlinear function; the nonlinear function is a polynomial regression function or an exponential regression function, specifically: Where p is the order of the polynomial, or f(x)=a exp(bx), and the loss function of the nonlinear function is:

[0030]

[0031] The objective function of the generative adversarial network is as follows:

[0032] The loss function of the initial generator network:

[0033] tem is the temperature vector of the input conditions, z is derived from the prior distribution p z is the noise vector sampled in; G(z;θ) is the output of the generator (generated samples); D(·;φ) is the discriminator function used to evaluate the authenticity of the input;

[0034] The loss function of the initial discriminator network is:

[0035]

[0036] θ is the parameter of the generator; φ is the parameter of the discriminator; z is the noise vector, drawn from the prior distribution p z G(z; θ) is the output of the generator (generated samples); D(·; φ) is the discriminator function used to evaluate the authenticity of the input; x is from the real data distribution p data Samples of is a sample from the linear interpolation between the true data distribution and the generator output; λ is the weight parameter of the gradient penalty.

[0037] include:

[0038] An air conditioning load generation module based on a physical model, the module is used to complete load prediction using the air conditioning physical model;

[0039] An air conditioning load generation module based on a generative adversarial network, the module is used to train the generative adversarial network to complete the generation of the load;

[0040] An air conditioning load generation module based on physics and generative adversarial network tuning is used to complete the final generation capacity of the load.

[0041] The air conditioning load generation module based on the generative adversarial network uses the trained discriminator network to discriminate the authenticity of the predicted load data according to the existing data. When the authenticity of the predicted load data meets the requirements of the discriminator network, the predicted load data is output as the final predicted load data.

[0042] The air conditioning load generation module based on physics and generative adversarial network tuning uses the data in the dataset to train the initial generator network and the initial discriminator network, including inputting the data in the dataset into the initial generator network to obtain generated data, inputting the generated data and the data in the dataset into the initial discriminator network, and using the initial discriminator network to discriminate the authenticity of the generated data, until the authenticity of the generated data meets the preset requirements, and obtaining the trained discriminator network.

[0043] A computer device includes a memory and a processor, wherein the memory stores a computer program, and is characterized in that the memory stores the computer program, and when the processor executes the computer program, an air conditioning load generation method that integrates the aforementioned air conditioning physical model and a generative adversarial network is implemented.

[0044] A computer-readable storage medium having a computer program stored thereon, characterized in that the storage medium includes a stored program, wherein when the program is running, the device where the storage medium is located is controlled to execute the aforementioned air conditioning load generation method that integrates an air conditioning physical model with a generative adversarial network.

[0045] The present invention provides a new air-conditioning load generation method, which can realize the dynamic generation capability of air-conditioning load. The generation capability can be changed according to the size, model, environment and other characteristics of different air conditioners, and dynamically adapt to the generation of air-conditioning load. The method supports the accuracy correction capability of air-conditioning load generation. The generation method using a generative adversarial network combined with a model-driven generation method can use the constructed physical model to predict the load on the one hand, and use a data-driven method to correct the model on the other hand, thereby ensuring the accuracy of air-conditioning load generation. The method combines the physical model with the data-driven model of the generative adversarial network, provides physical interpretability, and can generate unknown data, thereby better meeting the needs of air-conditioning load simulation, and can also better support the analysis needs of air-conditioning load characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of the method of the present invention;

[0047] Figure 2 The present invention generates a flow chart based on the air conditioning physical model load;

[0048] Figure 3 The present invention is a flow chart of generating a countermeasure network load based on air conditioning. DETAILED DESCRIPTION

[0049] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0050] The present invention relates to an air conditioning load generation method integrating an air conditioning physical model with a generative adversarial network, comprising the following steps:

[0051] 1) Constructing a basic air conditioning model, which includes a compressor thermal model, an evaporator thermal model and an environmental model of the air conditioning;

[0052] 2) Convert the basic air conditioning model into the polynomial form required for nonlinear regression to obtain the air conditioning load model;

[0053] 3) Using the data set to perform nonlinear optimization on the air conditioning load model to determine the polynomial parameters in the air conditioning load model and obtain the air conditioning load polynomial model;

[0054] 4) Generate forecast load data using the air conditioning load polynomial model;

[0055] 5) The data set is updated according to the predicted load data, and when the data in the updated data set meets the preset conditions, the updated data set is used as the air conditioning load generation data set.

[0056] When the data in the updated data set meets the preset condition, it includes: when the proportion of real data in the updated data set is greater than or equal to a first preset value, it indicates that the data in the updated data set meets the preset condition, otherwise it indicates that the data in the updated data set does not meet the preset condition.

[0057] In step 5), updating the data set according to the predicted load data refers to using the discriminator network of the training process to evaluate the authenticity of the predicted load data. When the authenticity evaluation result is true, the predicted load data is used to update the data set; otherwise, the air-conditioning load model is re-solved by nonlinear optimization using the data set, and the predicted load data is generated according to the air-conditioning load polynomial model obtained after the re-solution.

[0058] The method further comprises:

[0059] 1) Constructing a generative adversarial network for air conditioning load, wherein the generative adversarial model includes an initial discriminator network and an initial generator network;

[0060] 2) When the data in the updated data set does not meet the preset conditions, the updated data set is used to train the initial generator network and the initial discriminator network to obtain the trained discriminator network and the generator network, and the trained discriminator network is used to perform authenticity evaluation on the load data generated during the training process of the initial generator network. When the authenticity evaluation result is true, the load data generated during the training process of the initial generator network is used to update the data set.

[0061] The method further comprises:

[0062] When the data in the updated data set does not meet the preset conditions, the air conditioning load model is re-solved by nonlinear optimization using the updated data set, and the predicted load data is generated according to the air conditioning load polynomial model obtained after the re-solution.

[0063] Update the dataset with forecasted load data, including:

[0064] The existing data in the data set is merged with the predicted load data to form a new data set.

[0065] Update the dataset using the load data generated during the initial generator network training process, including:

[0066] The existing data in the dataset is merged with the load data generated during the initial generator network training process to form a new dataset.

[0067] The initial generator consists of several deconvolution layers, batch normalization layers, a ReLU activation function and a Tanh activation function;

[0068] The initial discriminator is composed of several convolutional layers, batch normalization layers, a LeakyReLU activation function and a Sigmoid function.

[0069] The basic model of air conditioning is: air =f(P comp ,Q evap ,W comp ),

[0070] Where W comp is the compressor power, W comp =Q evap +Q loss , Q evap is the heat absorbed by the evaporator, Q loss is the heat lost by the system;

[0071] Q evap is the heat transfer coefficient of the evaporator, Q evap =U evap ·A evap ΔT evap , ΔT evap is the heat transfer area of ​​the evaporator, ΔT evap is the temperature difference across the evaporator;

[0072] Where P comp is the environmental impact factor, P comp =α·T ambient +β·RH+γ·v wind , where T ambient is the ambient temperature, RH is the relative humidity, v wind is the wind speed, α, β, γ are model parameters, which quantitatively describe the influence of various environmental factors on the load of the air-conditioning system.

[0073] The air conditioning load model is composed of the P air The nonlinear regression transformation is as follows:

[0074] y=f(x)+∈

[0075] y is the output load value, x is the input feature vector, ∈ is the error term, and f(x) is a nonlinear function; the nonlinear function is a polynomial regression function or an exponential regression function, specifically: Where p is the order of the polynomial, or f(x)=a exp(bx), and the loss function of the nonlinear function is:

[0076]

[0077] The objective function of the generative adversarial network is as follows:

[0078] The loss function of the initial generator network:

[0079] tem is the temperature vector of the input conditions, z is derived from the prior distribution p z is the noise vector sampled in; G(z;θ) is the output of the generator (generated samples); D(·;φ) is the discriminator function used to evaluate the authenticity of the input;

[0080] The loss function of the initial discriminator network is:

[0081]

[0082] θ is the parameter of the generator; φ is the parameter of the discriminator; z is the noise vector, drawn from the prior distribution p z G(z; θ) is the output of the generator (generated samples); D(·; φ) is the discriminator function used to evaluate the authenticity of the input; x is from the real data distribution p data Samples of is a sample from the linear interpolation between the true data distribution and the generator output; λ is the weight parameter of the gradient penalty.

[0083] include:

[0084] An air conditioning load generation module based on a physical model, the module is used to complete load prediction using the air conditioning physical model;

[0085] An air conditioning load generation module based on a generative adversarial network, the module is used to train the generative adversarial network to complete the generation of the load;

[0086] An air conditioning load generation module based on physics and generative adversarial network tuning is used to complete the final generation capacity of the load.

[0087] The air conditioning load generation module based on the generative adversarial network uses the trained discriminator network to discriminate the authenticity of the predicted load data according to the existing data. When the authenticity of the predicted load data meets the requirements of the discriminator network, the predicted load data is output as the final predicted load data.

[0088] The air conditioning load generation module based on physics and generative adversarial network tuning uses the data in the dataset to train the initial generator network and the initial discriminator network, including inputting the data in the dataset into the initial generator network to obtain generated data, inputting the generated data and the data in the dataset into the initial discriminator network, and using the initial discriminator network to discriminate the authenticity of the generated data, until the authenticity of the generated data meets the preset requirements, and obtaining the trained discriminator network.

[0089] A computer device includes a memory and a processor, wherein the memory stores a computer program, and is characterized in that the memory stores the computer program, and when the processor executes the computer program, an air conditioning load generation method that integrates the aforementioned air conditioning physical model and a generative adversarial network is implemented.

[0090] A computer-readable storage medium having a computer program stored thereon, characterized in that the storage medium includes a stored program, wherein when the program is running, the device where the storage medium is located is controlled to execute the aforementioned air conditioning load generation method that integrates an air conditioning physical model with a generative adversarial network.

Claims

1. An air conditioning load generation method integrating an air conditioning physical model with a generative adversarial network, characterized in that: The steps include: 1) Constructing a basic air conditioning model, which includes a compressor thermal model, an evaporator thermal model and an environmental model of the air conditioning; 2) Convert the basic air conditioning model into the polynomial form required for nonlinear regression to obtain the air conditioning load model; 3) Using the data set to perform nonlinear optimization on the air conditioning load model to determine the polynomial parameters in the air conditioning load model and obtain the air conditioning load polynomial model; 4) Generate forecast load data using the air conditioning load polynomial model; 5) The data set is updated according to the predicted load data, and when the data in the updated data set meets the preset conditions, the updated data set is used as the air conditioning load generation data set.

2. The method for generating air conditioning load by integrating an air conditioning physical model with a generative adversarial network according to claim 1, characterized in that: The data in the updated data set satisfies the preset condition means that when the proportion of real data in the updated data set is greater than or equal to a first preset value, the data in the updated data set satisfies the preset condition, otherwise the data in the updated data set does not satisfy the preset condition.

3. The method for generating air conditioning load by integrating an air conditioning physical model with a generative adversarial network according to claim 1, characterized in that: In step 5), updating the data set according to the predicted load data refers to using the discriminator network of the training process to evaluate the authenticity of the predicted load data. When the authenticity evaluation result is true, the predicted load data is used to update the data set; otherwise, the air-conditioning load model is re-solved by nonlinear optimization using the data set, and the predicted load data is generated according to the air-conditioning load polynomial model obtained after the re-solution.

4. The method for generating air conditioning load by integrating an air conditioning physical model with a generative adversarial network according to claim 2, characterized in that: The method further comprises: 1) Constructing a generative adversarial network for air conditioning load, wherein the generative adversarial model includes an initial discriminator network and an initial generator network; 2) When the data in the updated data set does not meet the preset conditions, the updated data set is used to train the initial generator network and the initial discriminator network to obtain the trained discriminator network and the generator network, and the trained discriminator network is used to perform authenticity evaluation on the load data generated during the training process of the initial generator network. When the authenticity evaluation result is true, the load data generated during the training process of the initial generator network is used to update the data set.

5. The method for generating air conditioning load by integrating an air conditioning physical model with a generative adversarial network according to claim 2, characterized in that: The method further comprises: When the data in the updated data set does not meet the preset conditions, the air conditioning load model is re-solved by nonlinear optimization using the updated data set, and the predicted load data is generated according to the air conditioning load polynomial model obtained after the re-solution.

6. The method for generating air conditioning load by integrating an air conditioning physical model with a generative adversarial network according to claim 3, characterized in that: Update the dataset with forecasted load data, including: The existing data in the data set is merged with the predicted load data to form a new data set.

7. The method for generating air conditioning load by integrating an air conditioning physical model with a generative adversarial network according to claim 4, characterized in that: Update the dataset using the load data generated during the initial generator network training process, including: The existing data in the dataset is merged with the load data generated during the initial generator network training process to form a new dataset.

8. The method for generating air conditioning load prediction based on an air conditioning physical model according to claim 1, characterized in that: The initial generator consists of several deconvolution layers, batch normalization layers, a ReLU activation function and a Tanh activation function; The initial discriminator is composed of several convolutional layers, batch normalization layers, a LeakyReLU activation function and a Sigmoid function.

9. The method for generating air conditioning load by integrating an air conditioning physical model with a generative adversarial network according to claim 1, characterized in that: The basic model of air conditioning is: air =f(P comp ,Q evap ,W comp ), Where W comp is the compressor power, W comp =Q evap +Q loss , Q evap is the heat absorbed by the evaporator, q loss is the heat lost by the system; Q evap is the heat transfer coefficient of the evaporator, Q evap =U evap ·A evap ΔT evap , A evap is the heat transfer area of ​​the evaporator, ΔT evap is the temperature difference across the evaporator; Where P comp is the environmental impact factor, P comp =α·T ambient +β·RH+γ·v wind , where T ambient is the ambient temperature, RH is the relative humidity, v wind is the wind speed, α, β, γ are model parameters, which quantitatively describe the influence of various environmental factors on the load of the air-conditioning system.

10. The method for generating air conditioning load by integrating an air conditioning physical model with a generative adversarial network according to claim 9, characterized in that: The air conditioning load model is composed of the P air The nonlinear regression transformation is as follows: y=f(x)+∈ y is the output load value, x is the input feature vector, ∈ is the error term, and f(x) is a nonlinear function; the nonlinear function is a polynomial regression function or an exponential regression function, specifically: Where p is the order of the polynomial, or f(x)=a exp(bx), and the loss function of the nonlinear function is:

11. The method for generating air conditioning load by integrating an air conditioning physical model with a generative adversarial network according to claim 1, characterized in that: The objective function of the generative adversarial network is as follows: The loss function of the initial generator network: tem is the temperature vector of the input conditions, z is derived from the prior distribution p z is the noise vector sampled in; G(z;θ) is the output of the generator (generated samples); D(·;φ) is the discriminator function used to evaluate the authenticity of the input; The loss function of the initial discriminator network is: θ is the parameter of the generator; φ is the parameter of the discriminator; z is the noise vector, drawn from the prior distribution p z G(z; θ) is the output of the generator (generated samples); D(·; φ) is the discriminator function used to evaluate the authenticity of the input; x is from the real data distribution p data Samples of is a sample from the linear interpolation between the true data distribution and the generator output; λ is the weight parameter of the gradient penalty.

12. A load forecasting module of an air conditioning load generation method integrating an air conditioning physical model with a generative adversarial network, characterized in that: include: An air conditioning load generation module based on a physical model, the module is used to complete load prediction using the air conditioning physical model; An air conditioning load generation module based on a generative adversarial network, the module is used to train the generative adversarial network to complete the generation of the load; An air conditioning load generation module based on physics and generative adversarial network tuning is used to complete the final generation capacity of the load.

13. The load prediction module of the air conditioning load generation method of the air conditioning physical model and the generative adversarial network fusion according to claim 12 is characterized by: The air conditioning load generation module based on the generative adversarial network uses the trained discriminator network to discriminate the authenticity of the predicted load data according to the existing data. When the authenticity of the predicted load data meets the requirements of the discriminator network, the predicted load data is output as the final predicted load data.

14. The load prediction module of the air conditioning load generation method of the air conditioning physical model and the generative adversarial network fusion according to claim 12 is characterized by: The air conditioning load generation module based on physics and generative adversarial network tuning uses the data in the dataset to train the initial generator network and the initial discriminator network, including inputting the data in the dataset into the initial generator network to obtain generated data, inputting the generated data and the data in the dataset into the initial discriminator network, and using the initial discriminator network to discriminate the authenticity of the generated data, until the authenticity of the generated data meets the preset requirements, and obtaining the trained discriminator network.

15. A computer device, characterized in that: The invention comprises a memory and a processor, wherein the memory stores a computer program, and is characterized in that the memory stores a computer program, and when the processor executes the computer program, an air conditioning load generation method by integrating an air conditioning physical model with a generative adversarial network as described in any one of claims 1 to 10 is implemented.

16. A computer-readable storage medium having a computer program stored thereon, characterized in that: The storage medium includes a stored program, wherein when the program is running, the device where the storage medium is located is controlled to execute an air conditioning load generation method that integrates an air conditioning physical model and a generative adversarial network as described in any one of claims 1 to 10.

Images