Multi-energy load multi-task learning prediction method and device based on GAN and DQN

Abstract

The present invention discloses a multi-energy load multi-task learning prediction method and device based on GAN and DQN, and the method comprises the following steps: collecting multi-energy load data and performing normalization processing; establishing a GAN model, and combining the data generated in the GAN model with the multi-energy load data to form a multi-energy load data set; combining 4 different depths of CNN with Q learning method to establish 4 different depths of DQN models, and using the multi-energy load data set for training; using the DQN model to perform multi-task learning on different types of load prediction tasks, mining the coupling relationship between different types of loads, and establishing a load prediction model; inputting the multi-energy load data into the load prediction model, and obtaining the prediction results of cold, hot, electric and gas loads. The present invention can better adapt to the characteristics of areas with dispersed energy distribution, small multi-energy load demand, and insufficient multi-energy data collection, and improves the accuracy of multi-energy load prediction.

Description

Technical Field

[0001] The present invention relates to a multi-energy load multi-task learning prediction method and device based on GAN (generative adversarial networks) and DQN (Deep Q Network), belonging to the technical field of comprehensive energy load prediction. Background Art

[0002] An integrated energy system refers to a new integrated energy system that uses advanced physical information technology and innovative management models to integrate multiple energy sources such as coal, oil, natural gas, electricity, and thermal energy in a certain region, and achieves coordinated planning, optimized operation, collaborative management, interactive response, and mutual complementation among multiple heterogeneous energy subsystems. While meeting the diversified energy needs within the system, it is necessary to effectively improve energy utilization efficiency and promote sustainable energy development.

[0003] At present, the industries in economically developed areas are concentrated, the energy consumption is large, the load density is high and growing rapidly, the energy supply has high safety and reliability, and the energy distribution network wiring is complex. Different from these areas, there are many areas with small and scattered loads, non-concentrated electricity consumption, low power load density, long distance from the power supply, long pipes and lines with poor quality, relatively strong seasonality and time period of energy consumption, and abundant renewable resources such as wind, light, biomass and combustible waste. It can be seen that there are great differences in energy resources and demand in different regions. With the increase in people's demand for various types of energy, the integrated energy system has developed rapidly, integrating various heterogeneous energy sources such as electricity, heat, cold energy, etc. in the region, realizing diversified coordinated planning, coordinated management and optimized operation of energy, and multi-energy load forecasting is the premise for optimized operation and reasonable planning of the integrated energy system.

[0004] For multi-energy load forecasting, some scholars have tried to use nonlinear autoregressive models, support vector machines, radial basis function neural networks, long short-term memory, etc. to predict multi-energy loads, and have achieved high prediction accuracy. At the same time, some studies have used the Pearson correlation coefficient to prove that different types of loads in the integrated energy system have a strong coupling relationship. However, most of the current studies are single-task learning methods, which cannot perform feature mining on various types of loads with coupling relationships. In addition, the existing multi-energy load forecasting research focuses on the prediction of multi-energy loads with obvious changes and high energy consumption in economically developed areas. There is still a lack of research on multi-energy load forecasting in areas with irregular energy consumption, low energy consumption, and insufficient multi-energy load data collection. Therefore, it is very necessary to carry out research on high-precision prediction models for multi-energy loads in such areas. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a multi-energy load multi-task learning prediction method and device based on GAN and DQN, which can not only predict the multi-energy system, but also improve the accuracy of load prediction.

[0006] The technical solution adopted by the present invention to solve the technical problem is:

[0007] On the one hand, an embodiment of the present invention provides a multi-energy load multi-task learning prediction method based on GAN and DQN, comprising the following steps:

[0008] Collecting and normalizing multi-energy load data, wherein the multi-energy load data includes cooling, heating, electricity and gas load data and historical meteorological data;

[0009] Establishing a GAN model, inputting the multi-energy load data into the GAN model, generating data and the multi-energy load data to form a multi-energy load data set; the GAN model includes a generator model and a discriminator model;

[0010] Four different depths of CNNs were combined with Q-learning methods to build four different depths of DQN models, and they were trained using a multi-energy load dataset.

[0011] Four DQN models with different depths are used to perform multi-task learning on different types of load forecasting tasks, explore the coupling relationship between different types of loads, and establish a load forecasting model;

[0012] Input the multi-energy load data into the load forecasting model to obtain the forecast results of cooling, heating, electricity and gas loads.

[0013] As a possible implementation of this embodiment, the formula for normalizing the multi-energy load data is:

[0014]

[0015] Among them, x′ ij is the normalized value of the input variable, x ij is the original value of the input variable, is the minimum value of the jth input variable, is the maximum value of the jth input variable.

[0016] As a possible implementation of this embodiment, establishing a GAN model includes:

[0017] The GAN model is composed of the generator model G and the discriminator model D;

[0018] Input random noise z into the generator model G to generate data G(z);

[0019] The multi-energy load data and data G(z) are input into the discriminator model D for discrimination, so that the generator model G and the discriminator model D compete with each other and are iteratively optimized to generate data that obeys the distribution of the real multi-energy load data.

[0020] As a possible implementation of this embodiment, the four CNNs with different depths are combined with the Q learning method to establish four DQN models with different depths, and the DQN models are trained using a multi-energy load data set, including:

[0021] The multi-energy load dataset was input into four CNNs of different depths to extract the features of the multi-energy load;

[0022] The learned data features are input into the Q network respectively, and four DQN models with different depths are constructed through the Markov decision method;

[0023] The multi-energy load dataset is input into four DQN models for training respectively. The DQN model continuously tries and improves the state, action and reward until the loss function of the obtained action converges and outputs the Q value.

[0024] As a possible implementation of this embodiment, the state is multi-energy load data input into CNN, and the action is multi-energy load prediction data.

[0025] On the other hand, an embodiment of the present invention provides a multi-energy load multi-task learning prediction device based on GAN and DQN, comprising:

[0026] A data acquisition module, used to collect and normalize multi-energy load data, wherein the multi-energy load data includes cold, hot, electric and gas load data and historical meteorological data;

[0027] A GAN model building module is used to build a GAN model, input the multi-energy load data into the GAN model, and the generated data and the multi-energy load data form a multi-energy load data set; the GAN model includes a generator model and a discriminator model;

[0028] The DQN model building module is used to combine CNNs of four different depths with the Q-learning method to build four DQN models of different depths and train them using a multi-energy load dataset;

[0029] The load forecasting model building module is used to use four DQN models of different depths to perform multi-task learning on different types of load forecasting tasks, explore the coupling relationship between different types of loads, and establish a load forecasting model;

[0030] The load forecasting module is used to input multi-energy load data into the load forecasting model to obtain the forecast results of cooling, heating, electricity and gas loads.

[0031] As a possible implementation of this embodiment, the formula for the data acquisition module to normalize the multi-energy load data is:

[0032]

[0033] Among them, x′ ij is the normalized value of the input variable, x ij is the original value of the input variable, is the minimum value of the jth input variable, is the maximum value of the jth input variable.

[0034] As a possible implementation of this embodiment, the GAN model building module includes:

[0035] GAN model composition module, used to form a GAN model from a generator model G and a discriminator model D;

[0036] A data G(z) generation module is used to input random noise z into the generator model G to generate data G(z);

[0037] The discriminator discrimination module is used to input the multi-energy load data and data G(z) into the discriminator model D for discrimination, so that the generator model G and the discriminator model D confront each other and iteratively optimize to generate data that obeys the distribution of the real multi-energy load data.

[0038] As a possible implementation of this embodiment, the DQN model building module includes:

[0039] The feature extraction module is used to input the multi-energy load data set into four CNNs with different depths to extract the features of the multi-energy load;

[0040] The DQN model construction module is used to input the learned data features into the Q network respectively, and construct four DQN models of different depths through the Markov decision method;

[0041] The DQN model training module is used to input the multi-energy load data set into the four DQN models for training. The DQN model continuously tries and improves the state, action and reward until the loss function of the obtained action converges and outputs the Q value.

[0042] As a possible implementation of this embodiment, the state is multi-energy load data input into CNN, and the action is multi-energy load prediction data.

[0043] The technical solution of the embodiment of the present invention may have the following beneficial effects:

[0044] The present invention can better adapt to the characteristics of areas with dispersed energy distribution, small multi-energy load demand, and insufficient multi-energy data collection, improve the accuracy of multi-energy load prediction, and further guide the development and construction of different characteristic areas in my country.

[0045] The present invention proposes a multi-energy load forecasting method based on GAN network, deep Q network and multi-task learning. In the modeling process, in order to eliminate the defect of low prediction accuracy caused by insufficient data, GAN is used to simulate the existing multi-energy load data to generate similar multi-energy load data, and then a plurality of DQN models with different depths are built through the Markov decision method. Finally, the multi-task learning mechanism is used to further explore the coupling characteristics between multi-energy loads, establish a load forecasting model, and obtain a more accurate multi-energy load forecasting result. The novel multi-energy load forecasting method proposed by the present invention has important practical guiding significance for accelerating the planning and transformation needs of various regions. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flowchart of a multi-energy load multi-task learning prediction method based on GAN and DQN according to an exemplary embodiment;

[0047] Figure 2 is a schematic diagram of a multi-energy load multi-task learning prediction device based on GAN and DQN according to an exemplary embodiment;

[0048] Figure 3 is a schematic diagram of a load forecasting model according to the present invention according to an exemplary embodiment;

[0049] Figure 4 It is a training diagram of multi-task learning according to an exemplary embodiment. DETAILED DESCRIPTION

[0050] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0051] In order to clearly illustrate the technical features of the present solution, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings. The disclosure below provides many different embodiments or examples for realizing different structures of the present invention. In order to simplify the disclosure of the present invention, the components and settings of specific examples are described below. In addition, the present invention may repeat reference numbers and / or letters in different examples. This repetition is for the purpose of simplification and clarity, and does not itself indicate the relationship between the various embodiments and / or settings discussed. It should be noted that the components illustrated in the accompanying drawings are not necessarily drawn to scale. The present invention omits the description of known components and processing techniques and processes to avoid unnecessary limitations on the present invention.

[0052] like Figure 1As shown, an embodiment of the present invention provides a multi-energy load multi-task learning prediction method based on GAN and DQN, comprising the following steps:

[0053] Collecting and normalizing multi-energy load data, wherein the multi-energy load data includes cooling, heating, electricity and gas load data and historical meteorological data;

[0054] Establishing a GAN model, inputting the multi-energy load data into the GAN model, generating data and the multi-energy load data to form a multi-energy load data set; the GAN model includes a generator model and a discriminator model;

[0055] Four different depths of CNNs were combined with Q-learning methods to build four different depths of DQN models, and they were trained using a multi-energy load dataset.

[0056] Four different depth DQN models are used to perform multi-task learning on different types of load forecasting tasks to share features, explore the coupling relationship between different types of loads, and establish a load forecasting model, such as Figure 3 As shown;

[0057] Input the multi-energy load data into the load forecasting model to obtain the forecast results of cooling, heating, electricity and gas loads.

[0058] As a possible implementation of this embodiment, the formula for normalizing the multi-energy load data is:

[0059]

[0060] Among them, x′ ij is the normalized value of the input variable, x ij is the original value of the input variable, is the minimum value of the jth input variable, is the maximum value of the jth input variable.

[0061] As a possible implementation of this embodiment, establishing a GAN model includes:

[0062] The GAN model is composed of the generator model G and the discriminator model D;

[0063] Input random noise z into the generator model G to generate data G(z);

[0064] The multi-energy load data and data G(z) are input into the discriminator model D for discrimination, so that the generator model G and the discriminator model D compete with each other and are iteratively optimized to generate data that obeys the distribution of the real multi-energy load data.

[0065] As a possible implementation of this embodiment, the four CNNs with different depths are combined with the Q learning method to establish four DQN models with different depths, and the DQN models are trained using a multi-energy load data set, including:

[0066] The multi-energy load dataset was input into four CNNs of different depths to extract the features of the multi-energy load;

[0067] The learned data features are input into the Q network respectively, and four DQN models with different depths are constructed through the Markov decision method;

[0068] The multi-energy load dataset is input into four DQN models for training respectively. The DQN model continuously tries and improves the state, action and reward until the loss function of the obtained action converges and outputs the Q value.

[0069] As a possible implementation of this embodiment, the state is multi-energy load data input into CNN, and the action is multi-energy load prediction data.

[0070] The training process of multi-task learning of the present invention is as follows Figure 4 As shown, each prediction task has its own training set {x,y i}, i = 1, 2...m. In the multi-task training process, multiple training sets are merged and input into DQN models of different depths respectively, and the loss function is calculated using homoscedastic uncertainty.

[0071] like Figure 2 As shown, an embodiment of the present invention provides a multi-energy load multi-task learning prediction device based on GAN and DQN, including:

[0072] A data acquisition module, used to collect and normalize multi-energy load data, wherein the multi-energy load data includes cold, hot, electric and gas load data and historical meteorological data;

[0073] A GAN model building module is used to build a GAN model, input the multi-energy load data into the GAN model, and the generated data and the multi-energy load data form a multi-energy load data set; the GAN model includes a generator model and a discriminator model;

[0074] DQN model building module, which is used to combine CNNs of four different depths with Q learning methods to build four DQN models of different depths and train them using multi-energy load data sets;

[0075] The load forecasting model building module is used to use four different depth DQN models to perform multi-task learning on different types of load forecasting tasks to share features, explore the coupling relationship between different types of loads, and build a load forecasting model, such as Figure 3As shown;

[0076] The load forecasting module is used to input multi-energy load data into the load forecasting model to obtain the forecast results of cooling, heating, electricity and gas loads.

[0077] As a possible implementation of this embodiment, the formula for the data acquisition module to normalize the multi-energy load data is:

[0078]

[0079] Among them, x′ ij is the normalized value of the input variable, x ij is the original value of the input variable, is the minimum value of the jth input variable, is the maximum value of the jth input variable.

[0080] As a possible implementation of this embodiment, the GAN model building module includes:

[0081] GAN model composition module, used to form a GAN model from a generator model G and a discriminator model D;

[0082] A data G(z) generation module is used to input random noise z into the generator model G to generate data G(z);

[0083] The discriminator discrimination module is used to input the multi-energy load data and data G(z) into the discriminator model D for discrimination, so that the generator model G and the discriminator model D confront each other and iteratively optimize to generate data that obeys the distribution of the real multi-energy load data.

[0084] As a possible implementation of this embodiment, the DQN model building module includes:

[0085] The feature extraction module is used to input the multi-energy load data set into four CNNs with different depths to extract the features of the multi-energy load;

[0086] The DQN model construction module is used to input the learned data features into the Q network respectively, and construct four DQN models of different depths through the Markov decision method;

[0087] The DQN model training module is used to input the multi-energy load data set into the four DQN models for training. The DQN model continuously tries and improves the state, action and reward until the loss function of the obtained action converges and outputs the Q value.

[0088] As a possible implementation of this embodiment, the state is multi-energy load data input into CNN, and the action is multi-energy load prediction data.

[0089] The specific process of multi-energy load forecasting in the present invention is as follows.

[0090] (1) Collect real multi-energy load data, including cooling, heating, electricity, gas load data and related historical meteorological data, and normalize these data.

[0091] (2) Input the real multi-energy load data collected in step (1) into GAN. GAN consists of a generator model G and a discriminator model D, whose inputs are random noise z and real multi-energy load data x, respectively. G(z) is the data generated by G that approximates the characteristics of x as much as possible. Through D's continuous discrimination of the sources of x and G(z), G and D are made to confront each other and iteratively optimize until G learns and can generate data that obeys the distribution of real multi-energy load data, and together with the collected real multi-energy load data, it forms a multi-energy load data set.

[0092] (3) The multi-energy load data set in step (2) is input into four CNNs of different depths to extract the features of the multi-energy load, and then the learned data features are input into the Q network respectively, and then four DQN models of different depths are constructed through the Markov decision method. The DQN model continuously tries and improves the state (multi-energy load data input into CNN), action (multi-energy load prediction data) and reward until the loss function of the obtained action converges and outputs the Q value.

[0093] (4) Using the four different depth DQN models in step (4) to perform multi-task learning on different types of load forecasting tasks, we can share features and further explore the coupling relationship between different types of loads. Figure 3 The load forecasting model shown.

[0094] (5) Input the real multi-energy load data collected in step (1) into the load prediction model, and output the prediction results of cooling, heating, electricity and gas loads.

[0095] The present invention uses the four different depth DQN models in step (4) to perform multi-task learning training on different types of load forecasting tasks. Figure 4 As shown, each prediction task has its own training set {x,y i}, i = 1, 2...m. In the multi-task training process, multiple training sets are merged and input into DQN models of different depths respectively, and the loss function is calculated using homoscedastic uncertainty.

[0096] The present invention constructs a multi-energy load prediction method based on GAN, DQN and multi-task learning, which overcomes the disadvantage of insufficient multi-energy load data. Through neural networks of different depths, it adapts to loads with different characteristics and better mines the coupling characteristics between multi-energy loads, thereby improving the prediction accuracy of multi-energy loads, and thus guiding the development and construction of different characteristic regions in my country.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A multi-task learning prediction method for multi-energy load based on GAN and DQN, It is characterized in that The following steps are involved: Collecting and normalizing multi-energy load data, wherein the multi-energy load data includes cooling, heating, electricity and gas load data and historical meteorological data; Establishing a GAN model, inputting the multi-energy load data into the GAN model, generating data and the multi-energy load data to form a multi-energy load data set; the GAN model includes a generator model and a discriminator model; Four different depths of CNNs were combined with Q-learning methods to build four different depths of DQN models, and they were trained using a multi-energy load dataset. Four DQN models with different depths are used to perform multi-task learning on different types of load forecasting tasks, explore the coupling relationship between different types of loads, and establish a load forecasting model; Input the multi-energy load data into the load forecasting model to obtain the forecast results of cooling, heating, electricity and gas loads.

2. According to claim 1, the multi-energy load multi-task learning prediction method based on GAN and DQN, It is characterized in that The formula for normalizing multi-energy load data is: Among them, x' ij is the normalized value of the input variable, x ij is the original value of the input variable, is the minimum value of the jth input variable, is the maximum value of the jth input variable.

3. According to claim 1, the multi-energy load multi-task learning prediction method based on GAN and DQN, It is characterized in that The establishing of the GAN model comprises: The GAN model is composed of the generator model G and the discriminator model D; Input random noise z into the generator model G to generate data G(z); The multi-energy load data and data G(z) are input into the discriminator model D for discrimination, so that the generator model G and the discriminator model D compete with each other and are iteratively optimized to generate data that obeys the distribution of the real multi-energy load data.

4. The multi-energy load multi-task learning prediction method based on GAN and DQN according to any one of claims 1 to 3, It is characterized in that The four different depths of CNNs are combined with the Q learning method to establish four different depths of DQN models, and trained using a multi-energy load data set, including: The multi-energy load dataset was input into four CNNs of different depths to extract the features of the multi-energy load; The learned data features are input into the Q network respectively, and four DQN models with different depths are constructed through the Markov decision method; The multi-energy load dataset is input into four DQN models for training respectively. The DQN model continuously tries and improves the state, action and reward until the loss function of the obtained action converges and outputs the Q value.

5. According to claim 4, the multi-energy load multi-task learning prediction method based on GAN and DQN, It is characterized in that The state is the multi-energy load data input into CNN, and the action is the multi-energy load prediction data.

6. A multi-energy load multi-task learning prediction device based on GAN and DQN, It is characterized in that include: A data acquisition module, used to collect and normalize multi-energy load data, wherein the multi-energy load data includes cold, hot, electric and gas load data and historical meteorological data; A GAN model building module is used to build a GAN model, input the multi-energy load data into the GAN model, and the generated data and the multi-energy load data form a multi-energy load data set; the GAN model includes a generator model and a discriminator model; DQN model building module, which is used to combine CNNs of four different depths with Q learning methods to build four DQN models of different depths and train them using multi-energy load data sets; The load forecasting model building module is used to use four DQN models of different depths to perform multi-task learning on different types of load forecasting tasks, explore the coupling relationship between different types of loads, and establish a load forecasting model; The load forecasting module is used to input multi-energy load data into the load forecasting model to obtain the forecast results of cooling, heating, electricity and gas loads.

7. The multi-energy load multi-task learning prediction device based on GAN and DQN according to claim 6, It is characterized in that The formula for normalizing the multi-energy load data by the data acquisition module is: Among them, x' ij is the normalized value of the input variable, x ij is the original value of the input variable, is the minimum value of the jth input variable, is the maximum value of the jth input variable.

8. The multi-energy load multi-task learning prediction device based on GAN and DQN according to claim 6, It is characterized in that The GAN model building module includes: GAN model composition module, used to form a GAN model from a generator model G and a discriminator model D; A data G(z) generation module is used to input random noise z into the generator model G to generate data G(z); The discriminator discrimination module is used to input the multi-energy load data and data G(z) into the discriminator model D for discrimination, so that the generator model G and the discriminator model D confront each other and iteratively optimize to generate data that obeys the distribution of the real multi-energy load data.

9. The multi-energy load multi-task learning prediction device based on GAN and DQN according to any one of claims 6 to 8, It is characterized in that The DQN model building module includes: The feature extraction module is used to input the multi-energy load data set into four CNNs with different depths to extract the features of the multi-energy load; The DQN model construction module is used to input the learned data features into the Q network respectively, and construct four DQN models of different depths through the Markov decision method; The DQN model training module is used to input the multi-energy load data set into the four DQN models for training. The DQN model continuously tries and improves the state, action and reward until the loss function of the obtained action converges and outputs the Q value.

10. The multi-energy load multi-task learning prediction device based on GAN and DQN according to claim 9, It is characterized in that The state is the multi-energy load data input into CNN, and the action is the multi-energy load prediction data.

Images