Adaptive new energy ultra-short-term power prediction method and device based on reinforcement learning

Abstract

The present invention relates to an adaptive new energy ultra-short-term power prediction method and device based on reinforcement learning, the method comprising obtaining environmental variables, action space and reward function for constructing an adaptive intelligent agent; wherein the environmental variables are environmental variable indicators reflecting environmental characteristics, the action space is a set of action functions taken by the intelligent agent in decision-making, and the reward function is an evaluation result of the corresponding change of the environmental variables after the intelligent agent performs the action; constructing an adaptive intelligent agent according to the environmental variables and the reward function; processing the environmental variables, and training the adaptive intelligent agent using the processed environmental variables to obtain an adaptive prediction intelligent agent. The present invention will adaptively select the prediction results of a single type of prediction method that best fits the external environment according to the environmental variables, thereby maximizing the accuracy of the prediction results. The method provided in the present application has a simple implementation process and has a strong application prospect.

Description

Technical Field

[0001] The present invention belongs to the field of electric power technology, and specifically relates to an adaptive new energy ultra-short-term power prediction method and device based on reinforcement learning. Background Art

[0002] In recent years, new energy sources, mainly wind power and photovoltaic power, have developed rapidly and their scale in the power grid has been increasing. New energy power prediction technology is the basic data for power grid operation control. At present, my country has proposed to build a new power system with new energy as the main body. It can be foreseen that the scale of new energy in the power system will increase rapidly in the future. Against the background of the increasing scale of new energy, the impact of its prediction accuracy on the operation efficiency of the power grid has become increasingly prominent and has become the focus of current research.

[0003] There are a lot of studies on renewable energy power prediction technology. From the perspective of implementation, the current renewable energy power prediction technology can be divided into the following two categories:

[0004] (1) Single-type prediction method, that is, the prediction method used is based on a single technical principle to predict the power of new energy, such as time series extrapolation method, related factor prediction method, etc.;

[0005] (2) Comprehensive prediction method, which integrates the results of multiple single-type prediction methods to minimize the limitations of different single-type prediction methods in actual prediction and ensure that the final prediction results have the highest possible prediction accuracy and prediction stability.

[0006] Since the technical principle of a single-type forecasting method is generally unique, the scope of application of any single-type forecasting method is relatively limited. When the external environment changes, the accuracy and stability of a single-type forecasting method are difficult to guarantee. Therefore, compared with a single-type forecasting method, the accuracy and stability of a comprehensive forecasting method are generally better. The difficulty of a comprehensive forecasting method lies in the reasonable formulation of weight coefficients for different single-type forecasting methods to maximize the accuracy of the comprehensive forecasting results. At present, the proportional weighting method is the most commonly used comprehensive forecasting method, and its basic implementation steps are:

[0007] Select multiple single-class forecasting methods and evaluate their forecasting accuracy based on historical data;

[0008] Select a single type of forecasting method whose forecasting accuracy meets the requirements;

[0009] The prediction results are integrated with the same proportional coefficient as the prediction results of the comprehensive prediction method.

[0010] In summary, due to the single technical principle, the single-type prediction method has limited prediction applicability and a high probability of large prediction deviation in practical applications. The comprehensive prediction method integrates multiple single-type prediction methods with relatively high prediction accuracy, so it has a high prediction accuracy and good prediction result stability. However, it is precisely because the comprehensive prediction method is an integration of multiple single-type prediction methods that it is affected by the deviation of the single-type prediction method and it is difficult to obtain a higher prediction result. Summary of the invention

[0011] In view of this, the purpose of the present invention is to overcome the shortcomings of the prior art and provide an adaptive new energy ultra-short-term power prediction method and device based on reinforcement learning to solve the problem of low accuracy of new energy power prediction in the prior art.

[0012] To achieve the above objectives, the present invention adopts the following technical solution: an adaptive new energy ultra-short-term power prediction method based on reinforcement learning, comprising:

[0013] Obtaining environmental variables, action spaces, and reward functions for constructing an adaptive intelligent agent; wherein the environmental variables are environmental variable indicators that reflect environmental characteristics, the action space is a set of action functions taken by the intelligent agent in making decisions, and the reward function is an evaluation result of the corresponding changes in the environmental variables after the intelligent agent performs an action;

[0014] Constructing an adaptive agent based on the environmental variables and the reward function;

[0015] The environmental variables are processed, and the self-adaptive agent is trained using the processed environmental variables to obtain a self-adaptive prediction agent.

[0016] Furthermore, the step of obtaining the environment variables for constructing the adaptive agent includes:

[0017] Basic data related to the change of new energy power is obtained, and environmental variables are obtained according to the basic data.

[0018] Furthermore, the obtaining of the action space for constructing the adaptive agent includes:

[0019] The action space selects the situation state variables for each single-class prediction method; the set of action functions is

[0020] A={a1,a2,…a N}

[0021] Among them, A is the action function set, a1, a2, and a3 represent three different single-class prediction method selection situation state variables, and the value range of a1, a2, and a3 is 0 or 1. If the value is "1", it indicates that the single-class prediction method is selected as the prediction method under the characteristics of the environmental variable; if the value is "0", it indicates that the single-class prediction method is not selected as the prediction method under the characteristics of the environmental variable.

[0022] Furthermore, the reward function includes:

[0023] Prediction accuracy reward and action space non-compliance penalty; where the prediction accuracy reward refers to the level of the adaptive prediction agent's prediction accuracy indicator; and the action space non-compliance penalty is the penalty when the action space does not meet the requirements; the reward function is:

[0024] R=α 1 R 1 -α 2 P 2

[0025] In the formula, R is the reward function, R 1 is the prediction accuracy reward function, P 2 is the action space non-compliant penalty function, α 1 , α 2 are the weight coefficients of the prediction accuracy reward function and the action space non-compliance penalty function, α 1 , α 2 are all coefficients greater than 0, and α 1 >>α 2 .

[0026] Further, the prediction accuracy reward function is a new energy power prediction accuracy function;

[0027]

[0028] Among them, P F , P A They are the predicted power and actual power of new energy power respectively.

[0029] Furthermore, the problem of action space non-compliance refers to the fact that the selected single-class prediction method is not unique;

[0030] Introduce a penalty function for non-compliant action spaces into the reward function, so that in each round of prediction, only one prediction method is selected as the prediction result of the comprehensive prediction method;

[0031] The penalty function is:

[0032]

[0033] Among them, P2 is the penalty function, N is the number of single-class prediction methods in the action space, that is, the number of actions in the action space. When a single-class action space is selected, the penalty function takes the value of 0, and the reward function is only related to the prediction accuracy reward function; otherwise, a penalty function will be generated, which will affect the value of the reward function.

[0034] Furthermore, the processing of the environment variables includes:

[0035] Obtain an environmental variable set according to the environmental variables, and discretize the environmental variable set using a limit discretization method to obtain a discretized environmental variable set;

[0036] The environment variables are

[0037] E={t,v,p,P}

[0038] Among them, E is the set of environmental variables, t, v, p, and P are environmental variables, t is the temperature, v is the wind speed, p is the sunshine intensity, and P is the new energy power at the previous moment.

[0039] The discretized environmental variable set is

[0040]

[0041] in, is the discretized set of environmental variables, They are the temperature, wind speed, sunshine intensity, and the discretized values ​​of the new energy power at the previous moment.

[0042] Furthermore, the processing of the environmental variables to train the adaptive agent to obtain the adaptive prediction agent includes:

[0043] Inputting the discretized environmental variable set into the adaptive agent for training to obtain an adaptive prediction agent;

[0044] The adaptive prediction agent is used to output the state variables of a single-class prediction method, and takes the prediction result of the single-class prediction method with a value of 1 as the output result.

[0045] Furthermore, the environment variables include:

[0046] Temperature, wind speed, sunshine intensity and renewable energy power at the previous moment.

[0047] The embodiment of the present application provides an adaptive new energy ultra-short-term power prediction device based on reinforcement learning, comprising:

[0048] An acquisition module is used to acquire environmental variables, action spaces and reward functions for constructing an adaptive intelligent agent; wherein the environmental variables are environmental variable indicators that reflect environmental characteristics, the action space is a set of action functions taken by the intelligent agent in decision-making, and the reward function is an evaluation result of the corresponding changes in the environmental variables after the intelligent agent performs an action;

[0049] A building module for building an adaptive agent based on the environment variables, action space and reward function;

[0050] The training module is used to obtain historical data and use the historical data to train the adaptive intelligent agent to obtain an adaptive prediction intelligent agent.

[0051] The present invention adopts the above technical solution, and the beneficial effects that can be achieved include:

[0052] The present invention provides an adaptive new energy ultra-short-term power prediction method and device based on reinforcement learning. The invention is essentially still a comprehensive prediction method. Compared with the traditional weighted average comprehensive prediction method, the technical solution provided by this application constructs an adaptive selection layer based on reinforcement learning, which will adaptively select the prediction results of a single type of prediction method that best fits the external environment according to environmental variables, thereby maximizing the accuracy of the prediction results. The implementation process of this method is simple and has strong application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0054] Figure 1 It is a schematic diagram of the steps of the adaptive new energy ultra-short-term power prediction method based on reinforcement learning of the present invention;

[0055] Figure 2 A schematic diagram of the structure of the adaptive prediction agent provided by the present invention;

[0056] Figure 3 It is a structural schematic diagram of the adaptive new energy ultra-short-term power prediction device based on reinforcement learning of the present invention. DETAILED DESCRIPTION

[0057] To make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be described in detail below. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other implementation methods obtained by ordinary technicians in this field without creative work belong to the scope of protection of the present invention.

[0058] A specific method and device for adaptive new energy ultra-short-term power prediction based on reinforcement learning provided in an embodiment of the present application are described below in conjunction with the accompanying drawings.

[0059] like Figure 1 As shown, the adaptive new energy ultra-short-term power prediction method based on reinforcement learning provided in the embodiment of the present application includes:

[0060] S101, obtaining environmental variables, action space and reward function for constructing an adaptive intelligent agent; wherein the environmental variables are environmental variable indicators reflecting environmental characteristics, the action space is a set of action functions taken by the intelligent agent in decision-making, and the reward function is an evaluation result of the corresponding change of the environmental variables after the intelligent agent performs an action;

[0061] In some embodiments, basic data related to the change of new energy power is obtained, and environmental variables are obtained based on the basic data.

[0062] The purpose of this step is to construct the adaptive prediction agent environment variables. The environment variables refer to the environmental variable indicators that best reflect the characteristics of the external environment selected from the basic data according to the actual needs of new energy power prediction.

[0063] Preferably, the present invention selects four indicators, namely, temperature, wind speed, sunshine intensity, and new energy power at the previous moment, as environmental variables. By constructing an environmental variable set with the above four indicators, the environmental variables can be expressed as:

[0064] E={t,v,p,P} (1)

[0065] Where E is the set of environmental variables, and t, v, p, and P are temperature, wind speed, sunshine intensity, and new energy power at the previous moment, respectively.

[0066] Preferably, the four environmental variables are discretized using the limit discretization method to meet the requirements of reinforcement learning training. The so-called limit discretization method refers to the process of converting the data into discrete positive integers based on the data fluctuation range with its maximum and minimum values ​​as the limit values. This method is a common method in reinforcement learning agent training, does not affect the main innovative content of the present invention, and does not elaborate on its implementation process. It is stipulated that the discretized environmental variable set can be expressed as:

[0067]

[0068] In the formula, They are the temperature, wind speed, sunshine intensity, and the discretized values ​​of the new energy power at the previous moment.

[0069] S102, constructing an adaptive intelligent agent according to the environmental variables and the reward function;

[0070] S103, processing the environmental variables, and using the processed environmental variables to train the adaptive intelligent agent to obtain an adaptive prediction intelligent agent.

[0071] The working principle of the adaptive new energy ultra-short-term power prediction method based on reinforcement learning is as follows: the application first constructs the required conditions of the intelligent agent, including environmental variables, action space and reward function, and then constructs an adaptive intelligent agent according to the above conditions, and then processes the environmental variables, and uses the processed environmental variables to train the adaptive intelligent agent to obtain an adaptive prediction intelligent agent.

[0072] In some embodiments, obtaining an action space for constructing an adaptive agent includes:

[0073] The action space selects the situation state variables for each single-class prediction method; the set of action functions is

[0074] A={a1,a2,…a N} (3)

[0075] Among them, A is the action function set, a1, a2, and a3 represent three different single-class prediction method selection situation state variables, and the value range of a1, a2, and a3 is 0 or 1. If the value is "1", it indicates that the single-class prediction method is selected as the prediction method under the characteristics of the environmental variable; if the value is "0", it indicates that the single-class prediction method is not selected as the prediction method under the characteristics of the environmental variable.

[0076] Specifically, constructing an action space is essentially constructing a set of action functions that the agent can take after making a decision.

[0077] In some embodiments, the reward function includes:

[0078] Prediction accuracy reward and action space non-compliance penalty; where the prediction accuracy reward refers to the level of the adaptive prediction agent's prediction accuracy indicator; and the action space non-compliance penalty is the penalty when the action space does not meet the requirements; the reward function is:

[0079] R=α 1 R 1 -α 2 P 2 (4)

[0080] In the formula, R is the reward function, R 1 is the prediction accuracy reward function, P 2 is the action space non-compliant penalty function, α 1 , α 2 are the weight coefficients of the prediction accuracy reward function and the action space non-compliance penalty function, α 1 , α 2 are all coefficients greater than 0, and α 1 >>α 2 .

[0081] Specifically, the reward function proposed in the present invention includes two parts: prediction accuracy reward and action strategy non-compliance penalty. The prediction accuracy reward refers to the level of the adaptive prediction agent prediction accuracy index; while the action strategy non-compliance penalty is the penalty for the action strategy not meeting the requirements.

[0082] Preferably, the prediction accuracy reward function is a new energy power prediction accuracy function. In the present invention, the prediction relative accuracy is selected as the evaluation function, which can be expressed as:

[0083]

[0084] Where P F , P A They are the new energy power prediction and actual power respectively.

[0085] In some embodiments, the problem of action strategy non-compliance refers to that the selected single type prediction method is not unique. According to the design requirements of the present invention, in each round of prediction, one and only one prediction method is selected as the prediction result of the comprehensive prediction method. To this end, a penalty function for action strategy non-compliance must be introduced into the reward function to ensure that the final selection result meets the above requirements. In the present invention, the square value of the difference between the sum of the action strategy results and 1 is used as the penalty function for the action strategy non-compliance problem, which can be expressed as:

[0086]

[0087] Where N is the number of single-class prediction methods in the action strategy, that is, the number of actions in the action space. Under normal circumstances, if a single-class action strategy is selected, the penalty function value is 0, and the reward function is only related to the prediction accuracy reward function; otherwise, a penalty function will be generated, which will affect the value of the reward function.

[0088] After the above required conditions are determined, an adaptive intelligent agent is constructed according to the environmental variables, action space and reward function, and then the adaptive intelligent agent is trained.

[0089] In this application, historical data is used to train the adaptive prediction agent, wherein the historical data is historical data related to the impact of new energy power.

[0090] The training process is to discretize the four environmental variables in the historical data using the limit discretization method to meet the requirements of reinforcement learning training. It can be understood that the discretization method is a common method in reinforcement learning agent training, and this application will not repeat it here. The obtained adaptive prediction agent is as follows Figure 2 shown.

[0091] The application process of the adaptive prediction agent is as follows: the collected real-time environmental variables are obtained and then discretized. The four discretized environmental variables are used as the input data of the adaptive prediction agent, that is, the selection state variables of a single-class prediction method can be output; the prediction result of the single-class prediction method with a value of 1 is used as the output result of the entire prediction method, and the prediction is completed.

[0092] In some embodiments, Figure 3 As shown, the present application provides an adaptive new energy ultra-short-term power prediction device based on reinforcement learning, comprising:

[0093] The acquisition module 301 is used to acquire the environment variables, action space and reward function for constructing the adaptive intelligent agent; wherein the environment variables are environment variable indicators reflecting the characteristics of the environment, the action space is a set of action functions taken by the intelligent agent in decision-making, and the reward function is an evaluation result of the corresponding change of the environment variables after the intelligent agent performs the action;

[0094] A construction module 302, for constructing an adaptive agent according to the environmental variables, action space and reward function;

[0095] The training module 303 is used to obtain historical data, and use the historical data to train the adaptive agent to obtain an adaptive prediction agent.

[0096] The working principle of the adaptive new energy ultra-short-term power prediction device based on reinforcement learning provided in the embodiment of the present application is that the acquisition module 301 acquires the environmental variables, action space and reward function used to construct the adaptive intelligent agent; wherein, the environmental variables are environmental variable indicators reflecting the characteristics of the environment, the action space is a set of action functions taken by the intelligent agent decision, and the reward function is the evaluation result of the corresponding changes of the environmental variables after the intelligent agent executes the action; the construction module 302 constructs the adaptive intelligent agent according to the environmental variables, action space and reward function; the training module 303 acquires historical data, and uses the historical data to train the adaptive intelligent agent to obtain an adaptive prediction intelligent agent.

[0097] In summary, the present invention provides an adaptive new energy ultra-short-term power prediction method and device based on reinforcement learning, the method comprising obtaining environmental variables, action space and reward function for constructing an adaptive intelligent agent; wherein the environmental variables are environmental variable indicators reflecting environmental characteristics, the action space is a set of action functions taken by the intelligent agent in decision-making, and the reward function is an evaluation result of the corresponding change of the environmental variables after the intelligent agent performs the action; an adaptive intelligent agent is constructed according to the environmental variables and the reward function; the environmental variables are processed, and the adaptive intelligent agent is trained using the processed environmental variables to obtain an adaptive prediction intelligent agent. The present invention will adaptively select the prediction results of a single type of prediction method that best fits the external environment according to the environmental variables, thereby maximizing the accuracy of the prediction results. The method provided in this application has a simple implementation process and has a strong application prospect.

[0098] It can be understood that the method embodiment provided above corresponds to the device embodiment described above, and the corresponding specific contents can be referenced to each other and will not be repeated here.

[0099] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) that contain computer-usable program code.

[0100] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0101] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction method, which is implemented in the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0102] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0103] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. An adaptive ultra-short-term power prediction method for new energy based on reinforcement learning, characterized in that: include: Obtaining environmental variables, action spaces, and reward functions for constructing an adaptive intelligent agent; wherein the environmental variables are environmental variable indicators that reflect environmental characteristics, the action space is a set of action functions taken by the intelligent agent in making decisions, and the reward function is an evaluation result of the corresponding changes in the environmental variables after the intelligent agent performs an action; constructing an adaptive agent based on the environmental variables, action space, and reward function; Processing the environmental variables, and using the processed environmental variables to train the adaptive intelligent agent to obtain an adaptive prediction intelligent agent; The step of obtaining environment variables for constructing an adaptive intelligent agent includes: Acquiring basic data related to the change of new energy power, and acquiring environmental variables according to the basic data; the acquisition is used to construct an action space of the adaptive intelligent agent, including: The action space selects the situation state variables for each single-class prediction method; the set of action functions is in, is a set of action functions, , , Indicates the state variables of the three different single-class prediction methods. , , The value range is 0 or 1. If the value is "1", it indicates that the single-category prediction method is selected as the prediction method under the characteristics of the environmental variable; if the value is "0", it indicates that the single-category prediction method is not selected as the prediction method under the characteristics of the environmental variable; The reward function includes: Prediction accuracy reward and action space non-compliance penalty; where the prediction accuracy reward refers to the level of the adaptive prediction agent's prediction accuracy indicator; and the action space non-compliance penalty is the penalty when the action space does not meet the requirements; the reward function is: In the formula, is the reward function, is the prediction accuracy reward function, is the action space non-compliant penalty function, , are the weight coefficients of the prediction accuracy reward function and the action space non-compliance penalty function, respectively. , are all coefficients greater than 0, and ; The prediction accuracy reward function is a new energy power prediction accuracy function; in, , They are the predicted power and actual power of new energy power respectively; The problem of action space non-compliance refers to the fact that the selected single-class prediction method is not unique; Introduce a penalty function for non-compliant action spaces into the reward function, so that in each round of prediction, only one prediction method is selected as the prediction result of the comprehensive prediction method; The penalty function is: in, is the penalty function, N is the number of single-class prediction methods in the action space, that is, the number of actions in the action space. When a single-class action space is selected, the penalty function takes a value of 0, and the reward function is only related to the prediction accuracy reward function; otherwise, a penalty function will be generated, which will affect the value of the reward function; The processing of the environment variable includes: Obtain an environmental variable set according to the environmental variables, and discretize the environmental variable set using a limit discretization method to obtain a discretized environmental variable set; The environment variables are in, is the environment variable set, , , , For environment variables, For temperature, is the wind speed, is the sunlight intensity, is the new energy power at the previous moment; The discretized environmental variable set is in, is the discretized set of environmental variables, , , , They are the temperature, wind speed, sunshine intensity, and the discretized values ​​of the new energy power at the previous moment; The method of training the adaptive agent using the processed environmental variables to obtain an adaptive prediction agent includes: Inputting the discretized environmental variable set into the adaptive agent for training to obtain an adaptive prediction agent; The adaptive prediction agent is used to output the state variables of a single-class prediction method, and takes the prediction result of the single-class prediction method with a value of 1 as the output result.

2. The method according to claim 1, characterized in that: The environment variables include: Temperature, wind speed, sunshine intensity and renewable energy power at the previous moment.

3. An adaptive new energy ultra-short-term power prediction device based on reinforcement learning, characterized in that: include: An acquisition module is used to acquire environmental variables, action spaces and reward functions for constructing an adaptive intelligent agent; wherein the environmental variables are environmental variable indicators that reflect environmental characteristics, the action space is a set of action functions taken by the intelligent agent in decision-making, and the reward function is an evaluation result of the corresponding changes in the environmental variables after the intelligent agent performs an action; A building module for building an adaptive agent based on the environment variables, action space and reward function; A training module, used to obtain historical data, and use the historical data to train the adaptive agent to obtain an adaptive prediction agent; The acquisition module is further used to acquire basic data related to the change of new energy power, and acquire environmental variables according to the basic data; The action space selects the situation state variables for each single-class prediction method; the set of action functions is in, is a set of action functions, , , Indicates the state variables of the three different single-class prediction methods. , , The value range is 0 or 1. If the value is "1", it indicates that the single-category prediction method is selected as the prediction method under the characteristics of the environmental variable; if the value is "0", it indicates that the single-category prediction method is not selected as the prediction method under the characteristics of the environmental variable; The reward function includes: Prediction accuracy reward and action space non-compliance penalty; where the prediction accuracy reward refers to the level of the adaptive prediction agent's prediction accuracy indicator; and the action space non-compliance penalty is the penalty when the action space does not meet the requirements; the reward function is: In the formula, is the reward function, is the prediction accuracy reward function, is the action space non-compliant penalty function, , are the weight coefficients of the prediction accuracy reward function and the action space non-compliance penalty function, respectively. , are all coefficients greater than 0, and ; Wherein, the prediction accuracy reward function is a new energy power prediction accuracy function; in, , They are the predicted power and actual power of new energy power respectively; Among them, the problem of action space non-compliance refers to the fact that the selected single-class prediction method is not unique; Introduce a penalty function for non-compliant action spaces into the reward function, so that in each round of prediction, only one prediction method is selected as the prediction result of the comprehensive prediction method; The penalty function is: in, is the penalty function, N is the number of single-class prediction methods in the action space, that is, the number of actions in the action space. When a single-class action space is selected, the penalty function takes a value of 0, and the reward function is only related to the prediction accuracy reward function; otherwise, a penalty function will be generated, which will affect the value of the reward function; The training module is further used to obtain an environmental variable set according to the environmental variables, and to discretize the environmental variable set using a limit discretization method to obtain a discretized environmental variable set; The environment variables are in, is the environment variable set, , , , For environment variables, For temperature, is the wind speed, is the sunlight intensity, is the new energy power at the previous moment; The discretized environmental variable set is in, is the discretized set of environmental variables, , , , They are the temperature, wind speed, sunshine intensity, and the discretized values ​​of the new energy power at the previous moment; Inputting the discretized environmental variable set into the adaptive agent for training to obtain an adaptive prediction agent; The adaptive prediction agent is used to output the state variables of a single-class prediction method, and takes the prediction result of the single-class prediction method with a value of 1 as the output result.

Images