CN-122022257-A - MARL-based electric vehicle type virtual power plant response potential evaluation method

Abstract

A MARL-based electric vehicle type virtual power plant response potential assessment method relates to the technical field of virtual power plant operation and power system demand response and comprises the following steps of constructing a multi-main-body cooperative scene based on electric vehicle type virtual power plant participation power system demand response, constructing a charging power dynamic estimation model based on incomplete vehicle data, defining MARL multi-agent reinforcement learning elements, constructing a response potential pre-assessment model based on a fused CNN-GRU network and an attention mechanism, and carrying out iterative optimization on the response potential accurate assessment model by means of a MARL multi-agent reinforcement learning assessment framework. The method solves the problems of high parameter dependence, poor dynamic adaptability, insufficient evaluation precision and the like in the existing evaluation method, and realizes efficient and accurate evaluation of response potential.

Inventors

• WANG FEI
• ZHU TIANBO
• ZHANG LIJIE
• WANG RONGQIANG
• LIU QINZHE
• DAI HAONAN
• WANG YUQING
• ZHEN ZHAO
• LI GANG
• WANG YUCUI
• Fan Jiyang
• Mi Dabin
• Li Zhitan
• SHI ZHENJIANG
• WANG JIANFENG
• SUN RONGFU
• ZHANG ZHANG
• ZHANG ZHIYONG
• ZHANG JINGYU
• WANG GE

Assignees

• 华北电力大学（保定）
• 河北建投能源投资股份有限公司
• 国网冀北电力有限公司
• 国网河北省电力有限公司经济技术研究院
• 国网新疆电力有限公司营销服务中心

Dates

Publication Date

20260512

Application Date

20251226

Claims (8)

1. 1. MARL-based electric vehicle type virtual power plant response potential evaluation method is characterized by comprising the following steps of: constructing a multi-main-body cooperative scene based on the participation of an electric vehicle type virtual power plant in the demand response of a power system; constructing a charging power dynamic estimation model based on incomplete vehicle data; Defining MARL elements of multi-agent reinforcement learning; Constructing a response potential pre-evaluation model based on fusion of a CNN-GRU network and an attention mechanism; and carrying out iterative optimization on the response potential accurate evaluation model by means of MARL multi-agent reinforcement learning evaluation framework.
2. 2. The method for evaluating response potential of an electric vehicle type virtual power plant according to claim 1, wherein the multi-main-body collaborative scenario comprises three core main bodies of a power grid dispatching center, an electric vehicle type virtual power plant and an electric vehicle user; The method for constructing the multi-main-body collaborative scene comprises the following steps: In the day-ahead stage, the power grid dispatching center issues a demand response period and a capacity demand threshold value to the virtual power plant; The virtual power plant collects basic charging intention information of a user side, wherein the basic charging intention information comprises predicted charging duration and expected electric quantity replenishment proportion; in the daytime, a user vehicle is connected to a virtual power plant charging network, and the state and the position information of the battery are synchronized in real time; The virtual power plant dynamically schedules the charging behaviors of each vehicle through the multi-agent system, ensures that the estimated response potential is matched with the actual response capability, and meets the power grid scheduling requirement.
3. 3. The method for evaluating response potential of an electric vehicle type virtual power plant according to claim 1, wherein the method for constructing a dynamic estimation model of charging power comprises: Step 1, a piecewise function fitting method is adopted to respectively establish mapping relations between output power of a charger and charging duration and battery SOC in a constant-current phase charging mode, a constant-voltage phase charging mode and a trickle phase charging mode; step 2, calculating the switching node time length of different charging stages by knowing the rated capacity of the battery and the currently available SOC data; and 3, reasonably estimating the missing charging characteristic parameters according to the filling rule of the incomplete parameters, and generating a charging power curve in stages.
4. 4. The electric vehicle model virtual power plant response potential assessment method of claim 1, wherein the elements defining MARL multi-agent reinforcement learning include a state space, an action space, a reward function, and a MADDPG policy gradient algorithm defined for a multi-subject collaborative scenario.
5. 5. The method for evaluating response potential of an electric vehicle type virtual power plant according to claim 1, wherein the method for constructing the response potential pre-evaluation model by fusing a CNN-GRU network and an attention mechanism is characterized by comprising the following steps of: Step 1, data preprocessing comprises normalization and feature encoding of historical response data, meteorological data and holiday features, and a multidimensional input feature matrix is constructed; step 2, extracting space association features in a feature matrix by using a CNN network, wherein the space association features comprise charging load correlation of different areas and weather factor influence weights; Step 3, learning a time sequence dependency relationship through a GRU network, and capturing a trend change rule of historical response data; step 4, introducing an attention mechanism to strengthen the influence weight of the key time period and the core characteristics, and outputting a pre-evaluation response potential curve; and 5, adopting the mean square error and the average absolute percentage error as joint evaluation indexes, and optimizing model parameters.
6. 6. The method for evaluating response potential of an electric vehicle type virtual power plant according to claim 1, wherein the iterative optimization of the response potential accurate evaluation model achieves precision improvement by means of dynamic interaction and parameter iteration of MARL multi-agent reinforcement learning, and the iterative optimization method comprises: step 1, setting an evaluation period and a response period range, and determining a time boundary and a capacity requirement of power grid demand response; step2, the multi-intelligent system acquires incomplete parameters of the accessed vehicle, and generates an initial charging scheme through a dynamic charging power estimation model; Step 3, outputting a preliminary response potential predicted value by the pre-evaluation model as an initial input of the reinforcement learning model; step 4, the multi-agent executes charging action according to the real-time running state, and collects actual response data and evaluation deviation; step 5, based on the evaluation deviation and the result of the calculation of the reward function, iteratively updating the agent strategy and the pre-evaluation model parameters; And 6, repeating the iterative process until the deviation between the response potential evaluation value and the actual verification value meets a preset threshold value, and outputting a precise evaluation result.
7. 7. A non-transitory computer readable storage medium storing a computer program, characterized in that the computer program when executed by a processor implements the steps of the method of any of claims 1-6.
8. 8. A computer program product comprising a computer program, characterized in that the computer program, when being executed by a processor, implements the steps of the method according to any of claims 1-6.

Description

MARL-based electric vehicle type virtual power plant response potential evaluation method Technical Field The invention relates to the technical field of virtual power plant operation and power system demand response, in particular to a virtual power plant response potential evaluation method taking an electric automobile as a main aggregate resource, which is particularly suitable for an accurate evaluation scene based on multi-agent reinforcement learning. Background With the rapid development of new energy automobile industry, electric automobiles become high-quality demand side resources with dual properties of load and energy storage. The virtual power plant is used as a core carrier for aggregating distributed resources, a large number of electric vehicles are clustered and managed, and large-scale response capability is formed, so that the virtual power plant becomes an important force for participating in power grid demand response and optimizing energy configuration. Currently, response potential evaluation of an electric automobile type virtual power plant is a key premise of participation in electric power market trading and power grid dispatching. The traditional assessment method mainly depends on a statistical model or a single algorithm, and has the defects that the randomness and parameter incompleteness of the charging behavior of the electric automobile are difficult to process, so that the power estimation deviation is large, the assessment model lacks dynamic simulation of multi-main-body cooperative interaction, the actual response capability of a virtual power plant cannot be accurately reflected, the short plate exists in the fusion processing of the time sequence characteristics and key influence factors of the traditional deep learning model, the pre-assessment precision is limited, the effective iterative optimization mechanism is lacking, and the matching degree of the assessment result and the actual response capability is insufficient. In the prior art, part of researches adopt a single reinforcement learning algorithm to schedule virtual power plant resources, but a multi-agent cooperative evaluation frame is not formed, so that the management requirements of a large-scale electric vehicle cluster are difficult to adapt, and other researches focus on charge load prediction, but the charge load prediction is not deeply fused with response potential evaluation, so that the accurate conversion from load to response capability cannot be realized. Therefore, a comprehensive evaluation method capable of integrating multi-subject coordination, dynamic power estimation, deep learning pre-evaluation and reinforcement learning iterative optimization is needed, and accuracy and reliability of virtual power plant response potential evaluation of electric automobile type are improved. Disclosure of Invention The invention aims to provide a MARL-based electric vehicle type virtual power plant response potential evaluation method, which solves the problems of high parameter dependence, poor dynamic adaptability, insufficient evaluation precision and the like in the existing evaluation method and realizes efficient and accurate evaluation of response potential. The technical scheme includes that the electric vehicle type virtual power plant response potential assessment method based on MARL comprises the following steps of constructing a multi-main-body cooperative scene based on the electric vehicle type virtual power plant participating in power system demand response, constructing a charging power dynamic assessment model based on incomplete vehicle data, defining MARL multi-agent reinforcement learning elements, constructing a response potential pre-assessment model based on a fusion CNN-GRU network and an attention mechanism, and carrying out iterative optimization on the response potential accurate assessment model by means of a MARL multi-agent reinforcement learning assessment framework. The invention also provides electronic equipment, which comprises a memory and a processor, wherein the memory stores a computer program, and the processor realizes the steps of the electric automobile type virtual power plant response potential evaluation method based on MARL when executing the computer program. The invention also provides a non-transitory computer readable storage medium storing a computer program which, when executed by a processor, implements the steps of the MARL-based electric vehicle-type virtual power plant response potential assessment method. The method has the beneficial effects that the accurate assessment of the response potential of the electric automobile type virtual power plant is realized through the organic combination of multi-main-body collaborative scene construction, dynamic power estimation, multi-agent reinforcement learning framework design, deep learning pre-assessment and iterative optimization. The method can adapt to the actual scene with incomplete parame