基于确定性策略梯度深度强化学习和模仿学习的多源微电网经济优化调度策略

Economic Optimal Dispatch Strategy for Multi-source Microgrid Based onDeep Deterministic Policy Gradient and Imitation Learning

当前国家大力实施“双碳”战略,以新能源为主体的新型电力系统将呈现爆发式增长,风电、光伏等新能源机组逐渐成为主力电源,微电网得以提出并迅速发展,其调度与优化问题也成为研究的热点。传统的基于数学建模与求解的微电网有功优化调度方法计算量大且十分复杂,存在易陷入局部最优、模型修改困难等瓶颈问题,因此提出基于确定性策略梯度深度强化学习(DDPG)和模仿学习(XGBoost)的经济优化调度策略。首先构建微电网有功优化调度的统一预置数学模型,其次通过构建XGBoost模型学习得到初始宏观决策,最后再构建基于DDPG的微电网有功调度人工智能模型。通过进行XGBoost分类器和DDPG神经网络离线训练,得出基于确定性策略梯度深度强化学习和模仿学习的微电网优化调度在线决策模型,最后通过算例分析验证了模型和算法的有效性。

At present,the country is vigorously implementing the"double carbon"strategy.The new power system with new energy as the main body will show explosive growth,and wind power,photovoltaic and other new energy units will gradually become the main power supply.Microgrid has been proposed and developed rapidly,and its scheduling and optimization problems have also become a hot spot in research.The traditional active optimization scheduling method of microgrid based on mathematical modeling and solution is computationally intensive and complex,and there are bottlenecks such as local optimization and difficult model modification.Therefore,an economic optimization scheduling strategy based on Deep Deterministic Policy Gradient(DDPG)and imitation learning(XGBoost)is proposed.Firstly,a unified preset mathematical model for the active optimization dispatch of microgrid is constructed.Secondly,the initial macro decision is obtained by constructing the XGBoost model learning,and finally the DDPG-based microgrid active dispatching artificial intelligence model is constructed.Through the offline training of XGBoost classifier and DDPG neural network,an online decision-making model for microgrid optimal dispatch based on deterministic policy gradient deep reinforcement learning and imitation learning is obtained.Finally,the effectiveness of the model and algorithm is verified by example analysis.