Source-load stochasticity affects the safety, stability and economical operation of power grid, and makes it difficult to design optimal dispatch. Currently, due to the low accuracy of day-ahead forecast for wind power and load, dynamic optimal dispatch based on ultra-short-term wind power prediction is an effective way to reduce the power deviation in power system and to relieve the operating pressure on automatic generation control (AGC) unit. In this paper, by comprehensively considering the source-load stochasticity, the dispatch properties of unit, and the constraints of AGC regulating range, we establish a dynamic optimal dispatch model to minimize the running cost of the underlying power system, including the coal consumption cost of thermal unit, and the output adjustment cost of both the dispatchable unit and the AGC unit. Since each dynamic dispatch order has an impact, direct or indirect, on the subsequent system operation, we formulate the dynamic optimal dispatch problem as a Markov decision process (MDP) problem. In this framework, a deep Q-learning method is provided to optimize the dispatch strategy. Finally, the feasibility and the effectiveness of the proposed method is validated by means of numerical experiments on IEEE 30-node system. © 2021 Institute of Electrical Engineers of Japan. Published by Wiley Periodicals LLC.

中文翻译：

---

基于深度Q学习的电力系统日内动态优化调度

源负荷随机性影响电网的安全、稳定和经济运行，给优化调度设计带来困难。目前，由于风电负荷日前预报精度不高，基于超短期风电功率预测的动态优化调度是降低电力系统功率偏差、缓解电力系统运行压力的有效途径。自动发电控制 (AGC) 单元。本文综合考虑源负荷随机性、机组调度特性和AGC调节范围约束，建立动态优化调度模型，以最小化底层电力系统的运行成本，包括煤耗成本。热单元，以及可调度单元和 AGC 单元的输出调整成本。由于每个动态调度命令对后续系统运行都有直接或间接的影响，我们将动态最优调度问题表述为马尔可夫决策过程（MDP）问题。在该框架中，提供了一种深度 Q 学习方法来优化调度策略。最后，通过IEEE 30节点系统的数值实验验证了所提方法的可行性和有效性。© 2021 日本电气工程师学会。由 Wiley Periodicals LLC 出版。通过IEEE 30节点系统的数值实验验证了所提方法的可行性和有效性。© 2021 日本电气工程师学会。由 Wiley Periodicals LLC 出版。通过IEEE 30节点系统的数值实验验证了所提方法的可行性和有效性。© 2021 日本电气工程师学会。由 Wiley Periodicals LLC 出版。