基于DQN的梯级水电站实时负荷优化分配研究

DQN-based real-time optimized load distribution of cascade hydropower stations

摘要

【目的】流域梯级水电系统规模持续扩大与运行环境日趋复杂，传统优化调度方法难以适应流域复杂多样的调控要求，且其决策精度与求解效率均有限。【方法】以耗水量最小为主要目标，构建了兼顾电调-水调的梯级水电优化调度模型，并研发了基于深度强化学习(DQN)的高效求解方法。以大渡河中游梯级水电系统为研究实例，分别设置中等负荷，低负荷和高负荷三种工况，输入实际运行数据对模型进行训练，并结合耗水量、水位过程等角度对模型优化效果进行评估。【结果】结果显示：DQN算法可显著减少计算耗时，将计算效率提升约41.37倍；同时，DQN算法可以很好地平衡水位和流量等水调需求之间的冲突，相较于优化前，DQN可在将水位波动指数平均降低约0.058 m/min的同时将平均总耗水量减少1 158万m³;除此之外，提出的模型适用于多种工况，具有良好的稳定性。【结论】结果表明：基于DQN的负荷分配方法可有效增强系统运行稳定性与安全性，实现调度科学性与计算效率的双重突破；智能决策框架通过实时优化电站出力分配，显著降低水位波动与发电耗水，验证了电调-水调协同优化的可行性。该方法为梯级水电系统智能化调度与新时期复杂场景下的优化调控提供了新的技术路径。

Abstract

[Objective] With the continuous expansion of cascade hydropower systems and increasingly complex operational environments, traditional optimized scheduling method struggle to meet the complex and diverse regulation requirements of river basins, while their decision-making accuracy and computational efficiency remain limited. [Methods] To address these limitations, an optimized scheduling model was established for cascade hydropower systems that considered power generation scheduling and water resource regulation, with minimum water consumption as the primary objective. Additionally, an efficient solution method based on deep reinforcement learning(Deep Q-Network, DQN) was developed. Using the cascade hydropower system in the middle reaches of the Dadu River as a case study, three operating conditions(medium, low, and high load) were established. The model was trained using actual operational data and evaluated through water consumption and water level processes. [Results] The result showed that the DQN algorithm reduced computational time by approximately 41.37 times compared to conventional method. Furthermore, DQN effectively balanced conflicting water regulation demands(e.g., water level stability and flow control), achieving an average reduction of 0.058 m/min in the water level fluctuation index and a total water consumption decrease of 11.58 million m³ compared to pre-optimization. Notably, the proposed model exhibited good stability across diverse operating conditions. [Conclusion] The findings indicate that the DQN-based load distribution method enhances system operational stability and safety while achieving breakthroughs in both scientific scheduling and computational efficiency. By dynamically optimizing real-time power output distribution among stations, the intelligent decision-making framework significantly mitigates water level fluctuations and reduces water consumption in power generation, thereby validating the feasibility of coordinated power-water optimization. This method provides a novel technical approach for intelligent scheduling of cascade hydropower systems and their optimized regulation under complex operational scenarios in the new era.

关键词

负荷分配 / 梯级水电 / 深度强化学习 / 实时调度 / 动态规划 / 影响因素

Key words

load distribution / cascade hydropower / deep reinforcement learning / real-time scheduling / dynamic programming / influencing factors

基金资助

国家自然科学基金项目(52479013)