|
| 基于改进TD3算法的水风光蓄多能互补低碳协同调度策略 |
| Low-Carbon Economic Cooperative Dispatch Strategy for Hydro-Wind-Solar-Storage Systems Based on Im-proved TD3 Algorithm |
| 投稿时间：2026-03-20 修订日期：2026-06-10 |
| DOI： |
| 中文关键词: 多能互补调度 抽水蓄能改造 马尔可夫决策过程 TD3算法 阶梯式碳排放交易 |
| 英文关键词: multi-energy complementary dispatch pumped storage retro-fitting Markov decision process TD3 algorithm tiered carbon emission trading |
| 基金项目:国家电网公司科技项目(SGLNDK00DWJS1900039) |
|
| 摘要点击次数: 0 |
| 全文下载次数: 0 |
| 中文摘要: |
| 针对高比例可再生能源并网带来的不确定性与调度复杂性，提出一种基于改进TD3算法的水风光蓄多能互补低碳协同调度策略。构建涵盖水电、风电、光伏和抽水蓄能四单元的系统运行模型，引入阶梯式碳排放交易机制，建立综合运营成本最小化目标，并将联合调度问题建模为马尔可夫决策（Markov Decision Process，MDP）过程。该问题面临多源随机性、阶梯碳成本的非凸性及抽蓄模式切换带来的整数变量等挑战，传统随机规划与混合整数方法难以兼顾求解效率与在线实时性，而MDP与深度强化学习的组合可有效突破上述瓶颈。在标准双重延迟深度确定性策略梯度（Twin Delayed Deep Deterministic Policy Gradient，TD3）算法基础上，引入优先经验回放和自适应探索噪声两项改进，提升算法在连续调度空间中的收敛速度与策略鲁棒性。基于2024年中国东北区域电力系统实测数据的仿真实验表明：在风电等效出力系数0.34、光伏等效出力系数0.26的高VRE渗透典型日，系统单日综合运营成本降低3549.6万元，CO?减排1.31万吨，弃风弃光率从8.0%降至2.5%；抽蓄电站额定容量利用率达93.2%，调峰贡献率达62.3%；改进TD3算法收敛回合数较标准TD3减少24.6%，较深度确定性策略梯度（Deep Deterministic Policy Gradient，DDPG）减少52.4%。 |
| 英文摘要: |
| To address the uncertainty and scheduling complexity intro-duced by high-proportion renewable energy integration, this paper proposes a low-carbon cooperative dispatch strategy for hydro-wind-solar-storage systems based on an improved TD3 algorithm. A system operation model encompassing hydro-power, wind power, photovoltaics, and pumped storage is constructed. A tiered carbon emission trading mechanism is introduced, a comprehensive operating cost minimization objective is formulated, and the joint dispatch problem is modeled as a Markov decision process (MDP). The problem faces challenges such as multi-source stochasticity, non-convexity of tiered carbon costs, and integer variables arising from pumped-storage mode switching. Traditional stochastic programming and mixed-integer methods struggle to balance solution efficiency with online real-time perfor-mance, whereas the combination of MDP and deep rein-forcement learning effectively overcomes these bottlenecks. Based on the standard twin delayed deep deterministic policy gradient (TD3) algorithm, two enhancements—prioritized experience replay and adaptive exploration noise decay—are introduced to improve convergence speed and policy ro-bustness in continuous dispatch spaces. Simulation experi-ments using measured data from the Northeast China power system in 2024 show that on a typical high-VRE penetration day (wind capacity factor 0.34, solar capacity factor 0.26), the daily comprehensive operating cost is reduced by 35.496 million CNY, CO? emissions are cut by 13,100 tons, and the renewable energy curtailment rate drops from 8.0% to 2.5%. The rated capacity utilization of the pumped storage station reaches 93.2%, and its peak-shaving contribution rate reaches 62.3%. The improved TD3 algorithm reduces the number of convergence episodes by 24.6% compared to standard TD3 and by 52.4% compared to deep deterministic policy gradient (DDPG). |
|
View Fulltext
查看/发表评论 下载PDF阅读器 |
| 关闭 |