Scholarlink

The growing uncertainty and dimensionality of hydrothermal power systems challenge traditional stochastic dynamic programming (SDP) tools for long-term dispatch. This paper applies tabular Q-learning to a simplified Uruguayan system that aggregates four major hydroelectric plants into one equivalent reservoir within a Gymnasium-based environment. The algorithm is trained over a three-year, weekly horizon under both deterministic and stochastic inflows to approximate optimal dispatch decisions without requiring an explicit transition model. Compared with the industrial SDP benchmark (MOP), Q-learning achieves virtually identical results under perfect foresight and remains within about 2.19% of cumulative cost under stochastic inflows (2759 MUSD vs. 2700 MUSD), while reproducing key reservoir and thermal-use patterns. These findings show that even a simple approach can deliver competitive long-term policies and provide a baseline for extending Reinforcement Learning to larger state spaces with renewable energy and storage.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Q-Learning Approach for Long-Term Hydrothermal Dispatch

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Florencia Roque,
  • Matías Rama,
  • Ignacio Salas,
  • Mónica Carle,
  • Rodrigo Porteiro

摘要

The growing uncertainty and dimensionality of hydrothermal power systems challenge traditional stochastic dynamic programming (SDP) tools for long-term dispatch. This paper applies tabular Q-learning to a simplified Uruguayan system that aggregates four major hydroelectric plants into one equivalent reservoir within a Gymnasium-based environment. The algorithm is trained over a three-year, weekly horizon under both deterministic and stochastic inflows to approximate optimal dispatch decisions without requiring an explicit transition model. Compared with the industrial SDP benchmark (MOP), Q-learning achieves virtually identical results under perfect foresight and remains within about 2.19% of cumulative cost under stochastic inflows (2759 MUSD vs. 2700 MUSD), while reproducing key reservoir and thermal-use patterns. These findings show that even a simple approach can deliver competitive long-term policies and provide a baseline for extending Reinforcement Learning to larger state spaces with renewable energy and storage.