Scholarlink

The integration of renewable energy sources into modern electricity systems introduces high uncertainty, making energy related forecasts increasingly complex. This paper proposes a framework to forecast Iberian day-ahead market clearing prices using multiple forecasting techniques, from pure statistical to machine learning models. A reinforcement learning approach dynamically selects the best forecasting model for each negotiation period. The methodology was validated using market price data from Portugal and Spain. All models underwent sensitivity analysis to determine optimal parameters and were cyclically retrained to incorporate new data patterns. The case study focuses on the first half of 2024, during which the reinforcement learning model continuously adapted based on real prices and model predictions, identifying the most suitable approach at each iteration. The models used in the study include Ridge Regression, Stochastic Gradient Descent, Random Forest, Extreme Gradient Boost, Deep Neural Network, and Long Short-Term Memory. Results showed that the proposed method outperforms individual models, achieving performance improvements ranging from 7.27% to 88.38% compared to using any of the models alone.

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

Reinforcement Learning for Real-Time Price Prediction in Energy Systems

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • David Araújo,
  • Gabriel Santos,
  • Brígida Teixeira,
  • Pedro Faria,
  • Zita Vale

摘要

The integration of renewable energy sources into modern electricity systems introduces high uncertainty, making energy related forecasts increasingly complex. This paper proposes a framework to forecast Iberian day-ahead market clearing prices using multiple forecasting techniques, from pure statistical to machine learning models. A reinforcement learning approach dynamically selects the best forecasting model for each negotiation period. The methodology was validated using market price data from Portugal and Spain. All models underwent sensitivity analysis to determine optimal parameters and were cyclically retrained to incorporate new data patterns. The case study focuses on the first half of 2024, during which the reinforcement learning model continuously adapted based on real prices and model predictions, identifying the most suitable approach at each iteration. The models used in the study include Ridge Regression, Stochastic Gradient Descent, Random Forest, Extreme Gradient Boost, Deep Neural Network, and Long Short-Term Memory. Results showed that the proposed method outperforms individual models, achieving performance improvements ranging from 7.27% to 88.38% compared to using any of the models alone.