ISEP - DM – Engenharia de Inteligência Artificial
Permanent URI for this collection
Browse
Browsing ISEP - DM – Engenharia de Inteligência Artificial by Author "ARAÚJO, DAVID SILVA"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- Forecasting day-ahead prices in the Iberian electricity market using a multi-agent reinforcement learning system in real-timePublication . ARAÚJO, DAVID SILVA; Vale, Zita Maria Almeida do; Santos, Gabriel José Lopes dos; Teixeira, Brígida Constança CorreiaThe use of renewable energy sources in modern electrical grids introduces significant uncertainty due to the grid’s instant response and unexpected fluctuations in generation. While crucial for sustainability, they require the development of optimized energy operations and designs that are complex and robust. Forecasting generation, consumption, or prices is a crucial aspect of an electrical system, being essential to maximise energy efficiency and support strategic planning for both suppliers and consumers. Therefore, this dissertation presents a multi-agent framework for forecasting daily Portuguese prices in the Iberian market to aid bidding proposals. It employs several forecasting models, ranging from statistical to machine learning. A reinforcement learning (RL) methodology is used to select the most appropriate forecasting model for each moment. Forecasting models and RL methodology were subjected to tuning tests to determine hyperparameters. The methodology was tested using real electricity production data and market prices from Portugal and Spain, between 2022 and 2024. In the application of the RL methodology, the forecasting models were retrained monthly, and the RL methodology was continuously updated with actual prices and forecasts so that the optimal model was selected in each iteration. The forecasting models included in the dissertation are Ridge Regression, Stochastic Gradient Descent, Random Forest, Extreme Gradient Boosting, Deep Neural Networks, and Long Short-Term Memory. These models were integrated into a multi-agent system that manages requests from different users, making the system robust and scalable. It comprises a main agent that receives user requests, organises them, and allocates them to various tasks, creating a “Task Agent” for each task. All case studies were conducted to evaluate the proposed solution and improve the efficiency of the RL methodology. The final case study simulates real-world conditions and demonstrated a 10% lower performance compared to the others. It highlighted the importance of daily features that are not available in time for bidding proposals. However, the final case study employed context, based on information from each renewable energy production source, which improves the RL model’s performance by 3%. Overall, results indicated that the RL methodology surpassed individual forecasting models across all case studies, with gains between 4.66% and 35.71%. As the case studies progressed, it was observed that the average difference between predicted and actual market values was 0.56 €/MWh in the non-reality reflecting case study and 6.13 €/MWh in the (near) real-time scenario. It is further concluded that hydro and wind energy are the primary influences on Portuguese energy prices, with solar energy also playing a notable role. In contrast, nuclear energy in Spain has the least impact.