Presentations

Abstract

The power production of a wind farm is influenced by numerous factors and fluctuates due to variable weather conditions. Electricity generated by wind farms can be traded across multiple power and reserve markets. Such transactions are typically completed ahead of the actual power generation (e.g. in day-ahead markets). However, inaccuracies in weather forecasts can result in the wind farm being unable to meet its contracted energy output, or conversely, being able to produce more energy. Such deviations from the contracted energy may incur balancing costs or represent missed opportunities for additional revenue. In contrast, the power production of a wind farm can also be set deliberately in imbalance in order to reduce the overall grid imbalance and so to obtain extra revenue. Under certain market conditions, integrating a green hydrogen production unit can enhance the operational profitability of a wind farm. Surplus electricity can be diverted to hydrogen production avoiding farm curtailment and the extra flexibility of the hybrid plant enables the operator to offer more profitable power balancing services to the grid. In this work, we have developed a multiagent reinforcement learning (RL) system with two collaborating RL agents who, sequentially, recommend optimal day-ahead trading volumes and perform real-time power control of a wind farm and a green hydrogen production unit. The multi-agent RL system is trained using SCADA data from a real-world offshore wind farm, historical day-ahead prices, grid balancing data and local weather forecasts. Results demonstrate that the RL agents successfully learned to maximize jointly the overall profit of the hybrid plant, taking into account the impact of several parameters on the amount and uncertainty of the energy production of the wind farm, adapting their behavior in function of the day-ahead prices and with the hydrogen production unit contributing to an increased profitability.