رحمت اله ميرزائي

Stochastic nature of wind characteristics and also special structure of large bearing of the Wind Turbine (WT) yaw system make the yaw control much difficult. For this reason, studies in the literature deem different assumptions and limitation to design yaw system controller, such as considering mean wind speed values, specified wind direction range, long yaw system operation runtime to reduce fatigue loads, etc. These assumptions reduce controller performance significantly in real WT operation. In this study, a novel method based on Reinforcement Learning (RL) has been suggested for yaw system control which make the controller able to operate in real world. The suggested method optimizes the cumulative profits of generated power considering loss costs of yaw system fatigue loads and consumed power for each incoming wind. DQN, Double DQN, and Dueling DQN algorithms have been used as RL algorithm which are trained through the train dataset, and test dataset is used to eva‎luate the performance of these algorithms. The results show that DQN algorithms have the best performance with total gained profit of 96.13% of total maximum profit and total captured power of 99.84% of total maxium power. Furthermore, Dueling DQN algorithm also has the best performance with total gained profit of 62.46% of total maximum profit, and total captured power of 86.45% of total maxium power of validation dataset of different location. As a results, our novel method presents an effective solution for controlling the yaw system, enabling it to perform optimally under varying wind characteristics. The results highlight the potential of our approach to improve the performance of WT yaw control in real-world applications.

انرژي بادي , توربين‌هاي بادي محور افقي , كنترل سامانه انحراف , هوش مصنوعي , يادگيري تقويتي

Wind Energy , Horizontal Axis Wind Turbine , Yaw System Control , Artificial Intelligence , Reinforcement Learning

Rahmat Allah Mirzaei

Meisam Farajollahi