Modified fuzzy Q-learning based wind speed prediction

• Published in: Journal of wind engineering and industrial aerodynamics 2020-11, Vol.206, p.104361, Article 104361
• Main Authors: Sharma, Rajneesh, Shikhola, Tushar, Kohli, Jaspreet Kaur
• Format: Article
• Language: English
• Subjects: Empirical mode decomposition; k-Nearest Neighbor (k-NN); Modified fuzzy Q learning; Pearson’s correlation coefficient; Support vector regression (SVR); Wind speed prediction
• ISSN: 0167-6105; 1872-8197
• DOI: 10.1016/j.jweia.2020.104361

Renewable energy has taken a center stage in sustainable and environmentally safe power generation. In this work, a novel model free Reinforcement Learning based wind speed forecasting technique has been proposed. Our technique uses modified fuzzy Q learning (MFQL) framework to accurately predict 1-min ahead wind speed from the publicly available data online. Empirical Mode Decomposition (EMD) and Pearson’s correlation coefficient have been used in the pre-processing stages to identify seven most relevant intrinsic mode functions (IMF) from the raw wind speed data. These IMFs form the inputs to an MFQL based forecaster which accurately forecasts wind speed using a reward/punishment approach. MFQL predictor is put to test on wind speed data obtained from National Institute of Wind Energy and Wind Resource Assessment data portal for 10 Indian cities, i.e., Bhogat, Chandori, Kotada, Charanka, Gandhi Nagar, Jambua, Keshod, Sadodar, Surat and Vartej located in the state of Gujarat, India. Our predictor is able to achieve an accuracy of 96.23% for Gandhi Nagar, 94.84% for Bhogat, 94.12% for Kotada and similar results are obtained for other locations. MFQL approach has been compared with SVR and k-NN. Results show that MFQL approach has higher accuracy than SVR and k-NN techniques. •A novel self-learning MFQL based predictor for wind speed.•Requires no model of the system.•Unsupervised learning paradigm used.•Achieves high prediction accuracy.•Site independent and has high reliability.