Unit:
Department of Informatics and TelecommunicationsΠληροφορική
Supervisors info:
Ιωάννης Εμίρης, Καθηγητής, Τμήμα Πληροφορικής και Τηλεπικοινωνιών, ΕΚΠΑ
Original Title:
Wind Energy Prediction Using Deep Learning Architectures
Translated title:
Wind Energy Prediction Using Deep Learning Architectures
Summary:
This thesis investigates the utilization of various deep learning architectures for energy prediction, utilizing Weather Research Forecasting (WRF) data and actual measurements from two wind farms. The preprocessing pipeline is thoroughly outlined, followed by experimentation with different feature extraction techniques such as CNNs, Mean Vector, and Central Vector, along with LSTM, Attention, and Transformer Blocks as temporal models. Attention-based models are emphasized as a notable contribution. The results show that efficient preprocessing is crucial for optimal performance and that attention-based models perform comparably or even better than LSTMs as temporal models in certain cases. Furthermore the study raises questions about the essentiality of CNNs feature extractor in some cases. It also suggests that transfer learning between nearby parks is a promising approach to address limited data, and can also be used for new parks where data are not available. Finally, the study also highlights certain limitations we faced and proposes avenues for future work.
Main subject category:
Technology - Computer science
Keywords:
Machine Learning, Wind Energy Prediction, Neural Network, Attention, Transformer Encoder, WRF
