Unit:
Κατεύθυνση Ηλεκτρονικός Αυτοματισμός (H/A)Library of the School of Science
Author:
Venitourakis Georgios
Supervisors info:
Διονύσιος Ρεΐσης, Καθηγητής, Τμήμα Φυσικής, ΕΚΠΑ
Άννα Τζανακάκη, Αναπλ. Καθηγήτρια, Τμήμα Φυσικής, ΕΚΠΑ
Νικόλαος Βλασσόπουλος, Δόκτωρ, Τμήμα Φυσικής, ΕΚΠΑ
Original Title:
XceptionLSTM: Advanced Sequence-to-Sequence Neural Networks for Short-Term Weather Forecasting Applications
Translated title:
XceptionLSTM: Advanced Sequence-to-Sequence Neural Networks for Short-Term Weather Forecasting Applications
Summary:
Renewable Energy has been the main global focus of the last two decades, as the energy industry searches for ways to steadily replace fossil fuels with greener energy sources. This transition paved the way to new concepts and techniques for controlling the power production and distribution in renewable energy parks. For improved efficiency, the controller in the Smart Grid concept often acts proactively, where it predicts events in the near future and prepares the underlying infrastructure for the upcoming event.
In the case of photovoltaic parks, a smart system makes short-term weather forecasts about the global horizontal irradiance and the cloud cover in the park’s area of interest. This thesis introduces a neural network as a solution to the short-term weather forecasting problem. The proposed model is an image regression recurrent neural network in the form of a spatio- temporal encoder/decoder. The basis of the structure is the Xception layer, which utilizes depwthwise and pointwise convolutions to infer data. The Xception layer is combined with long short-term memory cells to create a recurrent neural network with improved forecasting capabilities. The proposed model is optimized for inference on the edge and is evaluated on the Archon - Athens, Greece dataset.
Main subject category:
Science
Keywords:
deep learning, ConvLSTM, irradiance forecasting, edge computing, photovoltaic parks, ground-based sky images
