TY - THES ID - 146388614 TI - Master's Thesis : Comparison of probabilistic forecasting deep learning models in the context of renewable energy production AU - Stassen, Théo AU - Ernst, Damien AU - Gemine, Quentin AU - Mathieu, Sébastien PY - 2020 PB - Liège Université de Liège (ULiège) DB - UniCat KW - artificial intelligence KW - machine learning KW - forecasting KW - time series forecasting KW - probabilistic forecasting KW - deep learning KW - forecasting models KW - renewable energy KW - renewable energy production KW - model comparison KW - Ingénierie, informatique & technologie > Sciences informatiques UR - https://www.unicat.be/uniCat?func=search&query=sysid:146388614 AB - This master thesis subject addresses the question of what is the best forecasting method to implement in the context of the prediction of renewable energy production, to protect assets from oversupply. The growing scientific field of Deep Learning has a great potential to be exploited to achieve this goal. This works is composed of different parts. First part introduces the goal that we want to be pursued. The second part interrogates what what are the tools needed to accomplish the goal and defines the context on which the comparison will be performed. The third part is a comparison of the models considering a default forecasting goal . The fourth part is a discussion on what might be the most relevant metric considering the main goal. From this we define two metrics, Coverage and MASE and we finally perform in fifth part a comparison using metrics and loss that have been introduced .
The answer to the question of what is the better forecasting model in the defined context between all the tested models, the model that provides the better results, in terms of Coverage and MASE, is definitely the model MQCNN, which outperforms for the two metrics considered all the other presented models. MQCNN model is followed by MQRNN, DeepAr and SimpleFeedForward. ER -