Forecsting per le Rinnovabili, Esplorazione di Modelli di Quantum Deep Learning

The purpose of this thesis work is to explore the possibility of performing local predictions of quantities relevant to renewable energy production. In particular, we focus on forecasting global horizontal radiation for time horizons of 1, 4 and 12 hours in the future. We first consider simple statistical models for forecasting time series, such as Auto-Regressive (AR) and Vector Auto-Regressive (VAR) models, and then later introduce more complicated ones, based on Deep Learning, such as Long-Short Term Memory (LSTM) and Transformer. Finally, Quantum-Hybrid architectures are realized, which involve the modification of ”classical” counterparts through the inclusion of Variational Quantum Circuits, allowing for the implementation of QLSTM and QTransformer models. We find out that, in general, simple AR model overperforms others in almost any horizon, although both VAR, LSTM and QLSTM may be considered suitable for the task explored. Transformer and QTransformer lead to worst performance on average. Comparing classical Deep Learning models with their Quantum realization, however, we notice some improvement for both QLSTM and QTransformer, reflected particularly in faster learning for the former, and better learning for the latter.