Deep learning technique to forecast energy production in photovoltaic systems.

The decrease of oil storage and the strong impact on the environment are more and more worrying, in this scenario photovoltaic systems are necessary. Nonetheless the intermittent nature of the photovoltaic energy makes the use of it very difficult, and in particular its prediction. Several works use the solar radiation values like input in order to make prediction. In this work, the inputs will be the power generated by photovoltaic system mounted in Polytechnic of Turin and exogenous values obtained by various sensors. In particular, our aim is to demonstrate that features are more useful for long-term solar radiation forecasting, while for short time, only the power is sufficient. Most relevant features are obtained through feature selection, they are UV index, global horizontal irradiance, temperature, dew Point, humidity, cloud cover, sunshine duration and hour. These features mixed with the power generated by PV system feed four different machine learning models: a Echo State Network (ESN), a 1D Convolution Neural Network (1D CNN), a Feedfoward Neural Network (FNN), a Long Short-Term Memory (LSTM).