Deep Learning-based Forecasting of Renewable Energy Sources for Locations with Scarce Data

Deep learning-based models are one of the most popular data-driven approaches for forecasting renewable energy sources (RES) with reasonable accuracy since they can handle sequential data, such as the time series from weather variables. However, these models demand a large amount of historical weather data for training, which in most cases is limited in locations selected for installing distributed energy resources (DERs). This work uses an Encoder-Decoder Sequence to Sequence (Seq2Seq) model along a transfer learning methodology to predict wind speed and solar radiation on a medium-long term horizon. The transfer learning strategy is incorporated in the Seq2Seq model to improve prediction accuracy for a target location (e.g., A DERs with insufficient data) by using data generated from similar source locations (e.g., weather stations with extensive data). The results show that the transfer learning approach enhances the forecasting performance of the Seq2Seq model against the case of only using data from source locations. The proposed methodology offers an opportunity for the widespread adoption of sustainable technologies in urban and rural communities even when there are insufficient on-site meteorological measurements.