Daily Forecasting of Photovoltaic Power Generation with Multi- Technological data Using Enhanced Long Short-Term Memory Networks

Abstract

The spotlight on Algeria's efforts to tap into its solar resources and enhance its photovoltaic capabilities
has sparked widespread interest. With current achievements totaling 567.1 MW, the country plans for
a surge to 3000 MW in the foreseeable future. This increasing reliance on intermittent solar energy
underscores the importance of precise PV power forecasting for ensuring grid flexibility and reliability.
Deep learning methods have demonstrated promising outcomes in handling intricate data and
understanding systematic biases, surpassing conventional approaches. This study explores the
effectiveness of LSTM in predicting PV power output across diverse PV technologies. Our
methodology involves training the LSTM model extensively on large-scale Poly-Silicon module data
and subsequently applying this pretrained model to forecast power output in regions with similar
climatic conditions but different PV technologies. Specifically, the model, initially trained on extensive
data from the Djelfa power plant (January 2018 to December 2019), is tested using data from the
Ghardaïa PV station (July to December 2014) on four other PV technologies with different structures,
including Cadmium Telluride (CdTe), Amorphous Silicon, Mono-crystalline and Poly-crystalline with
fixed structures, and Mono-crystalline and Poly-crystalline equipped by a stacker system. This approach
demonstrates the substantial benefits of applying a pretrained model to smaller datasets in similar
climatic regions, particularly when dealing with varying PV technologies. The performance of our
LSTM model, evaluated using metrics such as RMSE ≤ 0.2090, NRMSE< 21.36%, r ≤ 0.9475, and
MAE ≤ 0.1516, confirms its robust prediction capability across different technological setups,
highlighting its practical applicability in diverse PV forecasting scenarios