Volume 4: Innovative Solutions for Energy Transitions: Part III

Hourly Solar Irradiance Prediction From Satellite Data Using Lstm Pratima Kumari1, Durga Toshniwal



Solar irradiance prediction is an emerging area of research for various applications in renewable energy domain. So far, numerous physical models, statistical models and machine learning based techniques have been utilized to accomplish solar irradiance prediction. However, existing models are not good at learning long-term historical dependencies, lead to compromise in modeling non-linear solar irradiance patterns. In this paper, a novel prediction model (i.e. Long Short Term Memory, LSTM) from deep neural network family is used to predict hourly solar irradiance with enhanced prediction accuracy by considering long-term historical data dependencies. To provide an extensive and strong assessment of proposed model, present study employs National Solar Radiation Database (NSRDB) data for evaluating prediction accuracy at 7 locations of India having different climatic conditions. The proposed model is compared with Feed Forward Neural Network (FFNN), Extreme Gradient Boost (XGBoost) and Persistence model at broader coverage of geographical regions. Empirical outcomes suggest that proposed LSTM model outperforms different models with an average forecast skill of 50.72% over persistence model.

Keywords renewable energy, deep learning, long short term memory, clearness index and climatic condition

Copyright ©
Energy Proceedings