Data-driven deep learning model for predicting ambient temperature: environment and solar energy

Abstract This study proposes and evaluates a hybrid gated recurrent unit--long short-term memory (GRU--LSTM) deep learning (DL) model to forecast ambient temperature, a key factor influencing photovoltaic (PV) module temperature and efficiency. Ambient temperature varies due to a range of environmental and weather conditions and is not consistently predictable during daytime. However, it is often assumed to remain constant in PV module design and solar power production calculations. To address unsolved issue, the study exploits the power of DL by introducing a novel hybrid model to forecast ambient temperature based on historical data, focusing on one month of measurements. The proposed model demonstrates predictive solid performance, with a mean absolute error (MAE) ranging from 0.024 to 0.046, root mean squared error (RMSE) in the range of 0.032 to 0.061, and an R 2 value between 0.882 and 0.962. Temperature data from the Icelandic Meteorological Office (IMO) and the Danish Meteorological Institute (DMI) is used to test the model across two locations in two countries, highlighting the model's robustness. Model training and inference are run on the high-performance computing (HPC) DEEP-DAM system at the Jülich Supercomputing Centre. Accurate ambient temperature forecasting enhances the precision of solar power production predictions and aids in managing power generation in hybrid wind-solar power plants.