Optimized deep recurrent neural networks for high-resolution solar radiation modeling in Turkey

Accurate estimation of solar radiation (SR) is essential for renewable energy planning, particularly in high-potential countries such as Turkey. This study evaluates three deep recurrent neural network architectures—Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and Gated Recurrent Unit (GRU)—for long-term monthly SR estimation across Turkey. NASA POWER data covering 2004–2022 were used at 3600 spatial grid points, with latitude and longitude as input variables. Models were trained using the ADAM optimizer and assessed through Root Mean Square Error (RMSE), Mean Absolute Error (MAE), R-squared (R²), and Pearson Correlation Coefficient (PCC). All models demonstrated strong predictive capability. BiLSTM achieved the lowest test errors (RMSE = 0.095; MAE = 0.074), indicating superior accuracy. GRU produced the highest R² (0.9373) and PCC (0.968), reflecting robust generalization performance. Spatial error analysis using Inverse Distance Weighting (IDW) interpolation showed that BiLSTM and GRU provided more geographically consistent estimations, especially in central and western Turkey. The results highlight the effectiveness of BiLSTM and GRU architectures for high-resolution solar radiation estimation and support their applicability in climate-sensitive and data-scarce regions for sustainable photovoltaic planning. These findings demonstrate that deep learning (DL) models can be effectively extended from temporal forecasting tasks to spatial estimation problems with limited input variables.