Statistical evaluation of a diversified surface solar irradiation data repository and forecasting using a recurrent neural network-hybrid model: A case study in Bhutan

Accurate prediction of surface solar irradiance (SSI) and selection of appropriate data sources are challenging in regions with diverse topography like Bhutan, affecting solar energy management and climate modeling. This study evaluated various SSI data sources, developed a hybrid ARIMA-LSTM-attention mechanism (AM) model for SSI prediction, and compared its performance with ARIMA and various RNN models. Ground-based measurements were employed to assess the accuracy of satellite and DSM-derived SSI datasets, with NASA POWER identified as the most reliable source (RMSE: 36–130.7 W/m 2 ; Average Pearson ‘r’: 0.76). The proposed hybrid model integrates ARIMA for linear trend capture, LSTM for nonlinear pattern recognition, and AM for relevant input sequence identification, validated through K-fold cross-validation. This approach demonstrated superior performance across most SSI stations, with RMSE ranging from 5–8.45 W/m 2 , MSE from 26.42-72.10 W/m 2 , MAE from 3.71-7.06 W/m 2 , and MAPE from 2.07-4.46 %. The study also highlighted the efficacy of Bayesian optimization for site-specific hyperparameter tuning and L1/L2 regularization to mitigate overfitting in RNN and hybrid models. This research advances SSI prediction methodologies by highlighting the synergy between high-quality radiation data, hybrid modelling approaches, and advanced optimization techniques, enhancing prediction accuracy and facilitating solar energy utilization in diverse topographical contexts • Evaluated various satellite and digital surface model (DSM)-derived Surface Solar Irradiance (SSI) datasets, with a focus on their reliability and accuracy in rugged topographical regions. • NASA POWER achieved higher accuracy than other SSI data sources while comparing it with ground-based SSI. • Developed of ARIMA-LSTM-Attention Mechanism (AM), a hybrid model to predict SSI including tuning using Bayesian approach. • Required location specific hyperparameters tuning for the hybrid and RNN models. • Demonstrated superior performance of hybrid ARIMA-LSTM-AM from ARIMA and other variant of RNN models.