Uncertainty quantification in sequential hybrid deep transfer learning for solar irradiation predictions

Hybrid deep learning model with multi-frequency capabilities is presented for simulating solar irradiation. Utilizing hourly recorded solar irradiation and climate data, model employs Convolutional Neural Network (CNN) to capture spatial and Long Short-Term Memory (LSTM) network to capture temporal characteristics. Initially, local models are developed individually, trained on single datasets. Additionally, global models are trained collectively by fusing multiple datasets. Furthermore, the study delves into transfer learning. To evaluate the uncertainty of prediction, Prediction Intervals (PIs) are computed using the bootstrap method and its results are compared with those of Bayesian Neural Network (BNN). The findings indicate that hybrid CNN-LSTM model surpasses single models in point predictions up to 25% in test. Local models demonstrate efficacy when deployed within their respective regions but exhibit notable errors when data scarcity hampers training. Global models exhibit adaptability to individual locations at expense of training efforts. Pre-training models on comprehensive and diverse source dataset, followed by transfer to a target dataset, generally yields superior performance. Concerning the model's performance in terms of PIs, CNN-LSTM model demonstrates the highest efficacy with average Prediction Interval Coverage Probability (PICP) of 0.82 and average Normalized Mean Prediction Interval Width (NMPIW) of 0.043. Also, in strategies involving limited or no available data for training, CNN-based models exhibit superior performance with lower uncertainty.