Methodology Based on BERT (bidirectional Encoder Representations from Transformers) to Improve Solar Irradiance Prediction of Deep Learning Models Trained with Time Series of Spatiotemporal Meteorological Information