Evaluation of Different Deep Learning Methods for Meteorological Element Forecasting

Abstract

Deep Learning (DL) models can make short- and long-term predictions in just a few seconds, beyond the capabilities of traditional physical models. However, the capabilities of different DL models for meteorological element forecasting, are still yet to be comprehensively evaluated. Here, DL models were used to forecast multiple meteorological elements, including temperature (T), surface net solar radiation (SSR), soil moisture (SM), and evapotranspiration (ET). We compared the seven models in term of training performance, prediction accuracy, and the effects of parameters. We found that the training of RNN-based models (LSTM, GRU, and Bi-LSTM) was faster than others. However, with sufficient training epochs, Transformer-based models consistently achieve the lowest loss function. Among the Transformers, the Informer demonstrates the best prediction accuracy in most scenarios. Beyond the choice of DL model, the prediction performance is also influenced by the meteorological element itself. The MTGNN is comparable to Transformer-based models for T and SSR forecasting, but it does not perform as well as the Informer for SM and ET. The sliding window size and prediction time step have a slight impact on the performance differences between the models. The results can offer insights into applying DL models in meteorological element forecasting.

Keywords

• Meteorological elements forecasting
• deep learning
• RNN
• GNN
• transformer