Design and evaluation of adaptive deep learning models for weather forecasting

Deep learning has emerged as a promising tool in time-series prediction tasks such as weather forecasting, and adaptive models can deal with dynamic data more effectively. In this work, we first investigate how successfully meteorological features can be predicted by a deep learning model based on long-short-term memory (LSTM). Then, we endeavor to improve the prediction model's performance by employing various LSTM types and choosing a model type that provides robust and accurate results. After that, we extend the proposed model to deal with univariate and multivariate problems, and we compare them. Finally, we apply the adaptive learning concept to the selected model by retraining and updating the model periodically to improve its accuracy. The experimental findings demonstrate that applying adaptive learning in the bidirectional LSTM-based model decreases the prediction error by 45% compared to the baseline models. Moreover, the results reveal that exploiting only the univariate model leads to learning from fewer features; and thus, less time and memory consumption for the model construction and usage.