Abstract Weather forecasting is a complicated process that involves the analysis of giant meteorological data set in order to forecast what the weather will be in future stages. More recently, statistical and machine learning techniques have been widely applied as conventional forecasting models yet are likely to fail in instances where it comes down to the non-linear nature of relationships between multivariate data in context which are likely to have complex dependence. The recent advances in deep learning (in particular, transformer-based designs) showed state-of-the- art performance in time series forecasting, as they are more successful in covering long-range dependencies and intricate patterns. Multi-variate time series forecasting of Weather prediction through transformers is a burning subject and this paper will set out to discuss the necessary performance inducing techniques (temperature, precipitation and humidity) of the key variables as concerns time series pattern. To make the forecasting more accurate we consider the myriads of methods available starting with data preprocessing and feature selection and extending to model fine-tuning. Our work is based on the literature, and it proves better results of transformer models in comparison with classical methods (RNNs and TCNs). To address the computational problem we use unsupervised pre-training and spatiotemporal analysis whereby we aim to enhance prediction accuracy. Findings of this research provide a good basis to develop more reliable and quicker weather forecasting to give advancements in climate science and meteorology.Keywords: Transformer Models, Weather Forecasting, Multivariate Time Series, Deep Learning, Meteorological Prediction, CNN-LSTM Hybrid, Spatiotemporal Analysis, Time Series Forecasting.
Sumithra et al. (Fri,) studied this question.