A Comparative Study of Adaptive Time Series Models for Weather Forecasting in Indian Scenarios

Given the erratic changes in the weather and climate patterns across the globe and their ramifications on the livelihood, safety and security of human and animal communities, weather forecasting has evolved as a pressing task to ensure the sustainability and security of these communities. In this work, the temperature data of different places across India are considered for training different conventional and deep learning-based time series models; specifically, while ARIMA and SARIMAX models are trained in conventional settings, LSTM networks are trained in deep learning settings. The performance of these models is evaluated using AIC, BIC and HQIC metrics, followed by a systematic comparison of their modelling performances. Empirical results show that LSTM-based models capture long-term dependencies in the time series data better than conventional time series models; this improved modelling using LSTMs is quantified using AIC, BIC and HQIC scores for different places across India. The AIC, BIC, and HQIC values reported by the LSTM models are several orders of magnitude lesser than those reported by ARIMA and SARIMAX models; consequently, LSTMs are then deployed for time series prediction using temperature data of places considered in the study.