A hybrid approach to weather seasonality study and forecasting

In the recent years, climate change produces a profound impact on the seasonality shifting. For example, even a slight variation in temperature may trigger an early spring start. This study presents a hybrid application of machine learning and deep learning to analyze the shifts in seasons across contiguous years and to accurately forecast weather variables. The daily meteorological features (minimum temperature, maximum temperature, wind speed, precipitation, snow cover, etc.) collected for the City of Toronto (Canada) area over a 20-year period spanning from 1st January 2001 till 31st December 2020 are used in the study. The temporal data has been pre-processed to normalize and rescale into a [0,1] range. The temporal data points are clustered to form distinct seasons applying various clustering algorithms: K-Means, Agglomerative hierarchical, Mean-shift, Affinity Propagation and Gaussian Mixture. The performance of these algorithms is compared to determine the most accurate seasonal clusters. The results show that the K-Means and Agglomerative hierarchical with ward linkage perform significantly better than other clustering algorithms to group temporal weather data. Obtained clusters are then used to analyze the shifts in seasonality for the period of 20 years. Furthermore, cascaded multi-layered Long Short-Term Memory (LSTM) deep learning model is used to forecast weather variables. Different variants of LSTM (i.e., single layer, multi-layer, and bidirectional) are evaluated to produce computationally inexpensive, lightweight, and accurate model.