WeathSOM: A Multi-View Self-Organizing Map Architecture for Efficient Clustering and Prediction of Weather Time Series

A time series is a collection of finite and discrete real values that change dynamically. Time series analysis is the task of identifying patterns and trends with the aim of predicting future values or labels. In this work, a Self-organizing map (SOM) architecture named WeathSOM is developed to perform multi-view learning, find stable clusters for an ensemble of multivariate time series data, and explore the climate data. WeathSOM trains Dynamic Time Warping (DTW) SOMs independently to produce generalizable and stable clustering results. Then, it applies feature extraction methods to eliminate multicollinearity and reduce the number of features. Besides the architecture, an innovative model is proposed to predict the classification labels of the new observations. WeathSOM is trained with a partitioned form of temperature and humidity observations from the Szeged and Berlin historical weather datasets. Evaluation results, relying on the clustering metrics, reveal the superiority of the architecture compared to basic SOM. They signify the efficiency of this architecture in reducing extraneous computations and improving the execution time. In addition, the values of the prediction accuracy depict the effectiveness of the proposed model in predicting weather conditions.