Fuzzy time series clustering using autoencoders neural network

This study presents a novel approach that employs autoencoders (AE)—an artificial neural network—for the nonlinear transformation of time series to a compact latent space for efficient fuzzy clustering. The method was tested on atmospheric sea level pressure (SLP) data towards fuzzy clustering of atmospheric circulation types (CTs). CTs are a group of dates with a similar recurrent SLP spatial pattern. The analysis aimed to explore the effectiveness of AE in producing and improving the characterization of known CTs (i.e., recurrent SLP patterns) derived from traditional linear models like principal component analysis (PCA). After applying both PCA and AE for the linear and nonlinear transformation of the SLP time series, respectively, followed by a fuzzy clustering of the daily SLP time series from each technique, the resulting CTs generated by each method were compared to assess consistency. The findings reveal consistency between the SLP spatial patterns from the two methods, with 58% of the patterns showing congruence matches greater than 0.94. However, when examining the correctly classified dates (i.e., the true positives) using a threshold of 0.8 for the congruence coefficient between the spatial composite map representing the CT and the dates grouped under the CT, AE outperformed PCA with an average improvement of 29.2%. Hence, given AE's flexibility and capacity to model complex nonlinear relationships, this study suggests that AE could be a potent tool for enhancing fuzzy time series clustering, given its capability to facilitate the correct identification of dates when a given CT occurred and assigning the dates to the associated CT.