Distributed Evolutionary Hyperparameter Optimization for Fuzzy Time Series

Time series forecasting is an essential task in the management of Smart Cities and Smart Grids, becoming even more challenging when it needs to deal with big data time series. The development of highly accurate machine learning models is yet harder when considering the optimization of hyperparameters, which is an expensive computational task. To tackle these challenges this work proposes the Distributed Evolutionary Hyperparameter Optimization (DEHO) for the Weighted Multivariate Fuzzy Time Series method (WMVFTS), a simple and non-parametric forecasting method with high scalability and accuracy, comprising a sequential training and forecasting procedure and a MapReduce extension for distributed processing. The proposed methods were evaluated using a cluster with commodity hardware and two big time series, showing increasing speed up for training and test times as new CPU cores are added to cluster. Then the DEHO method was executed in the computational cluster, achieving fast convergence and feasible processing time and generating highly accurate WMVFTS models.