A robust kernel-based fuzzy clustering algorithm for time series forecasting

Abstract

This paper proposes a new forecasting model for time series based on the Kernel-based fuzzy clustering algorithm. In the proposed fuzzy clustering model, the algorithm performs according to two phases. The first phase determines the optimal number of clusters based on the clustering problem. This phase uses the similarity coefficient index (SCI) as the evaluative measure for the distance of elements. In the second phase, the algorithm establishes the fuzzy relationship among elements using the Kernel-based fuzzy c-mean algorithm, and subsequently performs de-fuzzification and forecasts the time series according to a proposed mathematical rule. The proposed model is evaluated on the various benchmark datasets such as the Enrollment dataset, the Outpatient dataset, the M3-competition dataset, and the Covid-19 datasets of infections and fatalities. The forecasting performance of this model is also compared to that of other models using two statistical assessment criteria: the mean absolute error (MAE) and the symmetric mean absolute percentage error (sMAPE). The models are verified on a range of forecasting horizons and the errors are averaged over multiple tests to have a better accurate performance evaluation. The experiment results show that the fuzzy clustering model outperforms the other models on all the experimented datasets