Developing a forecasting model for time series based on clustering and deep learning algorithms

Abstract

This study proposes a new forecasting model for time series based on the improvement and combination of the cluster analysis (CA) algorithm and deep learning with Convolutional Neural Network (CNN) and Bi-Long Short Term Memory (BiLSTM) model. The proposed model is considered pioneering in this research direction with significant contributions to three main phases. For the first phase, the original series is converted into the percentage change series and is divided into clusters of an appropriate number using the CA algorithm. The next phase involves extracting the features of the new series based on the CNN with suitable parameters and input data enhancement from the results of the first phase. In the final phase, the BiLSTM model is applied to the series established from the second phase, and the forecasting principle for the future is established. The proposed model is detailed in the implementation steps, proving convergence, illustrated by numerical examples, and can be applied to real series using a Matlab procedure. The effectiveness of the proposed model is quite impressive as it surpasses many strong forecasting models on reputable benchmark datasets , including the M3-Competition dataset with 3,003 series, and M4-Competition dataset with 100,000 series.