فهرست نشریات علمی دسترسی آزاد ایران

With the increasing generation of time series data, forecasting models trained on a set of time series, known as global forecasting models, outperform univariate forecasting models trained on individual series. However, the performance of global models may decrease when faced with heterogeneous data sets of time series with different lengths. In this study, a new method for localization of clustering-based global forecasting models is presented. The main steps of the proposed method include (1) extracting relevant features from each time series (2) clustering time series based on features extracted using K-Medoids and spectral clustering algorithms (3) implementing a global forecasting model using a combination of Temporal Convolution Network and its training for each cluster. To evaluate the prediction accuracy of the proposed approach, experiments were conducted on the M3 dataset that contains 1426 time series with unequal-length. The results of the experiments show the superior performance of the proposed clustering-based models compared to the baseline models and the benchmark models. The proposed model has 0.57 less error in terms of SMAPE metric.