مشخصات پژوهش

t In many industries, such as banking, retail, medical, and tourism, abundant time series data is generated. Global forecasting methods have emerged to train a single model by leveraging cross-series information. However, the performance of the global models may be decreased when applied to heterogeneous unequal-length time series datasets. To address heterogeneity in time series, this study contributes a comprehensive methodology for time series forecasting, incorporating three key components: (1) automatic feature extraction from unequal-length time series utilizing two novel autoencoder models based on long shortterm memory and convolutional neural network methods, (2) time series clustering, (3) and global forecasting. To evaluate the forecasting accuracy of the proposed methodology, we conduct comprehensive experiments across various time series datasets including six datasets of M3, Tourism, and Hospital datasets. Across all datasets with 2620 time series, attained the lowest mean symmetric mean absolute percentage error (sMAPE) of 15.08, surpassing the baseline mean sMAPE of 15.42. It exhibits enhancements of 0.19 and 0.60 in mean sMAPE relative to ETS and ARIMA, respectively. Furthermore, it demonstrates improvements of 1.10, 0.52, and 2.11 compared to DeepAR, N-BEATS, and Transformer, respectively. A comparative analysis of our proposed methodology against clustering-based and ensembling-based models in the existing literature highlights its superiority in global time series forecasting.