مشخصات پژوهش

Global forecasting models (GFMs) have become essential in time series prediction, as they enable cross-learning across multiple series. Although GFMs have consistently outperformed univariate approaches, their performance decreases when applied to heterogeneous time series datasets, such as those found in economic and financial applications. Clustering techniques have been used to create homogeneous time series clusters. However, the main limitations of current clustering-based GFMs are as follows: (1) employing handcrafted features instead of deep learning and (2) there is no guarantee that the resulting clusters are optimal in terms of prediction accuracy. To address these limitations, we propose a novel deep time series clustering model that jointly optimizes clustering and forecasting accuracy. The proposed method simultaneously optimizes the reconstruction, clustering, and prediction losses to ensure clusters are optimized for accurate forecasting. In addition, it employs a neighbor-aided autoencoder to capture cluster-oriented representations, leveraging neighboring time series to improve feature learning. Furthermore, we incorporate an evolutionary learning component, which iteratively refines clusters through crossover and mutation to find optimal clusters in terms of forecasting accuracy. We evaluate our proposed method on eight publicly available datasets considering various state-of-the-art forecasting benchmarks. Results indicate that across all datasets with 2620 time series, the proposed method obtains the lowest mean symmetric mean absolute percentage error (sMAPE) of 14.90, surpassing the baseline deep clustering (15.15). It exhibits enhancements of 1.28, 0.70, and 2.29 in mean sMAPE relative to DeepAR, N-BEATS, and transformer, respectively. Furthermore, it demonstrates improvements when compared to the existing clustering-based global models. The source code of the proposed clustering method is made publicly available at https://github.com/alinowsh