Abstract: To address the issues of inadequate data pattern mining and low efficiency in utilizing existing information, we propose the Basisformer Time Series Forecasting Model integrating Pattern Decomposition. First, a novel decomposition architecture is designed. This architecture breaks the convention of using sequence decomposition as a preprocessing step, integrating it as a basic internal block of the deep model, thereby enhancing the model's decomposition capability for complex time series. Then, the Basisformer model, which possesses powerful forecasting capabilities, is utilized to predict the time series patterns of different modes, yielding efficient and accurate predicted data. Finally, a linear layer module is designed to fuse the predicted data for output, consequently generating the final prediction result. This process improves the model's fitting ability and generalization ability, thus further enhancing the overall forecasting performance. The proposed model is validated using four public datasets covering energy, economics, and weather aspects. The experimental results demonstrate the superior forecasting performance of the proposed algorithm.
赵鑫浩, 王威娜. 融合模式分解的Basisformer时序预测模型[J]. 吉林化工学院学报, 2026, 43(1): 81-87.
ZHAO Xinhao, WANG Weina. Time Series Prediction Model Combining Basisformer and Pattern Decomposition. Journal of Jilin Institute of Chemical Technology, 2026, 43(1): 81-87.
