Traffic speed sparse time series prediction model integrating spatiotemporal periodic features

Sparse time series forecasting (SparseTSF) is a recently proposed lightweight multi-step forecasting model with advantages such as high computational efficiency and wide adaptability. However, the SparseTSF model also has some limitations, for example, it can only downsample a single main period, making it difficult to handle multi-period data. Based on the characteristics of traffic forecasting, this paper integrates weekly and spatial feature extraction modules to extract deep features that fuse daily, weekly, and spatial features. The SparseTSF model is then optimized to construct a multi-period spatial time series forecasting (MSTSF) model. This model is applied to traffic speed forecasting scenarios, aiming to reduce prediction error and achieve a balance between performance and parameter size. Experiments on the Guangzhou traffic and Performance Measurement System (PeMS) datasets show that the MSTSF model performs well in both prediction accuracy and model efficiency. Compared with mainstream deep learning models, this model achieves lower prediction errors. The MSTSF model has advantages such as small parameter size, high iteration efficiency, and fast inference speed, making it suitable for scenarios with limited computational resources. Our model achieves better results in traffic speed forecasting.