Spike-TCN with Depthwise-Separable Convolution and Attention for Efficient Time Series Prediction

Abstract Spiking Neural Networks, known for their event-driven and energy-efficient characteristics, offer promising potential in temporal modeling. However, existing SNN-based models often struggle to capture long-range dependencies and inter-channel interactions, limiting their performance in complex time series prediction. To address these challenges, we propose the Spiking Depthwise Separable Temporal Convolutional Network (SDSTCN), which replaces conventional convolutions with depthwise separable convolutions. This decoupling of temporal and channel-wise computations enables fine-grained feature extraction while significantly reducing parameter overhead and redundant updates in spike propagation. To further enhance representational capacity, we design the Spiking Global–Local Attention (SGLA) module—an attention framework that models global inter-channel dependencies and local spatio-temporal patterns in parallel. Experimental results demonstrate that the proposed method improves prediction accuracy while maintaining low computational complexity across multiple public benchmark datasets.