An Efficient Spike-Inspired Differential Attention Model for Traffic Flow Forecasting

Abstract

Traffic flow forecasting is a fundamental task in intelligent transportation systems, and accurate multi-step prediction depends critically on effective modeling of both temporal dynamics and spatial dependencies within sensor networks. On the one hand, forecasting models must possess strong spatiotemporal interaction modeling capabilities; on the other hand, computational complexity must be carefully controlled to satisfy practical deployment requirements. In recent years, spiking neural networks have shown considerable potential for efficient sequential modeling tasks due to their event-driven nature, sparse activations, and low-power computational characteristics. Transformer-based methods effectively capture temporal dependencies but suffer from high computational costs, while lightweight channel-independent and MLP-based approaches improve efficiency at the expense of limited cross-node interaction modeling. To address these limitations, this paper proposes a spiking spatiotemporal forecasting framework for traffic flow prediction, which integrates a temporal core fusion mechanism and a spike-driven differential spatial attention mechanism within a unified architecture. Specifically, along the temporal dimension, an “aggregate-then-redistribute” strategy is employed to compress and reconstruct temporal context, thereby enabling efficient global temporal interaction while avoiding the quadratic complexity of conventional attention. To capture spatial correlations, we further propose a Spatial Spike-Driven Differential Attention module, which combines multi-level spike quantization, softmax-free linear attention, and local differential priors to model inter-node dependencies with relatively low computational overhead. By integrating temporal embeddings, calendar-aware features, and adaptive node representations, ESDAM enables efficient traffic flow forecasting under complex spatiotemporal dynamics with a favorable balance between accuracy and efficiency. Experimental results show that ESDAM achieves competitive performance on commonly used forecasting error metrics, including MAE and WAPE. Moreover, evaluations based on information-theoretic criteria, including AIC, SBIC, HQIC, and $\mathrm {AIC}_{c}$ , further demonstrate that ESDAM achieves a favorable trade-off among fitting quality, model complexity, and parameter efficiency.