Multi-Modal Traffic Flow Prediction with Dynamic FPGA Reconfiguration for Efficient Edge Computing

Urban areas are in the face of persistent traffic congestion, which disrupts city running and hinders sustainability attempts. There is no doubt that accurate traffic paragon prediction is crucial for effective intelligent transportation systems to some extent. However, current methods are beset by two main challenges: inadequate modeling of intricate traffic dynamics and limited hardware capabilities at the edge.This study introduces MSTF-Net, a novel framework that integrates diverse neural network architectures using adaptive fusion techniques. We propose a two-layer filtering approach after thinking thoroughly: one layer leverages static geographic boundaries, while the other adapts to dynamic traffic conditions. In terms of the practical meaning, this method deeply enhances computational efficiency while preserving prediction accuracy. For the innovative part, we developed a variable-precision FPGA implementation for the BiLSTM component, expeditiously balancing computational accuracy and energy efficiency for real-world applications. Evaluations on the PEMSD4 dataset demonstrate that MSTF-Net outperforms existing approaches, achieving an RMSE of 28.2 for short-term predictions (15 minutes) and maintaining robust performance over longer times. Compared to traditional fixedprecision methods, the reconfiguration strategy for the BiLSTM module reduces energy consumption by 2.8 times. This optimized implementation achieves 5.6 times higher power efficiency, in contrast to equivalent GPU-based execution. This paper presents a comprehensive hardware-software co-design methodology, offering practical insights for developing efficient traffic systems, making it less affected by the constrained traffic resources.