A Regional Attention Based LSTM Rectifying Method for Accurate and Efficient Traffic Speed Prediction

The short-term traffic speed forecasting has its significant importance for the modern intelligent transportation systems. To address the persistent challenge in maintaining predictive accuracy and computational efficiency simultaneously in large-scale road network forecasting when using existing methodologies, this study proposes a novel RAPR (Regional Attention Prediction Rectifying) framework that innovatively improves traffic speed prediction accuracy through a hierarchical architecture with spatiotemporal rectifying module and parallel prediction framework. The core of this method is to dynamically adjust local predictions using regional patterns. Eventually, compared to conventional approaches like standalone LSTM, experimental results show that RAPR gives an average of 16.5% higher prediction accuracy improvement while reduces computational complexity by 64.7% or more.