Dynamic Time Warping and Spatial Attention Optimized Deep BiLSTM Model for Traffic Flow Prediction

Addressing the deficiencies in modeling temporal pattern similarities and deep nonlinear characteristics of traffic flow, we propose an optimized DSA-DBLSTM algorithm. This algorithm integrates an improved Dynamic Time Warping (FastDTW) algorithm and an attention mechanism into a Deep Bidirectional Long Short-Term Memory (DBLSTM) network, aiming to enhance the accuracy and efficiency of traffic flow prediction. We preprocess taxi GPS trajectory data using a distributed resilient dataset (RDD) and apply Kalman Filtering (KF) to remove noise. Then, we construct the DSA-DBLSTM algorithm, optimizing the DBLSTM neural network by incorporating the FastDTW algorithm and the attention mechanism. Experimental results on real taxi GPS trajectory data demonstrate that, compared with 13 benchmark algorithms including ARIMA, SVR, KNN, SAEs, BP, CNN, GRU, ANN, LSTM, BiLSTM, DBLSTM, DSA-LSTM, and DSA-BiLSTM across multi-scale datasets, the DSA-DBLSTM algorithm exhibits optimal predictive performance.