BiLSTM-KAN: A Time Series-based Traffic Flow Forecasting Model

Currently, the task of traffic flow prediction poses a significant challenge. Most existing works utilize spatial features for overall forecasting, without giving much consideration to the issue of extracting data features from individual time series data. In the absence of knowledge about road topology, extracting effective time series features for a single node becomes particularly important. The BiLSTM-KAN model proposed in this paper is an innovative sequence-to-sequence (Seq2Seq) forecasting model designed to significantly enhance the accuracy and efficiency of time series data processing. This model deeply integrates the advanced feature extraction capabilities of Bidirectional Long Short-Term Memory Networks (BiLSTM) with Kolmogorov-Arnold Networks (KAN), forming a unique and powerful data processing mechanism. With the application of the BiLSTM-KAN model, it will be possible to more accurately capture the complex dynamic changes and underlying patterns in time series data, thereby providing more reliable and precise forecasting results for traffic flow prediction and other related field sequence data processing tasks.