Short-Term Freeway Traffic Volume Prediction with Weather and Holiday Features: A Comparative Study of LSTM and Transformer Models

Abstract

The task of predicting short-term traffic flows accurately during abnormal conditions like adverse weather or holidays remains a significant problem in Intelligent Transportation Systems, particularly when using isolated single-link detectors. This paper presents a freeway congestion forecast integrating weather variables and infrequent calendar events into individual link sensor networks. Both Long Short-Term Memory (LSTM) and Timeseries Transformer models are tested on multivariate I-94 highway data. Results indicate that the LSTM-based architecture, with strong feature normalization, achieves a Mean Absolute Percentage Error (MAPE) of 11.76%, outperforming the transformer model on isolated univariate time series. Furthermore, scenario-level evaluation demonstrates a significant weakness in current prediction frameworks: although they are precise in normal weather conditions, both models show terrible performance deterioration in extreme weather situations (such as snow fall, extreme rainfall) where error rates exceed 16 %. These results suggest the weaknesses of pure temporal autoregression and emphasize the importance of multidimensional context encoding, but at the same time they point out the ongoing difficulties of modelling long tailed calendar anomalies (e.g. the rare holiday events) within efficient traffic control.