Highway traffic flow prediction method based on multiscale CNN-BiLSTM network

Precise traffic-flow forecasting underpins modern intelligent transport and traffic management. To counter the limited ability of current approaches to mine multi-resolution temporal patterns and bidirectional sequential relations, we introduce a hybrid highway-volume prediction architecture that couples multi-scale convolutional blocks with a bidirectional long short-term memory network. The pipeline embraces three stages: data cleaning, multi-scale CNN-BiLSTM representation learning, and multi-step-ahead forecasting. Parallel convolutions with kernel sizes 3, 5 and 7 first harvest localized dynamics at diverse temporal resolutions; a BiLSTM then encodes both past and future contexts, yielding accurate predictions for lead times of 15, 30, 45 and 60 min. Evaluations on real freeway detector data show robust performance across all horizons: average R² = 0.9329, RMSE = 13.20, MAE = 10.12 and MAPE = 13.75%. The 30-min forecast achieves the best fit, delivering R² = 0.9356, RMSE = 12.95, MAE = 9.85 and MAPE = 12.97%. Ablation tests confirm the individual value of the multi-scale CNN and BiLSTM modules, raising R² by 8.06% over a plain LSTM baseline. Relative to ARIMA, XGBoost and CNN-Transformer, the proposed model ranks first on every examined metric.