[Purposes] To address the limitations of existing traffic flow forecasting methods, which fail to adequately account for the merging behaviour of urban expressway ramps and the spatio-temporal dynamics of traffic flows. [Methods] A multi-input single-output traffic flow prediction model for convergence zones is proposed. The integration of an enhanced Arnold and Levy sparrow search algorithm (AL-SSA) with convolutional neural networks-bidirectional long short-term memory (CNN-BiLSTM) is considered to enhance predictive accuracy and adaptive capabilities. To mitigate the impact of hyperparameter uncertainty and manual tuning costs within the CNN-BiLSTM framework, the AL-SSA is designed to dynamically optimise three critical hyperparameters: the number of hidden units, the initial learning rate, and the L2 regularisation coefficient. The optimised parameters are subsequently employed to train the CNN-BiLSTM network. [Findings] Experimental validation on the Changsha real-world dataset demonstrates that the proposed ALSSA-CNN-BiLSTM model outperforms baseline models such as SSA-CNN-BiLSTM. At a 15-minute prediction horizon, it achieves reductions of 7.59%, 8.67% and 7.64% in MAE, RMSE and SMAPE respectively. [Conclusions] Compared to existing algorithms, this model demonstrates enhanced accuracy and stability. By providing more precise short-term traffic information, it can improve road traffic conditions, offering a forward-looking and practical technical approach for urban traffic flow forecasting.