In order to manage urban traffic and assist people in making wise travel decisions, timely and accurate traffic speed forecasts have become extremely important. In this study, a new Neural Network (NN) structure with a Chaotic Particle Swarm Optimization (CPSO) technique is introduced into the two-layer bidirectional long short-term memory network, termed CPSO-DBLSTM, with the aim of improving forecast accuracy of traffic speed. To accelerate the convergence of the Particle Swarm Optimization (PSO) algorithm, CPSO method was applied, which is an optimization strategy based on chaotic theory and PSO. Chaotic theory was applied to the PSO technique to balance the exploration and exploitation phases. Evidently, it was found that the CPSO algorithm is more effective than the PSO algorithm by rapidly deviating from local optima due to the incredible behavior of chaos and its good capability. In comparison to Long Short-Term Memory (LSTM) NN and single-layer bidirectional LSTM (Bi-LSTM) NN, it was found that Bi-LSTM neural networks with two layers perform better. To improve the prediction accuracy of two-layered Bi-LSTM model, it was combined with the chaotic PSO technique which reduces the impact of a random selection of hyperparameters on the prediction performance. The proposed model has been tested on the real traffic data which were collected from selected locations in Delhi. According to the study results analysis, proposed model was found superior as compared to other baseline models in terms of evaluation metrics.