The current conventional traffic flow prediction methods adopted in Intelligent Traffic Systems (ITS) are severely frustrated by their static model architectures, fixed hyperparameter values, and restricted spatiotemporal learning capabilities, which are unable to model the nonlinear traffic dynamics in the real traffic system, abrupt traffic changes, and different characteristics of traffic flows in urban traffic systems. The limitations are causing poor prediction under high-dynamic and non-stationary traffic conditions, leading to inefficient Adaptive Signal Control (ASC), additional vehicular congestion, excessive fuel consumption, delayed emergency response operations, and reduced operational efficiency in smart city transportation infrastructure. To overcome these challenges, the research proposes an Adaptive Pattern Computation (APC) framework that combines Swarm Optimization-based automated hyperparameter tuning using Particle Swarm Optimization (PSO) and the Grey Wolf Optimizer (GWO) with a hybrid Deep Learning (DL) architecture. It combines Depthwise Separable Convolutional Neural Networks (DW-CNNs) for efficient spatial feature extraction, Long Short-Term Memory (LSTM) networks stacked to model temporal dependencies, and multi-head self-attention to maximize contextual learning and enable adaptive traffic prediction. The PSO-GWO optimization engine is capable of dynamically exploring the hyperparameter space of the CNN-LSTM model, and convolutional filter dimensions, hidden neuron configurations, attention heads, and learning rate schedules, and converges 43.1% faster than a grid search. The experimental validation on benchmark datasets, PeMS Bay Area, METR-LA, Urban Traffic Data, revealed excellent results with 98.3% prediction accuracy, 1.87 MAE, 2.94 RMSE, 3.12% MAPE, and 23 ms prediction latency, which significantly surpassed the performance of the LSTM, GRU, ARIMA, and Transformer baseline models and also offered scalable, energy-efficient, and real-time intelligent traffic management solutions.