Real-Time Traffic Flow Prediction Model Using Deep Learning Models

This article derives and evaluates a real-time urban traffic-flow forecasting end-to-end system from the METR-LA loop-detector data set. After reviewing statistical time-series, conventional machine-learning, and deep-learning methods, we identify the need for one spatiotemporal approach that handles missing data, nonuniform sensor inputs, and deployment constraints. We then introduce a three-stage pipeline: (1) a preprocessing module that imputes sensor outages, applies per-sensor Z-score normalization, and augments each time step with cyclical time-of-day and day-of-week features; (2) a hybrid CNN–GCN–LSTM forecasting model that learns local spatial patterns using 1D convolutions, global network structure using graph convolutions, combines these representations using optional attention, and learns temporal dynamics using an LSTM decoder; and (3) an inference optimization suite that combines magnitude-based pruning and 8-bit quantization and exports the compressed model as a TorchScript artifact for sub-200 ms streaming prediction on edge and CPU hardware. Trained with sliding-window samples, Adam optimization, learning-rate scheduling, early stopping, and transfer learning from a larger PeMS-Bay dataset, the model reaches one-step RMSE of approximately 11.2 vehicles/5 min and MAPE of 8.7 %, significantly outperforming ARIMA, SVR, single LSTM, CNN–LSTM, and GCN–LSTM baselines, and demonstrating robustness to sensor dropouts and quick convergence. These results confirm that the framework we have introduced makes both theoretical contributions and contributions to real-world deployment of intelligent-transportation systems, and we consider improvements for the future including incorporation of exogenous data, dynamic graph adaptation, meta-learning for cold-start sensors, and federated on-device training.