Long-term and short-term traffic flow prediction with different weather conditions using optimized wheel-graph attention based neural network

Accurate traffic flow prediction remains paramount in intelligent transportation systems, particularly during adverse weather conditions. Conventional models usually fail to concurrently capture spatial-temporal dependencies and environmental impacts, thereby limiting performance. This research proposes a novel hybrid approach, the wheel-graph attention-based bidirectional long short-term memory network with enhanced Fick’s law algorithm, to address this issue. The approach combines sparse non-negative matrix factorization for feature extraction, a hybrid giant trevally optimizer for feature selection, and the skill kepler-based fusion strategy for the robust combination of traffic and weather data. Experimental results show that the proposed approach achieves a Mean Absolute Error (MAE) of 5.25%, Root Mean Square Error (RMSE) of 8.91, and an accuracy of about 95%, significantly outperforming baseline models, which typically report MAE values between 12 and 19. Overall, the method yields a 35-40% improvement in prediction accuracy over existing approaches. High-precision industrial systems set a target of 2% MAE or less, but the proposed approach provides great accuracy in the real-world concerning weather variations across short-term, medium-term, and long-term horizons. The research, therefore, offers a scalable, weather-wise solution for traffic predictions in intelligent transportation systems.