A Multi-Scale Traffic Flow Prediction Model Based on LSTM

With the accelerating pace of urbanization, traffic flow prediction faces challenges in adapting to dynamic environments. Existing studies exhibit significant limitations in integrating external factors and extending to multi-scale prediction scenarios. Targeting the domain of traffic flow forecasting, this paper focuses on optimizing the Long Short-Term Memory (LSTM) network model, with particular attention to the impact of weather conditions on prediction accuracy and the performance differences between hourly and daily multi-scale forecasting. While mainstream approaches have improved LSTM performance through algorithmic optimization (e.g., Particle Swarm Optimization, Bayesian Optimization), a systematic solution to the integration of external factors and adaptation to different time granularities is still lacking. This study proposes an LSTM architecture that incorporates temperature as an embedded parameter, constructing a multi-factor input model and designing a dual-scale prediction framework at both the hourly (24-hour window) and daily levels (7/30/90-day windows). Using traffic flow and meteorological data from Interstate 94 in Minnesota, USA (20122018), the research explores the trade-off between external factors and time scales in LSTM modeling. The results provide a refined optimization path for traffic flow forecasting under complex scenarios.