LSTM parameter sensitivity-driven traffic flow prediction model optimization

With the continuous increase in the number of motor vehicles in major cities, traffic flow prediction has become a critical issue for alleviating congestion and optimizing travel strategies. Traditional models, such as Autoregressive Integrated Moving Average (ARIMA) and Kalman filtering, often struggle to capture the complex temporal characteristics of traffic data, resulting in relatively low prediction accuracy. Current mainstream research focuses primarily on deep learning-based models, which, although improving predictive performance, often lack systematic analysis and configuration of parameters. Moreover, the high complexity of deep learning models can hinder deployment efficiency in practical applications. This study focuses on the Long Short-Term Memory networks (LSTM) model and employs hourly traffic flow data from the city of Chicago for the year 2016. Through controlled experiments, it investigates the impact of parameters such as time steps and hidden layer dimensions on prediction performance. The results show that the LSTM model effectively captures the features of traffic flow data, with certain parameter configurations showing critical threshold effects. Additionally, single-layer and double-layer LSTM models demonstrate similar predictive performance. The findings verify the efficiency and high adaptability of the LSTM model in this context, provide theoretical support for model lightweighting, and offer guidance for optimizing models in resource-constrained real-world applications.