Comparative Analysis of Neural Network and Machine Learning Models for Short-Term Traffic Flow Prediction on Shenzhen Expressway

Introduction. With highway congestion increasing, the efficiency of intelligent transportation systems depends on highquality short-term traffic prediction. Statistical methods do not adequately account for nonlinear and dynamic traffic changes. Long short-term memory (LSTM) and support vector machines (SVR) offer more promising solutions. However, they are not ranked in terms of accuracy, as there are no studies comprehensively comparing their adequacy for shortterm traffic flow prediction. The proposed study fills this gap. The research objective is to compare the accuracy of LSTM and SVR, and select the optimal approach for traffic flow prediction on Shenzhen Meiguang Expressway. Materials and Methods. Traffic detector data was collected on the Meiguan Expressway in June 2021. Data preprocessing methods were used, including weighted mean imputation and normalization. Autocorrelation analysis was used for feature extraction, along with the creation of an interaction variable between speed and detector occupancy. Models were trained and tested on data collected from detectors at 5-minute intervals. Results. LSTM performed 17.86% better in terms of root mean square error, 19.82% better in terms of mean absolute error, and 25.78% better in terms of mean absolute percentage error. In periods with the lowest flow rate prediction error, RMSE, MAE, and MAPE for the LSTM model were 36.5%, 34.3%, and 42.3% lower, respectively. In periods with the highest error, RMSE, MAE, and MAPE for the LSTM model were 73.2%, 65.4%, and 64.4% lower, respectively. The Wilcoxon signed-rank test <0.05 confirmed the statistical significance of the differences. Discussion. The superior predictive performance of LSTM stems from its architecture, namely, the combination of interaction variables and lag metrics. LSTM accounts better for flow time dependences, adapts to complex, long-term dynamic changes, and remains accurate even with significant fluctuations. The lower predictive performance of SVR stems from its weak, nonlinear approximation ability. Sudden flow changes increase significantly error rates. Conclusion. When choosing between a neural network and a machine learning model for short-term traffic flow prediction on an expressway, the neural network model, such as LSTM, should be preferred. These research results can be useful in predictive strategies for reducing congestion. Short-term prediction based on LSTM can serve as a basis for optimizing traffic management, reducing congestion and pollutant emissions, and for optimizing intelligent transportation systems. A promising direction is the development of hybrid architectures that integrate contextual data (weather, infrastructure, accidents) to improve real-time predictions.