Clustered Spatio-Temporal Traffic Prediction with K-Means and GCN-LSTM Neural Networks: A Case Study in Santiago, Chile

Urban traffic congestion remains a critical challenge for logistics and mobility worldwide, particularly in contexts where the rapid expansion of urban transportation demands more efficient and sustainable solutions. From a technical perspective, traditional forecasting models often overlook the intricate spatio-temporal interdependencies between geographic locations and traffic speed patterns, thereby limiting their ability to accurately predict real-world traffic dynamics. In this work, we address this limitation by implementing a hybrid architecture that combines Graph Convolutional Networks (GCNs) with Long Short-Term Memory (LSTM) networks. Vehicular speed data collected via GPS sensors are transformed into graph-structured representations using geospatial clustering from approximately 22 million GPS-based speed records. We present a comprehensive study that includes data preprocessing, imputation of missing values, and a comparative evaluation of different models. Our experiments show that the proposed GCN-LSTM approach significantly outperforms classical baselines such as linear regression and standalone LSTM models, yielding higher R 2 values, from 0.29 to 0.53 at 1-hour granularity, and reaching 0.64 at 3-hour granularity. The results highlight that incorporating spatial clustering structures through GCNs enables a more accurate modeling of interdependencies across urban zones, leading to substantial improvements in predictive performance, particularly in metrics such as MAE and R 2 . These findings pave the way for the development of graph-based intelligent traffic management systems with direct applications in urban logistics, transportation planning, and decision-support systems.