Scalable Traffic Flow Forecasting Using Floating Car Data, Clustering, and Deep Learning Models

Cities and urban areas often struggle with road network oversaturation events leading to longer travel times, increasing pollutant emissions, and operational inefficiencies. To help better manage these negative effects, traffic management organizations depend on constant access to traffic data. Fixed sensors can provide this kind of data but with high costs that often lead to limited coverage, while Floating Car Data (FCD) generally provides extensive coverage but with partial sampling. This paper proposes an innovative data-driven framework that makes use of FCD data integrated with fixed sensor data, clustering methods, and machine learning models with the purpose of estimating current and forecasting future short-term traffic flows. To accomplish this, an unsupervised clustering of road segments based on similar traffic patterns is performed to allow the use of specialized machine learning models. LightGBM models are applied to segments characterized by low-variability traffic patterns, while CNN-BiLSTM models are adopted for highly variable conditions to infer actual traffic volumes from FCD samples. Afterwards, a single-layer LSTM model is employed to perform forward forecast predictions on road segments without sensor coverage. Experimental results show high accuracy for the short-term traffic flow forecasting and its potential to support real-time traffic management and urban mobility policies.