This paper proposes a two-level machine learning approach for traffic flow prediction in urban areas. The method combines unsupervised clustering with supervised learning to forecast traffic patterns. The first level uses clustering to identify similar traffic patterns from sensor data, while the second level employs supervised models to predict traffic flow for each cluster. The approach allows traffic flow prediction without requiring sensors, using data from existing sensors. The method was tested on a real-world case study in Catania, Italy, using data from 21 traffic sensors over a year. The data were preprocessed to create a structured dataset, with timestamps at 1-hour intervals and 14 columns representing traffic flow for each lane. The clustering step used Time Series K-Means (TSkmeans) with Dynamic Time Warping (DTW) to group similar traffic patterns. The forecasting step involved training models on time series data from each cluster and testing their ability to predict traffic flow for new, unseen data. The models were evaluated using metrics such as MAE, SMAPE, MSE, and RMSE. The results showed that XGBoost and CatBoost outperformed LightGBM and Random Forest in most metrics, with XGBoost performing better during the Christmas week when traffic patterns were more variable. The approach demonstrated the potential of machine learning in traffic flow prediction, enabling accurate forecasts for urban traffic management. The study highlights the effectiveness of combining clustering and supervised learning for traffic forecasting, providing a robust solution for urban planning and traffic management.This paper proposes a two-level machine learning approach for traffic flow prediction in urban areas. The method combines unsupervised clustering with supervised learning to forecast traffic patterns. The first level uses clustering to identify similar traffic patterns from sensor data, while the second level employs supervised models to predict traffic flow for each cluster. The approach allows traffic flow prediction without requiring sensors, using data from existing sensors. The method was tested on a real-world case study in Catania, Italy, using data from 21 traffic sensors over a year. The data were preprocessed to create a structured dataset, with timestamps at 1-hour intervals and 14 columns representing traffic flow for each lane. The clustering step used Time Series K-Means (TSkmeans) with Dynamic Time Warping (DTW) to group similar traffic patterns. The forecasting step involved training models on time series data from each cluster and testing their ability to predict traffic flow for new, unseen data. The models were evaluated using metrics such as MAE, SMAPE, MSE, and RMSE. The results showed that XGBoost and CatBoost outperformed LightGBM and Random Forest in most metrics, with XGBoost performing better during the Christmas week when traffic patterns were more variable. The approach demonstrated the potential of machine learning in traffic flow prediction, enabling accurate forecasts for urban traffic management. The study highlights the effectiveness of combining clustering and supervised learning for traffic forecasting, providing a robust solution for urban planning and traffic management.