<p>Traffic congestion is a major challenge in modern transportation systems, leading to increased travel time, greater fuel consumption, and higher levels of harmful emissions. Accurate traffic congestion forecasting is crucial for effective traffic control, such as managing traffic lights. To generate reliable predictions, it is crucial to precisely detect abnormal traffic patterns in the data, particularly in densely populated urban areas. This study presents a novel approach that integrates anomaly detection and ensemble learning for traffic congestion forecasting. As an initial step in the learning process, we evaluate different anomaly detection techniques to identify unusual traffic patterns in different locations over time. After predicting the anomaly, the data pattern is cleaned accordingly, and a set of baseline learner models is trained as a secondary learning process. The top-performing models are chosen and undergo an ensemble process to combine their results, evaluating both stacking and voting ensemble methods as a third learning process. We evaluate the efficacy of the proposed strategy by employing a real-world traffic dataset and diverse evaluation metrics. The dataset undergoes several preprocessing techniques, including the windowing process with various settings, to convert the time series data into frequency patterns and produce a more generalized model. The results show that the multilevel learning approach improves prediction accuracy compared to baseline models, highlighting its effectiveness. This study highlights the utilization of anomaly detection and ensemble learning to enhance the precision of traffic congestion prediction, thereby promoting further exploration of this approach in intelligent transportation systems.</p>