DefinePK

Abstract

Predicting traffic flow accurately and in real-time is critical in managing traffic and reducing congestion in urban areas. Traditional machine learning algorithms do not always perform well in accurately capturing and predicting the complex, nonlinear, and dynamic nature of real traffic observations and patterns. To improve this aspect of performance, this research offers an ensemble-based machine learning approach that consists of multiple base learners, which improves prediction accuracy and generalizability by combining machine learning models. The ensemble model includes the combined strength of a Multi-Layer Perceptron (MLP), a Support Vector Classifier (SVC), and a CNN-LSTM model that has the capability of addressing both spatial and temporal feature representation from video-based identification of traffic data. The context of the traffic flow prediction model is improved through the model's integration of real-time object detection of traffic frames, as well as incorporating the current weather conditions. Each base learner's predictions are optimally combined through a meta-learner Logistic Regression. The model performance is assessed through multiple evaluation criteria, including Mean Squared Error (MSE), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R-squared (R²). The experimental results demonstrated that the ensemble-based model surpassed traditional machine learning algorithms, such as Linear Regression, K-Nearest Neighbors, Random Forest, Decision Tree, and as well as Support Vector Regression. The ensemble model achieved upwards of 98% prediction accuracy, which was significantly better than any of the traditional machine learning algorithms tested for performance as well. The study demonstrates that ensemble-based learning techniques and multi-source feature integration can produce stable solutions for real-time traffic flow predictions to guide intelligent traffic systems development