This paper presents a deep learning model for short-term traffic flow prediction, which combines a linear model with $\ell_1$ regularization and a sequence of tanh layers. The model is designed to capture the nonlinear spatio-temporal effects in traffic flow, particularly the sharp transitions between free flow, congestion, and recovery. The model is tested on real-world data from Interstate I-55, predicting traffic flow during two special events: a Chicago Bears football game and an extreme snowstorm. The results show that the deep learning model provides accurate short-term traffic flow predictions, even in the presence of sudden regime changes. The paper discusses the challenges of traffic flow prediction, including the need for real-time forecasting to manage transportation systems and the difficulty of modeling nonlinear and non-stationary relationships in traffic data. It also reviews existing methods for traffic flow prediction, including traditional neural networks, statistical models, and machine learning approaches. The paper highlights the advantages of deep learning in capturing complex spatio-temporal patterns and its ability to handle high-dimensional data. The paper introduces a deep learning architecture that uses a hierarchical sparse vector auto-regressive technique for predictor selection and dropout for regularization. The model is trained using stochastic gradient descent and is shown to outperform traditional methods in terms of accuracy and robustness. The model is tested on real-world data from 21 loop detectors installed on Interstate I-55, and the results show that the deep learning model provides more accurate predictions than traditional methods, especially in the presence of sudden changes in traffic conditions. The paper also discusses the importance of data preprocessing techniques, such as median filtering and trend filtering, in improving the accuracy of traffic flow predictions. The results show that combining deep learning with these preprocessing techniques leads to better performance. The paper concludes that deep learning provides a powerful tool for short-term traffic flow prediction, particularly in the presence of sudden changes in traffic conditions.This paper presents a deep learning model for short-term traffic flow prediction, which combines a linear model with $\ell_1$ regularization and a sequence of tanh layers. The model is designed to capture the nonlinear spatio-temporal effects in traffic flow, particularly the sharp transitions between free flow, congestion, and recovery. The model is tested on real-world data from Interstate I-55, predicting traffic flow during two special events: a Chicago Bears football game and an extreme snowstorm. The results show that the deep learning model provides accurate short-term traffic flow predictions, even in the presence of sudden regime changes. The paper discusses the challenges of traffic flow prediction, including the need for real-time forecasting to manage transportation systems and the difficulty of modeling nonlinear and non-stationary relationships in traffic data. It also reviews existing methods for traffic flow prediction, including traditional neural networks, statistical models, and machine learning approaches. The paper highlights the advantages of deep learning in capturing complex spatio-temporal patterns and its ability to handle high-dimensional data. The paper introduces a deep learning architecture that uses a hierarchical sparse vector auto-regressive technique for predictor selection and dropout for regularization. The model is trained using stochastic gradient descent and is shown to outperform traditional methods in terms of accuracy and robustness. The model is tested on real-world data from 21 loop detectors installed on Interstate I-55, and the results show that the deep learning model provides more accurate predictions than traditional methods, especially in the presence of sudden changes in traffic conditions. The paper also discusses the importance of data preprocessing techniques, such as median filtering and trend filtering, in improving the accuracy of traffic flow predictions. The results show that combining deep learning with these preprocessing techniques leads to better performance. The paper concludes that deep learning provides a powerful tool for short-term traffic flow prediction, particularly in the presence of sudden changes in traffic conditions.