A Deep Learning-Based Framework for Accelerating EVs Traffic Assignment with User Equilibrium

As the adoption of electric vehicles (EVs) continues to grow, power and transportation systems have established a coupled relationship via EVs charging stations. Understanding the distribution of charging loads becomes crucial for the efficient coordination. Given the demands of Origin-Destination (OD) pairs, the Traffic Assignment Problem with User Equilibrium (TAP-UE) model is typically solved to simulate steady distributions of traffic flow and charging loads. Despite TAP-UE being a convex problem in general, the widely-used iterative solution process and nonlinear terms in the objective function hinders the fast simulation due to prolonged solution times. This paper proposes a deep learning-based framework to accelerate the TAP-UE solution process. First, a group of neural networks are trained to identity feasible paths for OD pairs, thereby circumventing the need for an iterative process. On that basis, the convexity-preserving reformulation of TAP-UE transforms it into an unconstrained nonlinear optimization problem and employs a neural network for the accelerated solution. The Hessian matrix is incorporated into the gradient update of network parameters using the L-BFGS algorithm, enhancing the convergence speed of the training process. The efficacy of the proposed deep learning-based framework is demonstrated in case studies.