Real-Time Energy and Thermal Optimization in Electric Delivery Fleets via Edge-Based Forecasting and Deep Q-Learning

Abstract

As e-commerce has seen tremendous growth, it has put extreme pressure on last-mile delivery systems. Although electric vehicles (EVs) have several positive attributes from a sustainability standpoint, they have energy inefficiencies when dealing with dense urban environments. Due to the nature of EVs operating in urban areas, they run into frequent stop-and-go cycles (creating lower-quality, more inefficient operation of battery charging), and because recovery cycles due to regenerative braking happen in unpredictable bursts, the thermal stress that occurs on the battery during these operations accelerates battery degradation. Also, the latency associated with cloud routing creates high latency that makes it unreasonable for quick adaptations needed for real-time applications. This research study presents a novel hybrid edge-intelligent framework that combines the use of Seasonal Autoregressive Integrated Moving Average (SARIMA) forecasting with that of an edge-deployed Deep Q-Network (DQN) reinforcement learning agent capable of optimizing battery thermal behaviors and optimizing energy consumption. For example, the SARIMA data is largely used to predict traffic density around an EV and its battery usage pattern (use of energy and charging demand) over short durations, while the DQN can make real-time charging and routing decisions within an average of 50 milliseconds or less. A synthetic dataset created from SUMO (Simulation of Urban Mobility) applying physics-based battery thermal dynamics demonstrates that there were energy savings on average of 27% and increased projected battery cycle life of 19% over the energy consumption and battery cycle life achieved using a rule-based benchmark. Further investigations expand on the full architecture, mathematical modeling, and trade-offs in applying edge computing to achieve scalable optimization of EV fleets.