A Queueing-Predictive Task Offloading Framework for Joint Latency and Energy Optimization in Smart City Edge–Cloud IoT Networks

Smart-city Internet of Things (IoT) platforms run safety-critical alarms, near-real-time video analytics and delay-tolerant batch jobs on a shared edge–cloud infrastructure. Keeping enough edge capacity powered for peak demand wastes energy for most of the day, whereas aggressive consolidation pushes work to distant clouds and erodes latency guarantees. This paper presents PLEO, a queueing-predictive latency–energy offloading framework that couples two control loops. A slow loop forecasts zone-level demand with quantile gradient-boosted trees and sizes the powered capacity of every edge site by minimizing a joint latency–energy cost under an M/G/m waiting-time chance constraint, waking processing units ahead of demand and releasing them only beyond a break-even horizon. A fast loop predicts the completion time and marginal energy of every task at the local site, at neighbouring sites and in regional clouds, and selects the cheapest destination that meets the slice deadline. PLEO builds on a reference edge–cloud synergy architecture whose topology, slices and reactive QoS orchestration rules are implemented as a baseline. In simulations driven by the real PeMSD8 loop-detector dataset (170 detectors in 10 edge zones, 84 million tasks per week), PLEO lowers infrastructure energy per task by 9.2% relative to the reactive orchestration baseline, 11.0% relative to edge-only and 7.6% relative to cloud-only processing, meets every critical-slice deadline and completes 99.82% of all tasks on time. Ablations show that queueing-aware capacity planning, rather than forecast accuracy, is the main source of the gain.