Deadline and energy reshape time-series model choice

When a prediction arrives too late, the network cannot use it to prepare a handover; repeated predictions also add to a device’s energy use. Which time-series model best balances classification performance with these timing and energy constraints?

The authors introduce a protocol that filters time-series models by deadline, then ranks those that pass by measured full-pipeline energy.

The authors evaluated eight architectures on a simulated Santa Clara handover-prediction task, with a fixed data split and results averaged over five training seeds. They measured full-pipeline inference, including input preparation, on a controlled server-class CPU.

LightGBM had the highest F1 score among the eight models, at 0.974 ± 0.002. F1 is a 0–1 measure of classification performance.

Under burst-schedule measurement with idle pauses, LightGBM used 4,947 microjoules per inference, compared with 16,181 for CNN-1D, the best neural model on F1.

This comparison suggests that the model leading on classification performance also had lower measured energy in this setting, while the execution schedule affected the apparent size of that gap.

The evidence comes from one simulated Santa Clara urban scene, and energy was estimated from CPU utilization on a server-class processor rather than measured on handset hardware. The study therefore does not establish the same model ranking or absolute energy use in other environments or on phones.