A highly accurate energy-forecasting model can make an edge device less energy-efficient overall, according to a new analysis of what the authors call the accuracy-efficiency paradox. The problem appears when the power used for inference and the long-term effect of heat on a battery exceed the energy saved by better predictions.
The researchers propose a total-cost-of-ownership framework that treats battery degradation as a physical loss of future energy capacity, not only a maintenance expense. That changes model selection: a complex architecture may forecast demand more precisely while running hot enough, and often enough, to shorten battery life. In thermally sensitive edge environments, the paper finds that this combined loss can outweigh the forecast improvement.
The conclusion is especially relevant to remote and mission-critical equipment, where replacing a battery can be expensive or difficult. It also cautions developers against reporting accuracy without measuring the device that delivers it. The paper presents an analytical framework rather than a universal ranking of forecasting models, and the outcome will vary with hardware, workload, climate and battery chemistry. Teams deploying on-device AI should therefore compare net energy saved over the system’s lifetime, including inference and degradation, instead of choosing the model with the smallest prediction error in isolation.