A Power Management Approach Resilient to Energy Harvesting Prediction Errors on Battery-Operated Cyber-Physical Systems

Power Management in battery-operated Cyber-Physical Systems is challenging due to the distinct demands of different execution profiles and variable energy harvesting results. Existing approaches often rely on predictors to estimate energy harvesting. Nevertheless, such solutions are prone to errors in scenarios with variable behavior. This work proposes a novel approach inspired by the confidence attribution mechanism introduced by Scheffel et al., which assesses prediction reliability by estimating confidence levels based on the difference between predicted and observed values, statistics extracted from the training of the predictor, and application-specific hyper-parameters that define tolerance and the overall sensibility of the solution to errors. The proposed solution considers a simple Exponentially Weighted Moving Average predictor for energy harvesting, alongside battery charge and critical operation thresholds, to dynamically optimize the selection of execution profile and maximize the system's overall utilization by adjusting energy allocation based on the current confidence of the system about its overall power state. Results demonstrate the ability of the proposed solution under diverse environmental conditions and varying prediction errors. Furthermore, the time complexity analysis corroborates the suitability of the present approach to enhance Cyber-Physical Systems reliability and performance.