EMBEDDED SELECTIVE LOAD CONTROL FOR EDGE-DEPLOYABLE UNINTERRUPTIBLE POWER SUPPLY MANAGEMENT WITH MACHINE LEARNING–READY ARCHITECTURE

Abstract

Single-phase UPS installations fail in an all-or-nothing manner; even brief overloads or unexpectedly lengthy outages can trip the UPS or cause the batteries to fail, resulting in all circuits shutting off, regardless of the mission-critical nature of the circuits.  The paper formalizes and evaluates a low-cost microcontroller-based supervisory controller that monitors mains and UPS voltage, as well as branch-level currents, and operates two sets of relay banks to shed/re-energize branch circuits based on a declared priority tree operating on two-time scales. The first, fast-acting protective layer ensures that overload and over-discharge situations are mitigated. In contrast, the second, slow-restoration layer utilizes battery-aware hysteretic thresholds (calculated at 11.4V ("trim"), 11V ("keep only essentials"), and 10.5V ("protective cutoff”).  In Proteus step-stress experiments on three branches with a total capacity of 230V/2kW, operating against a 5.9kW UPS ceiling, the controller will complete the elimination of incipient overload by shedding the lowest-priority active branch while maintaining the highest-priority branch until the final shelf. Once the headroom is sufficient, the restoration process is smooth and non-oscillatory. Beyond these outcomes of empirical rules, architecture and literature review outline the rationale for autonomy-aware control, including embedded feasible state of charge and state of health estimators, such as data-driven and neural approaches. The controller itself is designed to ingest such estimates when they become available. The combined telemetry stack provided by MQTT, for example, provides the necessary observability to facilitate and justify this integration. Therefore, the proposed system converts binary UPS failures into graceful, priority-based degradation, paving a clear course for future machine-learning/AI augmentation without altering the underlying circuit-level actuation model.