Gestão de armazenamento de energia em sistemas fotovoltaicos híbridos com aprendizado por reforço

Abstract

The tightening of the Brazilian distributed-generation regulatory framework, particularly the progressive charging of the “Fio B” tariff component, makes minimizing grid usage increasingly advantageous for residential consumers. Commercial hybrid inverters, however, operate under pre-configured fixed-rule modes that cannot simultaneously reconcile the reduction of grid consumption with the preservation of an energy reserve to sustain the load during supply interruptions. This work proposes an autonomous storage-management system based on a Deep Q-Network reinforcement learning agent, supported by a statistical planner, that decides the battery charge and discharge cycles from real operational data. The system was validated in the field, on a real residential hybrid photovoltaic plant, through a comparative experiment between the conventional Backup mode and control by the agent. During the night and early-morning window, the fraction of the load supplied by the grid dropped from 80.2% to 28.5%, with the battery’s share of meeting that demand rising from nearly zero to 56.5%. The results show that the agent significantly reduces grid dependence while preserving the reserve function, establishing intelligent storage management as a viable and replicable advantage in the distributed-generation context.