Increasing the self-consumption of on-site PV generation in warehouses reduces the need to export surplus electricity, thereby minimizing grid losses and lowering the facility’s carbon footprint. Accordingly, an HVAC-driven load shifting methodology empowered by machine learning (ML) is proposed to dynamically shift the energy demand of a conditioned warehouse toward periods of surplus PV generation. Additionally, the impact of an optimally sized battery system on the self-consumption of the building is investigated. A simulation methodology captures the baseline energy behavior of the building and its thermal and load responses to HVAC setpoint modifications. Baseline import/export data are used to evaluate a range of battery sizes, and the optimal size is determined through detailed economic assessments considering the Italian electricity market, accounting for battery degradation and its impact on performance. Subsequently, ML pipelines are developed to forecast PV output, baseline demand, and load response to different possible setpoint adjustments. These forecasts are then provided to a decision-making agent to enable informed control actions. The agent will then use real-time measurements of the input parameters from the co-simulation environment, feed them to the developed pipelines, and dynamically determine and apply HVAC control actions (setpoint relaxation and preheating of the indoor environment). Lastly, improvements in self-consumption ratio (SCR) are assessed for the baseline case, battery-only integration, HVAC-only interventions, and their combined implementation. The results demonstrate enhancement in SCR by nearly 26% compared to the baseline in the joint scenario, underscoring the potential of the proposed approach to reduce grid dependency in warehouse operations.