GIS-AHP Renewable Energy Siting with Deep Reinforcement Learning and Digital Twin Battery for 24/7 Carbon-Free Dispatch in Sri Lanka

Abstract

Sri Lanka's impending retirement of the 900 MW Norochcholai coal power station demands a rigorously integrated transition framework extending beyond isolated siting or dispatch studies. This paper presents a unified three-stage pipeline addressing the complete pathway from spatial planning to physics-aware intelligent dispatch. First, a Geographic Information System-Analytic Hierarchy Process (GIS-AHP) multi-criteria analysis scanning 863,375 raster pixels identifies ten optimal renewable energy sites across five provinces under Sri Lanka's dual monsoon climate, enforcing a 35 km geographic diversity constraint. Second, physics-based energy yield estimation using NASA POWER solar irradiance and ERA5 wind reanalysis data quantifies a 2,010 MW portfolio generating 4,460.1 GWh annually approximately 28% of national electricity demand with lifecycle CO2 emissions reduced by 98.2% against the coal baseline. Third, a Soft Actor-Critic deep reinforcement learning agent operating within a digital twin lithium iron phosphate battery environment incorporating cycle and calendar degradation achieves a 76.3% carbon-free energy score, outperforming rule-based (72.5%) and random (62.8%) baselines while preserving 90.5% battery health without an explicit degradation penalty. The framework avoids 4.44 million tons of CO2 annually and provides a transferable methodology for coal phase-out planning in developing island grid systems.