With　the　increasing　penetration　of　renewable　energy,　power　systems　face　spatiotemporal　heterogeneity　in　supply-demand　dynamics,　challenging　traditional　centralized　control　methods　due　to　communication　delays　and　privacy　risks.　Decentralized　Smart　Grid　Control　(DSGC)　offers　a　promising　paradigm　but　requires　precise　coordination　between　time　delay　parameters　(τ)　and　control　gains　(g)　to　ensure　stability.　This　study　proposes　a　multi-machine　learning　framework　to　address　two　critical　challenges:　(1)　predicting　stability　under　nonlinear　parameter　coupling　and　(2)　dynamically　quantifying　the　critical　maximum　delay　time　(τmax).　Three　models-Random　Forest　(RF),　Backpropagation　Neural　Network　(BPNN),　and　Support　Vector　Machine　(SVM)-are　implemented　for　stability　prediction.　Experimental　results　demonstrate　RF＇s　superior　performance,　achieving　92.30%　accuracy,　0.8916　F1-score,　and　0.9126　AUC,　outperforming　BPNN　and　SVM.　Key　stability　rules　are　extracted　via　ensemble　learning,　revealing　thresholds　for　τ　(6.16s),　g　(0.64),　and　power　fluctuation　(3.75)　that　significantly　influence　stability.　Delay　analysis　highlights　nonlinear　relationships　between　parameters,　with　low-gain　scenarios　showing　a　51.6%　stability　improvement　over　high-gain　cases.　This　work　provides　a　data-driven　framework　for　optimizing　DSGC　parameters　and　enhancing　grid　resilience　in　decentralized　energy　systems.　©　2025　IEEE.