The power grid of the US is the largest interconnected machine on Earth, yet it remains broken. In 2021, Winter Storm Uri knocked out power for 4.5 million Texans, killed 246, and caused $130 billion in damage. With power interruptions doubling since 2013, accessible grid intelligence is urgent. GridGuardian is an AI-powered national energy digital twin that converts public energy data into a visualization and prediction platform. It ingests real infrastructure data from the EIA (13,370 power plants), the Homeland Infrastructure Foundation-Level Database (52,244 transmission lines), OpenStreetMap (45,427 substations), and seven Independent System Operators totaling 111,041 mapped elements across the continental United States. Two neural network models were created and validated. A Stacked Bidirectional LSTM with Multi-Head Self-Attention trained on 26,112 hours of PJM Interconnection load data with 28 engineered features predicts electricity demand for a region 24 hours in advance with R^2 = 0.9409 and MAPE = 3.94%. A feedforward neural network trained on EIA residential and commercial building survey microdata predicts building-level energy profile (Residential R^2 = 0.9730). Both models surpassed their accuracy thresholds. The system layers on NOAA weather alerts for risk correlation, physics-based power flow analysis in PyPSA, a DOE methodology-based economic outage impact calculator, and a natural-language AI copilot for scenario discovery. GridGuardian is proof that open data and AI can make the invisible visible in infrastructure and empower a range of actors from communities to policymakers to grid operators to identify vulnerabilities, quantify the cost of inaction, and chart a path toward a stronger grid.