Deep Learning-Based Downscaling of CMIP6 for Projecting Heat-Driven Electricity Demand and Cost Management in Chengdu

Rapid warming and expanding heat seasons are reshaping electricity demand in cities, with basin-type megacities like Chengdu facing amplified risks due to calm-wind, high-humidity conditions and fast-growing digital infrastructure. This study develops a Transformer-based, multi-model downscaling framework that integrates outputs from 17 CMIP6 global climate models (GCMs), dynamically re-weighted through self-attention to generate city-scale temperature projections. Compared to individual models and simple averaging, the method achieves higher fidelity in reproducing historical variability (correlation ≈ 0.98; RMSD < 0.05 °C), while enabling century-scale projections within seconds on a personal computer. Downscaled results indicate sustained warming and a seasonal expansion of cooling needs: by 2100, Chengdu is projected to warm by ~2–2.5 °C under SSP2-4.5 and ~3.5–4 °C under SSP3-7.0 (relative to a 2015–2024 baseline). Using a transparent, temperature-only Cooling Degree Day (CDD)–load model, we estimate median summer (JJA) electricity demand increases of +12.8% under SSP2-4.5 and +20.1% under SSP3-7.0 by 2085–2094, with upper-quartile peaks reaching +26.2%. Spring and autumn impacts remain modest, concentrating demand growth and operational risk in summer. These findings suggest steeper peak loads and longer high-load durations in the absence of adaptation. We recommend cost-aware resilience strategies for Chengdu, including peaking capacity, energy storage, demand response, and virtual power plants, alongside climate-informed urban planning and enterprise-level scheduling supported by high-resolution forecasts. Future work will incorporate multi-factor and sector-specific models, advancing the integration of climate projections into operational energy planning. This framework provides a scalable pathway from climate signals to power system and industrial cost management in heat-sensitive cities.