The critical role of humidity in modeling summer electricity demand across the United States

Accurate demand prediction is critical for electricity adequacy planning, yet extreme heat waves increasingly drive unanticipated spikes in load, threatening grid reliability. The climate-demand nexus—the climate-sensitive portion of electricity demand—depends on effective characterization of heat stress, which involves both temperature and humidity and is closely linked to morbidity and mortality during heat events. Despite extensive literature on electricity demand prediction, most prior work relies on air temperature or derived metrics (e.g., degree days) and overlooks humidity, even though climate science shows air temperature alone is an incomplete measure of surface heat content. This paper addresses that gap by assessing multiple heat-stress metrics (dew point, discomfort index, heat index, humidex, wet bulb temperature, and wet bulb globe temperature) for modeling residential electricity demand across the contiguous U.S. during the summer months (May–August), when most states experience peak loads. The central thesis is that temperature-only models underestimate climate-sensitive cooling demand. The study develops state-level models using observed electricity consumption (EIA, 1990–2016), population data (U.S. Census), and harmonized climate data (NARR), after detrending electricity to isolate climate effects. It evaluates air-temperature-only models versus selected-features models that include the most predictive heat-stress measures, using a non-parametric Bayesian ensemble-of-trees framework. The analysis then quantifies the underestimation of temperature-only models under current and future climate scenarios.

Data and scope: The study focuses on the climate-sensitive portion of residential electricity demand in the contiguous U.S. during summer (May–August). - Electricity data: Monthly, state-level residential electricity consumption (EIA Form EIA-861), 1990–2016. Consumption converted to kWh, normalized per capita using state-level population (U.S. Census), and detrended to isolate climate effects from techno-demographic trends. Detrending uses: (i) compute yearly average E(y) across summer months; (ii) compute adjustment factor F_adj = E(y)/Σ_m E(m,y); (iii) adjusted consumption E(m,y)_adj = E(m,y)/F_adj. - Observed climate data: Sub-daily NCEP NARR (1990–2016), aggregated monthly and population-weighted to state level. Variables: air temperature, dew point temperature, discomfort index (DI), heat index (HI), humidex, wet bulb temperature (Tw), and simplified wet bulb globe temperature (sWBGT). Formulas align with literature (e.g., dew point via Lawrence 2005; Tw per Davies-Jones; DI = 0.5 Tw + 0.5 Tc; sWBGT = 0.56 Tc + 0.3936 RH + 3.94; Humidex = Tc + 5(e−10); HI via Rothfusz equation). Variables combine temperature, humidity, and pressure. - Projected climate data: Five GCMs under RCP8.5 (GFDL-ESM2M, HadGEM2-ES, IPSL-CM5A-LR, MIROC-ESM-CHEM, NorESM1-M), ~0.5° resolution, 1950–2099. Derived same heat-stress metrics. Two 20-year periods: reference (1991–2010) and future (2031–2050). Sensitivity analysis uses GCM-derived inputs for both periods to assess relative changes, avoiding source-induced bias. Modeling framework: - Two model families per state: (1) Air-temperature-only model using monthly surface air temperature as the sole predictor; (2) Selected-features model using data-driven variable selection among the suite of heat-stress metrics. - Algorithm: Bayesian Additive Regression Trees (BART), a non-parametric Bayesian ensemble-of-trees method with priors to regularize individual trees and enhance predictive performance. Response: adjusted, per-capita residential electricity consumption; predictors: heat-stress metrics (population-weighted state-level monthly aggregates). - Variable selection: Rank predictors by frequency of use as splitting variables across trees and MCMC iterations; average over 100 model runs to stabilize selection; retain top predictors per state. - Validation and tuning: Hyperparameters tuned via cross-validation. Ten-fold cross-validation used for training/testing; in-sample and out-of-sample R² and RMSE computed. Out-of-sample RMSE used for comparison between model families. For robustness to extremes, analyses performed for median (50th quantile) and 90th quantile of monthly demand distributions. Sensitivity analysis with GCMs: - Train state-level models using GCM-derived inputs for reference (1991–2010) and future (2031–2050) periods; produce four sets: air-temp reference, air-temp future, selected-features reference, selected-features future. Compute relative increases over time within each model and compare ratios of selected-features vs. air-temp for average and 90th percentile demand. Planning horizon selected to align with energy infrastructure timelines (to ~2050). Code and pre-processed data available at the corresponding author’s GitHub.

- Including humidity-informed heat-stress metrics significantly improves predictive accuracy for cooling demand: 35 of 48 states show out-of-sample RMSE improvements for selected-features models over air-temperature-only models. - High-energy-consuming states benefit notably from humidity-aware metrics (dew point, wet bulb temperature, wet bulb globe temperature, heat index). In California and Texas, air temperature is not among the key predictors selected for the climate-sensitive portion of demand. - Quantified improvements (out-of-sample RMSE) for 50th (and 90th) quantiles: California ~8% (7.7%), Texas 8.5% (21.1%), Illinois 8.6% (26.1%), North Carolina 7.1% (9%). - Practical magnitude: In Texas, an 8.5% improvement in August 2016 corresponds to 1,498,968 MWh—enough to sustain Austin, TX residents for over four months. In California (August 2016), an 8% median difference could sustain nearly 1.5 million households. - Future climate scenarios (RCP8.5, 2031–2050): Selected-features models project larger increases in the climate-sensitive portion of demand than air-temperature-only models in most states (32 states show higher increases). Example relative increases (selected-features, future vs. reference): Texas ~12%, California ~8%, Florida ~10%. - Underestimation is most pronounced for higher demand (90th percentile) associated with intense heat stress; for top energy-consuming states (e.g., California, Florida, Texas), selected-features projections can be nearly twice those from temperature-only models for extremes. - Overall, relying on air temperature alone can underestimate cooling demand by about 10–15% in many high-consumption states under present and future climates.

Findings demonstrate that air temperature alone is a necessary but insufficient predictor of climate-sensitive residential cooling demand. Humidity substantially shapes perceived heat stress and human cooling needs, and thus electricity use. Incorporating humidity-based metrics enhances predictive accuracy across the majority of U.S. states and reveals systematic underestimation by temperature-only models—particularly under future warming and during extreme heat months when grid resilience is most at risk. This underestimation could lead to inadequate supply planning, greater outage risks, and increased morbidity and mortality among vulnerable populations during heat events. Accounting for humidity in demand projections is therefore crucial for effective capacity expansion planning, demand response design, and adaptation strategies under climate change, especially as summertime generation is also constrained by warming.

This study establishes the critical role of near-surface humidity—captured via heat-stress metrics—in modeling the climate-sensitive portion of summer residential electricity demand across the U.S. Data-driven selection of heat-stress predictors combined with a BART modeling framework improves predictive accuracy in 35 states and indicates that temperature-only models substantially underestimate cooling demand, with even larger underestimation for extreme demand conditions. Under future climate scenarios (to 2050, RCP8.5), selected-features models project markedly higher increases in cooling demand than temperature-only models, especially in high-consumption states. Future research should extend this framework to: (i) integrate socioeconomic, demographic, and technological trends to project total residential demand; (ii) assess other climate-sensitive sectors (commercial/industrial); (iii) evaluate adaptation options and demand response under extreme heat; and (iv) explore regionalized infrastructure planning considering humidity-driven demand shifts.

- The analysis isolates only the climate-sensitive portion of residential summer cooling demand; it does not project total demand including socioeconomic, demographic, and technological changes. - Focus is on summer months; climate effects outside May–August are not assessed. - State-level aggregation and population-weighting may mask intra-state heterogeneity. - Projections rely on a subset of CMIP5 GCMs under RCP8.5 and on bias-corrected, 0.5° resolution data; model and scenario uncertainties remain. - Results pertain to residential sector only and may not directly generalize to other sectors without sector-specific modeling.