Accurate prediction of electricity demand is crucial for reliable power system planning and operation. Climate change is expected to increase the frequency and intensity of heat waves, leading to unexpected surges in electricity demand, particularly for space conditioning and refrigeration. Existing models often rely solely on air temperature to predict demand, neglecting the significant role of humidity in determining heat stress. This omission can lead to substantial underestimation of cooling demand and potentially compromise the resilience of power systems. The study addresses this gap by examining the climate sensitivity of residential electricity demand using various heat stress measures, including air temperature and near-surface humidity parameters. The focus is on summer months (May-August) across the contiguous US, given the seasonal peak in demand and the occurrence of heat waves. The central hypothesis is that air temperature alone underestimates the climate-sensitive portion of cooling energy demand. The study uses monthly aggregated, state-level electricity consumption data from 1990-2016, population data from the US Census Bureau, and harmonized climate data from NCEP North American Regional Reanalysis (NARR). The data are carefully adjusted to remove trends unrelated to climate, isolating the climate's impact on energy demand. Two sets of state-level models are developed: air-temperature-only models and selected-features models, which use a data-driven approach to select the most relevant heat stress measures.

The literature extensively covers electricity demand prediction, employing climate, technological, and socioeconomic factors. However, most studies focus on air temperature as the primary predictor, neglecting humidity's role. While climate science emphasizes that air temperature alone is an insufficient measure of surface heat content, the existing research on the climate-demand nexus primarily uses air temperature or derived features like cooling and heating degree days. This study highlights a gap in considering multiple heat stress indicators beyond air temperature, acknowledging the need for a more holistic approach to model human comfort levels and their impact on cooling demand. The authors reviewed existing heat stress measures to select the most effective parameters for predicting the climate-demand nexus, including dew point temperature, discomfort index (DI), heat index (HI), humidex, wet bulb temperature, and wet bulb globe temperature.

The study employs monthly aggregated, state-level electricity consumption data from the Energy Information Administration (EIA) for 1990-2016. Population data from the US Census Bureau and climate data from NCEP North American Regional Reanalysis (NARR) are also incorporated. The electricity consumption data undergoes a detrending process to isolate climate effects from technological and demographic changes using a widely-used method that normalizes per-capita demand and applies an adjustment factor to account for yearly variations. The climate data includes various heat stress measures: air temperature, dew point temperature, discomfort index (DI), heat index (HI), humidex, wet bulb temperature, and wet bulb globe temperature. All climate variables use a combination of temperature, humidity, and pressure. Equations for calculating dew point temperature, wet bulb temperature, DI, wet bulb globe temperature (sWBGT), humidex, and HI are provided in the supplementary material. The key methodology is centered around the development of two types of state-level models: air-temperature-only models and selected-features models, both using a Bayesian Additive Regression Trees (BART) algorithm for prediction. The selected-features models use a data-driven variable selection process within BART to identify the most impactful heat stress measures for each state. The model performance is evaluated using 10-fold cross-validation, comparing root mean square error (RMSE) and R-squared (R²) values between the air-temperature-only and selected-features models. The models are also applied to projected climate data from five Global Circulation Models (GCMs) under the RCP8.5 scenario for the periods 1991-2010 (reference) and 2031-2050 (future). The reference period data from the GCM is used instead of historical data to remove bias induced by the use of data from different sources in the sensitivity analysis. The relative increase in projected demand between the reference and future periods is calculated for both model types to assess the extent of underestimation by the air-temperature-only models, particularly focusing on the 90th percentile of demand representing extreme events.

The study reveals significant improvements in prediction accuracy for cooling demand in 35 states when incorporating both temperature and humidity compared to air-temperature-only models. Figure 1 shows the key predictors identified by the selected-features models across the US, demonstrating substantial variability in the importance of different heat stress measures across different states. High-energy-consuming states like California and Texas, along with many southern states, benefit significantly from including humidity-related measures. Figure 1b illustrates the percentage improvement in predictive accuracy (based on out-of-sample RMSE) for states using selected-features models compared to air-temperature-only models. Significant improvements (e.g., 8% for California, 8.5% for Texas) are observed for several high-energy-consuming states. These improvements translate to substantial energy consumption differences (e.g., equivalent to sustaining millions of households for months). Analysis of the 90th quantile predictions (Supplementary Figs. 2 and 3) further reveals improvement for heat stress months. The sensitivity analysis under future climate scenarios (using data from five GCMs) demonstrates that air-temperature-only models consistently underestimate the climate-sensitive portion of demand, especially at higher demand values associated with extreme heat events. Figure 2 shows the projected increase in cooling demand, revealing significantly higher increases for the selected-features models. Figure 3 compares the relative increase in projected demand from selected-features models versus air-temperature-only models for the top seven energy-consuming states showing improvement in the historical period, highlighting the significant underestimation by air-temperature-only models, especially during high-demand periods.

The findings confirm the hypothesis that air temperature alone underestimates the climate sensitivity of residential cooling demand. Humidity is a critical factor influencing heat stress and thus cooling demand. Ignoring humidity leads to a substantial underestimation of future demand increases, potentially causing inadequate investments in energy infrastructure and demand response programs. The underestimation is particularly concerning for extreme heat events, where grid resilience is most challenged. The data-driven framework presented provides a robust method for quantifying this underestimation, highlighting the importance of a holistic approach incorporating multiple heat stress measures for accurate demand projections. The limitations of the study should be acknowledged; for example, the results pertain only to the climate-sensitive portion of residential cooling demand and do not encompass other factors affecting total demand (economic growth, technological advancements, and demographic changes).

This study demonstrates the crucial role of humidity in accurately modeling summer electricity demand. Using air temperature alone significantly underestimates the climate-sensitive portion of cooling demand, particularly during extreme heat events. The data-driven framework presented offers a superior approach to predicting demand, enabling more effective planning and investment in power system infrastructure and adaptation measures. Future research could explore the applicability of this methodology to other climate-sensitive sectors and investigate the interaction between humidity and other factors influencing electricity demand.

The study focuses solely on the climate-sensitive portion of residential cooling demand. Other factors such as economic growth, population changes, technological advancements, and policy interventions are not directly considered. The generalizability to other geographical regions or countries needs to be further tested. Furthermore, the reliance on GCM projections introduces uncertainty, as future climate scenarios are inherently uncertain.