Building and explaining data-driven energy demand models for Indian states

Abstract Accurate forecasts of energy demand are crucial for managing India’s rapidly growing energy needs as it continues to decarbonise its grid. In this study, we develop state-level data driven models to predict weather-driven energy demand across India using the eXtreme Gradient Boosting framework. The models use as input population-weighted meteorological variables averaged over various timescales. The models are trained on daily energy demand data, scraped from reports issued by Grid-India, which we correct for trends in population and economic growth. The models demonstrate high skill, with half having <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" overflow="scroll"> <mml:mrow> <mml:msup> <mml:mi>r</mml:mi> <mml:mn>2</mml:mn> </mml:msup> <mml:mo>&gt;</mml:mo> <mml:mn>0.8</mml:mn> </mml:mrow> </mml:math> , significantly outperforming traditional multivariate linear regression models. We explain model behaviour through Shapley analysis and find a strong sensitivity to day of the week and public holidays, with reductions in energy demand on Sundays and varying impacts during holidays. While important variables vary by state and season, daily minimum temperature and 30 d mean temperature consistently emerge as key predictors, reflecting nighttime air conditioning use and seasonal heating or cooling needs. We also identify threshold behaviours, indicating large increases in energy demand once temperatures pass certain values. Using reanalysis, we extend our models to estimate all-India energy demand from 1979–2023, calibrated to 2023 conditions. We confirm a pronounced seasonal cycle, with greatest demand during the pre-monsoon and monsoon onset (May–June) and lowest demand in the winter (November–December). Combining these results with timeseries of renewable energy production, we find the largest energy deficit (demand minus renewable generation) occurs during or after monsoon withdrawal (September–October). Extreme deficit days, posing a risk to the national grid, are associated with early monsoon withdrawal or late monsoon breaks, leading to low wind speeds and persistently high dewpoint temperatures and cloud cover. The demand dataset created here can be used for energy grid management, siting of future renewable energy generation, and to aid with ensuring security of supply.