Coastal home shows far stronger weather-linked electricity demand than inland house

Two ordinary houses in southern India, one in the inland city of Bangalore and one on the humid coast at Mangalore, turned out to respond to weather in strikingly different ways, according to a new study that tracked their electricity use hour by hour alongside on-site weather readings. The research, published in Energy Informatics, analyzed synchronized smart-meter and weather measurements from the two monitored dwellings and found that climate alone explained very little of the demand in the inland home, while the coastal home showed much stronger coupling to solar-related variables, modest humidity effects, and exploratory signs of delayed weather responses. The findings are case-specific, but they offer a rare dwelling-scale look at how residential electricity demand can diverge even within a single Indian state.

The study was led by M Parameshwari and Gnaneswari Gnanaguru and published on 27 July 2026 as an open-access article in Energy Informatics, volume 9, article number 80. The authors set out to address a gap they identify in the existing literature: most climate-informed demand studies rely on aggregated, system-level datasets that can mask dwelling-scale variability, and few explicitly examine lagged weather-demand relationships or validate whether empirically derived climatic features actually add value beyond raw weather inputs. To probe these questions, the researchers combined descriptive, correlation, multivariate, lagged, and forecasting-oriented feature analyses in a single interpretable workflow, while maintaining what they describe as a clear distinction between association and causation.

The two monitoring sites were chosen to provide contrasting climatic exposures within the same state. The Bangalore dwelling, located at 13.0041°N, 77.7851°E, sits in a semi-arid tropical savanna climate marked by high diurnal temperature variability and strong solar exposure. The Mangalore dwelling, at 12.9580°N, 75.2555°E, experiences a humid tropical monsoon climate with high humidity, heavy rainfall, and relatively stable temperatures. Both are single-storey independent houses of similar size: the inland home has roughly 116.4 square metres of floor area and nine occupants’ worth of living space measured as about 317.1 cubic metres of internal volume, while the coastal home has approximately 111.5 square metres and 305.8 cubic metres. Each has a reinforced cement concrete roof slab 152 millimetres thick, walls about 305 millimetres thick, and neither had air conditioning during the monitoring period, relying instead on natural ventilation and ceiling fans, with the coastal household adding table fans.

The households themselves differ in ways that matter for demand. The Bangalore home housed three people: an employed adult male typically out of the house for about ten hours a day, an adult female homemaker, and a school-going child aged about twelve. The Mangalore home also had three occupants, but included two older adults who remained at home for longer durations. The smart meters recorded whole-dwelling electricity consumption, and the on-site weather stations measured temperature, dew point, humidity, rainfall, wind speed, wind gust, solar radiation, and UV index. The inland dataset spans March 2024 to February 2026, roughly 23 months of hourly observations totalling 16,782 records, while the coastal dataset covers January 2025 to February 2026, about 13 months and 9,422 records. Preprocessing found no missing values, no duplicate timestamps, and no gaps in the hourly sequence; outlier screening flagged some rainfall observations, but these were treated cautiously as physically plausible extreme events in strongly zero-inflated distributions rather than errors.

The descriptive results already pointed in different directions. The inland dwelling showed higher mean demand and greater variability, with pronounced daily peaks during morning hours from 07:00 to 09:00 and evening hours from 19:00 to 22:00, patterns the authors associate with occupant activity and temperature-related cooling needs. The coastal dwelling exhibited smoother demand profiles with less pronounced peaks, suggesting lower sensitivity to short-term temperature swings and a comparatively greater association with persistent moisture-related conditions. Seasonally, inland demand increased during pre-monsoon periods alongside elevated temperatures, whereas coastal demand rose during monsoon periods, likely reflecting sustained humidity and rainfall that affect indoor comfort and appliance usage. The authors caution that these observations come from individual residential units and should be read as case-specific rather than generalized regional conclusions.

Correlation analysis, using rank-based measures because of non-normal distributions, quantified the contrast. In the inland dwelling, the strongest climatic association was with UV index at rho = -0.119, followed by temperature at rho = 0.109, both statistically significant at p < 0.001 but weak in magnitude; rainfall, wind speed, and dew point were statistically insignificant. The coastal dwelling showed a much stronger association with UV index at rho = -0.649, with humidity showing only modest positive association and rainfall remaining weak, while dew point was again insignificant. The authors emphasize that these are statistical associations, not causal relationships, and that the UV index result should not be read as a direct physical driver; UV index may act as a proxy for broader daytime weather conditions, solar availability, cloud cover, or occupancy-linked activity patterns, an interpretation they describe as provisional.

Cross-correlation analysis added an exploratory temporal dimension. In the inland dataset, temperature showed a delayed association with demand at a lag of approximately four days, which the authors suggest may reflect time-shifted cooling-related electricity responses. In the coastal dataset, selected moisture-related variables showed exploratory delayed associations at approximately five to seven days, though these patterns are modest. The authors are careful to frame these lag relationships as temporal associations rather than statistically established delayed causation, noting they may reflect cumulative environmental exposure, delayed behavioral adaptation, or building-related response dynamics that cannot be confirmed with the present dataset alone. An exploratory Granger-causality assessment using daily mean series and lags from one to seven days found statistically significant evidence only for temperature in the inland dwelling, strongest at a one-day lag with p = 0.0081, while the coastal dwelling showed stronger short-lag predictability from humidity at a seven-day lag with p = 0.0281, solar radiation at a one-day lag with p = 0.0386, and UV index at a one-day lag with p = 0.0111; temperature did not reach conventional significance there. These results, the authors note, do not control for occupancy, appliance scheduling, or other unobserved behavioral drivers.

Multivariate analysis reinforced the pattern. Correlation heatmaps indicated weak climate-demand coupling in the inland dwelling, where demand variability appears influenced by factors beyond climatic variables, while the coastal dwelling exhibited stronger interdependencies, particularly involving humidity and solar-related parameters. Partial correlation analysis suggested these relationships persist after controlling for inter-variable dependencies. Principal Component Analysis showed that in Bangalore the first principal component is dominated by humidity, temperature, and solar radiation, indicating a thermal-moisture-radiative contrast in which higher humidity opposes warmer and more solar-exposed conditions, with the second component led by wind speed, wind gust, and dew point. In Mangalore, the first component is dominated by temperature, humidity, and solar radiation, consistent with a thermal-radiative exposure axis moderated by moisture, and the second is led by precipitation, wind gust, dew point, and wind speed, indicating a moisture-wind regime. K-means clustering further identified distinct demand-climate regimes in both dwellings, with inland clusters associated mainly with temperature variability and coastal clusters more closely tied to humidity and rainfall patterns.

Baseline regression models quantified how much climate alone can explain. The inland dwelling exhibited very low explanatory power, with R² of approximately 0.03, indicating that climatic variables alone are insufficient to explain demand variability and pointing to unobserved factors such as occupant behavior, appliance usage, and building characteristics. The coastal dwelling showed moderate explanatory power at R² of approximately 0.33, indicating stronger weather-demand association, though the authors stress this is still insufficient to support strong claims about the dominance of climatic drivers. A supervised prediction experiment then tested four models, Ridge Regression, Random Forest, XGBoost, and LSTM, each trained on raw climatic variables alone and then on raw variables augmented with engineered lag, rolling, and interaction features, using a chronological holdout split with the first 80 percent of observations for training and the remaining 20 percent for out-of-sample testing. In the inland dwelling, engineered features improved performance across all four models: Random Forest rose from R² = 0.055 to 0.139, XGBoost from 0.054 to 0.138, Ridge Regression from -0.004 to 0.085, and LSTM from 0.106 to 0.120, with the strongest gains among the tree-based models.

In the coastal dwelling, raw climatic variables already performed comparatively well, so engineered features produced only marginal gains for the tabular models, with Ridge Regression improving slightly from R² = 0.222 to 0.225, Random Forest from 0.235 to 0.243, and XGBoost from 0.237 to 0.254. The LSTM, however, performed best with raw inputs at R² = 0.280 and deteriorated substantially with the engineered set, falling to R² = -0.136. The authors conclude that the value of climate-aware feature engineering depends not only on climatic context but also on model architecture, and that more complex sequence models do not necessarily benefit from manually engineered predictors in small dwelling-scale datasets. Because the two datasets cover unequal durations, a sensitivity analysis restricted to the overlapping window from 11 January 2025 to 7 February 2026 preserved the main comparative pattern while showing more moderate gains, particularly in the coastal case, suggesting some feature-engineering benefits are sensitive to sample period. Permutation feature importance from the Random Forest baseline showed solar radiation as the dominant predictor in Bangalore, followed by temperature and UV index, whereas in Mangalore UV index dominated by a large margin. The authors note the predictive experiments did not incorporate occupancy schedules, appliance-level operation, socio-economic characteristics, or autoregressive load-memory terms, so the models should be read as climate-sensitive baselines rather than comprehensive residential demand models. They also flag limited uncertainty quantification, unequal monitoring durations that may affect seasonal comparability, and the associative rather than causal nature of the findings, positioning the work as a transparent comparative testbed for climate-informed residential load modelling rather than a basis for broad policy inference.

Subject of Research: Dwelling-scale comparative analysis of climate-sensitive residential electricity demand using synchronized smart-meter and weather data from inland and coastal homes in Karnataka, India.

Article Title: Climate-informed residential electricity demand modelling: a comparative study of inland and coastal dwellings

Article References: Parameshwari, M., & Gnanaguru, G. (2026). Climate-informed residential electricity demand modelling: a comparative study of inland and coastal dwellings. Energy Informatics, 9(1), Article 80. https://doi.org/10.1186/s42162-026-00679-w

Image Credits: AI Generated

DOI: 10.1186/s42162-026-00679-w

Keywords: residential electricity demand, smart meter data, weather-demand correlation, climate-aware feature engineering, Bangalore, Mangalore, load forecasting, lagged weather effects, Granger causality, principal component analysis, Ridge regression, Random Forest