AI maps the exact temperatures that switch the world’s heaters and air conditioners on

Every building on Earth has an invisible switch. Below a certain outdoor temperature, the heating comes on; above another, the air conditioning kicks in. These thresholds, known as base temperatures, sit at the heart of one of the most widely used tools in energy modelling, the degree-day method, yet for most of the planet nobody has ever measured them. A team led by Xiujuan He and Yuyu Zhou of the University of Hong Kong, working with colleagues at KAIST, the University of Maryland and the Chinese Academy of Sciences, has now produced the first global dataset of these switching points, published in the journal Earth System Science Data. The work promises to sharpen the forecasts that governments and grid operators rely on as they plan the decarbonization of the building sector, which accounted for 36 percent of global end-use energy and roughly 39 percent of energy and process-related carbon dioxide emissions in 2018.

The degree-day method is deceptively simple. Heating degree days count how far the average daily temperature falls below the heating base temperature, while cooling degree days count how far it rises above the cooling base temperature. Because they require little more than a thermometer and a threshold, degree-day models have become the workhorse of regional energy assessments, climate impact studies and integrated assessment models. But the threshold itself is anything but universal. In the United States, modellers conventionally use 18.3 degrees Celsius; the United Kingdom uses 15.5 degrees for heating and 22 degrees for cooling; China typically uses 18 and 24 degrees. Applying a single number across a continent ignores the reality that base temperatures shift with local climate, building insulation, occupant habits and wealth. The authors note that a one-degree error in the base temperature can add or remove roughly twenty cooling degree days per month in places like San Luis Obispo, California, and can distort the relationship between degree days and actual energy use in nonlinear ways.

The conventional way to pin down a region’s true base temperature is the energy signature method, which plots energy consumption against outdoor temperature and identifies the inflection point of the resulting curve, or the closely related performance line method, which fits energy use against calculated degree days. Both approaches have produced valuable local results, such as median base temperatures of 18.3 degrees for cooling and 17.4 degrees for heating across sixty Indian cities, or values of 15.42 and 21.18 degrees in Kerman, Iran. The catch is data. These methods demand high-resolution energy consumption records matched with meteorological observations, which exist mainly in wealthy, data-rich countries. For most of Africa, South America and Central Asia, no such records are available, so global assessments fall back on harmonized, one-size-fits-all thresholds that wash out regional differences.

The new study attacks this bottleneck with machine learning. The researchers assembled electricity and natural gas demand time series from 172 regions worldwide, spanning building-level smart meter data at second, minute and hourly resolutions as well as regional daily and monthly statistics. From these they distilled training targets using segmented linear regression, a statistical technique grounded in the energy signature method that fits straight lines to different portions of the energy-temperature relationship and locates the breakpoints where heating or cooling behaviour begins. Rigorous quality control followed: time series without seasonal cycles were discarded, abnormal daily records were removed with density-based clustering, and only regressions achieving a coefficient of determination above 0.5 with physically plausible thresholds between 10 and 30 degrees were retained. The result was a labelled set of cooling and heating base temperatures for 131 geographic units across 24 countries.

Those labels then fed a Bidirectional Long Short-Term Memory neural network, or BiLSTM, equipped with an attention mechanism. LSTM networks are designed to preserve information across long sequences through gated memory cells, and the bidirectional variant processes weather records both forwards and backwards in time, capturing how past and future conditions jointly shape energy behaviour. The model ingested monthly meteorological variables from the ERA5-Land reanalysis, including two-metre air temperature, dew point temperature, wind speed and surface solar radiation, alongside annual socioeconomic data such as population density and economic density, plus geographic coordinates. An attention layer learned to weight the months most relevant to base temperature estimation, and stratified sampling based on geographic clustering ensured that training and test sets represented the full spread of world regions. An adaptive loss function allowed the model to train jointly on both thresholds even where only one was available for a given region.

The performance figures are strong. The model achieved root mean square errors of 1.39 degrees Celsius on the training set and 1.33 degrees on the test set, with Pearson correlation coefficients of 0.84 for cooling base temperature and 0.70 for heating. Benchmark comparisons against ridge regression, random forest and a unidirectional LSTM confirmed that the bidirectional architecture earned its place, particularly for the harder heating threshold, where it reached a cross-validated correlation of 0.63 against 0.52 or less for the alternatives. Applied globally, the trained model produced estimates for 3,385 first-level administrative regions, each roughly the size of a U.S. state or Chinese province. The predicted cooling base temperatures range from 19 to 25 degrees Celsius, with a mean of 22 degrees, while heating base temperatures span 14 to 18 degrees, averaging 16 degrees, values consistent with physical expectations and indoor comfort research.

The global map reveals telling patterns. North America and Europe, dominated by wealthy countries with well-insulated building stocks, show the lowest median thresholds for both heating and cooling, echoing earlier findings that affluent populations switch on air conditioning at lower outdoor temperatures. Asia, stretching across the widest latitudinal band and the largest economic disparities, exhibits the broadest range of values, with southern China and India showing higher thresholds. The predicted median heating base temperature for India of 17.05 degrees closely matches the 17.4 degrees derived independently from empirical energy signature analysis, a reassuring sign of internal consistency. Within Africa, the dataset even captures heterogeneity among countries such as Libya, Gabon, Algeria, Botswana and South Africa, which show relatively low heating thresholds compared with their neighbours.

Crucially, the team validated the dataset against 16 independent energy demand records from every major continent, none of which had been used in training. Direct comparison showed average prediction errors of 0.68 degrees for cooling and 1.08 degrees for heating. Indirect validation, in which degree days computed from the predicted thresholds were regressed against actual energy consumption, delivered the headline result: the average correlation between degree days and energy demand rose from 0.67 to 0.81, and the root mean square error of both cooling and heating demand models fell by approximately 10 percent compared with official or empirical base temperatures. In a South Australian home, using the predicted thresholds cut the cooling model error by 21.2 percent; in Tacoma, Washington, the regional heating model error dropped 12.85 percent against the official 18-degree standard. Sensitivity analysis confirmed that the predicted thresholds sit close to the maximum achievable correlation in nearly every dataset tested.

The authors are candid about limitations. The dataset rests on electricity and natural gas statistics, so heating from coal, diesel or biomass in energy-poor regions remains invisible, and building characteristics such as age, insulation and floor area are not differentiated. Uncertainty, quantified through 30 Monte Carlo dropout runs, ranges from 0.56 to 2.26 degrees Celsius, with the highest values concentrated in sparsely sampled regions of Africa, South America and western Asia, where training coverage is thinnest. Even so, relative uncertainty hovers around 10 percent for both thresholds, and the team shows the labels are robust to the choice of fitting threshold and to an independent UK Met Office-based breakpoint method, which agreed with their estimates to within 0.83 degrees on average across twenty British regions.

The implications reach well beyond energy statistics. Because degree-day models underpin projections of how climate warming will shift demand between heating and cooling, more accurate thresholds translate directly into better climate impact assessments, heat exposure and heatwave warning studies, and protection strategies for vulnerable populations. The dataset is also designed for coupling into Earth system models such as CESM and integrated assessment models such as GCAM and IMAGE, where it could reduce long-standing uncertainties in climate-energy interactions and sharpen the decarbonization pathways that policy makers depend on. Freely available on figshare, the first global map of the temperatures at which humanity reaches for the thermostat may prove one of those quiet datasets that quietly rewrites how the energy transition is planned.

Subject of Research: A global machine-learning dataset of heating and cooling base temperatures for residential building energy demand modelling

Article Title: A global base temperature dataset for residential building energy demand modelling

Article References: He, X., Eom, J., Yu, S., Liu, S., Xu, W., & Zhou, Y. (2026). A global base temperature dataset for residential building energy demand modelling. Earth System Science Data, 18(10), 7417-7454. https://doi.org/10.5194/essd-18-7417-2026

Image Credits: AI Generated

DOI: 10.5194/essd-18-7417-2026

Keywords: base temperature, degree days, building energy demand, BiLSTM, machine learning, heating and cooling, energy modelling, climate change, ERA5-Land, decarbonization, Earth System Science Data, thermal comfort