Hourly smart meter readings can now reveal far more than a building’s total electricity consumption. Researchers have developed COF-Tool, an open-source framework that takes an hourly electricity load profile plus local weather data and returns two things most meters never record: what kind of building the profile belongs to, and how much of its electricity is actually going into heating. The tool, described in the journal Energy Informatics, was trained and validated on real energy measurements from thousands of Norwegian buildings and is available both as a GitHub repository and as a web application.
The need for such a tool stems from a basic limitation of modern metering infrastructure. Norway has fully rolled out advanced metering, so hourly electricity data exist for virtually every building in the country. Yet smart meters measure only the total imported electricity. They say nothing about the building itself, nor how that electricity is divided among heating, hot water, ventilation, appliances, and electric vehicle charging. Because Norway’s building stock is heavily electrified, with electric heating long dominant in both homes and workplaces thanks to abundant hydropower, knowing how much electricity goes to heating is essential for grid planning, customer segmentation, pricing studies, energy-efficiency analysis, and the development of demand-side flexibility measures. Sub-metering every building to find out would be prohibitively expensive, which is where data-driven methods come in.
COF-Tool consists of two supervised machine learning modules. The first is a classification module built on a random forest classifier, trained on hourly data from 6403 buildings spanning 12 building categories, from apartments and single-family houses to schools, offices, hotels, and hospitals. Each building in the training data carries a label combining its category with its heating type, electric or non-electric, yielding 20 possible classes in total. The classifier extracts 180 features from each load profile, including physics-based domain features such as average and peak electricity use, energy-temperature curve features that capture how consumption tracks outdoor temperature, and new features describing the daily load shape. Given a new profile, the module outputs a probability vector over all 20 building types rather than a single definitive answer.
On the held-out test set of 1280 buildings, the classification module reached 73 percent top-1 accuracy and recall, with precision and F1 score at 72 percent. The Matthews correlation coefficient of 0.68, which accounts for class imbalance, indicates a moderate to strong correlation between predicted and true labels. Notably, this performance is lower than the 84 percent accuracy reported in the earlier study on which the module builds, which used 2724 buildings. The researchers attribute the drop mainly to the enlarged dataset: thousands of residential units were added, and apartments and houses are frequently confused with one another, particularly when distinguishing those with and without electric heating. Self-reported heating labels add uncertainty here, since residents may underreport electric appliances. Schools also became harder to classify because many were deliberately moved out of the training set to avoid information leakage with the disaggregation module.
The second module tackles a harder problem: disaggregation. It estimates the hourly electricity used for heating directly from the total meter reading, using a Categorical Boosting regressor trained on data from 323 buildings across 9 building categories. The central obstacle is the scarcity of labelled data. Norwegian regulations do not require sub-meters for electric heating equipment, so very few all-electric buildings have ground-truth measurements of how much power their heaters draw. The researchers’ solution is a cross-domain training trick. Buildings with district heating have separate meters for electricity and heat, so their combined energy use and measured heating demand can serve as a proxy for the total-energy-plus-heating relationship in all-electric buildings. Models trained this way on district-heated buildings can then be applied to all-electric ones.
This cross-domain approach was previously validated only for school buildings in a proof-of-concept involving 74 schools. The new work extends it to 323 buildings and shows that the method generalises beyond schools to apartment blocks. In the test set, all-electric buildings were available only for schools and apartment blocks, and for those categories the test data contained a mix of electric-heated and district-heated buildings. For the remaining categories, validation relied on treating the combined electricity and district heating use as the total heating demand and estimating the district heating component from that aggregated signal, so true cross-domain testing was not possible there. Overall, median disaggregation performance was good across all building categories, though some individual buildings showed poor fit with measured values.
How the training data are composed matters, but not uniformly. Training on buildings from all categories produced a clear improvement for universities, whose bootstrap 95 percent confidence intervals for the mean normalised mean absolute error did not overlap with those from same-category training, partly because several university buildings had low-resolution heating meters that limited the quality of same-category training data. For shops the improvement was less certain, with overlapping confidence intervals and few test buildings. For more homogeneous categories with more buildings, such as schools, apartment blocks, kindergartens, and nursing homes, the two training scenarios differed little, with confidence intervals overlapping substantially. The researchers caution that this comparison cannot separate the effect of larger sample size from the benefit of greater category diversity, since a controlled experiment with equal building numbers per category was not feasible.
Feature choices also shaped the disaggregation results. Adding classification probabilities or an extended feature set individually brought only marginal gains, but combining both feature sets into a full set of 288 features improved the median values and interquartile ranges of the normalised mean absolute error and normalised mean bias error, indicating better accuracy and reduced variability across buildings. For R-squared and the coefficient of variation of root mean square error, only minor changes appeared. Among the most influential features were yearly statistics, autocorrelation measures, rolling averages, energy-temperature curve features, and linear regression features linking total energy use to outdoor temperature. The highest-ranked classification probability feature, the probability that a building is an office with electric heating, ranked 36th overall.
The researchers are candid about the limitations. Some district heating meters record energy in coarse increments, producing abrupt jumps and plateaus that make the heating signal harder to predict. Sub-metered electric heating data can be incomplete or faulty, since unmonitored ventilation heaters, hot water tanks, or apartment-level appliances mean the recorded ground truth may underestimate actual heating electricity, so poor metrics for some all-electric buildings may reflect flawed measurements rather than model failure. Performance was estimated from a single train/test split rather than repeated cross-validation, no hyperparameter optimisation was performed, and the disaggregation module is intended for offline retrospective analysis, not real-time use, because its features include lead values and yearly statistics. It has not been trained or validated for single-family houses or individual apartments, and the trained models are specific to Norwegian buildings, though the underlying method could be retrained on local data elsewhere.
Even with those caveats, the tool fills a genuine gap. Existing open-source frameworks for load disaggregation were designed for high-resolution appliance-level data with distinct power signatures and do not handle heating disaggregation from hourly meter readings, while energy signature methods model typical load curves rather than extracting heating electricity from actual measurements. No openly available tool previously offered building classification and heating load disaggregation from hourly smart meter data in cold climates. COF-Tool’s web interface accepts a single building file and returns a plot of the uploaded electricity use, estimated class probabilities across all building types, and a downloadable hourly time series of estimated heating electricity. For grid operators, energy analysts, and anyone trying to understand where electricity actually goes in a cold, heavily electrified country, that is a meaningful step forward, and the modular design means both modules can be retrained as more labelled data become available.
Subject of Research: A supervised machine learning framework and open-source tool that classifies building types and estimates heating electricity use from hourly smart meter data of Norwegian buildings.
Article Title: COF-Tool: a supervised method and a tool for building classification and disaggregation of electricity for heating from hourly electricity load profiles from buildings
Article References: Lien, S. K., & Rajasekharan, J. (2026). COF-Tool: a supervised method and a tool for building classification and disaggregation of electricity for heating from hourly electricity load profiles from buildings. Energy Informatics, 9(1), Article 66. https://doi.org/10.1186/s42162-026-00681-2
Image Credits: AI Generated
DOI: 10.1186/s42162-026-00681-2
Keywords: smart meters, machine learning, building energy, electric heating, disaggregation, district heating, random forest, CatBoost, Norway, open-source tool, load profiles, energy informatics
News Source: Denise Maddox. (November 4, 2026). New open-source tool reads building types and heating loads straight from smart meters. Scienmag.
