The researchers, led by Jørgen Norø Sørensen with colleagues including Moon Keun Kim, drew on data collected through the COFACTOR project on coincidence factors and peak load in the Norwegian low carbon society. The heating dataset covered 26,302 hours of measurements from the office building, while the electricity dataset spanned 36,695 hours from the school building, both recorded hourly. The raw data was described as highly variable and noisy, and it was reduced during a data mining process before modeling began.

For inputs, the study deliberately avoided complex operational variables such as occupancy rates, HVAC schedules, and indoor conditions, which are difficult and expensive to collect. Instead, the models received six features: ambient air temperature in degrees Celsius, outdoor humidity in grams per kilogram, wind speed in meters per second, wind direction, a weekday index running from Monday as zero to Sunday as six, and a working hour index set to one between 06:00 and 17:00 and zero otherwise. The authors note that these two temporal indices may indicate times when there is a high probability of building activity and increased energy consumption, serving as a proxy for occupancy data that is often unavailable.

The eight algorithms comprised four artificial neural network models using backpropagation with distinct optimization techniques, namely Levenberg-Marquardt backpropagation, Bayesian regularization backpropagation, scaled conjugate gradient backpropagation, and quasi-Newton backpropagation, plus AdaBoost, XGBoost, support vector regression, and linear regression. The neural networks were built in Matlab 2021a, while the remaining models were constructed and simulated in Python. Each dataset was split with 80 percent of the data used for training and 20 percent reserved for testing, a common distribution in machine learning practice. Data was normalized using the Min-Max method to scale values to a defined range and give different features equal influence.

The authors caution that these metrics can occasionally give conflicting signals. For example, on the electricity dataset, the SVR1 model achieved a better R2 score but a worse MAPE score than SVR2, illustrating that the four measures together give an indication of how well models adapt to actual consumption rather than a definitive ranking.

The results for heating demand were disappointing across the board. The neural network models produced relatively inaccurate predictions on the office building dataset, predicting excessively high heating demand until December 10, 2021, and then suddenly predicting values that were too low from the following day onward, a pattern the authors suggest may indicate a delay between the actual heating demand and the independent variables. AdaBoost and linear regression failed to fit the fluctuating data, while XGBoost responded almost too much, adapting to fluctuations but failing to match actual values accurately. The best-performing model for heating was SVR3, with an R2 of 0.42, but the authors state that even the SVR models did not achieve the accuracy desired for further prediction use, and all models showed high MAPE values indicating predictions far from actual heating demand.

Electricity consumption in the school building told a different story. The neural network models learned the patterns of the electricity data and produced the best overall results, with R2 values around 0.73 to 0.74 and MAPE values near 35 to 37 percent, although they did not adapt well to peaks in consumption. XGBoost adapted to the peaks but fluctuated rapidly, and its overall accuracy was not as good as that of the neural networks. According to the study’s conclusions, the QNBP2 model yielded the most accurate electricity consumption predictions. The authors attribute the improvement to clearer patterns in the electricity dataset, noting that electricity demand typically demonstrates more consistent daily and weekly patterns compared to heating demand.

A correlation analysis helped explain the gap between the two outcomes. Wind direction and wind speed showed very low correlation with electricity consumption and were excluded from the training phase. Temperature and humidity ratio showed a relatively strong correlation with electricity usage, while the weekday index was only weakly correlated with either electricity usage or heating demand. The authors conclude that the suboptimal heating predictions indicate the chosen input variables were inadequate to represent the dynamic behavior of building heating systems, which is influenced not only by exterior weather but also by occupant behavior, HVAC operating schedules, thermostat set-point adjustments, internal heat gains, building thermal mass, and management strategies. Because these operational factors were absent from the datasets, the models could not account for the rapid changes observed in the office building, resulting in comparatively low R2 values and elevated prediction errors.

Hyperparameter tuning produced mixed effects across the models. The team tested three representative parameter configurations for each method, derived from widely accepted settings in the literature, and manually adjusted hyperparameters rather than employing systematic optimization methods such as grid search, random search, or Bayesian optimization, which were outside the scope of the project. In the neural network models, increasing the number of hidden-layer neurons generally did not lead to notable improvements in prediction accuracy, with exceptions observed for the Bayesian regularization model on the heating dataset and the quasi-Newton model on the electricity dataset. Adjustments to the learning rate showed no consistent pattern of improving or degrading performance, and the same hyperparameter settings did not consistently produce the most accurate predictions on both datasets. The authors emphasize that their results should be seen as a comparative evaluation of methods under typical hyperparameter configurations rather than a comparison of fully optimized models.

The study’s broader message is that prediction accuracy is predominantly constrained by the quality of input features rather than the complexity of the model. Weather variables coupled with basic temporal indicators offer only limited explanatory power for heating demand, and the authors recommend that future research incorporate operational building data, including occupancy schedules, indoor environmental parameters, HVAC performance metrics, solar radiation, lagged energy consumption, and thermostat set-point information, to improve forecasting. They also note that accurate forecasting models require high-quality data and strong correlations between independent and dependent variables. For building owners and energy planners, the findings suggest that weather-driven machine learning models can already deliver useful electricity consumption forecasts even without comprehensive operational data, but reliable heating demand prediction will likely require richer information about how buildings are actually operated.

Subject of Research: Comparing machine learning models for predicting heating demand and electricity consumption in two Norwegian buildings using weather data and temporal indices.

Article Title: Comparative analysis of machine learning models for predicting heating demand and electricity consumption using weather data: Two case studies in Norway

Article References: Sørensen, J. N., Liu, J., Kim, M. K., & Lee, S. (2026). Comparative analysis of machine learning models for predicting heating demand and electricity consumption using weather data: Two case studies in Norway. Energy Reports, 16, Article 109762. https://doi.org/10.1016/j.egyr.2026.109762

Image Credits: AI Generated

DOI: 10.1016/j.egyr.2026.109762

Keywords: machine learning, building energy prediction, heating demand, electricity consumption, weather data, artificial neural networks, XGBoost, support vector regression, hyperparameter tuning, Norway, district heating, energy forecasting