A Comparison of Machine Learning Prediction Models to Estimate the Future Heat Demand

This paper compares machine learning models for short-term heat demand forecasting in residential and multi-family buildings, evaluating model suitability, data impact on accuracy, computation time, and accuracy improvement methods. The findings are relevant for energy suppliers, researchers, and decision-makers in optimizing energy management and improving heat demand forecasting. The included models in the study are k-NN, Polynomial Regression, and LSTM with weather data, building type, and time index as input variables. Single-dimensional models (Autoregression, SARIMA, and Prophet) based on historical consumption are also studied. LSTM consistently outperforms other models in accuracy across different input variable combinations, measured using mean absolute percentage error (MAPE). The incorporation of historical consumption data improved the performance of k-NN and Polynomial Regression models. The paper also explores dataset volume impact on accuracy and compares training and prediction times. k-NN has the least prediction times, Polynomial Regression takes longer, and LSTM requires more time. All models exhibit acceptable prediction times for heat consumption. LSTM outperforms single-dimensional models in accuracy and has lower prediction times compared to AR, SARIMA, and Prophet models.