Residential Heating Demand Prediction Using Time-Series Clustering and Deep Learning

This paper presents a novel approach for predicting residential heating demand by combining time-series clustering with deep learning models. Using a dataset of 4,660 district heating records, households were categorized into six clusters based on distinct heat consumption patterns, derived from time-series K-Means clustering. Various machine learning and deep learning models were evaluated, including Linear Regression, Support Vector Regression (SVR), Random Forest, Gradient Boosted Regression, and Gated Recurrent Unit (GRU) networks. Surprisingly, Linear Regression outperformed many complex models on the aggregated dataset, achieving an RMSE of 1,209.52 kWh and an R2 of 0.88. However, the GRU model consistently outperformed other models in most cluster-specific evaluations, achieving an RMSE of 1,255.57 kWh and an R2 of 0.87 on the aggregated dataset, and significantly lower RMSE and MAPE values in smaller clusters. The study demonstrates the importance of model selection based on cluster-specific characteristics and highlights the potential of combining time-series clustering with advanced models to improve forecasting accuracy.