Leveraging Self-Supervised Transformers for Heat Pump Consumption Classification Using the HEAPO Dataset

Abstract

Heat pumps play a crucial role in reducing carbon emissions in residential heating; however, their electricity consumption varies significantly across households, making large-scale performance assessment challenging. Existing data-driven approaches often rely on fully supervised learning and require extensive labeled data, which are costly to obtain through on-site inspections. This paper investigates whether routinely collected smart meter and weather data can be leveraged to automatically classify household heat pump electricity consumption into three categories: low, normal, and rather high. We propose a self-monitoring transformer framework that combines self-supervised pretraining and supervised fine-tuning for time series classification. A transformer encoder is first pre-trained on multi-day electricity consumption windows from largely unlabeled households using a reconstruction-based objective. The learned representations are then fine-tuned with a lightweight fully connected classifier on a limited labeled subset of the HEAPO dataset. Experimental results demonstrate that the proposed approach consistently outperforms state-of-the-art baselines, including RNN-, CNN-, and transformer-based time series models, achieving approximately 93% classification accuracy, an F1 score above 0.9, and a recovery rate close to 91%. These results indicate that self-supervised representation learning on smart meter data provides an effective and scalable solution for monitoring heat pump electricity consumption and supporting inspection planning.