Heat pump detection and load disaggregation in low-resolution smart meter data with convolutional neural networks

The transition from fossil heating systems to heat pumps introduces new challenges for energy grid management, particularly due to their contribution to peak loads. Demand-side flexibility offers a promising solution, but requires detailed monitoring of individual devices. This paper addresses two key tasks using residential smart meter data: detection of installed heat pumps and disaggregation of heat pump load profiles from total household consumption. Multiple machine learning models are trained and evaluated for both tasks using a large real-world dataset comprising more than 7000 heat pumps in Switzerland. For heat pump detection, a rule-based approach achieves classification precision of over 89%, indicating that robust detection is possible even with low-resolution data. For load disaggregation, we develop a dedicated one-dimensional convolutional neural network with differential input and auxiliary features, achieving a root mean squared error below 0.18 kWh, outperforming reference models by more than 20%. The proposed disaggregation model further demonstrates robustness to variations in residual household load and benefits from increased training data, with performance gains saturating beyond approximately 200 heat pumps. Finally, the combined detection and disaggregation framework is evaluated in a neighborhood-scale case study. On a representative winter day, the proposed approach estimates peak heat pump demand with a relative error of 6%, highlighting its potential to support artificial intelligence-enabled demand response energy management applications. • Optimized CNN architectures for heat pump detection and load disaggregation • High accuracy results validated on real-world smart meter data • Large-scale study with over 7000 heat pumps • Fully automated heat pump monitoring through combined detection and disaggregation.