Data-Driven Domestic Flexible Demand: Observations from experiments in cold climate

Dirk Reinhardt, Wenqi Cai, Sebastien GrosarXiv 2024 · DOI · arXiv

Abstract

This paper presents a three-year experimental study on a Norwegian residential building to evaluate data-driven modeling for domestic heating as a flexible power grid resource. We demonstrate that linear subspace identification (SI) methods can accurately predict thermal response and energy demand using minimal training data (one week) and low-cost IoT hardware. Furthermore, we show that quantile regression is essential for managing the economic risks associated with non-linear peak-demand penalties in modern grid tariffs.

Introduction

As the global energy mix shifts toward non-deterministic renewable sources like wind and solar, the power grid faces a significant "matching" problem. To maintain stability, demand must become as flexible as production once was. In cold climates, residential space heating accounts for nearly 80% of total household energy use, making it the primary lever for demand-side response.

However, managing this flexibility is technically complex. Commercial heat pumps typically operate as "black boxes" with internal safety logic, meaning users can only influence power draw indirectly via temperature setpoints. Additionally, every home possesses unique thermal dynamics influenced by stochastic factors such as occupant behavior and local weather. This study investigates which data-driven modeling techniques best predict a home's response to help homeowners optimize consumption against spot market prices and peak-demand penalties.

Electricity price distributions and time series for Trondheim and Oslo (2021-2022), highlighting the volatility and regional variance that drive the need for flexible demand.

Methods

The research evaluated four primary data-driven modeling approaches using a dataset collected over three years from an experimental house in Trondheim, Norway:

  1. Subspace Identification (OSSI & SI): Indirect linear methods that build predictors from past input-output data. Multi-step SI was specifically optimized for 12-hour prediction horizons to prevent error accumulation.
  2. Fundamental Lemma (FL): A direct data-driven approach that predicts future behavior using linear combinations of past trajectories without an explicit model.
  3. Quantile Regression (QR): Employed to predict the probability of demand exceeding specific thresholds, rather than just the mean consumption.

The physical setup included a 290m² house with four air-to-air heat pumps, monitored via Sensibo IoT climate sensors and Tibber Pulse power meters. Market data was sourced from Nord Pool (NO3 region).

Key Results and Insights

The study yielded several critical insights for the deployment of smart home energy management:

  • Model Efficiency: Indirect Subspace Identification methods significantly outperformed the direct Fundamental Lemma approach. SI methods were computationally efficient (10–20 seconds) and maintained a standard deviation of error below 1°C over a 12-hour horizon. In contrast, the FL approach was computationally prohibitive for larger datasets.
  • Rapid Deployment: Surprisingly, one week of training data was sufficient to build an accurate thermal model. Performance gains from using a full month of data were marginal, suggesting these systems can become "smart" almost immediately after installation.
  • Managing Economic Risk: Standard regression is insufficient for cost management due to the non-linear nature of grid tariffs. Quantile Regression (targeting the 0.9 and 0.95 quantiles) allowed the system to manage the "tail risk" of triggering expensive peak-demand penalties.
Hourly kWh distribution and profile showing the relationship between consumption peaks and step-wise cost penalties (NOK).

Conclusion

Data-driven linear models, particularly Subspace Identification, provide a robust and computationally light framework for predicting domestic thermal dynamics. When combined with Quantile Regression to navigate complex tariff structures, these models enable effective demand flexibility. Future work will focus on integrating these predictors into Economic Model Predictive Control (EMPC) frameworks and testing their performance across different heating technologies like geothermal or air-to-water systems.

Summary updated

BibTeX
@misc{reinhardt2024data,
  author = {Dirk Reinhardt and Wenqi Cai and Sebastien Gros},
  title = {{Data-Driven Domestic Flexible Demand: Observations from experiments in cold climate}},
  year = {2024},
  eprint = {2407.16475},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2407.16475},
}
Figure