AI Building EnergyEvidence Atlas
Home / Topics
Research theme

Energy Disaggregation and Demand Response

The selected studies cover supervised deep-learning disaggregation at appliance level and unsupervised HVAC disaggregation at whole-building level. The first emphasizes temporal representation learning; the second connects disaggregated HVAC demand to flexibility and demand-response estimation.

TSILNet framework for non-intrusive load monitoring and appliance-level energy disaggregation.
Figure 1. Overview of the proposed TSILNet framework for data acquisition, preprocessing, feature extraction, learning, and model evaluation.
Figure excerpt from Zhu et al. (2024), Building Simulation. DOI 10.1007/s12273-024-1175-9
Source-attributed paper figure
Method for HVAC energy disaggregation and demand-response potential evaluation.
Figure 1. HVAC energy disaggregation and demand-response potential evaluation method based on time-frequency decomposition and regression modeling.
Figure excerpt from Zhang et al. (2025), Building Simulation. DOI 10.1007/s12273-025-1367-y
Source-attributed paper figure

Questions addressed

  • How can aggregate energy data be separated into end uses?
  • Which NILM models combine convolution, attention, and recurrence?
  • How can HVAC disaggregation support demand-response planning?
Source figures

Methods and applications represented in this theme

TSILNet framework for non-intrusive load monitoring and appliance-level energy disaggregation.
Figure 1. Overview of the proposed TSILNet framework for data acquisition, preprocessing, feature extraction, learning, and model evaluation.
Figure excerpt from Zhu et al. (2024), Building Simulation. DOI 10.1007/s12273-024-1175-9
Source-attributed paper figure
Method for HVAC energy disaggregation and demand-response potential evaluation.
Figure 1. HVAC energy disaggregation and demand-response potential evaluation method based on time-frequency decomposition and regression modeling.
Figure excerpt from Zhang et al. (2025), Building Simulation. DOI 10.1007/s12273-025-1367-y
Source-attributed paper figure

Comparison of selected papers

PaperResearch problemMethodReported evidence
Zhu et al. (2024)Non-intrusive load monitoring must recover appliance-level demand from aggregate household signals despite nonlinear and long-range temporal dependencies.TSILNet combines a two-stage improved temporal convolutional network, an improved efficient channel-attention mechanism, skip connections, and LSTM sequence modeling.On REDD and UK-DALE, the abstract reports an average 18.67% reduction in MAE and a 38.70% improvement in F1 score across appliances.
Zhang et al. (2025)Utilities often possess only whole-building energy data, making it difficult to isolate HVAC demand and estimate demand-response flexibility.An unsupervised MSTL-VMD time-frequency decomposition is paired with an improved equivalent thermal-parameter model for demand-response potential estimation.Using three years of data from 10 office buildings, the reported HVAC-disaggregation NRMSE ranges from 2.4% to 12.6%; 80% of regression cases exceed R² = 0.8, and a 1°C indoor-temperature increase yields 7.3%–18.4% summer energy savings.

Paper pages