[Paper Review] Data-Driven Load Modeling and Forecasting of Residential Appliances

Yuting Ji, Elizabeth J. Buechler+1 more|arXiv (Cornell University)|Oct 8, 2018

Smart Grid Energy Management4 citations

TL;DR

This paper proposes a conditional hidden semi-Markov model (CHSMM) for data-driven, high-resolution load modeling and short-term forecasting of residential appliances using 1-minute sub-metered power data. By modeling appliance states and state durations probabilistically and conditioning transitions on exogenous variables like temperature and time of day, the method enables accurate forecasting and anomaly detection, with NRMSE reduced to 0.050 for aggregated air conditioners using model refinements.

ABSTRACT

The expansion of residential demand response programs and increased deployment of controllable loads will require accurate appliance-level load modeling and forecasting. This paper proposes a conditional hidden semi-Markov model to describe the probabilistic nature of residential appliance demand, and an algorithm for short-term load forecasting. Model parameters are estimated directly from power consumption data using scalable statistical learning methods. Case studies performed using sub-metered 1-minute power consumption data from several types of appliances demonstrate the effectiveness of the model for load forecasting and anomaly detection.

Motivation & Objective

• To develop a scalable, data-driven load model for residential appliances that captures stochastic behavior at high temporal resolution.
• To enable accurate short-term load forecasting by learning probabilistic state transitions and duration distributions from real power consumption data.
• To incorporate exogenous variables such as outdoor temperature and time of day to improve model adaptability to behavioral and environmental factors.
• To support demand response and grid reliability by enabling forecasting at the individual appliance and aggregated levels.
• To demonstrate the model’s utility in detecting anomalous appliance behavior through prediction error analysis.

Proposed method

• The model uses a conditional hidden semi-Markov model (CHSMM) to represent appliance operation as a sequence of unobserved states with random durations.
• State transitions and emission probabilities are conditioned on exogenous variables like outdoor temperature and time of day to reflect environmental and temporal influences.
• Model parameters are estimated via scalable statistical learning, specifically multivariate logistic regression (MNLR), to handle high-dimensional state and duration spaces efficiently.
• A short-term load forecasting algorithm is developed based on the learned CHSMM, enabling predictions over a 60-minute horizon.
• Two refinements—weighted MNLR and state-specific MNLR—are applied to improve modeling of long-duration events and state-specific duration distributions, especially for air conditioners.
• Anomaly detection is performed by identifying large load prediction errors, indicating deviations from typical appliance behavior.

Experimental results

Research questions

• RQ1 Can a data-driven CHSMM effectively model the probabilistic, state-based behavior of residential appliances using high-resolution power data?
• RQ2 How does conditioning state transitions and durations on exogenous variables improve forecasting accuracy?
• RQ3 To what extent can model refinements such as weighted and state-specific MNLR reduce prediction error for specific appliances like air conditioners?
• RQ4 Can prediction error be reliably used to detect anomalous appliance behavior, such as extended charging interruptions in electric vehicles?
• RQ5 How well does the model generalize across different levels of appliance aggregation?

Key findings

• The CHSMM achieved a 1-hour ahead load forecasting NRMSE of 0.109 for aggregated air conditioners using the basic model, which was reduced to 0.050 with both model refinements applied.
• The weighted MNLR refinement significantly improved prediction accuracy for infrequent, long-duration ON states in air conditioners, particularly during high thermal load periods.
• The state-specific MNLR approach reduced prediction error by enabling better approximation of the duration distributions for ON and OFF states.
• Anomaly detection based on prediction error successfully identified electric vehicles with extended charging interruptions, such as one with a 16.3-day OFF state in testing versus 0.98 days in training data.
• The model demonstrated robust performance across various appliance types and aggregation levels, with low mean prediction error for most appliances.
• The integration of exogenous variables like temperature and time of day enhanced model adaptability and forecasting accuracy.