TL;DR
This paper introduces a data-driven Demand Response (DR) scheme that automatically learns user preferences for appliance scheduling from energy consumption data. By replacing intrusive surveys with adaptive, probabilistic "flex-offers," the system achieves significant energy cost savings (up to 42%) while maintaining high user acceptance and reducing "user response fatigue."
What Problem Does This Paper Address?
To balance the grid as we integrate more fluctuating renewable energy sources, we need Demand Response (DR)—shifting residential energy use to times when supply is high or demand is low. However, current DR programs face a major hurdle: User Response Fatigue.
Most existing systems require users to fill out tedious surveys about their preferences or manually schedule appliances every day. Furthermore, these models are often "hard-coded" and fail to adapt as a user’s habits change. If a system is too demanding or ignores a user's need for clean dishes by a certain time, the user eventually opts out, rendering the DR scheme ineffective.
Methods
The authors propose a multi-stage, automated pipeline for direct load-control of "wet devices" (dishwashers and washing machines):
- Device Activity Prediction: Using historical load data, the system predicts if a device will be used on a given day (Naive Bayesian) and when it will be used (Linear Regression).
- Probabilistic Flex-Offers: Standard "flex-offers" define a fixed window for shifting a task. The authors extend this into a Probabilistic Flex-Offer, which uses random variables to model the uncertainty of when a device is actually "ready" and when it must be finished, allowing the scheduler to account for prediction errors.
- Adaptive User Utility Modeling: This is the core innovation. Instead of asking the user how much they value a delay, the system models User Flexibility using an exponential distribution. It assumes that the probability of a user accepting a delay decreases as the delay increases.
- Online Learning: The model uses Stochastic Gradient Descent (SGD) to update the flexibility parameter ($\lambda$) based on user feedback (e.g., whether the user accepted the suggested schedule or manually overrode it).
- Optimal Scheduling: The scheduler selects a start time that maximizes the Expected User Utility, which balances potential financial gains (from Spot and Regulation markets) against the predicted probability of user acceptance.
Datasets
The study evaluated the scheme using two primary sources:
- Intrepid Dataset: A real-world dataset containing energy consumption time series for 11 devices (washing machines and dishwashers) across 11 households over one year.
- Genetx Dataset: A synthetic dataset used to simulate various household categories (e.g., single worker, family with children) to test the model's robustness across different lifestyle patterns.
- Market Data: Real pricing data from the Danish energy market (DK1 zone) was used for the financial evaluation.
Key Insights and Ideas
- The "Passive" User is the Goal: The authors argue that the more interaction a user is required to perform, the higher the chance they abandon the program. By moving to a "passive feedback" loop (learning from overrides), they minimize the cognitive load on the consumer.
- Flexibility is a Trade-off: A "stricter" model that always tries to please the user results in high acceptance but low financial gain. Conversely, a model focused only on profit will be rejected. The adaptive model finds the "sweet spot" by learning the specific tolerance of each household.
- Uncertainty Matters: By treating the "ready" and "latest start" times as probabilistic rather than fixed, the scheduler can hedge against prediction errors. For example, if the device has a 10-hour flexibility window, a 3-hour prediction error can be absorbed without the user even noticing.
Key Results
- Financial Savings: On the real-world Intrepid dataset, the scheme achieved 32% savings in the Spot Market and 38% in the Regulation Market.
- User Acceptance: The system maintained a 64% acceptance rate on real-world data. On the more regularized synthetic dataset, acceptance reached 79%, with cost reductions up to 42%.
- Adaptability: The adaptive model significantly outperformed a "uniform" model (which assumes all delays are equally acceptable). The adaptive approach increased user acceptance from roughly 36% to 64% by learning from feedback.
Limitations and Future Work
- Device Scope: The current work focuses exclusively on "wet devices." While these account for ~30% of household energy, extending the model to HVAC systems or Electric Vehicles (EVs) would require handling different types of flexibility (e.g., temperature ranges vs. time shifts).
- Simple Baselines: The authors used relatively simple prediction models (Naive Bayes/Linear Regression) as baselines. They acknowledge that more sophisticated deep learning models for short-term load forecasting could further improve accuracy.
- Single Device Optimization: The paper schedules devices individually. Future work aims to establish models that can coordinate multiple devices within a single household or across a neighborhood to prevent creating new "peak" demand periods.
Summary updated