Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Who Will Shift? Learning Household Flexibility from Behavioral Patterns in Electricity Usage

Tu, Jingi LU (2025) DABN01 20251
Department of Statistics
Department of Economics
Abstract
This thesis develops an application-oriented framework to identify and predict residential demand-side flexibility from the temporal structure of electricity use, rather than energy totals. We construct structure-based, price-aligned labels that track shifts between peak and off-peak hours, and train three complementary models: a next-day behavioral classifier (M1) for short-horizon operations, a 30-day trend classifier (M2) for medium-term targeting, and a 30-day trend-score regressor (M3) to quantify structural change. Using high-resolution interval data from London and Ireland, we show that Transformer encoders effectively capture cross-day patterns, and that models trained in London transfer well to Ireland with minimal or no... (More)
This thesis develops an application-oriented framework to identify and predict residential demand-side flexibility from the temporal structure of electricity use, rather than energy totals. We construct structure-based, price-aligned labels that track shifts between peak and off-peak hours, and train three complementary models: a next-day behavioral classifier (M1) for short-horizon operations, a 30-day trend classifier (M2) for medium-term targeting, and a 30-day trend-score regressor (M3) to quantify structural change. Using high-resolution interval data from London and Ireland, we show that Transformer encoders effectively capture cross-day patterns, and that models trained in London transfer well to Ireland with minimal or no fine-tuning—demonstrating the robustness of both the labels and modeling approach across regions. Finally, we translate the results into a practical guide: M1 supports day-ahead micro-interventions around behavioral turning points; M2 helps prioritize pilots and user segments; and M3 informs tiered incentives, contract design, and aggregator portfolio strategies. The framework is interpretable, privacy-preserving, and portable, offering electricity-market stakeholders a low-burden tool to operationalize demand-side flexibility. (Less)
Please use this url to cite or link to this publication:
author
Tu, Jingi LU
supervisor
organization
course
DABN01 20251
year
type
H1 - Master's Degree (One Year)
subject
keywords
demand-side flexibility, temporal usage structure, transfer learning, stakeholder handbook
language
English
id
9213256
date added to LUP
2025-11-06 12:21:50
date last changed
2025-11-06 12:21:50
@misc{9213256,
  abstract     = {{This thesis develops an application-oriented framework to identify and predict residential demand-side flexibility from the temporal structure of electricity use, rather than energy totals. We construct structure-based, price-aligned labels that track shifts between peak and off-peak hours, and train three complementary models: a next-day behavioral classifier (M1) for short-horizon operations, a 30-day trend classifier (M2) for medium-term targeting, and a 30-day trend-score regressor (M3) to quantify structural change. Using high-resolution interval data from London and Ireland, we show that Transformer encoders effectively capture cross-day patterns, and that models trained in London transfer well to Ireland with minimal or no fine-tuning—demonstrating the robustness of both the labels and modeling approach across regions. Finally, we translate the results into a practical guide: M1 supports day-ahead micro-interventions around behavioral turning points; M2 helps prioritize pilots and user segments; and M3 informs tiered incentives, contract design, and aggregator portfolio strategies. The framework is interpretable, privacy-preserving, and portable, offering electricity-market stakeholders a low-burden tool to operationalize demand-side flexibility.}},
  author       = {{Tu, Jingi}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Who Will Shift? Learning Household Flexibility from Behavioral Patterns in Electricity Usage}},
  year         = {{2025}},
}