Electrical and Computer Engineering ETDs

Publication Date

Summer 7-28-2026

Abstract

The growing complexity and uncertainty of residential energy use, driven by electric

vehicles and renewable technologies, demand more intelligent and robust

management systems. Traditional methods often fail when faced with unpredictable

electricity prices and user behavior. This dissertation addresses this gap by presenting

a novel personalized framework combining detailed household energy modeling with

a risk-aware reinforcement learning agent for appliance scheduling.

The first contribution is a probabilistic, bottom-up simulation model that captures

the interdependent behaviors of occupants, appliances, and electric vehicles to

generate realistic, high-fidelity load profiles. The second contribution is a lightweight,

tabular Distributional Q-Learning (D-QL) algorithm that schedules flexible loads.

By learning the full distribution of potential costs rather than a single average value,

the D-QL agent inherently manages the risk posed by inaccurate price forecasts.

Results show that our simulation framework accurately reproduces real-world energy

data and future load scenarios. Furthermore, the D-QL scheduler consistently

outperforms a standard Q-learning counterpart in head-to-head comparisons,

delivering superior cost savings while maintaining user comfort. This research

demonstrates that combining personalized modeling with distributional reinforcement

learning offers a powerful and practical solution for robust energy management in

modern smart homes.

Document Type

Dissertation

Language

English

Degree Name

Computer Engineering

Level of Degree

Doctoral

Department Name

Electrical and Computer Engineering

First Committee Member (Chair)

Manel Martinez-Ramon

Second Committee Member

Ramiro Jordan

Third Committee Member

Marios Pattichis

Fourth Committee Member

Jose Cerrato

Download

Share

COinS