Exploring Reinforcement Learning to Optimize Residential Energy Usage
(2026) MVKM05 20261Department of Energy Sciences
- Abstract
- This thesis explores the use of Reinforcement Learning (RL) to optimize residential energy usage in a system composed of photovoltaic (PV) panels, a battery, an electric vehicle (EV), household consumption, and the electricity grid. A Deep Q-Network (DQN) agent was developed to minimize the electricity bill by optimizing battery usage and EV charging using real residential data supplied by Siemens Energy. Model performance was evaluated against a baseline case without optimization, a Mixed-Integer Linear Programming (MILP) formulation with perfect information – theoretical optimization ceiling, and an alternative optimization method – MILP using forecasts. The results show that the RL agent was capable of learning economically meaningful... (More)
- This thesis explores the use of Reinforcement Learning (RL) to optimize residential energy usage in a system composed of photovoltaic (PV) panels, a battery, an electric vehicle (EV), household consumption, and the electricity grid. A Deep Q-Network (DQN) agent was developed to minimize the electricity bill by optimizing battery usage and EV charging using real residential data supplied by Siemens Energy. Model performance was evaluated against a baseline case without optimization, a Mixed-Integer Linear Programming (MILP) formulation with perfect information – theoretical optimization ceiling, and an alternative optimization method – MILP using forecasts. The results show that the RL agent was capable of learning economically meaningful behaviour, such as charging during low-price periods and discharging during high-price periods. The best RL model achieved a 18.9% yearly cost reduction, while MILP with forecasts achieved 44.0%. The multi-objective nature of the problem highlights limitations related to reward design, action space complexity, and training stability. Overall, the study demonstrates the potential of RL for residential energy management, motivating a more robust and in-depth exploration of its application in this context. (Less)
- Popular Abstract
- As more households install solar panels, home batteries, and electric vehicles, managing when and how electricity is used becomes increasingly important. Making smart decisions about storing energy, charging an electric vehicle, or buying/selling electricity from the grid can reduce energy costs. This thesis investigates whether artificial intelligence can help automate these decisions.
The study developed an artificial intelligence system that learns through experience, using real energy consumption data provided by Siemens Energy. The system was designed to decide when to charge or discharge a home battery and when to charge an electric vehicle in order to minimize a household's electricity bill. Its performance was compared with a... (More) - As more households install solar panels, home batteries, and electric vehicles, managing when and how electricity is used becomes increasingly important. Making smart decisions about storing energy, charging an electric vehicle, or buying/selling electricity from the grid can reduce energy costs. This thesis investigates whether artificial intelligence can help automate these decisions.
The study developed an artificial intelligence system that learns through experience, using real energy consumption data provided by Siemens Energy. The system was designed to decide when to charge or discharge a home battery and when to charge an electric vehicle in order to minimize a household's electricity bill. Its performance was compared with a household that did not use any optimization, with mathematical optimization methods that represent the best possible, and with a forecast-based optimization method.
The results show that the artificial intelligence learned effective strategies, such as charging the battery when electricity prices were low and using stored energy when prices were high. Over one year, it reduced electricity costs by nearly 19%, while the forecast-based mathematical optimization achieved a 44% reduction. Although the artificial intelligence did not outperform the mathematical methods, the study identified several challenges that affect its performance, including balancing multiple objectives and ensuring stable learning.
These findings demonstrate that artificial intelligence has strong potential for residential energy management and encourage further research to improve its performance in real-world applications. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/student-papers/record/9247307
- author
- Ayres Barreto Caldeira Jorge, Pedro LU
- supervisor
-
- Rixin Yu LU
- organization
- course
- MVKM05 20261
- year
- 2026
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Residential Energy Systems Energy Management System Photovoltaic Generation Battery Energy Storage Electric Vehicle Charging Reinforcement Learning Mixed-Integer Linear Programming Demand-Side Energy Management
- report number
- ISRN LUTMDN/TMHP-26/5710 -SE
- ISSN
- 0282-1990
- language
- English
- id
- 9247307
- date added to LUP
- 2026-08-13 14:20:57
- date last changed
- 2026-08-13 14:20:57
@misc{9247307,
abstract = {{This thesis explores the use of Reinforcement Learning (RL) to optimize residential energy usage in a system composed of photovoltaic (PV) panels, a battery, an electric vehicle (EV), household consumption, and the electricity grid. A Deep Q-Network (DQN) agent was developed to minimize the electricity bill by optimizing battery usage and EV charging using real residential data supplied by Siemens Energy. Model performance was evaluated against a baseline case without optimization, a Mixed-Integer Linear Programming (MILP) formulation with perfect information – theoretical optimization ceiling, and an alternative optimization method – MILP using forecasts. The results show that the RL agent was capable of learning economically meaningful behaviour, such as charging during low-price periods and discharging during high-price periods. The best RL model achieved a 18.9% yearly cost reduction, while MILP with forecasts achieved 44.0%. The multi-objective nature of the problem highlights limitations related to reward design, action space complexity, and training stability. Overall, the study demonstrates the potential of RL for residential energy management, motivating a more robust and in-depth exploration of its application in this context.}},
author = {{Ayres Barreto Caldeira Jorge, Pedro}},
issn = {{0282-1990}},
language = {{eng}},
note = {{Student Paper}},
title = {{Exploring Reinforcement Learning to Optimize Residential Energy Usage}},
year = {{2026}},
}