PePTM: An Efficient and Accurate Personalized P2P Learning Algorithm for Home Thermal Modeling

Karim Boubouh, Robert Basmadjian, Omid Ardakanian, Alexandre Maurer, Rachid GuerraouiEnergies 2023 · DOI

TL;DR

This paper introduces PEPTM, a novel peer-to-peer learning algorithm for training personalized thermal models on smart thermostats. PEPTM significantly reduces energy consumption and preserves privacy by enabling homes to collaboratively learn models without a central server, outperforming traditional centralized and federated learning approaches.

What Problem Does This Paper Address?

Smart thermostats are increasingly incorporating machine learning for personalized home climate control, aiming to enhance comfort and conserve energy. However, training accurate thermal models directly on these devices (on-device learning) has been challenging due to the lack of efficient learning schemes. Traditional methods like Centralized Learning (CL) require sending raw, potentially private, data to a central server, raising privacy concerns and demanding significant energy. Federated Learning (FL) offers some privacy by aggregating model updates instead of raw data, but still relies on a central server, which can be a single point of failure and still incurs energy costs. This paper addresses the need for a decentralized, privacy-preserving, and energy-efficient method for personalized thermal modeling on smart home devices.

Methods

The core innovation is the PEPTM (Personalized Peer-to-Peer) learning algorithm. PEPTM operates in two phases:

  1. Local Training Phase: Each home initially trains a basic thermal model using its own local sensor data. This is designed to be brief, typically one or two epochs, to accommodate the limited computational power of typical smart home devices.
  2. Collaborative Training Phase: Homes then engage in peer-to-peer communication with a select group of similar homes. They exchange model updates (gradients) and collaboratively refine their local models. A key aspect is the weighted averaging aggregation rule, where updates from more similar homes have a greater influence, promoting faster convergence.

To further enhance efficiency and enable training on resource-constrained devices, PEPTM employs two abstraction techniques:

  • Temporal Abstraction: Sensor data is downsampled to larger time intervals (e.g., 15 minutes, 30 minutes, 1 hour) to reduce data size and processing load.
  • Spatial Abstraction (Clustering & Similarity Matrix): Homes are grouped into clusters of similar characteristics (based on floor area and age). For PEPTM, a similarity matrix is used to create a sparse network graph, connecting homes with a small set of highly similar neighbors, thus reducing communication overhead compared to fully connected clusters.

Datasets

The study utilizes the Ecobee dataset, which comprises time-series data from 1000 US homes collected over 12 months in 2017. This dataset includes sensor readings (temperature, humidity) and metadata like floor area and age. The data has a 5-minute resolution and a total size of 34 GB. The authors did not introduce new benchmarks but rather used this existing dataset to evaluate their proposed method against established CL and FL approaches.

Key Insights and Ideas

  • Decentralization is Key for Privacy and Efficiency: The paper strongly advocates for decentralized learning (P2P) as the most effective way to address both privacy concerns and the energy footprint of smart home applications.
  • Personalization Matters: A one-size-fits-all global model is insufficient for diverse home environments. PEPTM's strength lies in its ability to create tailored models for each home.
  • Abstraction is Crucial for Resource-Constrained Devices: Temporal and spatial abstractions are not just optimizations; they are enablers for running sophisticated ML on everyday smart home hardware.
  • Similarity-Based Networking is More Efficient than Clustering for P2P: While clustering works for CL/FL, a sparse network based on a similarity matrix is more efficient for P2P communication, balancing accuracy with reduced bandwidth and energy.
  • Energy Savings are Dramatic: The energy savings achieved by PEPTM are not marginal; they are orders of magnitude greater than CL and FL, making it a truly sustainable solution.

Key Results

  • Energy Efficiency: PEPTM requires 695 times less training energy than CL and 40 times less training energy than FL. This is primarily due to its use of low-power mobile devices and the absence of a power-hungry central server.
  • Performance: PEPTM achieves comparable performance to CL and is superior to FL in terms of accuracy metrics like RMSE and MAE.
  • Convergence: The optimal configuration for PEPTM was found to be around 300 rounds of collaborative training.
  • Temporal Abstraction Impact: Using temporal abstraction (e.g., 15-minute intervals) significantly reduces training time (up to 93% less time compared to no abstraction) and energy consumption, while maintaining good accuracy.
  • Network Density: A network density parameter (p) of 0.05 (averaging about 10 neighbors per home) was found to be optimal, balancing accuracy and energy savings.

Limitations and Future Work

  • Assumed Honest Homes: The current filtering component in PEPTM is designed to mitigate simple attacks but assumes an overall setting of honest homes. More robust defenses against malicious peers could be explored.
  • Scalability: While PEPTM is designed for efficiency, its scalability to very large networks of smart homes and its ability to handle dynamic network changes (peers joining/leaving) are areas for future research.
  • Security and Privacy Challenges: Although PEPTM inherently enhances privacy, further investigation into addressing potential security and privacy challenges in the peer-to-peer setting is suggested.
  • Dynamic Network Support: The paper mentions the need to support dynamic networks, where neighbors can change, which is a natural extension for real-world scenarios.

Summary updated

Abstract

Nowadays, the integration of home automation systems with smart thermostats is a common trend, designed to enhance resident comfort and conserve energy. The introduction of smart thermostats that can run machine learning algorithms has opened the door for on-device training, enabling customized thermal experiences in homes. However, leveraging the flexibility offered by on-device learning has been hindered by the absence of a tailored learning scheme that allows for accurate on-device training of thermal models. Traditional centralized learning (CL) and federated learning (FL) schemes rely on a central server that controls the learning experience, compromising the home’s privacy and requiring significant energy to operate. To address these challenges, we propose PePTM, a personalized peer-to-peer thermal modeling algorithm that generates tailored thermal models for each home, offering a controlled learning experience with a minimal training energy footprint while preserving the home’s privacy, an aspect difficult to achieve in both CL and FL. PePTM consists of local and collaborative learning phases that enable each home to train its thermal model and collaboratively improve it with a set of similar homes in a peer-to-peer fashion. To showcase the effectiveness of PePTM, we use a year’s worth of data from US homes to train thermal models using the RNN time-series model and compare the data across three learning schemes: CL, FL, and PePTM, in terms of model performance and the training energy footprint. Our experimental results show that PePTM is significantly energy-efficient, requiring 695 and 40 times less training energy than CL and FL, respectively, while maintaining comparable performance. We believe that PePTM sets the stage for new avenues for on-device thermal model training, providing a personalized thermal experience with reduced energy consumption and enhanced privacy.
BibTeX
@inproceedings{boubouh2023peptm,
  author = {Karim Boubouh and Robert Basmadjian and Omid Ardakanian and Alexandre Maurer and Rachid Guerraoui},
  title = {{PePTM: An Efficient and Accurate Personalized P2P Learning Algorithm for Home Thermal Modeling}},
  booktitle = {Energies},
  year = {2023},
  doi = {10.3390/en16186594},
}
Figure