[Paper Review] FedTrees: A Novel Computation-Communication Efficient Federated Learning Framework Investigated in Smart Grids

Mohammad Al-Quraan, Ahsan Khan+4 more|arXiv (Cornell University)|Sep 30, 2022

Privacy-Preserving Technologies in Data4 citations

TL;DR

FedTrees proposes a computation- and communication-efficient federated learning framework using LightGBM for energy demand forecasting in smart grids. By leveraging ensemble learning and a delta-based early stopping mechanism, FedTrees achieves comparable accuracy to FedAvg with only 2% of the computation time and 13% of the communication rounds, significantly improving resource efficiency.

ABSTRACT

Smart energy performance monitoring and optimisation at the supplier and consumer levels is essential to realising smart cities. In order to implement a more sustainable energy management plan, it is crucial to conduct a better energy forecast. The next-generation smart meters can also be used to measure, record, and report energy consumption data, which can be used to train machine learning (ML) models for predicting energy needs. However, sharing fine-grained energy data and performing centralised learning may compromise users' privacy and leave them vulnerable to several attacks. This study addresses this issue by utilising federated learning (FL), an emerging technique that performs ML model training at the user level, where data resides. We introduce FedTrees, a new, lightweight FL framework that benefits from the outstanding features of ensemble learning. Furthermore, we developed a delta-based early stopping algorithm to monitor FL training and stop it when it does not need to continue. The simulation results demonstrate that FedTrees outperforms the most popular federated averaging (FedAvg) framework and the baseline Persistence model for providing accurate energy forecasting patterns while taking only 2% of the computation time and 13% of the communication rounds compared to FedAvg, saving considerable amounts of computation and communication resources.

Motivation & Objective

• To address privacy and scalability challenges in smart grid energy forecasting by avoiding centralised data collection.
• To reduce the high computational and communication costs associated with deep learning models like LSTM in federated learning.
• To develop a lightweight, efficient federated learning framework suitable for deployment at the substation level in smart grids.
• To improve training efficiency through an adaptive early stopping mechanism that halts training when no further performance gain is observed.
• To evaluate the framework’s performance against FedAvg and baseline models in short-term energy demand forecasting.

Proposed method

• FedTrees employs LightGBM (LGBM), a gradient-boosting decision tree algorithm, as the client-side model in a federated learning setup to ensure fast training and low memory usage.
• The framework uses a delta-based early stopping algorithm that monitors the change in validation loss across communication rounds and terminates training when improvement falls below a threshold.
• Communication rounds are dynamically reduced by stopping training once the model converges, avoiding unnecessary iterations.
• The global model is updated via FedAvg aggregation, but training is limited to only as many rounds as necessary for convergence.
• Feature selection is applied to focus on the top four most relevant features, further reducing computational load and improving training efficiency.
• The system is evaluated in a cross-silo FL setting, with substations as clients, enabling scalable and privacy-preserving energy forecasting.

Experimental results

Research questions

• RQ1 Can a lightweight ensemble model like LightGBM outperform deep learning models such as LSTM in federated energy forecasting while reducing computational and communication costs?
• RQ2 How effective is a delta-based early stopping mechanism in reducing training rounds without sacrificing prediction accuracy in federated learning for smart grids?
• RQ3 What is the impact of feature selection on the performance and efficiency of federated energy forecasting models?
• RQ4 How does FedTrees compare to FedAvg and the Persistence baseline in terms of prediction accuracy, communication efficiency, and computation time?
• RQ5 Can federated learning at the substation level achieve better scalability and privacy preservation than client-level federated learning in smart grid applications?

Key findings

• FedTrees achieved a mean absolute error (MAE) of 0.0173 and a mean absolute percentage error (MAPE) of 3.69%, outperforming the Persistence model and closely matching FedAvg in accuracy.
• FedTrees required only 65 communication rounds, representing 13% of the 491 rounds needed by FedAvg, significantly reducing communication overhead.
• Computation time for FedTrees was 26.2 seconds, amounting to just 2% of the 1356 seconds required by FedAvg, demonstrating exceptional efficiency.
• When using only the top four features, FedTrees reduced communication rounds to 50 and computation time to 8.8 seconds, while maintaining a MAPE of 3.54%.
• The delta-based early stopping algorithm effectively reduced training rounds by detecting convergence early, with optimal settings found at delta = 0.00001 and window size = 10 for LGBM.
• FedTrees showed improved performance with feature selection, whereas FedAvg’s performance slightly degraded, highlighting FedTrees’ robustness to input dimensionality.