PaperlayerView exact revision on arXiv

Source-linked AI summary

Electrical Load Forecasting Using Edge Computing and Federated Learning

Afaf Taik, Soumaya Cherkaoui

arXiv:2201.11248v1cs.CE

TL;DR

Household short-term load forecasting requires diverse data, but fine-grained consumption records are privacy-sensitive and centralized collection can expose household behavior. The paper evaluates edge computing with federated LSTM training, finding overall personalization improvements and promising federated models for household forecasting. Its feasibility remains constrained by the computational burden of local neural-network training on edge devices.

  • Problem

    Household short-term load forecasting needs large, diverse datasets for accurate deep learning, while detailed consumption data can reveal appliance use and occupancy.

  • Method

    The paper evaluates LSTM forecasting trained through federated learning on household edge equipment, with local training and server aggregation of model updates.

  • Results

    5.07% overall MAPE improvement was reported for model 1 after personalization on participating clients, alongside 4.78% on non-participating clients.

  • Takeaways & Limitations

    Global models can forecast a population using only a subset of users, while local retraining can produce more accurate personalized predictions and provide a starting point for new clients.

  • Takeaways & Limitations

    Local neural-network training may compromise edge-device performance, and model accuracy still varies by user even after personalization.

Abstract

from arXiv · show

In the smart grid, huge amounts of consumption data are used to train deep learning models for applications such as load monitoring and demand response. However, these applications raise concerns regarding security and have high accuracy requirements. In one hand, the data used is privacy-sensitive. For instance, the fine-grained data collected by a smart meter at a consumer's home may reveal information on the appliances and thus the consumer's behaviour at home. On the other hand, the deep learning models require big data volumes with enough variety and to be trained adequately. In this paper, we evaluate the use of Edge computing and federated learning, a decentralized machine learning scheme that allows to increase the volume and diversity of data used to train the deep learning models without compromising privacy. This paper reports, to the best of our knowledge, the first use of federated learning for household load forecasting and achieves promising results. The simulations were done using Tensorflow Federated on the data from 200 houses from Texas, USA.

I. INTRODUCTION

Household short-term load forecasting is difficult because consumption profiles are volatile and behavior-dependent, while deep models need diverse data but centralized collection threatens privacy. The paper evaluates edge-based federated learning with LSTM models to address these challenges.

  • Household short-term load forecasting is challenging because electrical demand is volatile and strongly correlated with stochastic resident behavior.
  • Deep neural networks can capture complex, nonlinear household-load patterns, but overfitting motivates combining larger and more diverse household datasets.
  • Centralized transfer of fine-grained household records exposes sensitive information about device usage and occupancy to interception or misuse.
  • Federated learning trains a shared model by sending models to clients, aggregating local weight updates, and keeping household data on devices.
  • The proposed system combines edge equipment, federated learning, and LSTM forecasting for geographically and structurally similar houses over a short period.

II. RELATED WORK

Prior work improves short-term load forecasting with neural architectures, complementary data, or pooled users, but centralized approaches create privacy concerns. This paper uses edge equipment and federated learning to address privacy and prediction accuracy together.

  • LSTM-based forecasting is widely studied, but desired RMSE and MAPE accuracy remains difficult to achieve across short-term load settings.
  • Forecasting accuracy often drops on new datasets, motivating complementary weather or appliance data to improve generalization.
  • Clustering and pooling users can reduce uncertainty or increase training-data diversity, but these approaches remain heavily centralized and privacy-prone.
  • Existing privacy methods include geographically shared identifiers and consumption-data aggregation, each imposing constraints on individual treatment or data handling.
  • The proposed edge-based federated scheme trains a global LSTM model using household data locally while addressing privacy and prediction accuracy together.

III. SYSTEM MODEL

The system combines LSTM-based short-term load forecasting with federated learning across household edge equipment. Clients train locally on consumption data, while a server aggregates updates into a global model.

  • B. Federated Learning: The proposed architecture uses a Multi-access Edge Computing server and household clients with smart meters or other home-area-network devices.The MEC server orchestrates training rounds, while clients use their own electrical consumption data to train a global LSTM-based model.
  • A. Time series forecasting using LSTM: LSTM forecasts future electrical load from previous observations and models temporal correlations in historical time-series data.Its memory cell and gates regulate information flow and address vanishing and exploding gradient problems common in recurrent networks.
  • B. Federated Learning: Federated learning distributes most training across devices called clients rather than centralizing the training data.The approach is suited to privacy-sensitive, large, distributed, or non-identically distributed datasets.
  • B. Federated Learning: Each federated round selects clients, sends them the current global model, computes local SGD updates, and aggregates those updates into a new global model.Client selection favors houses with enough records and sufficient consumption variation to enrich the training data.
  • B. Federated Learning: Personalization retrains the centralized model locally for a small number of epochs using each user’s data to build a user-specific model.This addresses the possibility that a centralized model may not fit all users’ electrical consumption.
  • B. Federated Learning: Federated learning reduces privacy risks relative to centralized server storage, but individual weight uploads must not be inspected or analyzed.The described procedure processes ephemeral updates in memory and discards them after aggregation; aggregated data may still be sent for billing.

C. Networking Load Gain

The paper defines centralized and federated network loads to quantify the networking-load gain achieved by federated learning.

  • C. Networking Load Gain: Centralized network load accounts for each client’s transmitted data size and the number of hops between that client and the server.The data size sent by client k is S_k−d, and d_k denotes the centralized-training hop count.

A. Dataset Pre-Processing and Evaluation Method

The evaluation uses a selected subset of household electricity records and measures forecasting error with RMSE and MAPE.

  • A. Dataset Pre-Processing and Evaluation Method: The study selects 200 clients from Pecan Street Dataport, whose dataset contains circuit-level electricity-use data from approximately 800 U.S. homes.The selected homes have similar properties, including detached-family housing and a common area.
  • A. Dataset Pre-Processing and Evaluation Method: RMSE quantifies prediction error in energy units, while MAPE expresses error as a percentage relative to the actual value.These metrics evaluate the model’s performance with regard to prediction error.
  • A. Dataset Pre-Processing and Evaluation Method: In the evaluation formulas, ŷ_i is the predicted value, y_i is the actual value, and P is the number of predicted values.

B. Simulations setup

The simulations run TensorFlow Federated on a laptop, while the study deliberately focuses on evaluating federated learning rather than extensive hyper-parameter tuning.

  • B. Simulations setup: The simulations use TensorFlow Federated 0.4.0 with a TensorFlow 1.13.1 backend on a laptop equipped with an Intel i7 processor, 16GB memory, and an NVIDIA GeForce GTX 1070.
  • B. Simulations setup: The work does not extensively tune hyper-parameters because its focus is evaluating the federated learning paradigm.The setup follows prior guidance that multiple layers and sufficiently many hidden nodes can provide performance insensitivity to some layer configurations.

1) Evaluated scenarios:

The study evaluates federated-learning scenarios by varying the number of participating clients and local-training epochs, then assesses the resulting global models with RMSE and MAPE. Larger client subsets generally help, while additional local epochs can compensate when communication is costly.

  • Evaluated scenarios: Each federated-learning round selects a subset of clients, and the experiments vary subset size and local-training epochs across 20 rounds.The evaluated scenarios are summarized in Table I.
  • Global-model evaluation: Global models are evaluated using RMSE and MAPE for participating and non-participating clients.Tables II and III report results for 180 participating and 20 non-participating clients.
  • Global-model evaluation: MAPE excludes null consumption values, and the reported accuracy is considered reasonable for one-hour, single-house forecasting.The authors relate the observed values to prior work on short-term forecasting.
  • Scenario effects: Selecting more clients per round is preferable, whereas additional local-training epochs can offset the difference when transmitting updates is more expensive.The same pattern is reported for clients that did not participate in training.

3) Behaviour of personalization:

The paper examines whether locally retraining global models improves household-level forecasts. Personalization generally improves accuracy and can better follow individual consumption profiles, while global models remain useful starting points for clients with insufficient local data.

  • Personalization procedure: Personalization retrains each client’s model locally for five epochs and is evaluated separately for participating and non-participating clients.Results are reported in Tables IV and V.
  • Personalization results: 5.07% MAPE improvement occurs for model 1 on participating clients after personalization.The corresponding improvement for non-participating clients is 4.78%.
  • Personalization results: 4.78% MAPE improvement occurs for model 1 on non-participating clients after personalization.The authors report an overall improvement for most models.
  • Limitations observed: Some clients do not improve after retraining because historical data quality is poor, and these clients are treated as outliers outside the study’s scope.Their high MAPE affects average results.
  • Practical implications: Personalized models follow consumption-profile curves more accurately, while global models provide a useful starting point for new clients lacking enough personalization data.The comparison uses example participant and non-participant clients.

V. REMARKS & FUTURE WORK

The approach remains promising, but its feasibility depends on edge-device training capabilities and accuracy still varies across users. Personalization improves forecasts for both participating and non-participating clients, while outlier handling and client clustering remain future work.

  • Limitations: Edge-device capabilities constrain the feasibility of performing local neural-network training.Although newer IoT devices can run complex models, training may compromise device performance; TensorFlow Lite is identified as a possible future implementation foundation.
  • Future work: Model accuracy varies by user even after personalization, motivating client clustering beyond geographical proximity.The paper also identifies investigation of outliers as future work.
  • Personalization: Local retraining for 5 epochs reduced client 4313’s RMSE from 0.55 kW to 0.388 kW.Client 4313 participated in training the global model 4.
  • Personalization: Local retraining for 5 epochs reduced client 8467’s RMSE from 0.8 kW to 0.72kW.Client 8467 did not participate in training the global model 4.
  • Conclusion: The study reports federated edge training as promising for high-performing population models with reduced networking load and preserved consumption-data privacy.Personalized retraining can improve fit for applications requiring high accuracy.
Loading 2201.11248v1…