Forecasting Energy Availability in Local Energy Communities via LSTM Federated Learning

TL;DR

The paper addresses privacy-aware forecasting of energy production and consumption in Local Energy Communities (LECs). It proposes a horizontal Federated Learning framework using Long Short-Term Memory (LSTM) networks to train predictive models locally while sharing only model updates. The approach is evaluated on a synthetic dataset representing 200 prosumers and 200 consumers, comparing Stand-Alone, Centralized, and Federated configurations and exploring multiple aggregation methods, including FedProx. Results indicate that Centralized learning yields slight accuracy gains over Federated learning, but Federated approaches preserve privacy and offer scalable, privacy-preserving forecasts for LEC optimization. The work demonstrates the practical viability of FL-LSTM for decentralized energy management and outlines paths toward greater scalability, real-time deployment, and integration with edge computing and blockchain.

Abstract

Local Energy Communities are emerging as crucial players in the landscape of sustainable development. A significant challenge for these communities is achieving self-sufficiency through effective management of the balance between energy production and consumption. To meet this challenge, it is essential to develop and implement forecasting models that deliver accurate predictions, which can then be utilized by optimization and planning algorithms. However, the application of forecasting solutions is often hindered by privacy constrains and regulations as the users participating in the Local Energy Community can be (rightfully) reluctant sharing their consumption patterns with others. In this context, the use of Federated Learning (FL) can be a viable solution as it allows to create a forecasting model without the need to share privacy sensitive information among the users. In this study, we demonstrate how FL and long short-term memory (LSTM) networks can be employed to achieve this objective, highlighting the trade-off between data sharing and forecasting accuracy.

Figures (7)

• Figure 1: A schematic representation of the Local Energy Community (LEC): multiple users with different energy profiles are connected in a community to share local resources. The goal of the community is to minimize access to the power grid by relying on local batteries (and production).
• Figure 2: This is our proposed FL architecture: on the left, there are several LSTM layers that represent different client instances, also called 'runners'. The clients train on their local data using a model shared by the server and send the parameters to it. The server then acts as a central aggregator and combines the parameters to obtain a better global model.
• Figure 3: A) Mean Squared Error (MSE) of the three approaches (i.e., Stand-Alone, Federated, Centralized learning), considered a small subset of 10 users and a 10 epochs training (for FL they are Federated Rounds). The plot shows the mean MSE of the different stores during testing phase. B) 10-clients LEC MSE fitting losses performance comparison over time, after a flow of 10 Federated Rounds.
• Figure 4: Federated aggregation methods MSE loss comparison including five of the most used aggregation algorithms such as FedAvg, FedMedian, FedProx, FedAdam and FedYogi on a 10-users small subset, these methods were tested in 10 Federated rounds.
• Figure 5: Seasonal Box-Plot of Mixed LEC which shows the interquartile ranges, whiskers, and outliers of a 50/50 LEC.