A comparative analysis of federated learning strategies for short-term photovoltaic power forecasting
Vasilis Michalakopoulos, Leonidas Spyropoulos, Loukas Ilias, Elissaios Sarmas, Vangelis Marinakis, Dimitris Askounis
Abstract
Research highlights the urgent need for efficient and secure methods to predict photovoltaic (PV) energy generation, given the increasing dependence on renewable energy sources and the sensitivity of the associated data. Ensuring data security while maintaining high accuracy in results is essential for fostering trust and compliance in energy management systems. Federated Learning (FL) presents a distributed Machine Learning solution that prevents the exchange of sensitive private data while preserving high performance in forecasts. Despite limited studies on the applications of FL in PV power forecasting, existing research often overlooks the effects of various aggregation algorithms and the implications of short-term scenarios. Identifying the gap, this is the first study presenting the analysis of performance of six different weight aggregation algorithms, using a stacked BiLSTM (Bidirectional Long-Short Term Memory) architecture, assessing them under short-term contexts. A real case study involving data from seven PV plants in Portugal, located in four different cities and each with different nominal capacities, as the FL clients, has been conducted. Our results outline that FL, particularly when employing proximal terms in the loss function, outperforms Centralized Learning (CL) and Localized Learning (LL) in 1-day (short-term) scenarios. Specifically, FL improves the normalized root mean squared error and R 2 by up to 10% compared to CL, reaching 25% when opposed to LL. Overall, this paper demonstrates that FL not only excels in privacy-sensitive contexts but also consistently delivers high-accuracy results. • Comparative study of six FL weight aggregation methods using a stacked BiLSTM model. • Analysis of PV plant data from seven clients across four cities in Portugal. • Evaluation of FL vs LL and CL models for short-term forecasts. • Ensured robustness with long testing periods for reliable performance assessment.