Empowering federated learning techniques for privacy-preserving PV forecasting

Machine learning techniques hold significant potential in providing precise forecasts for residential photovoltaic (PV) time series production. This advancement, however, introduces a critical concern: the need to prioritize and uphold privacy when dealing with sensitive personal data within these machine learning models. Preserving the confidentiality of individual information becomes crucial, and the challenge is amplified by the substantial size of the datasets making the process more intricate. This paper explores privacy-preserving federated learning models in the context of big data-driven residential PV production. By leveraging FL techniques, we propose a framework that enables collaborative model training across decentralized prosumer energy data without compromising sensitive information. Prior to the federated approach, an extensive hyperparameter tuning procedure is executed individually for each Long Short-Term Memory (LSTM) model trained on distinct household data. Afterward, a range of clustering algorithms utilizes the models hyperparameters as the input space, each representing the respective prosumer and they undergo evaluation leveraging well-established indexes as metrics. This method not only optimizes the individual LSTM models but also upholds privacy by ensuring that sensitive data remains decentralized throughout the process. The proposed solution aggregates the weights of locally trained LSTMs using the FedAvg algorithm, whilst integrating the differential privacy aggregator. Our solution underwent experimentation involving thirty energy prosumers, utilizing nearly four years worth of data. Various scenarios were chosen, encompassing both local learning and centralized learning methods.