Federated transfer learning with orchard-optimized Conv-SGRU: A novel approach to secure and accurate photovoltaic power forecasting
Peer reviewed, Journal article
MetadataShow full item record
Original versionSalman Bukhari, S. M., Raza Moosavi, S. K., Zafar, M. H., Mansoor, M., Mohyuddin, H., Sajid Ullah, S., Alroobaea, R. & Sanfilippo, F. (2023). Federated transfer learning with orchard-optimized Conv-SGRU: A novel approach to secure and accurate photovoltaic power forecasting, 48, Article 100520. https://doi.org/10.1016/j.ref.2023.100520
Accurate photovoltaic (PV) power forecasting is pivotal for optimizing the integration of RES into the grid and guaranteeing proficient energy management. Concurrently, the sensitive nature of data obtained from individual PV systems underscores paramount concerns regarding data privacy and security. In this manuscript, we introduce an innovative approach for PV power forecasting that addresses these concerns, deploying federated learning (FL) combined with TL. This is orchestrated via a hybrid deep learning model, denominated as Federated transfer learning (TL) Convolutional Neural Network with Stacked Gated Recurrent Unit (FL-TL-Conv-SGRU). To optimize the performance of the Conv-SGRU model, we employ the OA for hyperparameter tuning, a novel bio-inspired technique inspired by orchard gardening practices. This algorithm presents a distinctive interplay between exploration and exploitation in the hyperparameter space, potentially elevating the model’s performance. Our exposition covers eight disparate datasets from PV systems, which are judiciously split into two cohorts, safeguarding data privacy. Through the prism of FL, we ensure data security by orchestratively training the Conv-SGRU model over distributed datasets. This strategy allows tapping into the shared wisdom across the datasets, all the while ascertaining individual data remains localized, boosting model generalization and predictive prowess. Additionally, TL is invoked to benefit from pre-trained feature representations, facilitating effective knowledge transmission across diverse PV setups with unique characteristics and locales. The put-forth FL-TL-Conv-SGRU design amalgamates the essence of FL, TL, convolutional neural networks, and stacked gated recurrent units. This ensemble aids in deciphering spatial–temporal intricacies intrinsic to PV power generation. Through empirical analyses, we evince that our FL-TL-Conv-SGRU model transcends conventional forecasting paradigms, emphasizing its adeptness in delivering meticulous forecasts over a range of PV installations. Our results accentuate the bifurcated importance of the federated TL framework: a capability for collaborative training with an unwavering commitment to data privacy, and a proficiency in exploiting decentralized data. This strategy is particularly salient given the shifting regulatory milieu centered on data safeguarding and confidentiality. As we transition towards a world more reliant on renewable energy, our proposed stratagem promises to be a cornerstone for efficient, sustainable energy management, heralding a future replete with green energy.