A solar forecasting framework based on federated learning and distributed computing. (November 2022)

• Record Type: Journal Article
• Authors: Wen, Haoran; Du, Yang; Lim, Eng Gee; Wen, Huiqing; Yan, Ke; Li, Xingshuo; Jiang, Lin
• Is Part Of: Building and environment. Volume 225(2022)
• Publication Date: 2022-11
• Subjects: Solar forecasting -- Deep learning -- Federated learning -- Distributed computing -- Sustainable environment; Buildings -- Environmental engineering -- Periodicals; Building -- Research -- Periodicals; Constructions -- Technique de l'environnement -- Périodiques; Electronic journals; 696
• DOI: 10.1016/j.buildenv.2022.109556
• Languages: English
• ISSNs: 0360-1323

Abstract

Abstract: Solar forecasting is a crucial and cost-effective tool for better utilization of solar energy for smart environment design. Artificial intelligence (AI) technologies, such as machine learning (ML) and deep learning (DL), have gained great popularity and been widely applied in solar forecasting in recent years. However, conventional AI-based forecasting methods suffer from high variability and stochasticity of solar irradiation, making unreliable predictions due to heterogeneous solar resources. Moreover, the training process of DL models is less flexible and requires immense data. Even for a well-trained model, it can still yield deteriorated performances on other datasets of varying data distributions. To tackle the deficiencies of AI forecasting models, we present a flexible distributed solar forecasting framework based on a novel spatial and temporal attention-based neural network (STANN) in conjunction with federated learning (FL) technique, considering multi-horizon forecasting scenario from 5–30 min. The STANN model consists of a feature extractor and a forecaster, which can be respectively trained on various local datasets for better localization, and updated to further improve forecasting accuracy through global parameter aggregation under the proposed framework without data gathering. We evaluate effectiveness of the proposed method by conducting extensive experiments on real-world datasets and compare it to other popular forecasting models. The results demonstrate that our approach outperforms the other benchmarks with higher forecasting accuracy for all forecast horizons and better generalization on various datasets, achieving the highest forecast skill of 28.83% at 30 min horizon and an improvement of 11.2% compared with the centralized, localized, and conventional FL training methods. Highlights: Data heterogeneity affects solar forecasting performance at different locations. Spatial and temporal attentions improve both feature extraction and forecasting accuracy. The proposed framework localizes the feature extractors, while collaboratively training the forecasters in a federated way. Ten real-world datasets collected at different locations are rendered for training. Four training strategies are compared and the proposed one outperforms the others.