A data-driven multi-model methodology with deep feature selection for short-term wind forecasting. (15th March 2017)

• Record Type: Journal Article
• Authors: Feng, Cong; Cui, Mingjian; Hodge, Bri-Mathias; Zhang, Jie
• Is Part Of: Applied energy. Volume 190(2017)
• Page Start: 1245
• Page End: 1257
• Publication Date: 2017-03-15
• Subjects: Wind forecasting -- Machine learning -- Multi-model -- Data-driven -- Ensemble forecasting -- Feature selection; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2017.01.043
• Languages: English
• ISSNs: 0306-2619

Abstract

Graphical abstract: Highlights: An ensemble model is developed to produce both deterministic and probabilistic wind forecasts. A deep feature selection framework is developed to optimally determine the inputs to the forecasting methodology. The developed ensemble methodology has improved the forecasting accuracy by up to 30%. Abstract: With the growing wind penetration into the power system worldwide, improving wind power forecasting accuracy is becoming increasingly important to ensure continued economic and reliable power system operations. In this paper, a data-driven multi-model wind forecasting methodology is developed with a two-layer ensemble machine learning technique. The first layer is composed of multiple machine learning models that generate individual forecasts. A deep feature selection framework is developed to determine the most suitable inputs to the first layer machine learning models. Then, a blending algorithm is applied in the second layer to create an ensemble of the forecasts produced by first layer models and generate both deterministic and probabilistic forecasts. This two-layer model seeks to utilize the statistically different characteristics of each machine learning algorithm. A number of machine learning algorithms are selected and compared in both layers. This developed multi-model wind forecasting methodology is compared to several benchmarks. The effectiveness of the proposed methodology is evaluated to provide 1-hour-ahead wind speed forecasting at seven locations of the Surface Radiation network. Numerical results show that comparing to the single-algorithm models, the developed multi-model framework with deep feature selection procedure has improved the forecasting accuracy by up to 30%.