Reanalysis and Ground Station data: Advanced data preprocessing in deep learning for wind power prediction

• Published in: Applied energy 2024-12, Vol.375, p.124129, Article 124129
• Main Authors: Ding, Jun-Wei, Chuang, Ming-Ju, Tseng, Jing-Siou, Hsieh, I-Yun Lisa
• Format: Article
• Language: English
• Subjects: Component Kriging Interpolation; Deep Learning; Ground Station Data; Performance-Based Clustering; Reanalysis Data; Wind Power Forecasting
• ISSN: 0306-2619
• DOI: 10.1016/j.apenergy.2024.124129

Amidst the global transition to renewable energy, accurate wind power forecasting is becoming increasingly critical for grid integration. This study introduces a robust deep learning model that evaluates wind power generation using a novel approach that combines reanalysis and ground station data, aiming to improve prediction accuracy. To address potential discrepancies in raw data, preprocessing methods such as Boxplot and the DBSCAN algorithm are applied, effectively reducing outlier impact. A key innovation in our methodology is the component kriging interpolation, significantly enhancing the precision of wind speed and direction predictions. Moreover, we introduce two novel concepts: performance-based clustering and the integration of informational inputs. The clustering method organizes data by training performance, promoting a more efficient and accurate model training process. Informational inputs provide additional context, enabling the model to discern patterns across multiple wind farms, thereby improving generalizability. Employing a sophisticated four-layer Long Short-Term Memory (LSTM) network, coupled with a Multilayer Perceptron (MLP), the model uses a 48-hour lead time of combined meteorological and informational data to predict the capacity factor for the forthcoming hour. Our comprehensive testing across various wind farms yields an impressive R2 score of 0.95, demonstrating the model's exceptional predictive capabilities. This study not only advances wind power forecasting methods but also sets a new standard for the integration of diverse data sources in predictive modeling. The findings provide valuable insights for future research, potentially extending these innovative methodologies to broader applications in sustainable energy forecasting globally. •Developed a deep learning model integrating reanalysis and ground station data.•Introduced component Kriging interpolation to enhance wind power forecast accuracy.•Proposed performance-based clustering for refined predictive model training.•Demonstrated model efficacy with an R2 of 0.95 across multiple wind farms.•Advanced outlier removal methods significantly improved data quality.