Wind farm power forecasting by physical data fusion

Power forecasting is a fundamental research topic in the wind power industry. Existing wind power forecasting methods predominantly rely on either data-driven or physics-driven approaches, with few studies combining physical models and data-driven techniques despite their significant complementary potential. A data-driven model is established using K-means clustering, empirical mode decomposition, and parallel weighted long short-term memory networks. A novel integrated approach combining physics-driven and data-driven methods was developed for wind farm forecasting. Validation using real-world data from a Chinese wind farm demonstrated that the proposed integrated method achieved 21.67% higher prediction accuracy than data-driven methods and 35.17% higher accuracy than physics-driven methods. These results confirm the superiority and reliability of physics-data fusion methods in wind farm ultra-short power forecasting.