A widely applicable and robust LightGBM - Artificial neural network forecasting model for short-term wind power density

Abstract

Wind energy is a clean and renewable source that reduces greenhouse gas emissions. To smooth the impact of wind energy fluctuations on the power grid and power supply, much research has predicted the wind speed of wind farms to estimate power generation. However, most studies overlook the nonlinear relationship between wind speed and power generation, and data sources are usually limited to one or two wind farms. This study constructs a wind power density prediction model based on the LightGBM and artificial neural network to solve the above problems. Its data collection process does not require meteorological measurement equipment and has good universality, stability, and robustness. LightGBM is used to extract feature information and then train it using an artificial neural network, which has a high tolerance for data loss. The model performance was validated using data from six terrains from 2020 to 2022. While results showed that the average prediction error was 71.68 % less than 2 % and 82.188 % less than 6 %, with an average R2 of 0.9755 and an average correlation coefficient of 0.9875, proving the practical significance of the model that can be used to guide electricity trade.

Keywords

• Wind power density
• Short-term
• Next hour forecasting
• LightGBM
• Random forest
• Artificial neural network