Attention based Transformer coupled with convoluted neural network for ultra-short- and short-term forecasting of multiple wind farms

Abstract

ABSTRACT This research paper evaluates machine learning (ML) algorithms for wind power forecasting using three univariate data from wind farms over various periods and two different time intervals. The goal is to assess the accuracy and robustness of predictive models generated with larger datasets for ultra-short-term and short-term forecasting. A novel ML algorithm, the Transformer coupled with Convoluted Neural Network (CNN) based on attention mechanisms, is proposed and compared with existing models using the same datasets. The study utilizes three different time periods for ultra-short-term and short-term forecasting. The Maximal Information Coefficient (MIC) methodology is employed to optimize the number of inputs, aiding in identifying linear and nonlinear relationships between input and output parameters. The results demonstrate that the Transformer algorithm produces the most accurate forecasting. The combination of Transformer and CNN with attention mechanisms has proven effective in capturing complex patterns and dependencies in the wind power data, leading to superior forecasting accuracy. Using multiple datasets from different wind farms and time periods further strengthens the reliability and generalization capabilities of the proposed algorithm. These findings contribute to the advancement of ML-based wind power forecasting techniques and have significant implications for integrating renewable energy sources into the power grid.