Attention Based Short Term Wind Power Forecasting Using Sample Entropy And Correlation Analysis for Noise filtration

Abstract

Wind power forecasting is one of the emerging research topics. Due to the curse of intermittency, wind power is tough to study and predict. High precision forecasting of wind is a present urge in order to guarantee smooth operations and planning in power systems. Wind power depends on more than one factor and thus multivariate time series forecasting is the only way by which an attempt can be made to understand wind patterns completely. However, two big challenges that are associated with wind Multivariate time series forecasting (MTFS) is how to reduce the influence of noise components existing in wind and how to select features effectively. This work aims to give the answer to both these questions. To filter out noise and reduce its impact on wind series, Complete Ensemble Empirical Mode Decomposition using Adaptive Noise (CEEMDAN) decomposition techniques has been used. By using the recent breakthrough in deep learning, authors use Attention based approach to select the right feature. The entire method consists of three stages namely decomposition, prediction and ensemble. In the decomposition stage, sub series are selected based on their sample entropy and further correlation analysis is done on them. In the prediction stage, Genetic algorithm is applied to search for the optimal values of hyper parameters. Experiments show that sample entropy based Intrinsic Mode Function (IMF) selection in combination with further correlation analysis helps filter out noise to some extent. Attention based approach in combination with decomposition method performs better over non attention based models in terms of RMSE and MAE scores. However attention based approach is computationally heavy and so an increase in accuracy is obtained at the cost of computational efficiency.