Short Term Wind Energy Forecasting Using Sample Entropy Based Decomposition and Deep Neural Network

Wind suffers from the curse of intermittency. This intermittency gives wind energy a non stationary nature which makes correct wind power forecasting a challenge.In this work, two decomposition algorithms namely Ensemble Empirical Mode Decomposition (EEMD, recursive and noise assisted) and Variational Mode Decomposition (VMD,non recursive and noise free) are employed to make the wind time series more stationary before feeding it into the network for predictions. This decomposition tends to increase the computational burden and hence a sample entropy based Intrinsic Mode Function (IMF) selection process is carried out which filters the unwanted IMFs. Sample entropy based approach also helps in determining the appropriate number of decomposition’s when using VMD algorithm. Hyper parameter optimization is achieved via genetic algorithm. In order to better quantity the uncertainty associated with wind power, prediction intervals are constructed around every point prediction by using a simple Gaussian distribution approach. Comparison of various deep learning models demonstrating the effect of IMF selection approach is then given using two different wind power data sets from different geographical locations.The results show that sample entropy based selected IMFs in combination with hybrid models show better forecasting results with decreased computational load.