Hybrid Machine Learning Models for Improved Short-Term Wind Forecasting

Wind's intermittent nature adds complexity to its analysis and its integration to electrical systems. So, accurate wind forecasting is crucial for effectively incorporating wind energy into power systems. This paper proposes an improved hybrid model for short-term wind prediction. It utilizes Empirical Mode Decomposition (EMD), a data-driven de-noising technique, to break down signals into more interpretable components. EMD separates nonlinear non-stationary data into Intrinsic Mode Functions (IMF) and a residue, which are then input to a Convolutional Neural Network. EMD is prone to mode mixing issues, where a single IMF contains drastically different scales or a signal of a certain scale is spread across multiple IMFs, leading to signal intermittency. To address this, Ensemble Empirical Mode Decomposition (EEMD) is introduced, incorporating white noise ensembles to establish a consistent reference frame in the time-frequency domain. This added noise helps consolidate similar signal portions within a single IMF. The IMFs and residues obtained from EMD and EEMD are fed into a Convolutional Neural Network and compared with benchmark models such as Bi LSTM and LSTM. Performance evaluation of these Machine Learning Algorithms is done and compared with various performance metrics such as MSE, MAE, RMSE, MAPE, and Loss and testing time. The hybrid model exhibited superior performance over other machine learning models in enhanced wind speed forecasting.