Optimized Forecasting Model to Improve the Accuracy of Very Short-Term Wind Power Prediction

Abstract

This article proposes a novel framework to improve the prediction accuracy of very short-term (5-min) wind power generation. The framework consists of complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), monarch butterfly optimization (MBO) and long short-term memory (LSTM), called CEMOLS. The CEEMDAN is employed to extract complex hidden features of time-series data into intrinsic mode functions that are predicted using LSTM models with dropout regularization to retain long-term relationships between input and output data, while the optimization algorithm tunes the hyperparameters of the forecasting model. Data from four real wind farms in New South Wales are collected and preprocessed to train and test the forecasting models. Recently developed rival models are compared to identify the best-performing prediction model. The analysis demonstrates that the proposed CEMOLS with low computation time can improve forecasting accuracy on average by 32.96% in mean absolute error, 47.10% in root mean square error and 32.33% in mean absolute percentage error as compared to the benchmark Persistence model. It also demonstrates that sensitive and statistical analysis needs to be carried out to determine robust prediction models among rival models for practical application.