Multivariate wind speed forecasting based on multi-objective feature selection approach and hybrid deep learning model

Abstract

This study proposes an effective model for enhancing the short-term wind speed forecasting performance by considering the effect of multiple meteorological factors. (a) The filter-wrapper non-dominated sorting differential evolution algorithm incorporating K-medoid clustering (FWNSDEC) is designed to select key meteorological factors and generate multiple feature subsets. For each feature subset, the hybrid deep learning model is designed: (b) singular spectrum analysis (SSA) is used to decompose the meteorological factors and construct the three-dimensional input structure; (c) convolutional long short-term memory (ConvLSTM) network is then adopted to process the sample set of three-dimensional sequence, and the final forecasting result is the average prediction of all the built ConvLSTMs. To evaluate the effectiveness of FWNSDEC-SSA-ConvLSTM, three comparative experiments on four datasets collected from the National Renewable Energy Laboratory are implemented. Experiment results show that the average mean absolute percentage error over four datasets for 1-step-ahead, 2-step-ahead, and 3-step-ahead forecasting is 1.42%, 1.99%, and 3.28%, respectively, which are much better than the predictions using feature selection benchmarks, hybrid forecasting benchmarks with different deep learning networks and data decomposition methods, and other advanced forecasting systems. Extended model discussion in terms of Friedman test and parameters sensitivity analysis further verifies the potential of proposed model.