A hybrid structure of an extreme learning machine combined with feature selection, signal decomposition and parameter optimization for short-term wind speed forecasting

Abstract

Influenced by various environmental and meteorological factors, wind speed presents stochastic and unstable characteristics, which makes it difficult to forecast. To enhance the forecasting accuracy, this study contributes to short-term multi-step hybrid wind speed forecasting (WSF) models using wavelet packet decomposition (WPD), feature selection (FS) and an extreme learning machine (ELM) with parameter optimization. In the model, the WPD technique is applied to decompose the empirical wind speed data into different, relatively stable components to reduce the influence of the unstable characteristics of wind speed. A hybrid particle swarm optimization gravitational search algorithm (HPSOGSA) combining conventional PSOGSA with binary PSOGSA (BPSOGSA) is utilized to realize the FS and parameter optimization simultaneously. The PSOGSA is employed to tune the parameter combination of input weights and biases in ELM, while BPSOGSA is exploited to select the most suitable features from the candidate input variables determined by a partial autocorrelation function for reconstruction of the input matrix for ELM. The proposed forecasting strategy carries out multi-step short-term WSF using mean half-hour historical wind speed data collected from a wind farm situated in Anhui, China. To investigate the forecasting results of the hybrid model, a lot of comparisons and analyses are executed. Simulation results illustrate that the proposed WPD-ELM model with FS and parameter optimization can effectively catch the non-linear characteristics hidden in wind speed data and provide satisfactory WSF performance.

A Compound Wind Power Forecasting Strategy Based on Clustering, Two-Stage Decomposition, Parameter Optimization, and Optimal Combination of Multiple Machine Learning Approaches

“…Wavelet neural network is constructed based on feedforward back-propagation network (BPNN), namely, the transfer function, also named activation function, in hidden neurons nodes of BPNN is replaced by wavelet function [32][33][34] . The basic structure of WNN includes an input layer, hidden layer, and output layer.…”

“…The dimension numbers of input nodes, hidden nodes, and output nodes in ELM are set according to the works by the authors of [2,34], and the parameters in WNN are determined according to the work by the authors of [32]. The dimension in BSA-LSSSVM is set as 2.…”

“…In the further comparisons, the recently developed models including WPD-HPSOGSA-ELM [34], EEMD-HGSA-LSSVM [39], and TSD-HBSA-DAWNN [42] are constructed to illustrate the effectiveness of the proposed model. The forecasting results and their improved indexes are listed in the Tables 7 and 8, respectively.…”

A Multistep Wind Speed Forecasting System Considering Double Time Series Features

“…In [15], the author pioneered a model combining a wavelet soft threshold denoising algorithm and deep learning method and made some improvements to the learning algorithm. In [16], Sun et al proposed a prediction model combining wavelet packet decomposition, feature selection, and a parameteroptimized extreme learning machine. They implemented parameter optimization with a hybrid particle swarm optimization gravity search algorithm.…”

Multi-Objective Optimal Dispatching for a Grid-Connected Micro-Grid Considering Wind Power Forecasting Probability

“…[6], wavelet decomposition signal processing method was used to decompose the wind speed time series, then the decomposed results were employed as the inputs of artificial neural network for wind speed forecasting, then, adaptive probabilistic concept of confidence interval approach was applied to address the uncertainty problems of wind power. Most of WPF approaches yield deterministic forecasting [8], [9], no matter how much the forecasting accuracy of wind power deterministic prediction model is, there still exist some uncertainty components, thus cannot provide the uncertainty information about wind power [10], [11]. Thus, in recent years, wind power probabilistic forecasting models have been developed to provide uncertainty information for studying the optimal energy management [12], [13].…”