Comparison of Lavenberg-Marquardt, Scaled Conjugate Gradient and Bayesian Regularization Backpropagation Algorithms for Multistep Ahead Wind Speed Forecasting Using Multilayer Perceptron Feedforward Neural Network

Abstract: Wind speed forecasting is critical for wind energy conversion systems since it greatly influences the issues such as scheduling of the power systems, and dynamic control of the wind turbines. Also, it plays an essential role for siting, sizing and improving the efficiency of wind power generation systems. Due to volatile and non-stationary nature of wind speed time series, wind speed forecasting has been proven to be a challenging task that requires adamant care and caution. There are several state-of-the-art methods, i.e., numerical weather prediction (NWP), statistical, and hybrid models, developed for this purpose. Recent studies show that artificial neural networks (ANNs) are also capable of wind speed forecasting to a great extent.  In this paper, 3-layer perceptron feedforward neural network is employed for comparison of three different training algorithms, i.e., Lavenberg-Marquardt (LM), Scaled Conjugate Gradient (SCG) and Bayesian Regularization (BR) backpropagation algorithms, in the view of their ability to perform 12 multistep ahead monthly wind speed forecasting. Horizontal wind speed, absolute air temperature, atmospheric pressure and relative humidity data collected between November 1995 - June 2003 and July 2007 – April 2015 for city of Roskilde, Denmark is used for training, validation and testing of the network model. The performed experiment shows that for 12 multistep ahead wind speed forecasting, SCG algorithm has obvious preference in terms of prediction accuracy with mean absolute percentage error (MAPE) of 3.717%, followed by LM and BR algorithms with MAPE of 4.311% and 4.587% accordingly. As a result, within the scope of this study, SCG algorithm is found to be more suitable to build a multistep ahead wind speed forecasting model.