Accurate and efficient forecasted wind energy using selected temporal metrological variables and wind direction

Abstract

The aim of this work is to find the most efficient and suitable input features to be selected for forecasting monthly wind energy accurately. Machine learning is employed for a modular pipelined neural network, composed of time-delayed and feedforward networks with features of metrological variables such as atmospheric temperature, humidity, wind direction, and wind speed frequency distribution parameters. Logged data over a year’s period at a UAE site are analyzed on daily and monthly bases depending on their variation characteristics, in which standard Weibull probability distribution function is used for the feedforward neural network together with wind direction data, while daily average ambient temperature and humidity are attempted for the composite time delay networks. Different network abstractions of input features are compared, and it is found that wind direction data offer a better wind speed forecast. Wind energy is calculated based on monthly forecasting. A detailed adaptive probabilistic analysis is conducted to predict thresholds in variations of the forecast analysis. Error estimation tools are performed for adopting this method