Adaptive wind data normalization to improve the performance of forecasting models

Abstract

Wind speed forecasting, a time series problem, plays a vital role in estimating annual wind energy production in wind farms. Calculation of wind energy helps to maintain stability between electricity production and consumption. Deep learning models are used for predicting time series data. However, as wind speed is non-stationary and irregular, pre-processing of these data is necessary to get accurate results. In this paper, static normalization techniques like min–max, z-score, and adaptive normalization are used for pre-processing wind datasets, and further, their forecasting results are compared. Adaptive normalization increases the learning rate and gives better forecasting results than static normalization. The RMSE value was reduced by 9.18% for the NREL dataset when adaptive normalization was used instead of z-score normalization and by 23.58% for the Weather dataset. The datasets used are taken from National Renewable Energy Laboratory (NREL) and Kaggle’s Dataset.