Short-Term Wind Characteristics Forecasting Using Stacked LSTM Networks

Abstract

Onshore/offshore wind turbines play a vital role in addressing the increasing worldwide energy demand. Enhancing the wind power harnessing capability of turbines and extending the life expectancy of their components support further reductions in the final cost of wind energy. Data-driven techniques can complement existing physics-based approaches for complex problems such as wind farm wake modeling. In this paper, a deep learning model is developed to predict the local short-term wind characteristics. A data pre-processing pipeline that includes data cleaning and normalizing steps is developed to generate the training dataset. Time-series forecasting models based on long-short-term-memory (LSTM) and convLSTM are developed and trained for local short-term wind forecasting. The model is validated through experiments on three-year data from the National Renewable Energy Laboratory (NREL) database. The conducted experiments showed favorable performance based on root mean square error (RMSE) and R2 test scores. The R2 values for predicting 1-minute, 30-minute, and 1 hour, wind characteristics for both LSTM and convLSTM were above 0.92. The results are in agreement with the literature. They also demonstrate the effectiveness of the developed models for short-term wind forecasting compared to similar ones.