Wind Speed Forecasting toward El Nino Factors Using Recurrent Neural Networks

Abstract

Wind speed prediction is needed in various sectors, such as in industry. However, the weather conditions change all the time, so it is not easy to predict. The wind that blows is influenced by several climatic weather factors such as humidity and rainfall. Also, wind speed patterns change when there is a global climate change, El Nino. Therefore this research involved the element in predicting wind patterns one week ahead. Recurrent Neural Networks (RNN) and Long-short Term Memory (LSTM) methods are used for sequential data processing, such as climate data. Climate data for ten years used were wind speed, humidity, and rainfall provided by Meteorological, Climatological, and Geophysical Agency (BMKG) of a weather observation station, while Southern Oscillation Index (SOI), was obtained from the Australian Bureau of Meteorology (ABM). Data need to be pre-processed to solve missing data. Moreover, all variables were normalized and segmentation by overlapping to avoid data discontinuity. The results showed that the use of the amount of data, learning rate, epoch, and selection of the right optimization model could give good accuracy. The purpose of the proper configuration has a good performance, with accuracy reaching 88.34%. The results showed that the use of SOI factors improved correctness from 74.75% without SOI. The results also show that the model Adaptive Moment Estimation (Adam) provides better accuracy than the model Stochastic Gradient Descent (SGD), which gives an accuracy of only 71.84%. Meanwhile, the study also examined the effect of the learning rate and composition of training data and test data. The best accuracy is shown for the learning rate of 0.020 and 80:20% of training and test data comparison.