Development of Long Short-Term Memory (LSTM) Bayesian Network Method for Predicting Wind Power Potential in a Wind Power Plant in Indonesia

Abstract

The need for renewable energy has increased recently, along with the shortage of non-renewable energy sources such as petroleum, coal, uranium, crude oil, and others. One of the renewable energies whose technology has recently been developing is wind power; however, it still suffers from a drawback due to the fluctuations in energy production. Increasing wind energy potential requires a wind power prediction method that can predict the intermittent patterns of the prediction result from the generated wind power. In dealing with the frequent intermittent patterns that fluctuate frequently and have many variations, the Triple Exponential Smoothing Multiplicative LSTM (TES-MLSTM) model can read them and then predict with a short term few steps ahead. In this paper, LSTM Bayesian Network as another deep learning method is proposed and compared with the TES-MLSTM. This method uses the same LSTM base, enhanced with its hyperparameter tuning and run in a Bayesian Network. The model parameters are learned from the training data, and hyperparameters are tuned to get the best fit. The tuned hyperparameter will be processed using Bayesian Network. In the experiment, we used the 2013 dataset of Pandansimo wind power plant (PLTB) in Indonesia as the input data. The average wind power prediction errors (MSE) using the TES-MLSTM and LSTM Bayesian Network are 0.891 and 0.644, respectively. It can be concluded that the proposed LSTM Bayesian Network method is more accurate in predicting the wind power potential of a wind turbine than the TES-MLSTM method.