1. Introduction
Recently, renewable energy is increasingly being discussed with the aim of phasing out coal power and nuclear power generation to address changes in the internal and external conditions, such as new climate system, Fukushima nuclear power plant accident, increasing frequency of earthquakes, and serious air pollution. Korea is announcing the ’Renewable Energy 3020’ implementation plan, which aims to achieve a 20% share of renewable energy by 2030, and is expanding the capacity of wind power facilities from 1.2 GW in 2017 to 17.7 GW in 2030. Among different sources of renewable energy, wind energy, in which wind power is converted into electric power using a wind turbine [
1], has gained particular interest. However, the output of wind power varies through time and space due to various factors, such as wind speed, wind direction, and temperature. Therefore, when large-scale wind power is linked to the grid, instability of the grid arises, and effective connection of the grid becomes difficult. In other words, the maintenance of the generator, generator shutdown plan, economic dispatch, power supply plan (generator, transmission line, etc.), and electric power market bidding become important factors when wind energy is the central supply and demand generator in the power system. For this reason, it is very important to predict the output of the wind turbine considering the stability of the wind generator system, for which various studies have been conducted recently [
2,
3,
4].
Significant amounts of research have been conducted to predict wind power in the short, medium, or long term, and the corresponding methods can be mainly divided into three categories: physical methods, data-driven methods (statistical and artificial intelligence methods), and hybrid methods [
5].
(1) Physical methods predict wind power using mathematical modeling, considering weather data (air pressure map, jet stream, etc.) and environmental characteristics (temperature, humidity, topography, land use, etc.) [
6]. These methods require a large amount of data because the construction of the prediction model is complicated due to the mathematical modeling of many variables and the accuracy of prediction increases in proportion to the amount of data. Therefore, the mathematical structure of the predictive model is very difficult to achieve, the calculation process is complicated, and the calculation time is lengthy. The most widely used method is the numerical weather prediction (NWP) [
7]. The NWP is suitable for long-term prediction rather than short-term and medium-term prediction because of the large amount of computation.
(2) Data-driven methods predict the wind speed within a few hours through pattern analysis, which is trained based on past data. They are more suitable for predicting wind speed values within a short period of time, as the amount of past data is continuously increasing. These methods are divided into statistical methods and artificial intelligence methods. The former can achieve high accuracy in short-term prediction, but they have a disadvantage in that they cannot accurately predict the wind power due to accumulation of errors in long-term prediction. For example, among statistical methods, the auto-regression integrated moving average (ARIMA) model is a time series analysis model [
8], and various approaches have been suggested to overcome the disadvantage of ARIMA [
9,
10] which cannot accurately predict wind power due to the aforementioned error accumulation. The latter include neural network (NN) [
11], fuzzy inference [
12], particle swarm optimization (PSO) [
13], genetic algorithm (GA) [
14], support vector machine (SVM) [
15], and long short-term memory (LSTM) [
16]. The artificial intelligence method has superior performance for general purpose, but it has a disadvantage in that the relationship between model elements cannot be accurately explained.
(3) Hybrid methods employ the approach of prediction by applying statistical prediction after acquiring weather forecast data through physical methods and combining physical and statistical methods. In other words, existing physical methods are more suitable for long-term prediction than local prediction because of their large scale. Data-driven methods feature high accuracy in short-term prediction but accumulate errors in long-term prediction. Therefore, existing methods suffer from a large amount of errors, and other research is under way to improve the prediction method [
17,
18,
19,
20,
21]. Hybrid methods can be categorized into four types [
22,
23,
24]: weight-based combined approach [
25,
26,
27], data pre-processing technique based combined approach [
28,
29,
30,
31,
32], parameter selection and optimization technique based combined approach [
33,
34,
35,
36], and data post-processing technique based combined approach [
37]. Firstly, the strategy of the weight-based combined approach is to assign factors to models according to their performance. It is simple and easy to implement, suitable for a wide range of prediction times, and has the advantage of adapting to new data. However, it does not guarantee the best prediction along the prediction horizon, and has a disadvantage in that an additional model is necessary for determining the weight. Secondly, the purpose of the data pre-processing technique based combined approach is to forecast the subseries obtained by decomposition. The advantage of this method is that the performance is better than that of the abovementioned approaches, and it is robust against sudden changes of wind speed. However, it requires detailed mathematical knowledge of the decomposition model and has a disadvantage of slow response time to new data. Thirdly, the purpose of the parameter selection and optimization technique based combined approach is to optimize the parameters of the forecasting model. The parameters adopted for predicting wind power are meteorological factors such as temperature, humidity, precipitation, snowfall, cloud, sunshine, wind speed, and wind direction. This method features easier determination of parameters with a relatively basic structure than the two methods mentioned above. However, it is difficult to write and implement according to the knowledge of the designer. Finally, the data post-processing technique based combined approach forecasts residual errors caused by the forecasting model. Since this method considers residual errors from the model, it can provide more accurate predictions than the abovementioned three methods. However, it has a disadvantage in that the calculation time is lengthy because residual errors must be calculated.
This study adopts an approach similar to the third type of hybrid method. In other words, the input data required for wind power prediction are classified by models, which consist of multivariate models and the performance of each model is measured. We selected wind power, wind speed, and wind direction among the meteorological factors that have the greatest impact on the prediction of short-term wind power according to the characteristics of wind farms. For performance measurement, a modified LSTM was applied to the artificial intelligence technique. We analyzed the performance of the modified LSTM by model and then combined the superior models to predict the optimum wind power. Experimental data for verifying the proposed model were acquired from wind farms in Jeju Island, South Korea. The reason for selecting Jeju Island is to promote 2 GW offshore wind power for realizing ’Carbon free island Jeju by 2030’. Moreover, Jeju Island, which has diverse winds due to Jeju Island’s characteristics, has the best conditions for the verification test. Finally, in order to predict the short-term wind power efficiently, LSTM was designed using MATLAB and the prediction error was verified through root mean square error (RMSE) and mean absolute percentage error (MAPE).
The remainder of the paper is structured as follows.
Section 2 describes related works on recurrent neural network (RNN) and LSTM.
Section 3 explains traditional LSTM problems and solutions, proposing a modified LSTM as well as data sets and multivariate models.
Section 4 analyzes the experimental environment and results for wind farms A, B, and C in Jeju Island. Finally,
Section 5 presents conclusions and future research scope.