Hybrid Forecasting Model for Short-Term Wind Power Prediction Using Modified Long Short-Term Memory

Abstract

Renewable energy has recently gained considerable attention. In particular, the interest in wind energy is rapidly growing globally. However, the characteristics of instability and volatility in wind energy systems also affect power systems significantly. To address these issues, many studies have been carried out to predict wind speed and power. Methods of predicting wind energy are divided into four categories: physical methods, statistical methods, artificial intelligence methods, and hybrid methods. In this study, we proposed a hybrid model using modified LSTM (Long short-term Memory) to predict short-term wind power. The data adopted by modified LSTM use the current observation data (wind power, wind direction, and wind speed) rather than previous data, which are prediction factors of wind power. The performance of modified LSTM was compared among four multivariate models, which are derived from combining the current observation data. Among multivariable models, the proposed hybrid method showed good performance in the initial stage with Model 1 (wind power) and excellent performance in the middle to late stages with Model 3 (wind power, wind speed) in the estimation of short-term wind power. The experiment results showed that the proposed model is more robust and accurate in forecasting short-term wind power than the other models.

Support Vector Machine and Long Short-term Memory using Multivariate Models for Wind Power Forecasting

“…However, the most important factors in short-term wind power forecasting are the use of certain variables (e.g., wind power, wind direction, wind speed, and meteorological factors) and the application of learning algorithms (e.g., NWP, statistical algorithms, autoregressive integrated moving average, machine learning, and a deep neural network). Previous studies have considered two factors in short-term wind forecasting [11,12]. The present study compares the performances of existing research.…”

“…In [12], four multivariate models for wind power forecasting are used. M1 indicates the wind power, M2 indicates the wind power and direction, M3 is the wind power and speed, and finally M4 is the wind power, speed, and direction.…”

“…At this time, the learning and test data are all normalized, and the collection period, collection time, learning data, test data, and total data are applied for each region, as shown in Table 1. Table 2 shows the difference in performance between the SVM using a multivariate model [11] and LSTM using the hybrid model [12] for short-term wind power forecasting in areas A, B, and C of Jeju Island. In calculation complex, the SVM using a multivariate model [11] takes longer to compute than the LSTM using the hybrid model [12] because it uses k-means clustering to extract the wind characteristics of the wind direction and speed.…”

“…With the RMSE and MAD, the SVM using a multivariate model [11] seems to outperform the LSTM using a hybrid model [12]. With the RMSE, the SVM using a multivariate model is superior to LSTM using hybrid model by 0.6, 2.0, and 3.0 per region.…”

“…With the RMSE, the SVM using a multivariate model is superior to LSTM using hybrid model by 0.6, 2.0, and 3.0 per region. Specifically, the performances of the SVM using a multivariate model [11] and LSTM using the hybrid model [12] are not very different in region A, but the difference in performance is large in regions B and C. Figs. 7 and 9 show the performance comparison of the SVM using a multivariate model and LSTM using the hybrid model for the observed value.…”