Very Short-Term Wind Power Forecasting Using a Hybrid LSTM-Markov Model Based on Corrected Wind Speed

Abstract

A Markov chain (MC) model is a statistical method of predicting future outcomes using past experience. This study proposes a hybrid method that uses a long short-term memory (LSTM) and a MC method to produce very accurate short-term (10-min) forecasts for the power output from a wind turbine (WT). The proposed method has three stages. The first stage uses k-means clustering to partition the wind power data into several clusters. The second stage uses LSTM models to initially predict the wind power output for each cluster. The final stage uses a MC method to construct the transition probability matrix for every 10-mimute time period. Using the transition probability matrices, the final predicted value for the WT power output is estimated using the prediction results for each cluster in the LSTM. This article also suggests a wind speed correction approach to enhance the forecasted wind speed result achieved by applying the weather research and forecasting model in order to generate more accurate wind power forecasting results. The proposed method is tested using a 3.6 MW WT power generation system that is located in Changhua, Taiwan. The effectiveness of the proposed model is compared with support vector regression (SVR), random forest (RF), LSTM and bidirectional gated recurrent unit (Bi-GRU) methods.