A Hybrid Ultra-Short-Term and Short-Term Wind Speed Forecasting Method Based on CEEMDAN and GA-BPNN

Reliable ultra-short-term and short-term wind speed forecasting is pivotal for clean energy development and grid operation planning. During the wind forecasting process, decomposing the measured wind speed into data with different frequencies is a solution for overcoming the nonlinearity and the randomness of the natural wind. Existing forecasting methods, a hybrid method based on empirical mode decomposition and the back propagation neural network optimized by genetic algorithm (EMD-GA-BPNN), rely on partial decomposing the measured wind speed into data with different frequencies and subsequently achieving forecasting results from machine learning algorithms. However, such methods can roughly divide IMF signals in different frequency domains, but each frequency domain contains signals with multiple frequencies. The condition reflects that the method cannot fully distinguish wind speed into data with different frequencies and thus it compromises the forecasting accuracy. A complete decomposition of measured wind speed can reduce the complexity of machine learning algorithm, and has become a useful approach for precise simulations of wind speed. Here, we propose a novel hybrid method (CEEMDAN-GA-BPNN) based on the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) by completely decomposing the measured wind speed. The decomposition results are put into the back propagation neural network optimized by a genetic algorithm (GA-BPNN), and the final forecasting results are achieved by combining all the output values by GA-BPNN for each decomposition result from CEEMDAN. We benchmark the forecasting accuracy of the proposed hybrid method against EMD-GA-BPNN integrated by EMD and GA-BPNN. From a wind farm case in Yunnan Province, China, both for ultra-short-term forecasting (15 min) and short-term forecasting (1 h), the performance of the proposed method exceeds EMD-GA-BPNN in several criteria, including root-mean-square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R-2). The forecasting accuracy in decomposed components of low frequencies outperform components of high and middle frequencies. Fine improvement of the error metric (in percentage) in ultra-short-term/short-term forecasting is found by the complete decomposition method CEEMDAN-GA-BPNN: RMSE (7.0% and 8.6%), MAE (7.41% and 7.9%), MAPE (11.0% and 8.7%), and R-2 (2.2% and 11.0%), compared with the incomplete decomposing method EMD-GA-BPNN. Our result suggests that CEEMDAN-GA-BPNN could be an accurate wind speed forecasting tool for wind farms development and intelligent grid operations. Significance StatementNonlinearity and randomness of natural wind speed data are the limitations for short-term and ultra-short-term wind speed forecasting. By decreasing forecasting error in machine learning training process, data decomposition for the measured wind speed has become an effective method for overcoming this issue. Nonetheless, the normal incomplete decomposition method will compromise the extent of forecasting accuracy. We introduce a novel hybrid and complete decomposition method CEEMDAN-GA-BPNN (the complete decomposition method). Measured wind speed data from a wind farm in Yunnan Province, China, has been utilized. CEEMDAN-GA-BPNN outperforms EMD-GA-BPNN (the partial decomposition method) in forecasting accuracy both in the ultra-short-term and the short-term wind speed forecasting.