Wind speed forecasting using a new multi-factor fusion and multi-resolution ensemble model with real-time decomposition and adaptive error correction

Abstract

The chaos and non-stability of the wind pose challenges for obtaining reliable wind speed forecasting. Accurate and reliable wind speed forecasting is critical to the efficient and safe operation of wind power systems. In the study, a new hybrid multi-factor fusion and multi-resolution ensemble model is proposed for forecasting wind speed. The proposed hybrid forecasting model adopts real-time data decomposition and adaptive multiple error correction strategies and it consists of three phases. In stage I, the low-resolution and high-resolution data obtained by the averaging method and the Staked Auto-Encoder (SAE) feature extraction method, respectively. The data are used by Bidirectional Long Short-Term Memory (BiLSTM) for multi-step forecasting. The forecasting results are reasonably ensemble by Non-dominated Sorting Genetic Algorithm III (NSGA-III) to complete the multi-resolution ensemble phase. In stage II, Ljung-Box Q-Test (LBQ-test) is utilized to detect predictable components in the forecasting results. And Auto-Regression Moving Averaging (ARMA) completes adaptive multiple error corrections. In stage III, Multi-universe Optimization (MVO) is used to ensemble the wind speed forecasting results of multiple positions. At this point, the multi-position data fusion phase is completed. Wind speed data from Xinjiang, China are carried out to validate the efficiency of the proposed model. Experimental analysis shows: The performance of the proposed hybrid model is superior to other comparative models. The MAE error of 3-step wind speed forecasting from the four sets of experiments is only 0.4614 m/s, 0.4482 m/s, 0.6755 m/s, and 0.2912 m/s, respectively.