A multiple-output hybrid wind speed prediction model with accuracy self-assisted module

In recent years, wind power has emerged as a prominent renewable energy source, and the need for reliable wind speed prediction models has become paramount to ensure smooth and predictable wind power supply. This study proposes an accuracy self-assisted projection model that can forecast wind speed for the next 24 hours with a 6-hour forecast output step. The model building process commences with a random sampling approach applied to the wind speed dataset for dividing the training, validation, and test sets. The CEEMDAN (complete ensemble empirical mode decomposition with adaptive noise) method is then utilized to decompose the wind speed signal into IMFs (intrinsic mode functions) that are fed into the short-term forecasting module. The forecast results from the short-term forecast module are processed and fed back into the long-term forecast model as part of the input tensor. Controlled experiments and validations demonstrate that: (a) The random sampling approach for dataset partitioning is effective in avoiding seasonality effects; (b) The short-term prediction model output can assist the long-term prediction in signal extension and tensor fusion aspects; and (c) The transfer learning approach is effective in reducing computational and time costs in training multiple sub-models. The proposed model focuses exclusively on wind speed prediction; future extensions may integrate wind direction forecasting to enhance comprehensive wind energy management.