Prediction of the total solar irradiance based on the CEEMDAN-BiGRU-Attention model

Total solar irradiance (TSI) data play an important role in guiding production, but they are difficult to predict accurately because of nonlinearity and noise. In this paper, a hybrid model of complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and a bidirectional gated recurrent units network (BiGRU) that introduces an attention mechanism into adaptive noise is proposed. First, the CEEMDAN method is used to decompose the TSI. Then, the fuzzy entropy (FE) of each component is calculated, and the input sequence is synthesized by K-means clustering of the near waves. Finally, the combined sequences are trained in a BiGRU model with multiple hidden layers, and the attention mechanism acts on the time step. In this paper, the TSI at 5485 time points was predicted with MAE, MSE and MAPE values of 0.1122, 0.0259, and 0.00008, respectively. Compared with BiGRU, the hybrid model decreased by 11.74% on MAE, 5.29% on MSE and 0.009% on MAPRE. The CEEMDAN-BiGRU-Attention model has good application prospects in the field of TSI prediction.