Abstract: Based on the observed soil moisture content data from 10, 20, and 40 cm soil layers at Tianxing and Pojiao stations in Yunnan Province, a prediction model (TVFEMD-BSLO/AO/IVYA/EGO/PSO-FLN) was constructed by improving the time-varying filter empirical mode decomposition (TVFEMD) and fast learning network (FLN) methods to enhance the time-series prediction accuracy of soil moisture content. By comparing the performance of different optimization algorithms, a superior modeling approach was provided for soil moisture prediction. The results showed that the TVFEMD decomposition performance was primarily influenced by two key parameters: Bandwidth threshold and B-spline order. Optimizing these two parameters using the IVYA algorithm improved the time-series decomposition quality and further enhanced the model’s prediction performance. The TVFEMD-BLSO/AO/IVYA/EGO-FLN model demonstrated outstanding prediction performance on the training set, with a mean absolute percentage error (MAPE) of 0.002%~0.077% and a coefficient of determination (R²) of 0.999 7~1.000 0. The MAPE in the prediction set was 0.006%~0.459%, and R² was 0.996 6~1.000 0. Compared with the TVFEMD-PSO-FLN model, the TVFEMD-BLSO/AO/IVYA/EGO-FLN model showed significant improvements in both fitting performance and prediction accuracy. Optimizing FLN hyperparameters using BLSO, AO, IVYA, and EGO algorithms effectively improved model performance, with the IVYA algorithm exhibiting the most notable optimization effect.

Key words: time-varying filter empirical mode decomposition (TVFEMD), algorithm optimization, fast learning network (FLN), soil moisture content, prediction