Drought Prediction Method Based on an Improved CEEMDAN-QR-BL Model

In this study, a new broad learning (BL) model based on an improved complete ensemble empirical mode decomposition adaptive noise (CEEMDAN) is proposed to resolve the low accuracy, poor robustness, and long delay problems that are present in current drought assessments. First, the extreme delay method was applied to improve the CEEMDAN end effect. The improved CEEMDAN method was then used to decompose a series of non-steady-state signals from drought monitoring into multiple steady-state components. A BL model based on orthogonal trigonometry (QR) was then used to predict these multiple steady-state components, and the predicted components were further reorganised to obtain a high-precision drought sequence. On this basis, CEEMDAN was introduced into the orthogonal triangular broad learning (QR-BL), and a drought prediction model (CEEMDAN-QR-BL) combining CEEMDAN and QR-BL was proposed. Finally, the De Martonne aridity index was used to calculate the drought sequence results and determine the drought grades. To meet the real-time requirements of drought prediction, parallel computing was introduced into the CEEMDAN-QR-BL model, and a drought prediction method based on parallel CEEMDAN-QR-BL was constructed. The experimental results show that, when compared with a support vector regression model combined with an empirical mode decomposition, the reliability and accuracy of the CEEMDAN-QR-BL increases by 29.57% and 11.84%, respectively. In addition, when compared with only BL, the prediction efficiency of QR-BL improved by 62.29%.