Reconstruction of Runoff Series Under Variation

Without giving consideration to series variation, commonly used reconstruction methods may lead to series over-fitting or bring about too much uncertainty. Under the condition of abrupt changes and on the basis of Bayesian Decision and Generalized Linear Model, this paper puts forward a reconstructed model with the aid of empirical mode decomposition (EMD), Generalized Linear Regression, and support vector machine (SVM). Firstly, the model identifies the variation points of the series with Bayesian Decision to divide the series into segments. Secondly, EMD is performed on the series segment by segment. Finally, modeling and simulation are conducted by SVM and optimal subset regression. Taking the runoff from 1951 to 2010 in Huaxian Station of Weihe River as an example, the author acquires the EMD components and frequencies of the annual runoffs in the three stages (1951-1971, 1972-1984, and 1994-2010) before/after the 3 variation points (1971, 1984 and 1994), and reconstructs the historical runoff data and uses the SVR to predict each component of the EMD. The segmental optimal subset regression features small prediction error and high precision because each IMF component produced in EMD decomposition has strong regularity.