SELECTION OF PREDICTORS FOR PREDICTIVE MODELS OF THE AVERAGE DISCHARGE IN THE HYDROMETRIC SECTION OF THE OB RIVER NEAR BARNAUL DURING THE FLOOD PERIOD
A.V. IGNATOV
V.B. Sochava Institute of Geography, Siberian Branch, Russian Academy of Sciences, 664033, Irkutsk, ul. Ulan-Batorskaya, 1, Russia
ignatov@irigs.irk.ru
Keywords: формирование стока, стохастическая модель, операторы регрессии, условие выбора модели, компьютерный эксперимент, ансамблевый прогноз, runoff formation, stochastic model, regression operators, model selection condition, computer experiment, en semble forecast
Abstract
The problem of constructing a statistical predictive model and the related problem of selecting predictors for the forecasted variable are considered. This study is based on using data on the interannual variability in average discharge of the Ob in the hydrometric section of Barnaul during the flood period and on hydrometeorological characteristics having a potential influence on it. It is argued that the result from selecting predictors for the predictive model depends not only on the data used but also on the method employed in solving the problem. Such a method is determined by the mathematical operator used to approximate the simulated dependence, the optimality criterion of the model and the algorithm for selecting predictors. To study the influence of the modeling method on its result, a number of computer experiments were carried out, and each of them used different meth ods to find the best combination of predictors. It is shown that the best solution on the training sample is not always confirmed on independent data. To improve the sustainability of the simulation results, it is recommended that the criteria for selecting the optimal model should be used, which include assessments of its reliability. Use and comparison of different methods of construct ing models made it possible to identify the main predictors which explain most of the variance of the forecasted discharge. They are determined primarily by data on the object and have a physical interpretation. Their selection is less dependent on the method used to construct the model. The highest effectiveness was shown by the method of constructing the predictive model as an ensemble of partial models, each of which uses a limited number of non-intersecting predictors. Retaining the sustainability of the simulation result, make it possible to take into account, along with the main predictors, also the influence of secondary factors and, hence, to improve somewhat the quality of the predictive technique being developed.
|
