AGU indexes: 1860, 1869 STATISTICAL METHODS FOR RIVER’S RUNOFF FORECAST Pisarenko V.F. 1 , Lyubushin A.A. 2 , Canu S. 3 , Kanevsky M.F. 4 , Savelieva E.A. 4 , Demianov V.V. 4 , Bolgov M.V. 5 , Rukavishnikova T.A. 1 , Zaliapin I.V. 1 1 International Institute of Earthquake Prediction Theory and Mathematical Geophysics, RAS, Moscow; 2 Institute of Physics of the Earth, RAS, Moscow; 3 Institut National des Sciences Appliquees, Rouen, France; 4 Nuclear Safety Institute, RAS, Moscow; 5 Water Problem Institute, RAS, Moscow. ABSTRACT Statistical methods of data analysis are applied to monthly runoff values of rivers. The analysis is based on the theory of periodically correlated random processes. This approach takes into account both deterministic and stochastic components of seasonal variations of runoff data. Prediction formulae of runoff value one month ahead, based on 1-12 preceding monthly runoff values, are derived. This method was applied to 9 rivers with different hydrological regimes. The efficiency of prediction, which can be defined as ratio of residual variance to initial variance of runoff, varies from river to river in the range 0.1÷0.6. The suggested method of prediction was compared with Artificial Neural Networks. Factor analysis of correlation matrices of deviations of rivers runoff from the cyclic average allowed extracting the maximum number of common orthogonal factors. This number gives the measure of complexity of the runoff dynamics. Hydrological aspects of results of statistical analysis of runoffs are discussed. Key words: runoff forecasting, periodically correlated random processes, artificial neural networks. INTRODUCTION River’s runoff forecasting is a difficult hydrological topic, since underlying physical processes are complex and far from being considered as adequately described by a system of the corresponding equations. Therefore, statistical and “data mining” methods are widely used aiming to forecast rivers’ runoffs (Fernandez, Salas, 1974; Moss, Bryson, 1974; Vecchia, 1985; Loucks, Stedinger, Haith, 1988; Privalsky, Panchenko, Assarina, 1992). Several modifications of statistical forecast based on the autoregression time-series analysis combined with harmonic trend modeling were suggested in (Kashyap, Rao, 1976). But they all have an essential deficiency: the stochastic component is modeled by stationary autoregression whereas it exhibits a distinct seasonal periodicity. The reliable runoff forecast is necessary for a large number of applications connected both with water use (water-supply, power, navigation, etc.) and with protection from flooding. Certainly, the reliability of operative runoff forecast depends in many respects on existing monitoring systems providing necessary data for prediction, and such systems should be constantly developed. The problem described in this article consists in developing forecast methods based only on past runoff data. Hydrological experience shows that in such formulation long-term forecasts are hardly possible. In contrast, forecast of one time step ahead (one month ahead) is quite realistic which permits more effective use of regulating capacities of water reservoirs. Theoretical results of the paper include conclusions on type and complexity of stochastic runoff model with sampling interval one month (in other words - model with seasonal course). Such models are of particular interest in investigations of complex water economic systems using simulation methods. Examples of models with seasonal course are known in hydrology (see e.g. (Ratkovich, Bolgov, 1997) and references