Метод нейронных сетей для климатического прогнозирования водности Чернореченского водохранилища

Abstract

New method for precipitation forecasting at the Ai-Petri region is proposed in this work. This method includes a model based on artificial neural networks. A set of global oceanic and meteorological indices were used as the input parameters of the model. SST and SLP data sets from NCEP / NCAR and HadISST re-analyses in 1950-2020 were used for indices calculation. A feature of the proposed model is the decomposition of the predicted series into two orthogonal signals, their independent modeling and subsequent addition of the calculated model values. The sum of model calculations signals was verified. The model was verified in the period 2007–2020. The possibility of forecasting average monthly precipitation amounts with a lead time of up to 6 months is shown. It is found the possibility of the model to predict precipitation in the winter and summer seasons, September and October, which is 70% of the average long-term annual precipitation. It is shown that the best forecast of precipitation for the winter season can be made in November, and with a higher quality of the forecast – in December. The average absolute deviation in the control sample was 28% and 23%, respectively. Taking into account that the maximum precipitation in the mountains occurs in the cold half of the year, and the fact that the main volume of water content of the Chernorechensk reservoir is formed in the cold period of the year, then the forecast of precipitation for the winter season is of the greatest importance. The forecast of precipitation for the summer period and September-October can be made in April (the average absolute deviation is 22%). In addition, the work tested and confirmed the possibility of a climate forecast of atmospheric pressure. The obtained results can be useful for early assessment of the level of filling of the Chernorechensk reservoir.