Abstract: In this work, we consider the difference in external environments on which the forecasts are based and the difference in forecast errors under different scenarios, and work out a method for division of the prediction scenarios by the key influencing factors, such as rainfall and forecast period. Then, the data of historical forecast errors can be used to obtain their distribution trends and thus determine the confidence interval at the 90% level for different scenarios. Considering prediction scenarios, prediction errors, and their trends, we construct a new multidimensional and multi-attribute method for correcting inflow prediction, using the variable mode decomposition and a long-short term memory neural network model. Through a case study of the Three Gorges reservoir, we find the average relative error of inflow predictions is reduced from 8.32% to 6.36%, with a reduction rate of 23.6%. And other indicators are improved to varying degrees, such as average absolute error, root mean square error, and coefficient of determination. This shows our method has achieved a significant improvement on the accuracy of inflow prediction models through increasing the effective information input to the correction model.

Key words: prediction scenario, scenario division, inflow prediction, prediction correction, long short-term memory neural network model, Three Gorges Reservoir