Probabilistic Hydrology Forecasting Based On NAR Dynamic Neural Network Posterior Information

Hydrological processes on the current level of knowledge was a random evolution in view of time and space. In fact, in the real world we did not have an exact description. The best way to solve the problem was to use a mathematical model. Because humans on weather systems(such as precipitation) knowledge of how to run was not complete, therefore, long-term hydrological forecasting often changed a lot of errors.Naoli River Basin was located in cold temperate continental monsoon climate zone, the average annual rainfall was 534.5 mm according to statistics for many years, the distribution during the year was very unequal. Rainfall in June-September accounted for 72% of the all year. As we all know, the main source of irrigation was precipitation, and rainfall largely reflected the trend of disasters, so the precipitation forecasting for the promotion of social and economic sustainable and stable development was of great significance. If we can accurately forecast precipitation in Naoli River Basin, make decisions or take measures to regulate the river basin water resources and scheduling according to the results, irrigated agricultural production would have greatly improved.This article studied the data of Naoli River Basin Caijuzi hydrological station from 1991 to 2011 the monthly and annual precipitation,and then used the probability forecasting method to forecast June 2012 and 2012 annual precipitation. The main research methods and conclusions are as follows:(1)The priori information of probability precipitation forecasting to determineWe analysised the inherent random feature of precipitation data, and then screened the probability distribution pattern suitable to determine the parameters of the prior distribution and its probability density distribution. In this article, it multipurpose used non-linear data fitting priori information technology, mathematical statistics methods. Their distribution pattern mainly conform to Normal, Lognormal and Weibull.(2) The posterior information of precipitation forecasting to determineWe selected the appropriate deterministic hydrological forecasting models. Then calibrated the parameters of the model and obtained the random feature of their distribution. This article selected NAR dynamic neural network, built lots of it to forecast on condition that conforming to fitting precision and autocorrelation of errors,and then get posterior information after obtaining forecasting assemblage.(3) Probabilistic precipitation forecast methodAccording to the probabilistic forecasting methods based on Bayesian theory, we will determine the prior distribution and posterior information by combining results after posterior distribution. Later the largest posterior probability distribution of precipitation occurs as forecast values and calculated the probability of its occurrence, probabilistic hydrologic forecasting was built.(4) The monthly and annual precipitation forecasts compared with actual values and contrast their respective probability of occurrence was found the probability of 2012’s monthly and annual actual precipitation less than the forecasting precipitation, even the actual probability of one month of the flood season is much less than the forecasting precipitation. After analysis of 2012’s actual monthly and annual rainfall data and found that the situation is much different from many years of annual precipitation statistics.(5) The applicability of the probabilistic forecasting system was stronger than traditional deterministic forecasting system,to avoid the dangers of the single predicted value,therefore, there are more practical significance for decisions and measures.