The Studies On Prediction Technique Of Time Series Based On RBF Neural Networks

At the time of the development of science and technology, higher and higher demandsof the accuracy for the meteorological forecast are made. This is not only related to thedevelopment of agriculture, but also related to the government who can take measures inadvance to the inclement weather, people’s food, clothing, travels and other aspects.Tongling and Chizhou city, which are situated in the middle and lower areas ofChangjiang, were choosed as our study area, and the meteorological data which areobtained from meteorological station of Tongling and Chizhou were used as our object.The chief tasks were made as follows:(1)Using statistical analysis for the data of monthtemperature time series of Tongling from1960to2008, a seasonal ARIMA model wasestablished;(2) Using statistical analysis for the data of month temperature andprecipitation of Chizhou from1959to2010, according to the seasonal effect, SeasonalProduct Model was established to predict the month data of2010;(3)On the basis of timeseries analysis, using the neural networks theory, combined with the seasonal effect of timeseries and the Radial Basis Function(RBF) neural networks method, the temperature andprecipitation of Chizhou from January to December,2010, were forecasted respectively.According to the above methods, the main conclusion of the study were obtained asfollows:1. The temperature series analysis in1960to2007and the prediction situation oftemperature in2008of Tongling:In Tongling, the annual average temperature from1960to2007is16.4℃. In recentyears, temperature first fell slightly, then later began to rise slightly from1990s. Weakertrend but stronger cyclicity, the annual fixed cyclicity was contained. After the seasondifference, smoothing, parameter estimation, model order determination and predict themonth temperature in2008. The prediction result shows: the mean absolute error is0.875;the trend of the prediction and actual value is same.2. The meteorological data analysis from1959to2009and the data predictionsituation in2010of Chizhou:(1) The analysis and prediction based on Time Series Model:In Chizhou, the annual average temperature from1959to2009is16.1℃, The annualaverage precipitation is1457.9mm, warm and humid climate with four distinct seasons.Over the years, the month temperature changed random in a range. In recent years, temperature first fell slightly, then later began to rise slightly from1980s. Weaker trend butstronger cyclicity, the annual fixed cyclicity was contained. The high precipitation datashowed a strong volatility, while the low precipitation value showed steady. There are lessrain in spring and winter while heavier rainfall in the summer and autumn, namely theprecipitation time series data show obvious seasonal variations, which maximumprecipitation is247.7mm in June.The mean absolute error of month temperature prediction is1.5, the average relativeerror is13.3%; the mean absolute error of month precipitation prediction in2010is7.15,the average relative error is6.97%. The trend of the prediction is the same as the actualvalue, but the relative error of January and December is rather higher.(2) The analysis and prediction based on RBF neural networks model:According to the RBF neural networks theory, combined with the cyclicity of timeseries, after train and test, we establish a kind of forecasting model based on RBF neuralnetworks with three layers. We can get the month temperature prediction and the monthprecipitation prediction in2010using the prediction model. The results show: The meanabsolute error of month temperature prediction is1.13, the average relative error is7.72%;the mean absolute error of month precipitation prediction in2010is6.74, the averagerelative error is6.97%.In this paper, the methods for dynamic data modeling and time series predictiontechnology were offered.