Forecast Method Research Of Nonlinear System Time Series Based On Neural Network

Modern forecasting technology, born at the beginning of the20th century，is a subjectforecasting and assessing the state of the future development. That is speculating unknown fromknown、speculating the future by the past. This subject is quickly used in almost all fields andhas achieved fruitful results, especially in weather, economic，scientific research and otherfields.Usually the objects forecasted are often influenced by many uncertain factors. So theresulting time series also shows strong nonlinear characteristics. The effect of traditional linearprediction model based on statistical theory is not good for this type of sequence. Theemergence of intelligent prediction algorithm represented by artificial neural network algorithmprovides an effective way to solve this problem.At present, for solving time series predictionproblems of nonlinear system based on neural network algorithm. Scholars at home and abroadhave done a lot of meaningful work, and results have been achieved. But there are still someproblems to be solved：1、The determination of the structure of the network mostly still rely on experience orexperimental method. And there is little research on determining the topological structurecombined with sequence characteristics. At the same time, there is a lack of a dedicated neuralnetwork model for seasonal time series prediction.2、In response to the diversity of forecasting object, the combination forecasting method isproposed. However, in the practical application, it is difficult to choose the appropriatecandidate model combination according to the combination forecasting theory fornon-professionals. How to the forecasting model’s self-adaptive selection is an urgent issue.This paper’s application background is bird’s early warning platform of province naturalfund project. According to these problems that exist in the current study, and combined with theproject actual needs, related research work has been done. In this paper, the main work andinnovations are as follows:1、In view of time series of nonlinear systems whose seasonal features are obvious,combined with seasonal analysis methods in the field of statistics, A neural network prediction model which is called DSBPANN for seasonal time series forecasting is put forward. Comparedwith traditional methods, the pertinence of the method for the seasonal sequence is stronger.Modifying the traditional BP neural network from two aspects of the network structure andhandling the training sample. According to the seasonal features, the patterns of networkstructure and the training sample are determined. So that converting the network fitting for allsamples to fitting for each season model fitting. In this way, the pertinence of the network isstronger, and precision is increased effectively.2、In view of the difference and diversity of time series of different nonlinear systems, theself-adaptive combination forecasting model is proposed.To overcome the disadvantages thatthe general combination forecast theory requires selecting candidate model accurately, Themethod that adjusting the blend weight of different models adaptively using neural network isput forward.Giving different submodels appropriate blend weight to replace the model selectionprocedure.The experimental result shows that self-adaptive blending forecasting algorithm hasbetter adaptability for different sequences, and reduce the demands for the specialty of theforecasters in the context that assuring the prediction accuracy close to the best blending. Sothat the combination forecasting technology can be applied to bird’s early warning project.Based on the above two theory’s innovation work, combined with the t actual needs of thebird’s early warning project, Bird’s early warning prototype system is designed and realized.