Research Of Multi-Model Integration Weather Forecast Based On Neural Network

Weather forecast is regarded as the important reference for people to arrange their travel plan. It is the security for cities and towns to prevent climate disaster as well. As the progress of modern meteorological technology, weather forecast has more strict requirements in high accuracy and timeliness, integrated weather forecast technology become a focus in the study. Integrated forecast is a mathematical model which is designed to aggregate and analysis a variety of independent model forecast results, the main purpose of the method is getting a more ideal and accurate integrated conclusion as the result. The research on medium-term and short-term integrated forecast operation is scanty at present time. The purpose of this project research is establishing integration forecast algorithm model with artificial neural network(ANN). Next step is developing software for city’s observation sites accurate weather forecasting. The software can provide integrated forecast data as fixed-point of time and quantitative result. It can reduce the forecast error of numerical models and avoid subjective decision by forecasters.For now, most integration algorithm is based on ANN by using BP network, but it has slow training process and it gets into local minimum easily. The radial basis function(RBF) neural network is characterized by high learning speed, convergent and strong timeliness. So selecting RBF-ANN is good for integrated forecast problem. This thesis uses a new type of integration course:choosing several models numerical data as the input of RBF-ANN, through the fully trained network, the output is the unitary integration result.The main work in this paper is:First, analyzed characteristics of Tianjin observation site numerical model data and selected several reliable numerical models as integration members, built correction method to correct error of the temperature data; then set up integrated forecast model based on RBF-ANN, developed multi-model integration system for temperature forecast at the same time. The system can run the objective integration forecast operation for every6hours, the maximum and the minimum temperature of all day. The system had realized automatic integration, manual integration on specified conditions and historical data query. The result shows that integration data is better than any single model which takes part in the integration on error and accuracy comparison. RBF has a shorter run time than Particle Swarm Optimization and Genetic algorithm. As a consequence, this system is characterized by good stability, high accuracy and efficient work time. It has preferable application prospect in the meteorological operation.