时问序列动态学习率神经网络模型及其初步试验

神经网络在气象上的应用往往是采用固定学习率的BP算法建模，学习过程易出现振荡现象和网络存在冗余连接等缺陷，基于此对神经网络进行了改进。利用时间序列分析方法对样本数据进行处理，用改进后的神经网络对时间序列样本数据进行训练预测，创建了时间序列动态学习率神经网络模型。最后用库车县1997—2007年四季的平均气温值作样本数据进行训练，其训练精度和拟合度都达到很高的标准，用该模型预测了库车县2008年的气温。通过实例证明这个模型在气象预测领域有一定的实用价值。

Air Temperature Predication Using a Dynamic Learning Rate Neural Network Model Based on Time Series

Application of neural network in weather for forecasting was often adopt BP （Back Propagation） algorithm with a fixed learning rate to modeling the multi-level model, and many limitations are presented such as vacillation phenomena and redundancy link in network, so this paper improve neural network model. First the example data is disposed by time series method, next the neural network model is modified, and then the time series sample is trained by using a modified model, after above steps this paper set up a dynamic neural network with time serial model. Finally based on the observed temperature data of Kuche in 1997-2007 is trained by using a dynamic model and the result shows the training precision and fitting accuracy to be a high standard, and to predict 2008 temperature of Kuche with this model. Through this example prove this model to be a utility value for the domain of meteorological prediction.