热带气旋路径人工神经元预报方法对比试验研究

AN EXPERIMENTAL STUDY ON ARTIFICIAL NEURAL NETWORK FORECASTING MODELS OF TROPICAL CYCLONE PATH

分别对具有动量项BP、LM、RBF人工神经网络建立36、48、60、72小时的热带气旋路径预测模型,各用100个独立样本进行预测检验,分析了网络“学习好,预报差”的原因,解决这一问题的关键是选择合适的网络结构参数、相应的学习算法和合适的预报因子,并总结了合理应用人工神经网络建立预测模型的经验。针对人工神经网络模型不具有自动选取因子的功能,给实际应用造成困难,提出了基于RBF的逐步选取因子的算法,并进行了对比试验,表明该方法具有较高的实用价值。

Based on BP, LM, RBF artificial neural network with a term of momentum, forecasting models for tropical cyclone path of 36, 48, 60 and 72 hours are set up, and run with 100 independent samples. The results show that the models with good fitting generally produce bad forecast. The keys to avoid this embarrassing situation are proper parameters for network structure, corresponding algorithm and suitable predictors. In view of the fact that artificial neural network models lack the mechanism of automatic predictor adaptation, an algorithm of stepwise predictor adaptation of RBF model is proposed in this study. The comparison with the other models shows that the suggested algorithm is worth to be tried in routine forecasting operation.