利用人工神经网络模型预测西北太平洋热带气旋生成频数

Through the correlation analysis of the tropical cyclone (TC) frequencies in the summer and autumn (from June to October) and the large-scale environmental variables in the spring (from March to May) in 60 years (from 1950 to 2009), eight pre-forecasting factors with relatively high correlations were selected to establish an artificial neural network (ANN) model. The TC frequencies in the summer and autumn of 8 years from 2010 to 2017 were hindcast, and the hindcast results were compared and analyzed with those obtained by the traditional multiple linear regression (MLR) method. The results show that the ANN model has a high fitting accuracy for the historical data of 60 years, with a correlation coefficient as high as 0.99 and an average absolute error as low as 0.77. In the 8-year hindcast, the correlation coefficient of the ANN model is 0.80 and the average absolute error is 1.97; while the correlation coefficient of the MLR model is only 0.46 and the average absolute error is 3.30. The performance of the ANN model in fitting historical data and hindcasting is significantly better than that of the MLR model, and it can be considered for application in actual operational forecasting in the future.

通过对60年(1950~2009年)北半球夏、秋季(6~10月)热带气旋(TC)频数与春季(3~5月)大尺度环境变量的相关分析,挑选出8个相关性较高的前期预报因子建立人工神经网络(ANN)模型,对2010~2017年8年夏、秋季TC频数进行回报,并将回报结果与传统多元线性回归(MLR)方法所得结果进行对比分析。结果表明,ANN模型对60年历史数据的拟合精度高,相关系数高达0.99,平均绝对误差低至0.77。在8年回报中,ANN模型相关系数为0.80,平均绝对误差为1.97;而MLR模型相关系数仅为0.46,平均绝对误差为3.30。ANN模型在历史数据拟合和回报中的表现都明显优于MLR模型,未来可考虑应用于实际的业务预测中。