两种方法在不同海拔地区计算湿球温度的对比

选取1990~1999年贵州省3个国家基准站（威宁、贵阳、三穗）气象观测数据，评估了逼近法在贵州不同海拔地区计算湿球温度的效果，对比了BP（Back Propagation）神经网络模型和逼近法在计算湿球温度方面的优劣。结果表明:（1）3站逼近法计算值与观测值之间平均绝对误差分别为0.059℃、0.046℃、0.042℃，误差<0.1℃的数据比例为83.91%、91.52%、92.76%；当气温低于0℃时，误差>0.2℃的频率呈增长趋势，其原因可能是逼近法中对结冰的判别和实际情况存在差异导致高海拔地区的计算效果差于低海拔地区。（2）3站BP神经网络模型计算精度比逼近法分别提高60.71%、57.45%、57.78%，误差<0.1℃的数据比例提高到97.38%、97.18%、97.44%，有效地解决了高海拔地区气温低于0℃频率较高而导致逼近法计算误差偏大的问题，在低海拔地区的计算结果也优于逼近法。（3）BP神经网络模型计算湿球温度需要对各测站进行单独拟合，在低海拔地区针对大量站点且计算精度要求不高时可用逼近法，反之则用BP神经网络建立单站模型。 

Comparison of Two Methods for Calculating Wet Bulb Temperature at Different Altitudes

By using data from Weining, Guiyang, and Sansui meteorological stations from 1990 to 1999, the effect of the approximation method in calculating the wet bulb temperature at different altitudes in Guizhou was evaluated, and the advantages and disadvantages of the BP neural network model (BPNN) and the approximation method in calculating the wet bulb temperature were compared. The results are as follows: (1) From the comparison of the wet bulb temperature between calculated by approximation method and observed at stations, the average absolute errors of the Weining, Guiyang, and Sansui stations were 0.059℃, 0.046℃ and 0.042℃, respectively. The proportions of data with errors less than 0.1 °C were 83.91%, 91.52%, and 92.76%, respectively. When the temperature was lower than 0℃, the frequency with the error greater than 0.2 ℃ showed an increasing trend, it is believed that there was a certain difference in the judgment of icing with the approximation method, which made the worse calculation at high altitude than that at low altitude. (2) Compared with the approximation method, the accuracy of BPNN for predicting wet bulb temperature was improved by 60.71%, 57.45%, 57.78%, respectively. The proportions of data with errors less than 0.1 °C increased to 97.38%, 97.18%, 97.44%, respectively. It effectively solved the problem of calculation error of approximation method caused by the high frequency of temperature below 0 ℃ in high altitude areas. The calculation results in low altitude from BPNN were also better than those from the approximation method. (3) The BPNN needed independent fitting for calculating the wet bulb temperature. The approximation method could be used when the requirement of the calculation accuracy was not high. Otherwise, a single station model should be established by the BP neural network.