Abstract: Based on deep long short term memory（LSTM） neural network, the estimation models （LSTM₅ and LSTM₁） of precipitable water vapor （PWV） in Huaihua distinct are proposed by using multi（air temperature, air pressure, dew point temperature, relative humidity, water vapor pressure and hourly precipitation） and single factor（water vapor pressure） separately, besides the estimation precision of the models are analyzed. As the result shown, the LSTM models have high estimation precision, whose coefficients of determination are both greater than 0.94, mean values of root square error, mean absolute error and mean absolute percentage error are below 1.158 1 mm, 0.709 9 mm and 4.54% respectively. The precision of LSTM models improves more than 70% compared to linear estimation model or quadratic polynomial estimation based on water vapor pressure, and the precision of LSTM₁ model slightly better than that of LSTM₅. The distribution of estimation errors relates to PWV value, which is first increased and then decreased along with the increasing of PWV. Furthermore, there is negative correlation between the estimation precision of the models and altitude of the stations.

Key words: LSTM, GPS/MET, atmospheric precipitable water vapor, estimation model, Huaihua area