In recent years, it has been found that the Precipitable Water Vapor (PWV) time series derived from ground-based GPS measurements can be used in to forecast precipitation in different regions. However, it is inevitable to consider the impact of several meteorological parameters such as temperature, pressure, relative humidity, water vapor pressure, total cloud cover and day of year (doy) besides PWV on rainfall prediction. In order to predict the precipitation at Tehran station, two types of Artificial Neural Network (ANN), including Multi-Layer Perceptron (MLP) and Nonlinear Auto-Regressive with Exogenous Inputs (NARX) were employed based on mentioned parameters. At first, these neural networks were trained under various circumstances (i.e. with and without PWV) with the help of collocated meteorological and GPS data from years 2007–2010 and then the networks were utilized to forecast different intensities of precipitation over 2011. The results showed that deletion of PWV values from input data will reduce the precision of MLP predictions for the range of rainfalls less than 6 mm. For the range of precipitation above 3 mm, the use of PWV has a positive impact on the output of the NARX model. In addition, the effect of the length of training data on the performance of the proposed models was investigated in terms of Mean Bias Error (MBE), Root Mean Square Error (RMSE) and False Alarm Ratio (FAR) statistics. The best results in the study region were achieved from 4years trained MLP and 2years trained NARX models. Comparing the outputs of the MLP and NARX models with the Global Forecasting System (GFS) 6h forecasts as a standard meteorological forecast showed that the efficiency of the NARX model is higher than the MLP and GFS, especially in moderate and strong rainfall classes. Also, seasonal comparison of these errors showed that both models underestimate rainfall values higher than 3 mm. In almost all seasons, the underestimation of the NARX model was less than MLP. With all pros and cons, the NARX model showed greater performance than MLP for both non-rainfall and rainfall events.

近年来，已经发现，基于地面GPS测量的可降水量水汽（PWV）时间序列可用于预测不同地区的降水。但是，除了PWV之外，不可避免地要考虑温度，压力，相对湿度，水蒸气压力，总云量和年（doy）等几个气象参数对降雨预测的影响。为了预测德黑兰站的降水，基于上述参数，采用了两种类型的人工神经网络（ANN），包括多层感知器（MLP）和带有外来输入的非线性自回归（NARX）。首先，这些神经网络是在各种情况下进行训练的（即 借助并置的2007-2010年气象和GPS数据，然后利用这些网络来预测2011年的不同降水强度。结果表明，从输入数据中删除PWV值会降低精度。 MLP预测的降雨范围小于6毫米。对于3 mm以上的降水范围，PWV的使用对NARX模型的输出有积极影响。此外，还根据均值偏差误差（MBE），均方根误差（RMSE）和误报率（FAR）统计数据研究了训练数据长度对所提出模型的性能的影响。在研究区域中，最好的结果是通过4年训练的MLP和2年训练的NARX模型获得的。将MLP和NARX模型的输出与全球天气预报系统（GFS）的6h预报作为标准气象预报进行比较，结果表明，NARX模型的效率高于MLP和GFS，尤其是在中等强度和强降雨等级中。此外，这些误差的季节性比较显示，两个模型都低估了高于3 mm的降雨值。在几乎所有季节中，NARX模型的低估幅度都小于MLP。不管有什么利弊，NARX模型在非降雨和降雨事件中都表现出比MLP更好的性能。这些误差的季节性比较表明，两个模型都低估了高于3 mm的降雨值。在几乎所有季节中，NARX模型的低估幅度都小于MLP。不管有什么利弊，NARX模型在非降雨和降雨事件中都表现出比MLP更好的性能。这些误差的季节性比较表明，两个模型都低估了高于3 mm的降雨值。在几乎所有季节中，NARX模型的低估幅度都小于MLP。不管有什么利弊，NARX模型在非降雨和降雨事件中都表现出比MLP更好的性能。