Preliminary Analysis on the Effect of Temperature on LSTM-Based PWV Forecasting

Precipitable water vapor (PWV) can be applied in meteorological and climatological studies as it is a reliable indicator of the state of the atmosphere. However, the availability of PWV is unreliable. Having a more complete repository of PWV data allows for climatological studies as well as provides another avenue for rainfall prediction. Hence, there is a need to forecast and impute data to improve its robustness. To do this, the use of machine learning (ML) techniques such as Long Short-Term Memory (LSTM) can be used as it specializes in forecasting time-series values such as PWV. The effect of adding other meteorological parameters was analyzed and compared to a baseline model. One-step forecasting performance was similar throughout all of the models tested, having a spread within 0.3 mm with three outliers for the PWV model given by 1.9560, 1.9565, 1.9684; four outliers for the PWV-Temperature model given by 1.9573, 1.9521, 1.9523, 1.9569; and one outlier each for the PWV-Time and PWV-Time-Temperature models given by 1.9765 and 1.9815, respectively. It can be also noted that for data imputation, and the addition of other meteorological parameters had influenced the behavior of the imputation to capture the intricacies of PWV. The median and mean values for the PWV model, PWV-Time model, PWV-Temperature model, and PWV-Time-Temperature model are as follows: 10.869, 14.553; 15.661, 14.400; 11.724, 16.685; and 9.9262, 14.176, respectively. No outliers were observed for the data imputation models. It is recommended that further studies look into inputting other meteorological variables into the ML model. Moreover, it is also recommended that optimizations regarding the inputs and outputs (hyperparameters, lead and lag times) be done in order to improve the forecasting results.