Dynamic Sliding Window-Based Long Short-Term Memory Model Development for Pan Evaporation Forecasting

Abstract

Accurate and precise forecasting for pan evaporation (EPm) is one of the essential processes for multiple purposes of watershed decision making and sustainability. This is because EPm is highly influenced by climate change and thus it is a highly stochastic weather process. In this research, long short-term memory (LSTM) and Gaussian process regression (GPR) models were used to forecast monthly EPm data. The EPm data series were used to cover the period (1980 to 2020) and belong to two meteorological stations located in the United States of America. One station has a tropical climate in Hialeah, Florida (station 51) and the other has a Mediterranean climate in Markley Cove, California (station 9). The analysis was carried out using the optimum window sizes for the LSTM and GPR methods, which were 4 and 7, respectively. The modeling results indicated that the LSTM model reported maximum correlation value (R = 0.965) and (R = 0.701) for Station 9 and 51 respectively whereas minimum root mean square error (RMSE = 0.611) and (RMSE = 0.916) for station 9 and 51 respectively which outperformed the GPR model. Based on the relative error diagram, 78% of the LSTM model's results ranged between 25%. In addition, the LSTM model's normal distribution of errors showed that the mean error and standard deviations for station 9 are -0.0104 and 0.614 whereas for station 51 it was -0.026 and 0.921. For the same stations, the GPR model mean error and standard deviation for station 9 are -0.016 and 0.753 whereas for station 51 were -0.026 and 0.921 indicating that the LSTM model outperforms the GPR model in forecasting monthly EPm. In general, the development of the LSTM model as a dynamic soft computing model reported a robust and reliable tool for EPm forecasting at different climate characteristics.