Harnessing Neural Networks for Accurate Potential Evapotranspiration Forecasting: A Case Study in Semi-Arid Aurangabad District of Maharashtra State Using the Thornthwaite Method

Abstract

Potential evapotranspiration (PET) is a crucial parameter for effective water resource management, particularly in semi-arid regions like Aurangabad district. PET was estimated for nine stations in this region using the Thornthwaite method, which is a widely-used approach for determining this important variable. Time series analysis revealed significant autocorrelation and stationarity, confirming the suitability of the data for modeling. A feed-forward neural network with 12 input, 4 hidden, and 1 output neuron was implemented for PET forecasting. The FFNN (Feed Forward Neural Network) demonstrated robust accuracy, with training mean absolute error of 7.86–11.52 and validation MAE of 11.21–16.83. RMSE ranged from 10.32 to 15.12 during training and 14.94 to 23.06 during validation. Percent bias remained below 2.2%, and Nash-Sutcliffe efficiency exceeded 0.97 in training and 0.92 in validation. Vaijapur performed best, achieving a validation RMSE of 14.94, MSE of 223.35, and R² of 0.95. Conversely, Sillod recorded the highest validation errors, highlighting regional variability. The FFNN effectively captured PET dynamics with minimal over-fitting, supporting its application in optimizing water use and agricultural planning. Localized calibration is recommended for stations with higher errors to enhance accuracy.