Improving prediction accuracy of daily pan evaporation using meteorological variables with a novel explainable computational intelligence framework based on the Kolmogorov–Arnold network

Reliable quantification of pan evaporation (Epan) is essential for the effective design of irrigation systems and the sustainable management of agricultural water resources. This study represents the first comprehensive effort to introduce the Kolmogorov–Arnold Network (KAN) framework as a novel predictive approach for Epan modeling at two meteorological stations in the Kurdistan region of Iraq, providing a systematic evaluation of its predictive capability under this specific climatic context. The comparative evaluation of the predictive models indicates that KAN achieved the best numerical validation performance among the evaluated algorithms at both study sites. At the Chamchamal station, KAN achieves superior accuracy, reflected by a root mean square error (RMSE) of 1.152 mm, a mean absolute error (MAE) of 0.854 mm, and a correlation coefficient of 0.971. Extremely randomized trees (ERT) model ranked second, with slightly higher error values (RMSE = 1.176 mm, MAE = 0.868 mm) and a comparable correlation coefficient (R = 0.970). The random forest regression (RFR) model followed, exhibiting an RMSE of 1.20 mm, an MAE of 0.864 mm, and an R-value of 0.969. Similar performance trends were also observed at the Darbandikhan station. KAN again demonstrated the highest predictive capability, achieving an RMSE of 1.018 mm, an MAE of 0.776 mm, and an R-value of 0.962. The gradient boosting regression tree (GBRT) model ranked second, with an RMSE of 1.127 mm, an MAE of 0.851 mm, and an R of 0.953, followed by the ERT model, which recorded an RMSE of 1.138 mm, an MAE of 0.900 mm, and an R value of 0.954. The findings confirm the robustness and transparency of the proposed advanced KAN framework for estimating Epan. This modeling approach provides an effective tool for supporting agricultural water resource monitoring, reservoir operation planning, and the assessment of drinking water supply systems.HighlightsKAN framework proposed for accurate Epan prediction at two monitoring stations.Six distinct input feature sets were used to estimate Epan.Among the nine models, the KAN achieved the highest accuracy in Epan prediction.SHAP identified the most influential variables contributing to the model predictions of Epan.