Next-generation intelligent framework for pan evaporation prediction: introducing Chebyshev polynomial-based...

Area of Science:

• Hydrology
• Agricultural Science
• Water Resource Management

Background:

• Pan evaporation (Epan) is vital for hydrology and agriculture but challenging to predict due to climate dependencies.
• Accurate Epan prediction requires reliable models for time series forecasting.

Purpose of the Study:

• To propose and evaluate a Chebyshev Polynomial-Based Kolmogorov-Arnold Network (CKAN) for Epan prediction.
• To compare CKAN performance against established deep learning (LSTM, GRU, TFR) and machine learning (CART, XGBoost) models.

Main Methods:

• Developed a CKAN model for Epan prediction at Perth and Sydney, Australia.
• Implemented LSTM, GRU, TFR, CART, and XGBoost for comparative analysis.
• Utilized SHAP and LIME for input feature importance analysis and K-fold cross-validation for model generalizability.

Main Results:

• The proposed CKAN model demonstrated superior performance in Epan prediction compared to all other evaluated methods at both stations.
• SHAP analysis identified solar radiation (Perth) and minimum temperature (Sydney) as key predictors.
• LIME analysis highlighted mean temperature (Perth) and relative humidity (Sydney) as influential factors in selected samples.

Conclusions:

• The CKAN model offers a highly effective and generalizable approach for Epan prediction.
• Understanding key meteorological drivers through interpretable AI enhances hydrological and agricultural management.
• The study validates CKAN's potential for accurate time series forecasting in environmental applications.