Akademik Veri Yönetim Sistemi
EGE 14. ULUSLARARASI UYGULAMALI BİLİMLER KONGRESİ, İzmir, Türkiye, 23 - 29 Aralık 2025, ss.502, (Özet Bildiri)
This study aims to determine the most suitable statistical and machine learning approaches for short- to medium-term groundwater level (GTL) prediction by examining the groundwater level (GTL) dynamics in the Erzincan–Brastik groundwater observation well for the period 1987– 2022. Monthly GTL, precipitation, temperature, humidity, and flow data, along with their lagged values, are used from DSI and MGM. Variable relationships and the GTL estimation equation are established using mutual information, Spearman correlation, and multivariate regression analyses. The findings show that the most significant component on GTL is the GTL (t–1), the groundwater level one month prior, followed by temperature components lagged by 2–3 months, while precipitation and maximum precipitation variables have a weaker than expected effect. When the Exogenous Autoregressive Integrated Moving Average Model (ARIMAX) and the Multivariate Vector Autoregressive Time Series Model (VAR) models were compared for time series-based forecasting, the ARIMAX model significantly outperformed the VAR model (RMSE=1.183, MAE=0.918, R²=NSE=0.465) with lower error (RMSE=0.724, MAE=0.530) and higher explanatory power (R² and NSE=0.799). The RF and SVM models optimized with grid search yielded similar accuracies during the testing phase; R²=0.844 was obtained for RF, R²=0.849 for SVM, and the RMSE values for both models were approximately 0.63. The SHAP explainable artificial intelligence analysis showed that the most dominant variable in the RF model's predictions was YSS(t–1), with lagged temperature components playing a secondary role and humidity, current, and precipitation variables having limited effects.