AUTHOR=Tran Yen Binh , Arias-Rodriguez Leonardo F. , Huang Jingshui TITLE=Predicting high-frequency nutrient dynamics in the Danube River with surrogate models using sensors and Random Forest JOURNAL=Frontiers in Water VOLUME=4 YEAR=2022 URL=https://www.frontiersin.org/journals/water/articles/10.3389/frwa.2022.894548 DOI=10.3389/frwa.2022.894548 ISSN=2624-9375 ABSTRACT=

Nutrient dynamics play an essential role in aquatic ecosystems. Despite advances in sensor technology, nutrient concentrations are difficult and expensive to monitor in-situ and in real-time. Emerging data-driven methods may provide surrogate measures for nutrient concentrations. In this work, we use 4-years of water quality data with high-frequency (15-min) intervals acquired at 2 automatic stations in the German Danube River to train data-driven algorithms and build surrogate measures for nitrate (NO3--N), ammonium (NH4+-N), and orthophosphate (PO43--P). Pre-processing of the data included removing outliers and filling missing values by linear interpolation. Multiple Linear Regression (MLR) and Random Forest (RF) are trained, cross-validated, and tested using dissolved oxygen (DO), temperature (Temp), conductivity (EC), pH, discharge rate (Q), and chlorophyll-a (Chl-a) as input futures. Additionally, we used time-series data to develop cyclical features to test improvements in the underlying relationship between data. This work presents a thorough description of the modeling workflow, including intermediate steps for feature engineering, feature selection, and hyperparameter optimization. In total, 12 surrogate models (2 algorithms ^* 3 constituents ^* 2 stations) are compared with R² and RMSE as error metrics. The results show that RF outperforms MLR when adding at least three predictors for all the surrogate models. The MLR models give R²-values for NO3--N 0.67 and 0.89, NH4+-N 0.39 and 0.40, PO43--P 0.34 and 0.54 of Pfelling station and Jochenstein station, respectively. RF models produce accurate predictions and low error performances for all the targets NO3--N (R² = 0.99 and 0.99), NH4+-N (R² = 0.98 and 0.99), PO43--P (R² = 0.96 and 0.96). The percentage improvement of RMSE for RF compared to MLR in prediction nutrients ranges from 73 to 92%. This work demonstrates the usefulness of surrogate models using the RF algorithm when reproducing nutrient dynamics and serving as soft sensors for monitoring nutrient concentrations.