Novel simulation of aqueous total nitrogen and phosphorus concentrations in Taihu Lake with machine learning.

Abstract

This study demonstrates the utility of internal nutrient loads as an additional parameter to improve the performance of machine learning models in predicting the temporal variations of aqueous TN and TP concentrations in Taihu Lake, a large shallow lake. Internal loads, as a potential input parameter for machine learning models, were estimated using a mass balance calculation. The results showed that between 2011 and 2018 the maximum monthly internal loads of nitrogen and phosphorus in Taihu Lake were 4200t and 178t, respectively. Monthly changes in the aqueous TN and TP concentrations of Taihu Lake did not correlate significantly with inflow loads whereas the correlations with estimated internal loads were positive and significant. Long short-term memory (LSTM), random forest (RF), and gradient boosting regression tree (GBRT) models were built, and for all of them the inclusion of internal loads in the input parameters improved their performance. LSTM model III, whose input parameters included both inflow loads and internal loads, had the best performance, based on a testing root mean square error of 0.11mgTN/L and 0.017mgTP/L. A 28% decrease in the annual aqueous TP concentration in Taihu Lake in 2018 simulated by LSTM model III was achieved by lowering the average water level from 3.29m to 2.99m, suggesting a possible strategy to control the TP concentration in the lake. In summary, our study showed that aqueous TN and TP concentrations in shallow lakes can be simulated using machine learning, with LSTM models outperforming RF and GBRT models; in these models, internal loads should be included as an input parameter. Additionally, our study identified the water level as an important factor affecting the aqueous TP concentration in Taihu Lake.