An improved machine learning-based model to predict estuarine water levels

Abstract

The areas around estuaries are typically densely populated and economically developed. Therefore, robust flood risk assessment in these areas is critical. One of the key elements of flood risk assessment is the accurate prediction of estuarine water levels. However, the nonlinear interactions between riverine (i.e., upstream river discharge) and marine (i.e., tides) forces complicate the prediction of estuarine water levels. Traditional physics-based and data-driven models have made significant progress in predicting estuarine water levels, but they require upstream river discharge data as inputs. Considering the lack of such data, the development of new approaches is crucial. This study investigated a machine-learning-based light gradient boosting machine (LightGBM) framework for predicting estuarine water levels using historical water levels as the only inputs. Two prediction models based on the LightGBM framework, denoted as LightGBM1 and LightGBM2, are developed. The LightGBM1 model constructs only a single regression model and uses a recursive approach to generate multidimensional outputs. The LightGBM2 model constructs multiple regression models between the same inputs and outputs in each dimension. The LightGBM1 and LightGBM2 models were applied to the Yangtze estuary as a test case. The results demonstrate that both models are effective at predicting short-term (within 48 hours) estuarine water levels, but the statistical performance of LightGBM2 is better overall. For 24-hour prediction, the root-mean-squared errors of the LightGBM1 and LightGBM2 models are in the ranges of 0.14–0.17 m and 0.12–0.15 m, respectively.