Improving Lake Level Prediction by Embedding Support Vector Regression in a Data Assimilation Framework

Abstract

Data-driven models are widely used in the field of water level prediction due to their generalizability and predictive abilities. In long-series prediction, however, data-driven models degrade rapidly due to the uncertainty and constraints of model data and parameters. To address the problem of inaccurate continuous water level prediction, this study introduced a data assimilation technique, the unscented Kalman filter (UKF), and embedded support vector regression (SVR) into the framework and applied it to Dongting Lake, the second largest freshwater lake in China. The results demonstrated that the assimilation model is significantly better than the non-assimilation model in predicting water levels and is not affected by the characteristics of lake level changes, with the R2 increasing from 0.975–0.982 to 0.998–0.999 and the RMSE decreasing from 0.436–0.159 m to 0.105–0.042 m. The prediction lead time also increased with the increase of continuous assimilation data. Further analysis of the assimilation model showed that when there was an assimilation cycle, the prediction remained stable for successive sets of two or more assimilated data, and the prediction lead time increased with successive assimilated data, from 4–8 days (one successive assimilation data) to 9–12 days (five successive assimilation data). Overall, this study found that the data assimilation framework can improve the prediction ability of data-driven models, with assimilated models having a smaller fluctuation range and higher degree of concentration than non-assimilated models. The increase in assimilated data will improve model accuracy as well as the number of days of prediction lead time when an assimilation cycle exists.