Using a seasonal and trend decomposition algorithm to improve machine learning prediction of inflow from the Yellow River, China, into the sea DOI Creative Commons
Shuo Wang, Kehu Yang, Hui Peng

et al.

Frontiers in Marine Science, Journal Year: 2025, Volume and Issue: 12

Published: May 9, 2025

The Yellow River is the largest inflow into Bohai Sea, and its changes directly affect ecological environment marine health of Sea. Therefore, accurate prediction crucial for maintaining balance Sea protecting resources. Time decomposition algorithms, combined with machine learning, are effective tools to enhance capabilities models. However future data leakage from items was ignored in many studies. It necessary develop right method operate time avoid leakage. In this study, sea predicted based on a learning model (light gradient boosting machine, LightGBM) algorithm (seasonal trend using loess, STL), different ways STL were evaluated. results showed that overall performance STL–LightGBM better than LightGBM model. took historical 8 days as input, average NSE next 1–7 would reach 0.720. Even when forecast period 7 days, (NSE: 0.549 7-day lead time) 0.105 higher 0.444 time). We found pretreatment entire test set overestimated true STL–LightGBM. recommended preprocesses each sample study can provide help water resources management offshore environmental management.

Language: Английский

Using a seasonal and trend decomposition algorithm to improve machine learning prediction of inflow from the Yellow River, China, into the sea DOI Creative Commons
Shuo Wang, Kehu Yang, Hui Peng

et al.

Frontiers in Marine Science, Journal Year: 2025, Volume and Issue: 12

Published: May 9, 2025

The Yellow River is the largest inflow into Bohai Sea, and its changes directly affect ecological environment marine health of Sea. Therefore, accurate prediction crucial for maintaining balance Sea protecting resources. Time decomposition algorithms, combined with machine learning, are effective tools to enhance capabilities models. However future data leakage from items was ignored in many studies. It necessary develop right method operate time avoid leakage. In this study, sea predicted based on a learning model (light gradient boosting machine, LightGBM) algorithm (seasonal trend using loess, STL), different ways STL were evaluated. results showed that overall performance STL–LightGBM better than LightGBM model. took historical 8 days as input, average NSE next 1–7 would reach 0.720. Even when forecast period 7 days, (NSE: 0.549 7-day lead time) 0.105 higher 0.444 time). We found pretreatment entire test set overestimated true STL–LightGBM. recommended preprocesses each sample study can provide help water resources management offshore environmental management.

Language: Английский

Citations

0