Comparison of strategies for multistep-ahead lake water level forecasting using deep learning models

Accurate forecasting of multistep-ahead lake water level is valuable for extreme disaster prevention and eco-environmental protection. However, existing studies mainly focus on promoting forecasting accuracy from model techniques but neglect the importance of appropriate selection of multistep-ahead forecasting strategies. In this study, the most comprehensive summary of current multistep-ahead forecasting strategies was presented. Then, different strategies were used to predict water level of Poyang Lake for multiple horizons based on LSTM and BiLSTM, and four metrics were applied for evaluating the prediction performance of strategies and models. The results showed that all of the strategies obtained satisfactory accuracy for short-term forecasting of lake water level. However, for long-term forecasting, the Rec strategy and DirRec-R strategy significantly outperformed other strategies, the NSE, R2, MSE and MAPE of DirRec-R strategy at XZ station for 180-day ahead forecasting can be improved by 37%, 32%, 58% and 38% compared to widely used Dir strategy. Moreover, Rec strategy and DirRec-R strategy can capture flood peak values while other strategies performed unsatisfactory. Meanwhile, the BiLSTM achieved better performance than LSTM in 72% of the evaluation results for long-term forecasting, but the performance of BiLSTM and LSTM for short-term and medium-long-term forecasting did not exhibit significant difference. This study can provide a reference paradigm for future studies of multistep-ahead forecasting of lake water level or other hydrological variables.