Daily water-level forecasting for multiple polish lakes using multiple data-driven models

Water level in lakes fluctuates frequently due to the impact of natural and anthropogenic forcing. Frequent fluctuations of water level will impact lake ecosystems, and it is thus of great significance to have a good knowledge of water-level dynamics in lakes. However, forecasting daily water-level fluctuation in lake systems remains a tough task due to its non-linearity and complexity. In this study, two deep data-driven models, including gated recurrent unit (GRU) and long short-term memory (LSTM), were coupled with attention mechanism for the forecasting of daily water level in lakes for the first time. Daily water-level times series in five lowland lakes in Poland were used to evaluate the models. Root mean squared error (RMSE) and mean average error (MAE) were used for the evaluation of model performance. The modelling results were compared with the traditional feed-forward neural networks (FFNN), GRU, LSTM, and zero-order forecast. The modelling results showed that sequential deep learning models do not outperform feed-forward models in all cases. In most cases, LSTM with attention mechanism (average RMSE = 0.88 cm, average MAE = 0.69 cm) outperforms GRU with attention mechanism (average RMSE = 1.00 cm, average MAE = 0.81 cm). However, attention mechanism did not help to improve the accuracy of the GRU and LSTM for most cases. Based on the average performance in different lakes, GRU performs the best among the deep learning models (average RMSE = 0.84 cm, average MAE = 0.66 cm). Zero-order forecast models perform better than deep learning models for predicting tomorrow (average RMSE = 0.71 cm, average MAE = 0.39 cm), while deep learning models perform better as the horizon of prediction increases.

Automated summary

This study combines GRU and LSTM deep learning models with attention mechanisms for the first time to forecast daily water level in lakes. The results show that GRU performs the best among the deep learning models in different lakes, and the performance of deep learning models improves as the prediction horizon increases.