Evaluating long-term forecasting performance decay and seasonal effects in deep learning-based water level forecasting models across multiple northern lakes

Abstract

Advancements in deep learning methods have attracted researchers to integrate deep learning into hydrological prediction and modelling. The performance of these data-driven approaches to the lake environment can be affected by the choice of lead times, meteorological factors and the hydrological nature of the lakes, which need exploration. We used long short-term memory and gated recurrent unit regional models to forecast the daily water levels of 30 Finnish lakes from 1 day ahead to 15 days ahead with historical water level and meteorological observations as inputs. We observed that both forecasting models performed well for short lead times and their performance decayed monotonically, with the decay rates varying among different lakes, making long-term forecasting suitable for some but not all lakes. In addition, the long short-term memory model outperformed the gated recurrent unit model at longer lead times when trained on the combined data from all 30 lakes, whereas the gated recurrent unit model showed greater robustness under our leave-one-lake-out evaluation, where the target lake’s data was excluded from training. We also found that the coefficient of variation of water levels and the Richards–Baker flashiness index had positive correlations with the speed of increase in root mean squared error. Seasonally, we found Finnish lakes tended to have poorer performance metrics from April to June, which might be associated with increased uncertainty during the spring–summer transition, when meteorological conditions could change rapidly. Our results highlight the potential of deep learning methods for short-term lake water level predictions to supplement or back up traditional monitoring methods and reveal the models' performance from the temporal and seasonal perspectives. Overall, this study shows how some hydrological characteristics are associated with the performance of deep learning water level forecasting models and can help experts identify deep learning models’ capabilities and limitations, while encouraging researchers to develop strategies to mitigate their shortcomings and further develop the hydrological modelling toolkit to facilitate the understanding and response to hydrological processes.