Applying deep learning models to predict short-term water levels in Lake Vesijärvi in Lahti, Finland

Abstract

The water level of a lake impacts both its ecological systems and surrounding human societies. With climate change intensifying extreme weather events, fluctuation patterns of lakes’ water levels could be affected and force changes in water resource management and natural disaster control. Therefore, accurate and stable water level forecasting models can greatly help regions where a lake is a significant part of societal activities. The emergence of machine learning technologies and, particularly, deep learning technologies, offers a great opportunity to enhance estimation accuracy and shorten development time with reasonable resources. Current literature indicates limited adoption of deep learning models for lake water level forecasting in Finland as of this writing. Since the deep learning models’ performance can differ greatly due to the lakes’ unique hydrological characteristics, a universal forecasting model may not be realistic. Therefore, conducting systematic experiments is necessary for developing accurate models of a particular lake. However, it is feasible to create a research methodology that can be readily replicated in other lakes and even applied to different fields where deep learning-based time series forecasting is relevant. In this study, Long Short-Term Memory and Gated Recurrent Unit were chosen to perform 1-day, 3-day, and 7-day predictions for the water level of Lake Vesijärvi in Lahti, Finland. The results of this research were promising, all trials achieved a Nash-Sutcliffe Efficiency above 0.95 and a Root Mean Squared Error below 0.025. Long Short-Term Memory and Gated Recurrent Unit showed similar performance whereas the best models for each experiment group are based on single-layer bidirectional Gated Recurrent Unit structures. Thus, the methodology developed in this study proves effective and holds potential for various time series forecasting applications.