Time series prediction of reservoir water level using deep learning: the case of Bui Dam in Ghana

Accurate reservoir water level monitoring is crucial for hydro dam operations, impacting intervention plans and activities. Ghana faces climate change, population growth, and increased power and water consumption, causing frequent power outages and water shortages. Therefore, there is a need to implement a reliable, simple and cost-effective method to simulate water level fluctuations in the dam reservoirs. This paper examines deep learning algorithms with attention mechanisms to predict short- and long-term reservoir water levels. Two models— attention-based LSTM (AT-LSTM) and Transformer are compared against the standard long short-term memory (LSTM) for univariate water level forecasting. Historical daily water level data from 2014 to 2023 is used, with missing data accounted for through linear interpolation. The results indicate that the Transformer model outperformed  AT-LSTM and LSTM based on these statistical indexes (i.e., R2, NSE, KGE, RMSE, MAPE, and MB) and hypothesis testing. Given the Bui reservoir’s role in hydroelectric power generation, accurate water level prediction could improve revenue and energy production through better water management. The Transformer model’s precision makes it suitable for estimating the water level of the Bui reservoir and likely applicable in other settings.