Proactive dam operation based on inflow prediction by modified long short-term memory for improving resilience

Abstract

The operation of dams can have a significant downstream impact, causing major flood damage. For efficient dam operation, it is essential to accurately predict the inflow using deep learning such as a long short-term memory (LSTM). The modified LSTM (MLSTM) with new optimizers combined with a metaheuristic optimization algorithm was suggested. The MLSTM were trained using rainfall, water level, and discharge of the Yongdam Dam from 2010 to 2019 as input data and the inflow of the Daecheong Dam as output data. Normalization and time lagged cross correlation (TLCC) were applied as data preprocessing. The inflow of the Daecheong Dam in 2020 was predicted using rainfall, water level, and discharge of the Yongdam Dam. Compared to LSTM, MLSTM showed an accuracy improvement of about 63.85–83.34% based on the mean square error (MSE), which is a loss function. The results of proactive dam operation were calculated based on the predicted inflow considering TLCC. Compared to the existing operation, the peak water level of the reservoir was reduced from 77.54 m to 77.17 m, and the resilience was improved from 0.9225 to 0.9623. New dam operation can maintain low water levels of reservoir and improve resilience to prevent flood damage.