Enhancing Dam Safety and Management: Long Short-Term Memory Based Predictive Models for Accurate Alert Forecasting

Abstract

Dam management and early alert systems are critical for effective water resource management. Accurate prediction of dam alert signals facilitates proactive decision-making, thereby aiding in the effective management and reduction of potential risks linked to dam operations. Within this research, Long Short-Term Memory (LSTM) networks are utilized to forecast dam alert signals issued from the dam by leveraging daily parameters, including temperature, dew point, humidity, and other pertinent factors. The study utilizes a dataset of the Malampuzha Dam spanning 10 years, comprising various inputs and the corresponding alert levels. Our objective is to demonstrate the effectiveness of LSTM models in accurately predicting multi-level alert classifications. This is the first application of LSTM for multi-tiered dam alert classification in the Indian context. The LSTM model was trained using optimizers such as Adam, RMSProp, Stochastic Gradient Descent, Adagrad, and Nadam, using learning rates of 0.01, 0.001, and 0.0001, as well as epochs of 50, 100, and 500, and gradient clipping values of 0.5 and 1.0. Evaluation metrics including RMSE (Root mean square error), NSE (Nash-sutcliffe Efficiency), R-squared, and accuracy are employed to assess the model's performance. The LSTM model using the Nadam optimizer achieved high accuracy (99.13%). It was also observed that as the learning rate decreased, the model's accuracy decreased. An appropriate gradient clipping value is found to be 0.5 for the LSTM model.