2025, Vol. 6, Issue 1, Part A

Accurate long-term prediction of reservoir water levels is critical for sustainable water management in monsoon-dominated regions. This study presents a hybrid machine learning framework integrating XGBoost for feature optimization and Long Short-Term Memory (LSTM) networks for sequential forecasting. The reservoir water balance is modeled as

V_(t )= V_(t -1)+ I_t - O_t - E_t - S_t,h_t=(f-1)(V_(t )/A_(t ) ) (1)

Trained on monthly data (January 2010-December 2024) from the Public Health Engineering Department (PHED), Ranchi, the model achieves a validation MSE of 0.0078 and forecasts water levels to December 2032 with a seasonal pattern correlation of 0.95. The hybrid approach outperforms standalone ARIMA, LSTM, and ANN models in capturing non-linear monsoon dynamics, with RMSE = 0.28 ft-in and NSE = 0.96. Results demonstrate stable long-term levels (~2121.75 ft-in) and reliable extreme event prediction, supporting flood and drought mitigation at Gonda Dam.