Performance evaluation of the LSTM Model forecasting daily inflow into dams reservoirs

Real-time forecasting of daily inflows to reservoirs with a prediction horizon that extends several steps into the future is crucial for water resource planning and management. Despite numerous studies on inflow prediction using machine learning methods, few studies have investigated the predictive capabilities of these approaches with long lead time (several steps ahead) or gained insights through systematic comparisons of model predictive performance in the short term. In this study, the daily inflow to the Seimareh reservoir was predicted for the next 7 days using the Adaptive Neuro-Fuzzy Inference System (ANFIS) and the Long Short-Term Memory (LSTM) network. For this purpose, daily data on precipitation, temperature and inflow to the Seimareh reservoir from 2012 to 2018 were used for modeling. The results showed that the performance of the LSTM model was better than that of ANFIS in the daily forecast in several steps. Specifically, the maximum and minimum values of the Nash coefficient in the forecast horizon for the next seven days were 0.971 and 0.628 for the LSTM model and 0.858 and 0.393 for the ANFIS model, respectively. The optimal setting of the parameters, including the number of neurons in each layer, the number of epochs and the stack size in the LSTM model, is the key to the model's high potential to predict the inflow for the next seven days. Finally, the performance of the LSTM model in predicting the inflow to Seimareh during the 2019 flood was evaluated and it was found to predict flood discharges with acceptable accuracy up to the forecast horizon of the next seven days. These results indicate that the LSTM model is suitable for forecasting daily inflow and can help make strategic decisions in water resource management, especially under flood conditions.