A Hybrid CNN-LSTM Approach for Monthly Reservoir Inflow Forecasting

Reservoir modeling and inflow forecasting has a vital role in water resource management/controlling. Hydrological systems’ complex nature and problems in their application process have prompted researchers to look for more efficient reservoir inflow forecasting methods; hence, the development of artificial intelligence-based techniques in recent years has caused the hybrid modeling to become popular among hydrologists. To this end, effort has been made in the present study to develop a hybrid model that combines a Long-Short Term Memory (LSTM) algorithm—a special recurrent neural network—with a Convolutional Neural Network (CNN) algorithm for the reservoir inflow forecasting. To forecast the flow data, use was made of the support vector machines (SVM), Long Short-Term Memory (LSTM) algorithm, adaptive neuro-fuzzy inference system (ANFIS), Variable Infiltration Capacity (VIC) and autoregressive integrated moving average (ARIMA) model plus the data collected from the flow measurement stations of Doroodzan Dam reservoir in “Kor”—an important river in Fars Province, Iran. The model estimation results were evaluated by the RMSE, MAE, MAPE, MSE and R2 statistical criteria and showed that the hybrid CNN-LSTM method was the most successful model by achieving R2 ≈ 0.9278 (the highest).