Comparison of N-BEATS with Standalone and Hybrid Deep Learning Models in Monthly Inflow Forecasting to the Aras Dam Reservoir: A Feature Selection Analysis

Reservoir dams play a pivotal role in water resource management. Accurate prediction of inflow to reservoirs significantly enhances operational performance. While standalone artificial intelligence methods have recently been frequently used to predict inflow, hybrid models have shown quite more satisfactory success. In this study, various deep learning models, including MLP, GRU, LSTM, CNN, CNN-MLP, CNNGRU, CNN-LSTM, CNN-GRU-MLP, and CNN-LSTM-MLP, were utilized to predict the monthly inflow to the Aras reservoir in the Azerbaijan-Iran region. The results were compared with the Neural Basis Expansion Analysis for Time Series Forecasting (N-BEATS) model for univariate forecasting and the NBEATSx model for multivariate forecasting using a monthly inflow time series dataset. To enhance prediction accuracy, the hyperparameters of the models were optimized. Additionally, to evaluate the impact of feature selection on model performance, five different scenarios were developed as combinations of input variables for forecasting one future time step. The evaluation metrics revealed that among the scenarios, Scenario 5 (comprising lagged inflows at months 1, 11, and 12; lagged average monthly precipitation in the upstream basin at months 1 and 12; the solar month counter; and a three-month moving average of monthly inflow) yielded the best results. Among the models, the hybrid CNN-LSTM-MLP demonstrated the highest prediction accuracy. Specifically, the performance metrics for this model and the best scenario included MAE, RMSE, PBIAS, R², KGE, and NSE, which were 8.78 m³/s, 12.95 m³/s, 1.5%, 0.89, 0.91, and 0.89, respectively. Conversely, the NBEATSx model exhibited suboptimal performance, with reduced accuracy as the number of input features increased, although the N-BEATS model performed well in univariate forecasting. This study highlights the high potential of hybrid deep learning models in accurately forecasting reservoir inflows and underscores their utility in enhancing water resource and reservoir operation management.