Objective:

Forecasting agricultural water demand in the Hirmand Basin is essential for water-resources planning because of its arid climate, inflow fluctuations, and dependence on agriculture. This study compared the performance of ANN, BP, LSTM, and RNN models and identified the main factors influencing agricultural water demand in the Sistan region.

Method:

Agricultural water-demand data, climatic and hydrological variables, area, and their lagged values were used in univariate and multivariate models. Model performance was evaluated using forecasting accuracy metrics, and input-variable importance was assessed using the SHAP method.

Results:

The multivariate BP model effectively represented the nonlinear behavior of agricultural water demand by integrating climatic, hydrological, area, and temporal-lag variables. Forecasts through 2030 indicated a 24.35% increase in demand at the beginning of the forecast period, with demand remaining above the baseline thereafter. The highest demand was predicted during active agricultural months and peak-consumption periods. SHAP analysis showed that maximum wind speed, agricultural land area, evaporation, and lagged water-demand variables collectively accounted for 84.5% of total input-variable importance.

Discussion:

The multivariate BP modelcan support water-allocation planning and the identification of high-risk demand periods. The findings also highlight the importance of climatic conditions, area, and temporal dependence in regional water-resources management.