Investigation of the effect of climatic parameters in machine learning algorithms for streamflow predicting in Hamoon Helmand Catchment, Iran

Abstract

The objective of this study is to identify the effective input parameters for estimating streamflow using an M5 model tree and Genetic Algorithm (GA) and Long Short-Term Memory (LSTM) and to propose a dependable model. These methods were chosen for their ability to model complex relationships, high prediction accuracy, and efficient input optimization capabilities. The study utilized monthly data of rainfall, temperature, evaporation, and streamflow for the latest month (Q_t-1) and the preceding 2 months (Q_t-2) in the Hamoon Helmand catchment. Five scenarios were employed in M5 and the linear and nonlinear models of GA and LSTM. The models’ performance was assessed using statistical parameters such as RMSE, MAE, R, and NSE. In the initial scenario where all five parameters were considered, M5, the linear and nonlinear GA models and LSTM produced the most accurate results, with RMSE values of 9.27, 6.87, 5.54, and 5.55, respectively. The second scenario (rainfall; maximum, minimum, and average temperatures; Q_t-1 (discharge for 1 month before) and evaporation) revealed that the linear and nonlinear GA models, with RMSE values of 7.21 and 6.55, respectively, were more accurate than M5 and LSTM with an RMSE value of 8.58 and 6.78, respectively. In Scenarios 3 (rainfall; average temperatures; evaporation; and Q_t-1, Q_t-2 (discharge for 1, 2 months before)) and 4 (rainfall; average temperatures; Q_t-1, Q_t-2 (discharge for 1, 2 months before)), LSTM demonstrated superior performance. The results obtained from Scenario 5 (rainfall; maximum, minimum, and average temperatures and evaporation) indicate that in the absence of sufficient runoff data in a basin, there is no necessity to employ a nonlinear model; instead, modeling with an M5 model tree yields sufficiently accurate results. This research demonstrates the global significance of optimizing water resource management models in arid and climate-sensitive regions and contributes to the resilience and sustainability of resources in the face of climate change.