Over the last two decades, Artificial Neural Networks (ANNs) have been increasingly used to predict watershed responses due to their effectiveness in modeling complex precipitation-runoff phenomena. In this research, a multi-layer ANN model named GANN, based on watershed geomorphology, was created for long-term daily runoff simulation in a watershed upstream of a reservoir dam. A continuous genetic algorithm (CGA) was utilized for model training. To showcase the GANN model's performance in simulating the precipitation runoff process, the well-known hydrological model HEC-HMS was calibrated as a reference model using daily recorded time series data from the study basin. Geomorphological characteristics of the basin, such as the number of stream orders, average upstream area, and average stream length, were directly incorporated into the ANN structure. The model was trained using a sequence of previous time-step precipitation and runoff as input variables and current flow as the output. The GANN model was applied to predict daily runoff in the Golpayegan watershed, situated in a semi-arid region of Iran. The number of potential flow paths in the watershed determined the number of neurons in the hidden layer of the GANN structure, which remained constant during modeling. Furthermore, the calculated stream path probabilities in this watershed were used as connection weights between the hidden layer and the output in the GANN structure. The findings indicate that utilizing the current time-step of rainfall (Pt) and flow at the previous time-step (Qt-1) as input variables for GANN yields the best performance in simulating daily flow compared to other GANN patterns and the HEC-HMS model. Integrating an ANN model with the geomorphological features of the watershed and an efficient metaheuristic optimization algorithm (e.g., CGA) offers computational efficiency and is suitable for daily runoff simulation in semi-arid regions.