Drilling optimization is essential in constructing an enhanced geothermal system (EGS) and can be facilitated through predicting the rate of penetration (ROP). The ROP evolution along the depth was forecasted by considering the current and previous ROP values as input to a gated recurrent unit (GRU)-based deep learning model. Drilling data was obtained from two geothermal wells in Pohang, South Korea. Multiple data configurations for training and testing were designed from both wells. The proximity of the training section to the target results in improved accuracy in prediction (MAPE smaller than ~ 3%). Furthermore, larger depth spans of ROPs used for training resulted in better prediction outcomes. The model trained with the entire dataset from an adjacent well exhibited well-predicted ROP values for a new drilling hole (MAPE smaller than 5–10%). From the multiple-step forecasting analysis, the error tended to sharply increase as the number of predicted ROP values increased despite a large number of the input sequence (MAPE larger than 20%). Incorporating other drilling data besides ROP evolution did not improve the prediction. By predicting ROP evolution along the depth, the GRU-based model can assist operators in optimizing drilling processes and preparing for upcoming scenarios. The model can serve as a valuable tool for enhancing drilling efficiency and effectively managing operational uncertainties.