Abstract: In order to solve the problems of multiple factors, complex mechanism, nonlinear, strong coupling and other characteristics of rate of penetration in the process of complex deep drilling, this paper proposes a method of ROP prediction based on multi-dimensional time series long short-term memory neural network. Based on the actual logging data of Well X-1 in Xinjiang, this paper selects engineering parameters, hydraulics parameters, lithology parameters and other parameters as model inputs through the feature correlation analysis and the physical meaning of traditional ROP equation, and analyzes the ROP prediction effect of multi-dimensional features under different parameter combination modes. The results show that ROP is positively correlated with rotational speed, torque and riser pressure, but negatively correlated with vertical depth, hook load, specific weight on bit, drilling fluid conductivity, outlet drilling fluid temperature, outlet displacement and PDC drill ability, while weakly correlated with pump stroke and drilling fluid density; different characteristic parameter combinations have different prediction accuracy on ROP, among which the optimal parameter combinations are specific weight on bit, rotational speed, torque, hook load, riser pressure, drilling fluid conductivity, outlet drilling fluid temperature, outlet displacement, vertical depth and PDC drillability extreme value. MAE of ROP prediction is 0.30 m/h, MAPE is 11.35% and R² is 0.93; The optimal hyperparameter combination scheme of the model was determined by orthogonal experiment. The R² was increased by 0.06.

Key words: ultra-deep well, rate of penetration, LSTM, hyperparameter combination, orthogonal experiment