Abstract: For hematite grinding processes, it is not only difficult to describe the dynamics between grinding particle size (GPS) and control loop outputs by using mathematical formulas, but also impossible to measure the GPS online. Moreover, the process is sensitively influenced by the frequent fluctuation of ore compositions and properties. The above practical problems make the existing control methods unable to be applied. In this paper, a data-driven optimal operational control approach, which consists of loop pre-setting optimizer, performance index estimator, optimum setpoints critic, and GPS soft-sensor, is proposed by using neural network. In this approach, the GPS is estimated by the GPS soft-sensor first, then the loop pre-setting optimizer generates the optimum loop pre-setting values by making performance index estimate close to its optimal value. Finally, via the optimum setpoints critic the optimal setpoint is obtained. With the help of the control loops, the GPS can be maintained inside its desired range, and close to its target value as much as possible. Experiments using a developed research platform of optimal operational control are carried out with real time data to show the effectiveness of the proposed method.