To improve tensile force control performance by keeping a parameter of a feedforward compensator in a proper value by a neural network containing an inverse model of a tensile force system.

A dynamic characteristic of a tensile force control loop is expressed by a discrete-time mathematical model, and an inverse model of a ternsile force control system is found by using the mathematical model, and a neural network is learnt by deviation between a tensile force target value 19 and output and output of a tensile force feedback compensator 18. A feedforward compensating element to adjust respective parameters of the inverse modle is arranged in a loop connecting the tensile force desired value 19 and a speed control part of a driving roller, and tensilt force of a web between two rollers 2 and 3 is controlled.