The invention provides a reservoir production real-time optimization method based on an improved random disturbance approximation algorithm and belongs to the field of reservoir numerical simulation and optimal control. The method includes (1) giving initial control variables, boundary constraint and economic parameters; (2) supposing an iterative number k = 0, and performing numerical simulation computation to acquire a current NPV value; (3) generating Gauss type perturbation vectors according to spherical model covariance arrays, and computing the random disturbance gradients and an average value thereof; (4) updating the control variables; (5) performing reservoir simulation computation according to the updated control variables to acquire an updated NPV value; (6) judging whether the updated NPV value is larger than the current NPV value or not, if so, continuing the step (7), if not, halving the iteration step Lambda k, and returning to the step (4); (7) judging whether the convergence condition is met or not, if so, continuing the step (8), if not, setting the K = k+1, and returning to the step (3); (8) outputting the optimal control variable, and stopping the computation.