The dynamic characteristics of the natural gas engine combustion process are difficult to describe, highly nonlinear, and subject to unknown disturbances, resulting in the low control precision of feedback control variable of combustion process. To address the above issue, an improved model-free adaptive controller (MFAC) was proposed. Firstly, a natural gas engine system model was established and its effectiveness was verified, and the combustion midpoint (CA50) was selected as the feedback control variable of the combustion process. The traditional MFAC was obtained by establishing a discrete partial form dynamic linearization model, using the system input and output data to estimate the pseudo partial derivatives of the ignition timing and air-fuel ratio during the combustion process. The MFAC continuously tracked the expected state updated online to realize the tracking control of CA50. Furthermore, in order to take into account the dynamic convergence rate and tracking accuracy of CA50, a fast convergence term based on the differential of the tracking error was introduced into the traditional MFAC, and strict stability proof and convergence performance analysis were provided for the closed loop system. Finally, two test experiments were conducted to compare the proposed controller with the PID controller, and the results showed that the proposed improved MFAC made significant improvements in robustness, convergence rate and tracking accuracy.