An SF6 arc extinction sensor (AES) has the advantages of wide measurement, high sensitivity, and strong anti-interference ability, and has a wide range of applications in high-voltage substations. To effectively monitor and control SF6 gas in substation mechanical equipment, we have designed an SF6 AES based on non-dispersive IR (NDIR). However, in the actual measurement, temperature and air pressure differences in the environment affect the detection accuracy of the device, and an appropriate method of eliminating the measurement error caused by changes in the environment is required. In this paper, we propose the use of a gray wolf optimization-radial basis function (GWO-RBF) neural network to compensate for the measurement error caused by temperature and pressure changes. The experimental results show that the SF6 concentration error after the GWO-RBF algorithm is ±15 ppm in the concentration range of 0–2000 ppm and the full-scale error is 0.75%. Compared with uncompensated data and radial basis function (RBF) compensation methods, the proposed GWO-RBF algorithm effectively enhances the measurement accuracy and stability of the AES, allowing its volume and cost to be reduced.