Aiming at the fact that the measurement accuracy of nondispersive infrared SF6 gas sensor is susceptible to ambient temperature and air pressure, PSOBP neural network prediction model was used for realtime compensation for the measurement deviation caused by changing ambient temperature and air pressure. This method was then compared with other compensation approaches. The experimental results show that the relative measurement error of this SF6 gas sensor is 1.2%, the measurement accuracy is less than ±15×10-6 and the measurement resolution is 1×10-6 at the gas concentration of 0~1 000×10-6, the temperature of 10~40 ℃ and the air pressure of 100~120 kPa. It could effectively eliminate the nonlinear effect of fluctuating ambient temperature and air pressure. Compared with the empirical formula method and RBF neural network compensation method, this approach has higher measurement accuracy and stability, and dispenses with more circuit control module. So, it can help to reduce the volume and the cost of sensor.