The accuracy of sensor measurement has an important influence on energy efficiency and control characteristics of air conditioning systems. However, the sensor will inevitably have errors due to sensor installation location or aging in the practical system. In previous studies, the sensor virtual in-situ calibration (VIC) method is proposed to diagnose and calibrate sensor faults based on physics-based methods with plenty of sensors. However, the basic sensors are only installed due to the cost and limited installation space in the practical system, which restrains this method. To address this problem, the autoencoder (AE) based on the internal relationship between sensor variables is applied to establish the sensor model to achieve calibration. This paper studies the applicability of the AE that constructs sensor models with limited sensors, and the accuracy of the VIC method under different working conditions. The results show that the AE can well realize the sensor model construction with high accuracy, which lays a foundation for the implementation of the VIC method. Meantime, the accuracy for different scenarios of the single fault is more than 90.75% after calibration, the concurrent fault can also reach more than 88.5%. The supply air temperature calibration will make the continuous regulating valve keep the normal opening to maintain the air volume and ensure the stability of indoor temperature. The static pressure calibration will increase the energy efficiency of the supply fan by 17%, make the VAV box adjust normally, and avoid the disorder of the control loops. The differential pressure sensor calibration will increase the energy efficiency of the water pump by 21%, and ensure normal water flow. Meantime, the calibration of three sensors reduces the time to reach a stable state by more than 20.6%. The VIC method based on AE makes the application scenario of sensor fault calibration more extensive and universal.