In recent years, the analysis and improvement of temperature characteristics of Si-based capacitive accelerometers has received considerable research attention in the field of Microelectromechanical system (MEMS) sensors. Generally, the influence of temperature on the accelerometers can be mitigated by optimizing the structural design and compensating the output signal. Herein, the output characteristics of an accelerometer designed with asymmetrically arranged combs were analyzed under various temperatures. The purpose of this paper is to improve the temperature drift of scale factor (TDSF) of MEMS capacitive accelerometer, using the asymmetric layout structure to improve the TDSF fundamentally, and the least square method to achieve temperature compensation efficiently. The variations in the TDSF were compared for the symmetric and asymmetric structures. In addition, we modeled the accelerometer with an asymmetric structure for simulations to analyze the errors resulting from the electrostatic torsion phenomenon induced by the asymmetric structure. Moreover, a temperature compensation model was developed for the scale factor of the accelerometer, which was validated and verified with the data obtained from simulations and experiment. Furthermore, an accelerometer based on silicon on insulator was fabricated and tested to verify the simulation results and the compensation effects. According to the results, the scale factor of the studied accelerometer was 171.83 mV g−1 and the average value of the TDSF was 83.56 ppm °C−1 Overall, the experimental results were almost consistent with the simulation results. Under the asymmetric layout, the scale-factor stability improvement of the accelerometer could reach up to 86.96%, and the error caused by electrostatic torsion was ∼2.93%, which is relatively negligible. After compensation, the range and standard deviation of the scale factor of the accelerometer with respect to temperature were reduced by 94.46% and 95.69%, respectively, and the average value of TDSF was reduced by 95.90%, which verified the effectiveness of the compensation model.