Abstract: The development and application of high-precision gas sensors in industrial production has become an important development direction in the field of environmental monitoring technology. However, the performance of existing CO₂ sensors is significantly affected by ambient pressure under actual working conditions. In this study, we focused on two core issues by constructing a pressure compensation structure, carrying out experimental verification, and designing algorithmic compensation: the variation law of sensor detection accuracy with ambient pressure and the optimal pressure compensation CO₂ volume fraction prediction method, aiming to effectively improve the detection accuracy and operational stability of CO₂ sensors. First, a pressure compensation test platform for CO₂ sensors was built, and the test procedures as well as corresponding analysis methods were proposed. Second, CO₂ with volume fractions ranging from 2% to 18% was tested under ambient pressure of 80-120 kPa. The results show that for CO₂ gas with the same volume fraction, the greater the deviation of ambient pressure from the standard atmospheric pressure (101 kPa), the larger the deviation between the measured and true volume fractions. Meanwhile, the measured volume fraction increases slowly with the rise of ambient pressure, and the measured value is closest to the true value at 100 kPa. Third, the wavelet neural network-long short-term memory (WNN-LSTM) model and the horned lizard optimization algorithm-least squares-wavelet neural network-long short-term memory (HLOA-LS-WNN-LSTM) model were adopted to correct the test data before pressure compensation. The results indicate that after compensation by the optimized model, CO₂ gas with the same volume fraction can resist the influence of ambient pressure variation, and the measured volume fractions are highly consistent with the true values. In addition, it is difficult to directly judge the performance of the two pressure compensation algorithms only through error analysis, the ambient pressure, measured volume fraction, standard gas volume fraction and models were further integrated to compare their prediction performance. The results show that the HLOA-LS-WNN-LSTM model features higher prediction accuracy, better stability and is less affected by changes in ambient pressure. In summary, the variation law of CO₂ sensor detection accuracy with ambient pressure is clarified, and an optimal pressure compensation concentration prediction method is established. Using test data of 10% CO₂ standard gas, CO₂ high-precision concentration prediction is realized through small-sample training, which effectively guarantees the detection accuracy and stability of CO₂ sensors and provides technical support for accurate CO₂ volume fraction monitoring.