In this study, a multi-layer perceptron model (EL-NEP-MLP) combining ensemble learning and noise-enhanced prediction is proposed to improve the prediction accuracy and stability of gas concentration in photoacoustic spectroscopy (PAS) under complex noise environments. In practical applications, PAS signals are susceptible to various noises, especially in low-concentration gas detection. The multi-layer perceptron (MLP) is used for photoacoustic signal feature extraction, ensemble learning and noise-enhanced prediction to improve system robustness and generalization capability. The experimental results show that EL-NEP-MLP can predict the gas concentration quickly and accurately under complex noise conditions.