Abstract: The active noise reduction technology for low-frequency noise of induced draft fans in power plants is based on a simple noise model assumption. The collected noise signal is used as input to reduce the original noise. However, the complex and variable noise characteristics result in less noise reduction in the model. Therefore, research is being conducted on active noise reduction technology for low-frequency noise of induced draft fans in power plants based on generative adversarial networks. Given the complex sound field of induced draft fans in power plants, low-frequency noise exhibits non stationarity and frequency band concentration characteristics. It is necessary to deploy a multimodal sensor network to obtain signals and preprocess them through frequency domain selective filtering and multi-scale denoising to output high-quality data. Based on generative adversarial networks to extract noise features, a conditional deep convolutional generative adversarial network is constructed. The generator adopts an encoder decoder structure, and the discriminator is a convolutional neural network. Through adversarial training, the generator learns to generate high fidelity noise features under given operating conditions. Using a trained generator to generate matched inverse sound waves for active noise reduction, a long short-term memory network module is introduced to track the slow time-varying characteristics of noise. The cancellation signal is input into a multi-channel filtering minimum mean square algorithm adaptive controller to drive the secondary sound source to emit cancellation sound waves, which interfere with the primary noise in the target area. The experimental results show that in multi condition testing, the noise reduction of this method reaches 24.7dB, 20.1dB, and 16.5dB in conditions A, B, and C, respectively, far exceeding the comparison methods; After applying this technology, the peak sound pressure level of low-frequency noise in the induced draft fan of the power plant decreased from nearly 90dB to around 25dB, proving that the method has significant noise reduction effect and broad application prospects.