Aiming at the problem of outliers, noise and irregular disturbances prevailing in the monitoring parameters of the thermal system, a noise reduction method for monitoring parameter of the thermal system based on median regression empirical mode decomposition (MREMD) and variational mode decomposition (VMD) is proposed. The purpose is to enhance the accuracy of monitoring system regulation and the level of system operation management, while minimizing noise and disturbances in the monitoring parameters, all while preserving as much of the original data’s effective information as possible. The method firstly performs MREMD of the monitoring parameters to obtain a number of intrinsic mode functions (IMF). Secondly, chaotic time series analysis is applied to filter out the IMF components containing noise using permutation entropy, reconstructing them as the noise portion of the original data. Then the noise part is decomposed by VMD, and the optimal envelope entropy of the IMF obtained by the decomposition is used as the fitness function. The northern goshawk optimization (NGO) algorithm is used to optimize the VMD decomposition parameters, yielding the IMF with the lowest envelope entropy within the optimization range, which contains the effective information of the noise portion. Finally, this part was reconstructed by summing with the low frequency IMF component and residual component obtained by MREMD decomposition which both are contained trend information, to obtain the monitoring signal after noise reduction. The results demonstrate that through case studies, the modal double-decomposition noise reduction method proposed in this paper has highest signal-to-noise ratio and lower information entropy and power spectral entropy compared to mainstream wavelet threshold denoising methods and moving average filtering techniques.