Data Enhancement for Data-Driven Modeling in Power Plants Based on a Conditional Variational-Adversarial Generative Network

Abstract

The quality of data directly affects the performance of data-driven models that are developed to monitor the operation of power plants. Constrained by external instructions and other factors, the operational data of a power plant generally present an irregular distribution, which leads to data imbalance. In this paper, a novel data generation method called a conditional variational autoencoder and generative adversarial networks (CVAE-GAN) is proposed. The original samples are first mapped to the latent variables by the encoder network in the CVAE. On combining with the conditional variables, the latent variables are then mapped to a new sample space by the generation network of the proposed model. The CVAE-GAN model trained with the original imbalanced dataset can generate a uniformly distributed dataset under the guidance of incremental L 2 -discrepancy. A numerical simulation case is given to validate the superiority of the proposed model over other common generative models. Then, the proposed model is applied for NO x emission prediction in a coal-fired power plant. The results demonstrate that the accuracy of NO x emission predictions is improved by training the model with the dataset generated by the proposed method.

RegGAN: A Virtual Sample Generative Network for Developing Soft Sensors with Small Data

“…Furthermore, the effectiveness of the generative network’s training progress is evaluated through its loss function, with a decrease in loss corresponding to an increase in generative capabilities. Meanwhile, instead of weight clipping, which can limit performance and reduce stability, WGAN introduces a gradient penalty to enforce Lipschitz constraints . With the inclusion of the gradient penalty, the loss function of WGAN is updated as follows L = E x ∼ P g false[ F ( x ) false] − E x ∼ P r false[ F ( x ) false] + λ E x ∼ P x̃ false[ false( ∥ normal∇ x F ( x ) ∥ 2 − 1 false) 2 false] where λ serves as the weighting factor for the gradient penalty term, ∇ x F ( x ) is the gradient of the discriminative network, and P x̃ signifies the probability distribution between the virtual and original data.…”

“…Meanwhile, instead of weight clipping, which can limit performance and reduce stability, WGAN introduces a gradient penalty to enforce Lipschitz constraints. 40 With the inclusion of the gradient penalty, the loss function of WGAN is updated as follows where λ serves as the weighting factor for the gradient penalty term, ∇ x F ( x ) is the gradient of the discriminative network, and signifies the probability distribution between the virtual and original data. The inclusion of the gradient penalty technique effectively controls the gradient, thereby mitigating problems associated with gradient vanishing and exploding.…”

“…Meanwhile, instead of weight clipping, which can limit performance and reduce stability, WGAN introduces a gradient penalty to enforce Lipschitz constraints. 40 With the inclusion of the gradient penalty, the loss function of WGAN is updated as follows…”

Digital Restoration of Historical Buildings by Integrating 3D PC Reconstruction and GAN Algorithm

“…It was used to solve the data imbalance caused by irregular distribution of power plant operation data. The research results indicated that utilizing the dataset from the model for model training could improve the accuracy of NOx emission prediction [15].…”