This paper proposes a semisupervised deep neural regression network with embedding manifold (SSE-DNN) for soft sensor modeling that integrates manifold embedding into deep neural regression networks. Manifold embedding is imposed on the hidden layer of the deep neural regression network to form a semisupervised deep neural regression network. Manifold embedding exploits the local neighbor relationship among industrial data and utilizes unlabeled data effectively to improve the performance of the deep neural regression model. The SSE-DNN model exploits the global information and local manifold among industrial large data simultaneously and implements implicitly multimodal models of industrial process. The soft sensor model based on the SSE-DNN is applied to the estimation of total Kjeldahl nitrogen (TKN) in a long-term complicated wastewater treatment process. The experimental results demonstrate that the SSE-DNN model has a better performance than other soft sensors and provides an effective method for soft sensor modeling of complex industrial processes.