Municipal wastewater treatment processes are non-stationary, leading to degradation of the performance of trained soft measurement models. To address this problem, a deep online transfer learning (OTL) framework is developed in this article. First, offline modeling is carried out by long short-term memory networks (LSTM), and the temporal information in the data is fully extracted. Second, first determine whether the model needs to be reconstructed in the online phase through multiple kernel variants of maximum mean discrepancy (MK-MMD), then determine whether the model needs to be fine-tuned through prediction error, and complete the online update of the model. Then, the final prediction results are obtained by combining the models with different knowledge through ensemble learning. Utilizing data from BSM1, simulation outcomes demonstrate the superiority of the introduced methodology over prevailing approaches, evidencing its capability to achieve elevated precision in forecasting within the context of the erratic wastewater treatment process.