In the process of wastewater treatment, carbon sources have an important impact on denitrification, so accurate addition of carbon sources is of great significance to improve the quality of effluent from wastewater treatment, reduce costs and increase efficiency. In the past, most of the studies on carbon source dosing considered traditional mathematical calculation methods or manual experience, which lacked accuracy. In recent years, the carbon source dosing system of wastewater treatment has been gradually studied based on intelligent algorithms, but the established neural network has problems such as low efficiency, poor robustness and complex structure. Therefore, in order to solve the problem of low-carbon nitrogen ratio of influent water quality of domestic wastewater treatment plants in China, the Gate Recurrent Unit model and channel attention mechanism were incorporated into the carbon source dosing prediction model, and a new carbon source dosing prediction system for wastewater treatment plants was proposed. In this study, carbon source amount of Anaerobic-Anoxic-Oxic process of Rizhao Third Sewage Treatment Plant is taken as the research object, and the training verification was carried out based on the artificial neural network on the basis of multiple real-time monitoring data of influent water quality. The feasibility and robustness of this method are verified through the case study, which provides theoretical support for the carbon source dosing process of wastewater treatment.