The prediction and warning of total ammonia nitrogen (TAN) in aquaculture are crucial. Current optimizations of TAN prediction models primarily rely on improved algorithms. However, various management strategies, such as feeding strategies, may be implemented during the aquaculture production process, potentially affecting the prediction performance of these models. This study aims to model and predict TAN concentrations in recirculating aquaculture system (RAS) and investigate the impact of different feeding strategies on the prediction performance of TAN models. We measured 7 water quality parameters in the tanks of Oplegnathus punctatus within an RAS. The data were pre-processed using principal component analysis (PCA) or Pearson correlation analysis, resulting in three datasets: the original dataset (OD), the Pearson dataset (PearsonD) and the PCA dataset (PCAD). These datasets were then integrated with random forest (RF), back propagation neural network (BP), gated circulation unit (GRU), and long short-term memory (LSTM) to forecast TAN concentrations under two distinct feeding strategies. The performance of the models was evaluated using root mean square error (RMSE), mean square error (MSE), mean absolute error (MAE), and R-square value (R²-score). The RF modle exhibited the poorest prediction performance. The GRU and LSTM models demonstrated acceptable and stable prediction accuracy, while the accuracy of the BP model varied. Optimal prediction models differed under the two feeding strategies: under artificial feeding, Pearson correlation analysis combined with BP yielded higher accuracy, wheras under automatic feeding, Pearson correlation analysis combined with GRU performed better. Pearson correlation analysis combined with LSTM or GRU, and PCA combined with LSTM, showed superior performance throughout the experimental period. Compared to the automatic feeding strategy, under artificial feeding, the RMSE, MSE and MAE of the models built with Pearson correlation analysis and LSTM decreased by 0.007 2, 0.001 9 and 0.003 6 respectively, while the R²-score increased by 0.107 5. Similarly, the RMSE, MSE and MAE of the models built with Pearson correlation analysis and GRU decreased by 0.003 0, 0.000 8 and 0.003 0 respectively, and the R²-score increased by 0.082 6. Feeding strategy significantly influences the prediction accuracy of TAN models, and Pearson correlation analysis combined with GRU or LSTM could be employed to predict TAN in RAS effectively. This study provides a reference for the optimization of ammonia nitrogen prediction technology in RAS.