Abstract: To address the challenge of accurate real-time effluent ammonia nitrogen measurement stemming from the complex dynamics of urban wastewater treatment, a prediction method based on a hybrid fuzzy long short-term memory network-polynomial neural network (FLSTM-PNN) architecture is proposed. First, an FLSTM deep fusion model is constructed to integrate the uncertainty-handling capability of FNN with the temporal modeling advantages of LSTM, adaptively extracting fuzzy nonlinear dynamic features. Second, a two-stage ensemble modeling strategy is designed: In Stage I, the Elite Roulette Wheel Selection (ERWS) strategy is employed to screen a diverse set of models with superior predictive performance, effectively suppressing overfitting while retaining optimal individuals; In Stage II, prediction results of the selected models are utilized as input features for high-order nonlinear mapping via polynomial neural network (PNN), and ERWS is reapplied to select the optimal PNN output as the final prediction, fully leveraging complementary characteristics among models. Finally, the FLSTM-PNN model is validated based on real wastewater treatment plant data. Experimental results indicate that the proposed method achieves higher prediction accuracy than traditional methods, with a mean absolute percentage error (MAPE) of 7.52%.