To address the challenges of severe outlier interference and temporal modeling difficulties in predicting NOx concentrations within cement denitrification systems, this paper proposes a composite prediction method integrating SSAD-Z-score anomaly detection with LFLSTM deep learning models. First, based on NOx generation and SNCR reaction mechanisms, core input variables-including kiln tail flue gas chamber NOx concentration and SNCR ammonia injection frequency—are selected using correlation coefficients. To handle complex anomalies like backwashing, the SSAD-Z-score algorithm is designed. It combines sliding window slope changes with global Z-score criteria to achieve adaptive detection and correction for multiple anomaly types, including sudden spikes, trends, local disturbances, and backwashing data. Furthermore, multi-step lag features are introduced into the LSTM model to fully leverage historical time-series information, enhancing the model’s ability to capture dynamic changes in NOx concentration. Experimental results demonstrate that the SSAD-Z-score method significantly outperforms traditional Z-score methods, Isolated Forest, LOF, and Wavelet Transform in metrics such as anomaly detection coverage (47.67%) and recall (0.844). Additionally, SSAD-Z-score exhibits exceptional performance in corrected standard deviation, achieving the lowest value among all comparison methods with an average correction magnitude of 8.81. Parameter sensitivity analysis indicates that under moderate parameter settings—sliding window size 10, slope threshold 0.15, and Z-score threshold 1.65-SSAD-Z-score achieves optimal balance in terms of corrected standard deviation (35.65) and average correction magnitude (8.81). Inputting the corrected data into the LFLSTM model for NOx concentration prediction yielded an average single-sample prediction time of just 4.88 milliseconds-significantly below the 1-second data acquisition frequency. The model achieved an R2 of 0.884 5, RMSE of 2.026 5, and MAE of 1.798 5, demonstrating markedly superior predictive performance compared to mainstream models like Transformer, Informer, and GRU. Ultimately, the LFLSTM model combined with SSAD-Z-score correction demonstrated the best overall system performance, achieving an R2 of 0.824 097, an RMSE of 4.305 981, and an MAE of 3.297 620. These metrics significantly outperformed prediction models using either LFLSTM alone or Z-score-LFLSTM. The research results validate the effectiveness of the proposed method in enhancing anomaly detection accuracy and NOx concentration prediction precision, providing a robust data foundation and methodological support for intelligent control of cement denitrification systems.