Abstract: Soft sensors can indirectly quantify key parameters that are difficult to measure directly by establishing mathematical models. This technique is particularly useful in addressing the difficulty of measurement of critical effluent parameters in wastewater treatment process. However, with the increasing diversity and complexity of data, traditional neural networks become prone to parameter redundancy, computational inefficiency, and limited generalization capability. To address this issue, a soft sensor based on the sparse neural network (SPNN) is proposed. This model integrates the regularization-based sparsity constraint and periodic pruning strategy to reduce the number of non-zero parameters in the network, thereby building a more concise and efficient model structure. Besides, combined with preprocessing methods such as feature selection and data standardization, the prediction performance and generalization ability of SPNN are further enhanced. Experimental results on the University of California, Irvine (UCI) wastewater treatment dataset demonstrate that the SPNN significantly outperforms comparative models, including partial least squares (PLS), support vector machines (SVM), long short-term memory networks (LSTM), and deep neural networks (DNN). Specifically, compared with the best comparative model LSTM, the proposed method significantly improves the prediction accuracy with 88.87% reduction in root mean square error (RMSE), 75.82% reduction in mean absolute error (MAE), and 6.28% improvement in coefficient of determination (R²), which verifies the excellent accuracy and robustness of the SPNN in modeling complex wastewater data.