Effluent Quality Prediction Using Adaptive Sparse Deep Belief Network with Stochastic Configuration Learning Scheme

The accurate prediction of effluent water quality is essential for monitoring the operational status of wastewater treatment plants and supporting process regulation. To enhance feature representation and prediction accuracy in data-driven effluent quality modeling, this paper proposes an adaptive sparse deep belief network with stochastic configuration learning (ASDBN-SCL). First, an adaptive learning-rate strategy based on gradient-direction consistency is introduced during restricted Boltzmann machine pretraining to dynamically regulate the parameter-update step size. Second, a sparsity constraint based on the average activation probability of hidden neurons is incorporated to suppress redundant feature responses and promote compact representations. Third, stochastic configuration learning is employed as the supervised output module, in which hidden nodes are incrementally generated under a supervisory constraint and the output weights are analytically determined using least squares. The proposed method was evaluated using hourly monitoring data collected from a municipal wastewater treatment plant in Huai’an, China. After data preprocessing, 4500 valid samples were retained and chronologically partitioned into training, validation, and independent test sets. For effluent chemical oxygen demand prediction, the proposed method achieved a test RMSE of 0.3868±0.0174, an MAE of 0.3037±0.0142, and an R2 of 0.9954±0.0004. For effluent total nitrogen prediction, the corresponding values were 0.2239±0.0094, 0.1756±0.0082, and 0.9849±0.0013, respectively. These results indicate that the proposed framework provides accurate and reliable effluent water quality predictions for the investigated wastewater treatment process.