Fine-Temporal Filtration Turbidity Prediction in Drinking Water Treatment Using a Domain-Aware Deep Learning Framework

Short-term prediction of turbidity fluctuations during filtration is important for stable operation and water safety in drinking water treatment plants. This study presents a fine-temporal prediction framework for filtration-stage turbidity using 15-minute interval time-series data, focusing on domain-aware feature design. A parallel-to-convergence positive and negative convolution perceptron–convolutional neural network (P2C-PC) model was implemented to combine variable-wise feature extraction with CNN-based temporal integration. The model used nine water-quality and operational variables, including the settling-basin turbidity change rate (ΔSB), which was introduced to represent short-term upstream turbidity changes. The dataset consisted of 33,401 preprocessed records from a single drinking water treatment plant. Under the evaluated single-plant, low-turbidity operating conditions, the P2C-PC model achieved a test RMSE of 0.0088 NTU, showing lower prediction error than the baseline, comparison, and ablation models. Although ΔSB showed negligible linear correlation with FS turbidity (r = 0.01), the ablation result indicated that including ΔSB improved prediction performance. Additional interpretability analyses suggested that its contribution was associated with multivariate and time-dependent feature interactions. The framework provides a practical basis for decision-support applications aimed at operator awareness and short-term filtration-stage monitoring.