Enhancing combined sewer flow prediction in data-limited urban areas using a semi-supervised learning framework

Combined sewer network outfalls spanning 16 ungauged urban areas in Algiers, the capital and the most populated city in Algeria. Precise real-time combined sewer flow (CSF) prediction is essential, particularly during extreme rainfall events and varying wastewater flow patterns. This study proposes a novel semi-supervised learning framework that enables knowledge refinement and transfer to enrich the database relevant for targeted data-driven modeling. The approach leverages self-organizing maps (SOM) for clustering flow patterns across different networks and under diverse operational conditions, with cluster-specific artificial neural networks (ANN) to provide the required predictive modeling for the given network node and operation instance. We benchmark our proposed approach against standalone supervised models using real-world data from Algiers’ combined sewer network. At long-record sites, the framework delivered modest improvement (median KGE from 0.70 to 0.81 and R² from 0.74 to 0.76). However, at the short-record sites, where traditional ANN models often fail to deliver reliable forecasts, the framework demonstrated clear advantages, reducing RMSE by up to 30 % (median RMSE from 47 L/s to 34 L/s and KGE rose from 0.59 to 0.86). Our findings confirm the framework’s ability to deliver reliable predictions while offering interpretable hydrological insights, supporting scalable wastewater management, flood mitigation, and improved operational efficiency of wastewater treatment plants in developing cities with sparse monitoring infrastructure. • Predict flows across a combined sewer network in sensory/data-scarce regions. • Unreliable model performance by supervised learning approaches. • Improving results by proposing a semi-supervised learning framework. • Identify the key modeling aspects influencing prediction performance. • Achieve a balanced training dataset for each new prediction instance.