Development and Evaluation of Deep Learning Models for Forecasting Water Levels in Sewage Pumping Stations

Forecasting is a critical component of real-time control, maintenance planning, and emergency handling in hydraulic networks. Sewage Pumping Stations (SPS) are sensitive nodes in these networks: blockages, pump malfunctions, and sudden inflow surges can lead to operational failures or floodings, which may affect both infrastructure and nearby populations. Despite this, many SPS still operate without predictive support, relying on static threshold-based alerts and the site-specific experience of operators. The introduction of data-driven forecasting tools offers an opportunity to transition from reactive to proactive management. This work investigates the use of encoder-decoder deep learning architectures for forecasting water levels in SPS within the Puglia Aqueduct, one of Europe's largest wastewater networks. Recent digitalization efforts across the system have made high-frequency SCADA measurements available, enabling in turn the development of predictive pipelines for the management of the network. In collaboration with Fore Data, this study develops a complete end-to-end workflow that handles engineering, preprocessing, and modeling tasks, to evaluate the performance of water level forecasting models in pumping sites. A total of 48 model configurations with variations of architectures, feature sets, and forecast windows are compared to identify suitable candidates for multi-step prediction. Among all tested models, Sequence-to-Sequence architectures with Gated Recurrent Units (GRU) consistently delivered the best performance for medium- and long-horizon forecasts. The results demonstrate the viability of deep learning for SPS forecasting and highlight future directions, particularly the development of regime-aware models capable of distinguishing between dry and wet operational conditions in SPS.