Answers are grounded in this paper and cite their sources.
Study enhances accurate water level forecasting in urban drainage systems, suggesting improved flood management.
Accurate multi-step water level forecasting in urban drainage systems (UDSs) is essential for flood early warning and adaptive drainage management. However, conventional recurrent neural network-based models, such as gated recurrent units (GRUs), often suffer from substantial performance degradation over extended forecasting horizons due to recursive error accumulation and limited spatial awareness. To address these limitations, this study proposes a novel deep learning framework with three key innovations: (1) autoregressive sequence-to-sequence learning to mitigate error accumulation (GRU-AR), (2) multi-point synchronous prediction to capture spatial dependencies (GRU-AR-M), and (3) hybrid integration of one-dimensional convolutional neural networks with GRU for spatiotemporal feature extraction (CNN-GRU-AR). The framework was evaluated on a campus-scale drainage system and validated with field data from a Swiss combined sewer network. Results consistently demonstrate the effectiveness of all three enhancements. In the numerical case, CNN-GRU-AR increased the average Nash efficiency coefficient (NSE) from 0.74 to 0.94 and reduced the mean absolute error (MAE) by 59.1% compared to the standard GRU over a 60-minute forecasting horizon. In real-world validation, the model improved average NSE from 0.81 to 0.91 and reduced MAE by 21.5% over a 90-minute horizon. Notably, in the two cases, NSE at the 60- and 90-minute lead times improved from 0.48 to 0.89 and 0.45 to 0.81, respectively, confirming significantly enhanced long-term performance. The model also maintained high robustness under complex rainfall events, spatially distributed conditions, and noisy rainfall forecasts. These findings confirm the potential of CNN-GRU-AR for real-time, accurate, and scalable water level forecasting in operational UDSs.
No takes yet. Share an insight, caveat, or question.
Huang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: