Lag-aware LSTM forecasting of 5-minute stormwater inflow in a sponge city community: a cross-correlation-driven feature framework

Community-scale sponge city retention tanks need accurate short-term inflow forecasts. The forecasts support automated pump scheduling and overflow prevention. Most existing data-driven models were built for natural catchments or city-scale combined sewers. Community-scale sites differ. They show persistent dry-weather baseflow and compound detention from distributed low-impact development (LID) facilities. These features limit how well existing frameworks transfer. This study builds a lightweight, lag-aware framework for 5 min-ahead inflow prediction. A cross-correlation function (CCF) analysis extracts one physically interpretable rain-fall–inflow lag. The lag is combined with autoregressive inflow terms, recent rainfall, and multi-scale cumulative rainfall to form a 12-feature input. The framework was tested on 100 days of 5-min rainfall and inflow records from a 51484.6 m 2 LID-equipped residential community in City P, China. The site includes permeable pavements, depressed green spaces, and bioretention cells. The CCF identified a representative rainfall–inflow response timescale of about 70 min, consistent with the compound detention behavior of the LID facilities. Five models were compared under the same features. A dual-layer LSTM reached the highest test-set NSE of 0.7436. A single-layer LSTM (0.7403), SVR (0.7398), and dual-layer GRU (0.7348) fell within a 1% NSE band. A Random Forest baseline reached only 0.6724. The NSE ceiling reflects the baseflow-dominated variance of the inflow series, not a weakness of the models. The ablation study shows that the CCF-derived lag feature gives the largest single rainfall-related accuracy gain. Forecasts are near-unbiased at the 5 min step. The compact feature space lets operators swap the model to match on-site computing limits, which makes the framework easy to deploy at other sponge city communities.