Optimized Physics-Guided ARIMAX–LSTM for IoT-Enabled Urban Flood Forecasting

Source abstract

Urban flooding remains a significant challenge for sustainable city management because of nonstationary rainfall and limited observational data. The study proposes an IoT-enabled optimized Physics-Guided Transfer Learning (PGTL) ARIMAX–LSTM framework. The framework integrates the linear trend prediction ability of ARIMAX along with the nonlinear learning of residuals through LSTM networks. The framework further incorporates a physics-based discharge-consistency constraint and a transfer-learning strategy to support model adaptation under data-scarce basins. Evaluation of the proposed model structure is carried out using a hydro-meteorological dataset of 15 years collected from the Sukkur Barrage region. This dataset comprises both historical hydrological records and recent Internet of Things (IoT)-based rainfall and discharge observations. Performance was compared with six representative forecasting paradigms: Hydrological, ARIMAX, Random Forest (RF), Linear Regression (LR), Support Vector Regression (SVR), and LSTM, under identical preprocessing and evaluation conditions. The evaluation of the forecasting backbone on the Sukkur dataset achieved the lowest RMSE (121.8 m3/s) and the highest NSE (0.81) among all evaluated models. The improvement is consistent with the complementary roles of the ARIMAX trend estimator, the nonlinear residual learning capability of LSTM, and the physics-guided constraint that suppresses unrealistic discharge fluctuations. The results demonstrate the potential of the proposed model for IoT-enabled urban flood forecasting under variable hydrological conditions.