Hybrid Approach for Flood Forecasting in Urban Innovation Districts (HIDS–Unicamp): Integrating PCSWMM, Neural Networks, and Explainable Artificial Intelligence

Flood intensification due to climate variability and urbanization necessitates advanced forecasting tools, particularly in regions undergoing rapid transformation where drainage infrastructure data is often scarce. This study presents a methodological framework for flood forecasting in the Ribeirão Anhumas watershed (Campinas, Brazil), specifically applied to the International Hub for Sustainable Development (HIDS–Unicamp). As an innovation district currently under implementation, HIDS represents a unique opportunity to integrate predictive modeling into early-stage urban planning.

The methodology addresses data scarcity by integrating physical modeling with machine learning. We have established a simulation environment using PCSWMM to replicate hydrological behavior under distinct infrastructure scenarios. These simulations, driven by high-resolution precipitation (10-min) and geospatial data (1 m DTM), generate the necessary synthetic training data for regions where sensor networks are yet to be deployed. The proposed architecture is designed to perform a binary classification of flood occurrence (Flood/No Flood), utilizing a multi-model approach: Recurrent Neural Networks (RNNs) for temporal dynamics, Convolutional Neural Networks (CNNs) for spatial patterns, and Graph Neural Networks (GNNs) to explicitly model the hydrological connectivity of the watershed.

In this contribution, we present the complete data processing pipeline and the defined model architecture. The study focuses on evaluating the comparative performance of these architectures using classification metrics (accuracy, precision, recall, F1-score, and ROC curve). Furthermore, to ensure the model is transparent for decision-makers, we outline the application of Explainable AI (XAI) techniques, specifically SHAP and LIME. These are intended to identify the contribution of input variables to flood predictions, bridging the gap between "black-box" deep learning and interpretable hydrological processes. The final results aim to demonstrate how hybrid modeling can support the strengthening of early warning systems and resilience strategies in developing urban territories.