Introduction
Flash floods in the Spanish Mediterranean present a high-stakes challenge for emergency management. Intense precipitation leads to rapid flood peaks in areas with complex land use, requiring decision-makers to act quickly regarding dam operations and evacuations. However, traditional hydraulic models are often too computationally expensive for real-time use during a crisis, and human experts often struggle to translate their intuition into the structured rules required for automated systems.
This paper proposes a hybrid approach: a real-time expert system that uses "artificial experience" generated from pre-run simulations to provide rapid decision support. By combining rigorous mathematical models with human expert logic, the system bypasses the computational bottlenecks of traditional simulation.
System Architecture: The Hydrographic Subsystem
The methodology bridges the gap between raw data and actionable advice by structuring the system into two main components: the Hydrographic Subsystem and the Social-Territorial Subsystem. The Hydrographic Subsystem models the physical watershed, including drainage networks and floodable areas.
As illustrated in Figure 1, this subsystem identifies "floodable pockets" and "temporary courses" that are critical during a crisis. By pre-calculating how these specific areas react to various inflow levels, the system can provide immediate insights into which regions are at risk without having to re-run heavy mathematical integrations during an unfolding event.
The Inference Engine and Artificial Experience
The core innovation of this research is the use of Artificial Experience. Instead of relying solely on subjective interviews with experts, the researchers use established hydraulic models (such as the Stanford Watershed Model IV or SWMM) to run a vast array of "what-if" scenarios. These results are then analyzed to create a set of simplified inference rules.
The inference engine operates across three lines of reasoning: the present state (what is happening now?), future states (what will happen in the coming hours?), and control advice (what should be done?).
As shown in Figure 2, the system ingests primary data from real-time information systems to deduce facts at both controlled and uncontrolled points using Situation Rules (SR). It then applies a set of rules about the future—including Rain, Inflow, and Flood Level Rules—to predict impacts at various time intervals ($T+T_1, T+T_2, \dots$). This logical flow ultimately generates proposals for decisions based on civil defense requirements and dam operating rules.
Data Sources and Calibration
The system's reliability is heavily dependent on historical calibration and real-time sensor integration.
- Historical Data: Data from the devastating autumn 1982 floods in the Spanish Mediterranean regions were used for initial calibration.
- Real-time Data: The system is designed to ingest rain intensity and water level data from control points as the event occurs.
- Model-Generated Data: The "artificial experience" layer is derived from extensive runs of the Storm Water Management Model (SWMM), providing the precision of numerical integration within a rule-based framework.
Conclusion and Future Outlook
While the system provides a robust architecture for disaster response, it faces notable limitations. Its effectiveness is tied to the quality of the initial model calibration; if the underlying "artificial experience" is based on a flawed model, the expert system's advice will be similarly flawed. Furthermore, significant changes in land use or infrastructure require re-running the simulations to update the rule set.
Despite these challenges, the framework offers a vital blueprint for moving from improvised crisis management to a calculated, model-informed response, synthesizing the precision of mathematical equations with the speed of expert heuristics.
Summary updated