A Surrogate Modeling Framework for Compound Flood Risk and Optimization Analysis

Abstract

This research develops a surrogate modeling framework to efficiently analyze and optimize a proposed pump and gate system designed to mitigate compound flooding in the Clear Lake region, Texas. Traditional numerical hydraulic models are often computationally expensive for large number of simulations. As probabilistic assessments can require $O(10^3)$ to O(10^5) runs, a cheaper alternative would be necessary for robust probabilistic assessment. This study addresses this limitation by developing a deep-learning surrogate to approximate the complex hydrodynamic behavior.

The methodology involved three main stages. First, a 1D HEC-RAS model of the Clear Lake system was adapted to serve as the physics-based "ground truth" generator. Second, this model was used to generate a training dataset of 2,400 simulations. This was achieved by systematically sampling key infrastructure design parameters (gate width $W_g$, number of pumps $n_p$, and activation levels $h_{on}$) alongside a wide range of synthetic compound flood forcings (inflow hydrographs and downstream storm surge boundaries).

Third, three distinct Long Short-Term Memory (LSTM) network architectures (Models A, B, and C) were developed to compare different data encoding strategies. Model A, a direct sequence-to-sequence (seq2seq) model, was provided with all dynamic inputs, including the known pump discharge time series ($Q_{pump}$). Model B tested the model's ability to infer dynamics by replacing the $Q_{pump}$ time series with static design parameters ($n_p$, $h_{on}$). Model C used an autoregressive structure, feeding its own past water level predictions back as inputs to dynamically infer the pump response.

The results demonstrate that the fully-informed LSTM (Model A) can successfully learn and reproduce the governing hydrodynamic processes with very high accuracy. However, models that attempted to infer dynamic behavior from static design parameters (Models B and C) show reduced performance. These models particularly struggled to capture the sharp, transient effects of pump (de)activation, leading to overly smoothed predictions. This study concludes that while LSTMs are capable of learning the physical patterns of the system. The main challenge lies in feature encoding, specifically, enabling the model to capture complex, dynamic responses from static inputs. The framework demonstrates the potential of LSTMs, but emphasizes that how the data is represented is the key factor in developing a surrogate model suitable for design optimization.