Research map / Architectures and efficiency / RNNs and normalization
FIDLAR: Forecast-Informed Deep Learning Architecture for Flood Mitigation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FIDLAr, a two-network deep learning system (a Flood Manager that outputs gate/pump pre-release schedules and a pre-trained Flood Evaluator that forecasts water levels) where the frozen Evaluator's gradients train the Manager, applied to flood mitigation in South Florida.
Problem
Flood mitigation in coastal river systems requires choosing gate and pump pre-release schedules before storms; rule-based methods give excessive or inadequate releases, and physics-based iterative optimization (GA with HEC-RAS) is too slow for real-time use.
Approach
Train a differentiable DL evaluator to predict water levels, then freeze it and backpropagate a flood/wastage threshold loss through it to train a DL manager that outputs control schedules, avoiding the need for ground-truth optimal schedules; a custom Graph Transformer Network (GTN) backbone is used for both modules.
Key results
- GTN Evaluator achieves MAE 0.040 ft and RMSE 0.056 ft at t+1 for S1, lower than HEC-RAS (0.174 / 0.222) and the other DL baselines.
- As Manager, GTN reduces over-threshold timesteps at S1 from 96 (rule-based) to 22 and over-threshold area from 14.82 to 2.23.
- DL-based managers run in seconds at test time (GTN 4.90 s for mitigation), described as several orders of magnitude faster than HEC-RAS (45 min per prediction) and GA (estimated 30 h).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GTN rather than RCNN as the Manager | FIDLAR with GTN Evaluator · South Florida test set, S1 flood mitigation | Over Area: 3.28 -> 2.23 | Helped |
| Use GTN rather than the strongest baseline, TCN | FIDLAR with GTN Evaluator · South Florida test set, S1 flood mitigation | Under Timesteps: 306 -> 299 | Helped |
| Use GTN rather than Transformer as the Flood Evaluator | Flood Evaluator · Test set, S1 at t+1 | MAE (ft): 0.050 -> 0.040 | Helped |
| Use GTN rather than Transformer as the Flood Evaluator | Flood Evaluator · Test set, S1 at t+1 | RMSE (ft): 0.066 -> 0.056 | Helped |
| Add CNN (ablation) | GTN Manager · Flood mitigation test set, S1 | Over Timesteps: 37 -> 22 | Helped |
| Add Transformer (ablation) | GTN Manager · Flood mitigation test set, S1 | Over Timesteps: 32 -> 22 | Helped |
| Add GNN (ablation) | GTN Manager · Flood mitigation test set, S1 | Over Timesteps: 56 -> 22 | Helped |
| Add LSTM (ablation) | GTN Manager · Flood mitigation test set, S1 | Over Timesteps: 35 -> 22 | Helped |
| Add Attention (ablation) | GTN Manager · Flood mitigation test set, S1 | Over Timesteps: 32 -> 22 | Helped |
Surprising findings
- The Manager's ground-truth-free gradient training does not uniformly beat baselines on wastage: the GTN Manager has higher under-threshold timesteps (299 vs 306/291 for some baselines) than some DL managers, so improvement is mainly on flooding.
Limitations
- Evaluated on a single coastal river system in South Florida with one 11-year dataset.
- Wastage performance is not uniformly best; GTN wins mainly on flood metrics.
- Manager and Evaluator are trained on historical data that may not cover extreme events well.
Experimental setup
Datasets: SFWMD DBHydro hourly water level, rainfall, tide, gate and pump data (2010-2020)
Benchmarks: SFWMD DBHydro South Florida test set
Metrics: MAE RMSE over-threshold timesteps over-threshold area under-threshold timesteps under-threshold area running time
Baselines: Rule-based method Genetic Algorithm with HEC-RAS evaluator Genetic Algorithm with GTN evaluator MLP RNN CNN GNN TCN RCNN Transformer
Builds on: genetic algorithm HEC-RAS
Closest papers by meaning
- RiverMamba: A State Space Model for Global River Discharge and Flood Forecasting (NeurIPS 2025 · similarity 0.88)
- Pluvial Flood Emulation with Hydraulics-informed Message Passing (ICML 2024 · similarity 0.87)
- CoastalBench: A Decade-Long High-Resolution Dataset to Emulate Complex Coastal Processes (ICML 2025 · similarity 0.85)
- Physics-Informed Graph Neural Networks for Water Distribution Systems (AAAI 2024 · similarity 0.84)
- Early Detection of Extreme Storm Tide Events Using Multimodal Data Processing (AAAI 2024 · similarity 0.83)
- The Merit of River Network Topology for Neural Flood Forecasting (ICML 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.