Sparse Incident-Cluster Learning for 12-hour Port Flood Pre-warning in Digital-Twin Analytics

Jie Zhang; Qiang Ni; David Windridge; Huan X. Nguyen

cs.LGSep 5, 2026

Sparse Incident-Cluster Learning for 12-hour Port Flood Pre-warning in Digital-Twin Analytics

Authors: Jie Zhang, Qiang Ni, David Windridge, Huan X. Nguyen

Organizations: Lancaster University, Lancaster, United Kingdom · Middlesex University, London, United Kingdom

Abstract

Port flood digital twins require analytics that warn operators before disruption, but official warning incidents are often few and adjacent observations are temporally dependent. Row-level classification can therefore overstate performance by placing windows from the same event in both model-development and evaluation data. We formulate 12-hour port flood pre-warning as an incident-cluster learning problem and evaluate a digital-twin analytics module using eight-point water-level histories, prediction-time contextual covariates, and interpretable short-window dynamics. The protocol combines fold-specific sparse feature selection, warning-cluster grouping, negative-label controls, 100-repeat random top-k controls, and alert-episode evaluation. Liverpool is the primary four-cluster case study, with harmonised Humber/Hull-proxy and Wessex South data used for protocol-transfer checks. Across the Liverpool folds, the top-10 ElasticNet model achieves mean F2 = 0.696, compared with 0.633 without top-k truncation and 0.681 for full-feature weighted XGBoost. It is the strongest ElasticNet variant, remains competitive with the nonlinear reference using only ten predictors, and exceeds the repeat-level 95th percentile of broad and same-family random subsets. Contextual covariates provide a strong prediction-time anchor, complemented by physically interpretable local dynamics. Historical replay converts risk scores into alert episodes and measures alert duration and false-episode burden. The result is an offline-evaluated analytics and validation module designed for integration into a port digital twin.

Figures & tables

Source	Provider	Variables used	Role in this paper	Ref.
Flood monitoring archive	Environment Agency	Timestamped observed water levels, monitoring-site descriptors, and measurement-category records	Constructs the eight-point local water-level windows for Liverpool and additional port-region stress tests	[ 22 ]
Real-time flood monitoring API	Environment Agency	Station metadata, river/sea level station descriptors, and observation metadata	Supports site mapping, measurement-category definition, and reproducible access to level observations	[ 23 ]
Flood warning information service	Environment Agency	Public flood warning and alert records, warning dates, and warning categories	Defines warning-event dates, incident clusters, and 12-hour pre-warning labels	[ 24 ]
Processed tide products and tide-gauge records	BODC and NTSLF	Astronomical tide, observed tide, and derived residual candidates where available	Used in exploratory checks of tide-phase and surge-residual features; not retained as the main reported sparse signature	[ 25 ]

Table 1: Data sources and their roles in the study. Links are provided through the cited data-source records rather than as raw URLs in the main text.

Dataset	Windows	Clusters	Positives	Pos. rate	Stations	Role
Liverpool primary	2067	4	393	0.190	2	Main case study
Liverpool common-feature view	2067	4	393	0.190	2	Cross-site baseline
Humber/Hull proxy	4676	8	766	0.164	5	Protocol stress test
Wessex South	2139	6	327	0.153	3	Protocol stress test

Table 2: Warning-aligned datasets used in the experiments. Clusters are independent warning-event clusters used for grouped validation.

Figure 1: Digital-twin analytics loop for port flood pre-warning. The evaluated module connects water-level stream ingestion, digital state update, sparse risk inference, alert episode formation, and operator-facing decision support. The study evaluates this loop through offline historical replay rather than live deployment.

Figure 2: Sparse pre-warning learning protocol. The protocol converts water-level streams into eight-point histories, prediction-time contextual and physical descriptors, nested sparse feature selection, risk scores, and alert episodes under incident-cluster validation.

Figure 3: Incident-cluster validation design. Outer folds hold out assigned warning-event clusters, and feature ranking does not use outer-test rows. A source-day audit found no training–test overlap but found one validation–test source-day overlap in folds 1–3; threshold-dependent metrics are therefore reported with this caveat.

Figure 4: Foldwise historical replay design. Held-out water-level windows are scored in timestamp order and converted into alert episodes. The replay preserves the order of the evaluated stream but is distinct from a strict latest-event temporal holdout.

Stage	Operation
Input	Timestamp-ordered held-out water-level histories, prediction-time contextual covariates, trained risk model $f_{\theta}$ , validation-selected threshold $\tau$ , and a 30-minute episode-merging gap.
State update	At replay time $t$ , update the digital state using only observations available at or before $t$ .
Risk inference	Compute a pre-warning risk score $s_{t}=f_{\theta}(\mathbf{x}_{t},\mathbf{c}_{t})$ for the current replay state.
Episode formation	Aggregate simultaneous scores by their maximum, apply $\mathcal{E}_{\tau,\Delta}$ , and merge alerts separated by at most 30 minutes.
Event matching	Compare completed alert episodes with official warning times using watch-window, warning-window, and any-pre-event hit rules.
Output	Report warning hit rate, watch hit rate, false warning episodes per day, median lead time, missed incidents, and mean alert duration.

Table 3: Offline twin replay protocol for evaluating alert episodes from held-out historical streams.

Model	AP	Val.- $F_{2}$	Event recall	False/day	Inputs
Weighted XGBoost (full)	$0.577\pm 0.202$	$0.681\pm 0.125$	1.000	33.41	71
ElasticNet (all candidates)	$0.592\pm 0.160$	$0.633\pm 0.160$	0.708	17.76	71
ElasticNet (fold-specific top-10)	$0.567\pm 0.169$	$\mathbf{0.696}\pm 0.089$	0.833	24.72	10
XGBoost (fold-specific top-10)	$0.553\pm 0.171$	$0.650\pm 0.136$	0.917	37.56	10
ElasticNet (context only)	$0.417\pm 0.183$	$0.679\pm 0.119$	0.729	25.73	4
ElasticNet (physical only)	$0.383\pm 0.056$	$0.587\pm 0.160$	0.813	28.54	36

Table 4: Liverpool comparison under grouped incident-cluster evaluation. Values are mean

\pm

standard deviation across four folds. “All candidates” denotes input variables before ElasticNet shrinkage; “top-10” denotes fold-specific truncation after ElasticNet ranking; “weighted XGBoost (full)” corresponds to Current12hBinary.

Figure 5: Liverpool mean validation-threshold

F_{2}

and row-level false-alarm burden. Horizontal and vertical error bars are fold standard deviations. Marker area encodes the number of candidate inputs.

Fold	Test cluster	Threshold	AP	Val.- $F_{2}$	Event recall	False/day	XGB $F_{2}$
1	C1: 7–11 Sep.	0.35	0.796	0.826	1.000	20.57	0.864
2	C2: 20 Sep.	0.55	0.441	0.624	0.500	19.17	0.631
3	C3: 6–8 Oct.	0.05	0.592	0.668	0.833	28.82	0.649
4	C4: 14 Nov.	0.25	0.438	0.664	1.000	30.33	0.581

Table 5: Fold-level Liverpool results for the top-10 ElasticNet model, with the full weighted XGBoost

F_{2}

shown as a reference.

Control	AP mean $\pm$ SD	$F_{2}$ mean $\pm$ SD	$F_{2}$ p95	Runs $\geq$ selected
All candidates	$0.360\pm 0.033$	$0.553\pm 0.049$	0.624	0%
No cyclic time	$0.336\pm 0.017$	$0.531\pm 0.043$	0.586	0%
No station identity	$0.364\pm 0.037$	$0.550\pm 0.055$	0.631	0%
No time or station	$0.335\pm 0.017$	$0.543\pm 0.041$	0.596	0%
Physical candidates only	$0.347\pm 0.012$	$0.528\pm 0.032$	0.580	0%
Context candidates only ( $k=4$ )	$0.417\pm 0.000$	$0.685\pm 0.000$	0.685	0%

Table 6: Repeat-level random-subset controls (

n=100

complete four-fold repetitions per row). The selected top-10 reference has AP 0.567 and validation-threshold

F_{2}

0.696.

Figure 6: Repeat-level random top-10 controls. Points show mean validation-threshold

F_{2}

, horizontal bars show the 5th–95th percentile range, and the dashed line is the selected top-10 result (0.696).

Model	Real-label AP	Maximum control AP	AP difference
Full weighted XGBoost	0.580	0.406	0.173
Top-10 ElasticNet	0.567	0.367	0.200
Top-10 XGBoost	0.553	0.379	0.174

Table 7: Real warning labels versus the strongest synthetic negative control for each model.

Signature	Inputs	Val.- $F_{2}$	AP	Event recall
Cyclic time only	2	0.678	0.418	0.792
Station identity only	1	0.618	0.253	1.000
Physical level only	11	0.554	0.350	1.000
Physical shape only	25	0.594	0.380	0.813
All physical, no context	36	0.587	0.383	0.813
Fold-specific hybrid top-10	10	0.696	0.567	0.833

Table 8: Context-versus-physics decomposition using ElasticNet logistic regression. Values are four-fold means.

Figure 7: Context-versus-physics decomposition. Points show four-fold means and error bars show fold standard deviations. AP and validation-threshold

F_{2}

are displayed as distinct metrics and should not be compared as if they shared the same decision threshold.

Predictor	Family	Selected folds
Cyclic time-of-day cosine	Temporal context	4/4
Cyclic time-of-day sine	Temporal context	4/4
Monitoring-location identity	Site context	4/4
Within-window directional reversal count	Physical shape	4/4
Maximum absolute 15-minute level increment	Physical shape	4/4
Net change over the local window	Physical shape	3/4

Table 9: Recurring predictors in the fold-specific top-10 ElasticNet lists. Names describe the scientific quantity rather than the implementation column.

Feature	Normal mean	Pre-warning mean	SMD	Cliff’s delta	Direction consistency
Maximum absolute level increment	0.3134	0.4236	0.8290	0.4215	0.75
Local amplitude range	1.6028	2.1678	0.5976	0.3117	0.75
Mean absolute level increment	0.2332	0.3151	0.6319	0.3316	0.75

Table 10: Physical sanity-check predictors comparing normal and pre-warning windows. Standardized mean difference and Cliff’s delta summarize whether interpretable water-window variables separate the two states.

Site	Full XGBoost	Port-nested top-10	Liverpool-fixed top-10	Context only	Physical sanity
Liverpool	0.381	0.481	0.457	0.558	0.509
Humber/Hull proxy	0.522	0.526	0.527	0.542	0.465
Wessex South	0.503	0.527	0.514	0.528	0.422

Table 11: Cross-site protocol stress test using the common 67-feature view. Values are mean validation-threshold

F_{2}

across site-specific warning-cluster folds.

Model setting	Warning hit	Watch hit	False warn. /day	Median lead (h)	Mean duration (h)	Missed
Context-only SparseLogistic	1.000	1.000	0.564	15.28	14.75	0
Full 12-h XGBoost	1.000	0.813	0.524	15.99	28.29	0
Nested top-10 SparseLogistic	1.000	1.000	0.658	14.73	13.49	0
Nested top-10 XGBoost	1.000	0.813	0.465	16.29	29.92	0
Physical-only SparseLogistic	0.813	0.250	0.738	15.85	18.01	3

Table 12: Foldwise historical replay with maximum-score aggregation and stored validation thresholds. Lead time and duration are reported in hours.

Figure 8: Operational burden under stored validation thresholds. Marker position compares false warning episodes per day with mean alert duration; marker area represents warning hit rate.

Flood Forecasting

Flood forecasting predicts river, coastal, and urban water levels or inundation to support earlier warnings and flood response.

See all 18 papers

Sep 30, 2026

C-STRIDE: An Observation-Driven AI Digital Twin for Predicting Basin-Wide Flood Fields from Sparse Stream-Gauge Histories

Emergency managers need to know where floodwater is, how deep it is, and how it will change over the coming hours across an entire river basin. During a flood, however, real-time measurements come from only a handful of stream gauges, and high-resolution hydrodynamic models are too costly to rerun each time new data arrive or to run as large ensembles. We present C-STRIDE, an observation-driven AI digital twin that turns short records from a few stream gauges, together with terrain and rainfall, into basin-wide maps of water depth and extends these predictions up to a day ahead. It is trained on simulations from a calibrated two-dimensional hydrodynamic model and needs no separate data-assimilation step. In the Des Plaines River basin near Chicago, six gauges inform predictions over 4.2 million 30-m grid cells. Terrain improves the predictions most, rainfall keeps errors from growing over longer horizons, and together they reduce errors by about 40% compared with gauge records alone. When future rainfall is known, errors remain near 15% one day ahead, compared with nearly 40% without rainfall. Given real instead of simulated gauge records, the model shifts its predictions toward the observed hydrographs at three of six gauges without retraining, and it runs about 150 times faster than the hydrodynamic model. These results show how sparse gauges, terrain, and rainfall can be combined into fast, continuously updated flood predictions, a step toward operational flood digital twins that still requires testing with real-time data and rainfall forecasts.

Yanjie Tong, Phillip Si, Yuan Qiu +1

School of Computational Science and Engineering, Georgia Institute of Technology, 756 West Peachtree Street Northwest, Atlanta, 30308, GA, USA

Surrogate Modeling Flood Inundation Mapping Digital Twins Spatiotemporal Forecasting Flood Forecasting Hydrology

Ask PDF

May 13, 2026

Toward AI-Driven Digital Twins for Metropolitan Floods: A Conditional Latent Dynamics Network Surrogate of the Shallow Water Equations

AI-driven flood digital twins demand fast hydrodynamic surrogates for ensemble forecasting and observation assimilation. Yet even GPU-accelerated two-dimensional shallow water equation (SWE) solvers still require

\sim 55

minutes per

96

-hour run on a

\sim 4.2

-million-active-cell metropolitan basin (the DesPlaines River basin at $30\,\mathrm{m}$ resolution), making such workloads prohibitive at native resolution. We present the Conditional Latent Dynamics Network (CLDNet): a low-dimensional latent neural ODE driven by rainfall, paired with a coordinate-based decoder conditioned on static terrain (elevation, slope, Manning roughness) that reconstructs depth and discharge at arbitrary query points. Pointwise decoding decouples memory from grid size and handles irregular watersheds natively, enabling metropolitan-scale training on a single compute node and direct queries at exact gauge coordinates without raster snapping. We evaluate CLDNet on a synthetic $250{,}000$ -cell Texas benchmark and on a new DesPlaines case study of

114

real-rainfall Stage~~IV storms whose reference simulator we validate against United States Geological Survey (USGS) gauges at the April~~2013 flood-of-record (Nash--Sutcliffe efficiency

0.57

0.94

on mean-recentered water-surface elevation). CLDNet roughly halves the relative root-mean-squared error of an unconditional baseline, outperforms regular-grid VAE--ConvLSTM and FNO baselines on the Texas benchmark (both presuppose a Cartesian grid and do not apply to the irregular Des~Plaines watershed), reaches a critical success index of

\approx 86\%

at the

0.5\,\mathrm{m}

inundation threshold, and produces a full

96

-hour basin-wide forecast in

\sim 29

seconds -- a

\sim 115\times

speedup.

Phillip Si, Yuan Qiu, Omar Sallam +4

School of Computational Science and Engineering, Georgia Institute of Technology, 756 West Peachtree Street Northwest, Atlanta, 30308, GA, USA · Environmental Science Division, Argonne National Laboratory, 9700 S Cass Ave, Lemont, 60439, IL, USA

Digital Twins PDE Surrogate Modeling Flood Forecasting Neural Surrogate Modeling Spatiotemporal Forecasting Neural ODEs+1

Ask PDF

Jun 3, 2026

Data-efficient flood depth prediction through domain-aware coreset selection and tabular foundation models

Near-real-time flood depth prediction demands surrogate models that are accurate, fast, and transferable across watersheds. Supervised surrogates can match physics-based simulators in accuracy but need millions of training rows per watershed and cannot extrapolate beyond their original mesh. We propose a domain-aware coreset construction pipeline that conditions a tabular foundation model at inference time. The pipeline stratifies storms by return period and most-affected watershed, then samples hexagons with a target-aware spatial selector. With 0.7% of the per-watershed training pool, the model attains a mean

R^2

of 0.663 across nine Houston-area watersheds, within 98.5% of the supervised reference (

R^2

= 0.673). It transfers to held-out watersheds without task-specific retraining, staying ahead of a coreset-trained supervised baseline. On real storms it exceeds the supervised reference on a far out-of-distribution case and trails it on a mostly in-distribution one. Domain-aware coreset construction lets tabular foundation models deliver data-efficient, watershed-transferable flood predictions without per-watershed training.

Lipai Huang, Adithi Srinath, Manas Singh +2

Urban Resilience.AI Lab, Zachry Department of Civil and Environmental Engineering, Texas A&M University, College Station, TX, USA · Department of Computer Science and Engineering, Texas A&M University, College Station, TX, USA · Resilitix Intelligence LLC, Houston, TX, USA +1

Surrogate Modeling OOD Generalization Coreset Selection Spatiotemporal Forecasting Flood Forecasting

Ask PDF

Abstract

Flood Forecasting

Explore similar work

C-STRIDE: An Observation-Driven AI Digital Twin for Predicting Basin-Wide Flood Fields from Sparse Stream-Gauge Histories

Toward AI-Driven Digital Twins for Metropolitan Floods: A Conditional Latent Dynamics Network Surrogate of the Shallow Water Equations

Data-efficient flood depth prediction through domain-aware coreset selection and tabular foundation models