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

Task framing

• The target is whether an official flood warning starts within 12 hours of a local water-level window, not a future water-level value.This makes the task warning-aligned and lead-time-specific.Source: §1 · pp. 2–3

Digital-twin evaluation

• A digital-twin warning model should be assessed by whether risk scores become alert episodes with usable lead time and acceptable false-alert burden.Row-level classification scores alone do not measure these operational quantities.Source: §1 · p. 2

Validation challenge

• In Liverpool, 19.0% positive windows come from only four warning clusters, and overlapping adjacent windows can share hydrometeorological context.Randomly splitting related observations across development and evaluation can underestimate predictive error; feature ranking before holding out an incident can also leak information.Source: §1 · pp. 2–3

Results

• Liverpool’s top-10 ElasticNet reaches mean F2 = 0.696, versus 0.633 without top-k truncation and 0.681 for full-feature weighted XGBoost.The study presents it as the strongest ElasticNet variant and a compact, auditable representation for the alert-oriented objective.Source: §1 · p. 3

Approach

• The protocol embeds sparse feature selection in incident-cluster validation and tests it against dense baselines and random-subset controls.It also connects window-level scores to alert episodes through historical-stream replay.Source: §1 · p. 3

2.1 Data Sources

• Primary modelling data come from Environment Agency water-level and warning records; the final sparse signature uses prediction-time water-window descriptors and contextual covariates.Additional tide and external-driver sources were explored during the broader study.Source: §2.1 · p. 3, Table 1 · p. 4

2.2 Warning-Aligned Datasets

• Liverpool is the primary case, while Humber/Hull-proxy and Wessex South are protocol-transfer checks rather than proof of universal geographic generalization.Liverpool has four stored warning clusters; the cross-site analysis uses a 67-feature common view, unlike the 71-candidate Liverpool analysis.Source: §2.2 · p. 4, Table 2 · p. 4

2.3 Eight-Point Water-Level Windows

• Each inclusive history ends at the current observation and contains L water-level observations; with 15-minute sampling, the eight-point window spans 105 minutes.All descriptors use observations available at or before prediction time.Source: §2.3 · pp. 4–5

2.4 Target Definition and Incident Clusters

• A positive label means an official flood warning starts within the 12 hours after the window endpoint, with the earliest qualifying warning defining the associated event.The label is warning-aligned rather than defined by the water level at the endpoint.Source: §2.4 · p. 5
• Positive windows inherit their stored warning-event cluster as the validation group, and the study does not infer a new clustering threshold from outcomes.Training excludes windows assigned to the held-out cluster, but validation and test subsets share a source day in folds 1–3, so threshold-dependent results are not fully temporally isolated.Source: §2.4 · pp. 5–6

2.5 Feature Taxonomy

• The model input combines contextual covariates, observed levels and magnitude summaries, and descriptors of short-window water-level dynamics.Context includes time-of-day, monitoring location, and measurement category; dynamics include increments, trend, amplitude, contrasts, extrema positions, and reversals.Source: §2.5 · p. 6
• The physical descriptors include local trend and second-order increments, while reversal counts and half-window contrasts summarize direction changes and level shifts.The slope is a signed local trend per sampling step; q_i counts reversals and h_i contrasts mean levels across the window halves.Source: §2.5 · pp. 6–7
• The digital-twin workflow links water-level ingestion and digital-state updates to sparse risk inference, alert-episode formation, and operator-facing decision support.The loop is evaluated through offline historical replay rather than live deployment.Source: Figure 1 · p. 7

3 Proposed Sparse Pre-warning Analytics Module

• The protocol combines prediction-time feature filtering, fold-specific selection, incident-cluster grouping, random compact-feature controls, and alert-episode evaluation rather than proposing a new classifier.Its contribution is the evaluation design and decision mapping.Source: §3 · p. 7

3.1 Module Interface in a Port Digital Twin

• In the digital-twin workflow, water-level observations update recent history and context, the analytics module produces risk scores, and the policy layer converts scores into alert episodes.Source: §3.1 · p. 7

3.2 Feature Construction

• Chronological eight-point histories are paired with prediction-time features computed from current and previous observations, while leakage filters remove warning-status, future-event, and label-derived predictors.The feature-construction path is summarized in Fig. 2.Source: §3.2 · pp. 7–8, Figure 2 · p. 8

• Elastic-net logistic regression estimates the probability that a window belongs to the 12-hour pre-warning class.Source: §3.3 · p. 8
• The objective uses an ℓ1 component to encourage sparsity and an ℓ2 component to reduce instability among correlated window descriptors.Categorical variables use most-frequent imputation and one-hot encoding; numerical variables use median imputation, without an additional numerical scaling stage.Source: §3.3 · p. 8
• Within each fold, C and α are selected by validation AP, and the top-10 list is ranked from non-zero absolute coefficients using only the model-development subset.Thresholds maximize validation F2 subject to a precision floor of 0.25 when feasible.Source: §3.3 · p. 8
• Sparse selection is used to make the warning signature compact, reproducible, and easier to inspect under sparse incident-cluster validation.The paper does not present elastic-net as a novel model.Source: §3.3 · p. 8

• One full-feature reference is ElasticNet logistic regression fitted to all 71 Liverpool candidate variables without top-k truncation.Source: §3.4 · pp. 8–9
• Outer folds hold out assigned warning-event clusters and feature ranking excludes outer-test rows, but folds 1–3 have a validation–test source-day overlap.Threshold-dependent metrics are reported with this caveat.Source: Figure 3 · p. 9
• Current12hBinary is a full-feature weighted XGBoost classifier with a binary 12-hour target; a top-10 XGBoost model tests nonlinear interactions on the same compact representation.Current12hBinary is an established nonlinear reference, not a controlled sparsity ablation.Source: §3.4 · p. 9
• The full-versus-top-10 ElasticNet comparison isolates top-k truncation within one model family, while comparisons with Current12hBinary are reference comparisons rather than causal evidence for sparsity.All models use the same 12-hour target, outer warning-cluster assignments, validation-selected threshold rule, and evaluation metrics.Source: §3.4 · p. 9

3.5 Negative and Random-Subset Controls

• Warning-label controls use day-group permutations, temporally shifted warnings, and pseudowarning days, while eight random feature-subset schemes are repeated 100 times through the four-fold grouped protocol.The random subsets include broad, feature-family, and matched-alert-budget controls.Source: §3.5 · p. 9

3.6 Nested Grouped Validation

• The outer loop holds out one warning cluster, another cluster supplies model selection and threshold tuning, and remaining clusters supply training rows; validation–test overlap makes threshold-derived scores descriptive rather than leakage-free bounds.Feature ranking records confirm that outer-test rows were not used.Source: §3.6 · p. 9

3.7 Metrics

• AP evaluates ranking without choosing a threshold; validation-threshold F2 weights recall more heavily, while episode metrics report hit rates, lead time, missed warnings, and alert duration.Row-level false alarms per day are also reported.Source: §3.7 · p. 10

3.8 Baselines and Controls

• Dense full-feature baselines test whether non-sparse learning solves the task with all safe predictors.Context-only and physical-only models isolate signal sources, while random top-k controls compare selected features with arbitrary compact subsets.Source: §3.8 · p. 10
• Random top-10 controls vary feature composition, including subsets matched to the nested selected signature.Negative controls also test real warning labels against weakened label or feature constructions.Source: §3.8 · p. 10

3.9 Alert Episode Policy

• Consecutive alerting windows are merged into an episode, with the policy evaluating three hit rules.One specified rule counts any pre-event hit during [Tw −24h, Tw).Source: §3.9 · p. 10
• A high-recall model may still be less useful for decision support if its alert episodes are very long.Episode evaluation separates model ranking from operational burden.Source: §3.9 · p. 10
• The replay preserves the held-out stream’s order before converting its scored windows into alert episodes.It is distinct from a strict latest-event temporal holdout.Source: Figure 4 · p. 11

4 Offline Twin Replay Evaluation

• Each held-out fold is replayed in timestamp order as an experimental protocol, not as live deployment, hydrodynamic simulation, or strict forward-chaining evaluation.Model development may include incidents later than the held-out cluster; only its held-out score sequence is replayed chronologically.Source: §4 · p. 11
• The trained model scores ordered water-level histories and contextual covariates, then a stored validation-selected F2 threshold forms episodes by merging alerts no more than 30 minutes apart.Scores are aggregated by their maximum across simultaneous station–measurement windows; episodes are matched to official warning timestamps after the held-out sequence is generated.Source: §4 · p. 11
• The replay is closer to a digital-twin decision interface than shuffled row evaluation, but does not establish prospective temporal generalisation or operational readiness.The evaluated object is an operator-facing alert episode rather than only a row-level score.Source: §4 · p. 11

5.1 Main Predictive Performance and Cluster Sensitivity

• 0.696 mean validation-threshold F2 for top-10 ElasticNet versus 0.633 without top-k truncation and 0.681 for full weighted XGBoost.The top-10 representation reduces 71 candidate variables to 10 predictors; the table reports means and standard deviations across four Liverpool folds.Source: Table 4 · p. 12, §5.1 · p. 12
• C2 has the lowest AP and detects only half its represented warning timestamps, while top-10 is strongest on C1 and fold thresholds range from 0.05 to 0.55.The four-cluster mean therefore does not capture the full fold-level variation.Source: §5.1 · p. 12, Table 5 · p. 13, Figure 5 · p. 13

5.2 Selection and Label Controls

• The selected result exceeds every random-control repeat-level 95th percentile, although 1% of matched-family repetitions equal or exceed it.The repeat-level comparison aggregates the same four-fold evaluation; the selected predictors are not a unique physical law.Source: §5.2 · p. 13, Table 6 · p. 14, Figure 6 · p. 14
• Real warning timing outperforms the strongest synthetic label controls; 24- and 48-hour shifts provide additional checks but do not tune the real-label model.Table 7 compares real-label AP with the largest AP from day-group permutation or pseudowarning controls.Source: §5.2 · p. 13, Table 7 · p. 14

5.3 Composition and Stability of the Sparse Signature

• The selected signature combines contextual anchors with physically interpretable short-window dynamics, while the physical component’s separation after context adjustment remains unresolved.Cyclic time gives a strong contextual baseline, and physical descriptors retain predictive value without contextual covariates.Source: §5.3 · p. 15, Table 8 · p. 15
• Cyclic time, site identity, directional reversal count, and maximum absolute increment recur in every fold’s top-10 list.Pre-warning windows also show larger local changes and ranges, but marginal checks do not account for overlapping windows or establish causality or conditional independence from context.Source: §5.3 · pp. 15–16, Table 9 · p. 16, Table 10 · p. 16

5.5 Offline Replay and Alert Burden

• 13.5–29.9 hours mean alert duration and 0.47–0.66 false warning episodes per represented day characterize policies that hit every represented warning instance.The replay’s burden coordinates supplement window-level AP or F2 when choosing an operational threshold.Source: Table 12 · p. 17, §5.5 · p. 17, Figure 8 · p. 18

5.4 Cross-Site Protocol Stress Test

• Context-only models are strong at all three sites, while locally selected top-10 models remain close to full common-feature XGBoost in Humber/Hull and Wessex South.The harmonised 67-feature experiment tests protocol repetition, not unchanged Liverpool parameters; SparseLogistic denotes elastic-net sparse logistic regression.Source: §5.4 · pp. 16–17, Table 11 · p. 16

6.3 Summary of Evidence

• Mean validation-threshold F2 rises from 0.633 to 0.696 with top-10 ElasticNet, which remains competitive with full weighted XGBoost using ten predictors.The top-10 model has the strongest mean F2 among ElasticNet variants.Source: §6 · p. 18, §6.3 · p. 19

6.1 Context, Tide, and Physical Window Dynamics

• Strong cyclic-time and station context anchors the signature, while local amplitude, increments, and directional reversals provide physically plausible supporting evidence.Because verified astronomical phase and surge residual are not retained, the contextual component cannot be assigned uniquely to tidal physics.Source: §6.1 · p. 19

6.2 Validation Scope and Deployment Path

• Liverpool's four-cluster validation remains bounded: folds 1–3 share one validation–test source day, and replay alert durations span 13.5–29.9 hours.These conditions inform the next stages of deployment-oriented validation.Source: §6.2 · p. 19
• Humber/Hull is a proxy-region protocol test, not evidence of universal transfer; site-specific calibration is valuable when contextual baselines are strong.The common-feature experiment shows the workflow can be reconstructed at additional regions.Source: §6.2 · p. 19

6.3 Summary of Evidence

• The module updates short water-level state, estimates 12-hour risk, and converts scores into alert episodes, quantifying warning hits, false episodes, and time under alert.It is an offline analytics and evaluation protocol rather than a complete live digital twin.Source: §6 · p. 18, §6.3 · pp. 19–20
• Negative-label and complete-repeat random controls associate the result with real warning timing, and the selected model exceeds the 95th percentile of tested random schemes.At the individual-fold level, context-rich random subsets have broad high tails.Source: §6 · p. 18, §6.3 · p. 19
• Larger independent incident archives, verified tide and surge-residual inputs, strict forward temporal validation, and operator-tested alert policies are required before operational deployment.The current evidence is scoped to offline evaluation.Source: §6.3 · p. 20