Ashpond_system

AshPond Early Warning System is an AI-powered continuous monitoring platform designed to predict coal ash pond dyke failures before they occur. Coal ash ponds at thermal power plants retain millions of tonnes of toxic fly ash slurry behind earthen embankments. When these dykes fail, the consequences are catastrophic — groundwater contamination, agricultural destruction, and loss of life. The April 2020 Sasan UMPP breach in Singrauli is a documented case where no predictive system existed. AshPond is built to prevent the next one. The system monitors coal ash ponds across three high-density thermal clusters — Singrauli, Madhya Pradesh (Sasan UMPP, Vindhyachal STPS, Singrauli Super TPS, Rihand STPS), Korba, Chhattisgarh (NTPC Korba, CSEB Korba East, CSEB Korba West), and Bilaspur, Chhattisgarh (NTPC Sipat). Ponds are not manually catalogued — they are automatically detected. A satellite pond detection module queries Sentinel-2 imagery via Microsoft Planetary Computer, computes NDWI grids across each plant's search radius, applies threshold masking, runs scipy connected component labeling, filters components by area (5–500 hectares), extracts centroid coordinates via UTM-to-WGS84 projection, and deduplicates overlapping detections. This produces the live pond inventory without manual input. Each detected pond is assigned a Dyke Instability Index (DII) scored 0–1, classified as STABLE, ELEVATED, or CRITICAL, with a trend arrow (↑↓→) tracking movement between refresh cycles. The DII fuses five signals: Sentinel-2 NDWI for surface seepage detection, Sentinel-2 NDVI for embankment vegetation stress, Sentinel-1 SAR backscatter for subsurface moisture, Open-Meteo 72-hour batched rainfall forecasts, and historical breach proximity scores encoding site-specific structural vulnerability. Each signal is normalised and combined via calibrated weights summing to 1.0. A dominant top_risk_factor is identified per pond and surfaced on the map and dashboard. A Cox Proportional Hazards survival model — trained on 48 documented breach and near-miss events across Indian thermal plants, augmented 20x via Gaussian noise injection — converts the feature vector into a 30-day breach probability. Bayesian Monte Carlo sampling runs 1000 perturbation iterations per pond to produce 95% confidence intervals on every DII score, quantifying sensor uncertainty explicitly. Mahalanobis distance detects ponds whose sensor profiles deviate statistically from the safe historical baseline, flagging dangerous anomalous combinations that individual metrics may miss. SHAP KernelExplainer decomposes each breach probability into per-feature contributions, making predictions fully auditable. Retrospective validation reconstructs the 2020 Sasan UMPP event. PELT change point detection applied to the historical DII timeseries identifies a structural behavioural shift on April 7 — 13 days before the April 20 breach — validating the system's early warning capability against a real failure.