Probabilistic Graph Neural Inference for sustainable aquaculture monitoring systems under real-time policy constraints

Description and use

A researcher built a probabilistic graph neural inference system that models spatio‑temporal dependencies among IoT water‑quality sensors in aquaculture ponds and integrates real‑time policy constraints to enable uncertainty‑aware compliance decisions.

Problem

Per‑sensor IoT models missed cross‑pond dependencies and could not incorporate real‑time policy constraints, causing alert fatigue and non‑compliant decisions.

Outcome

The integrated system delivers uncertainty‑aware multi‑step dissolved oxygen forecasts, respects graph structure, and enables real‑time decisions that meet discharge, feeding, and antibiotic policy limits, reducing false alerts.

Workflow

• Deploy IoT water‑quality sensors across multiple fishponds to collect real‑time DO, temperature, pH, ammonia, etc.
• Model the pond network as a spatio‑temporal graph where each sensor is a node and hydrological connections (flow direction, rate, distance) are edges with attributes.
• Implement a gated graph convolution layer (HydroGatedConv) that aggregates messages using edge attributes and prevents domination by high‑degree nodes.
• Introduce latent variables per node with variational inference to obtain a posterior distribution over future sensor states.
• Integrate real‑time policy constraints (e.g., discharge limits, feeding quotas, antibiotic restrictions) into the inference pipeline to guide decisions based on the distribution.
• Deploy the trained model on edge controllers for real‑time operation, achieving compliance and reduced false alerts.

Evidence

The problem wasn't prediction accuracy on any single sensor — it was that ponds are not independent.

Source

A shared water channel meant that an ammonia spike in pond A could propagate to pond B within hours.

Source

My per-sensor model kept missing these cross-pond dependencies, and when I tried to bolt on a 'global' feature vector, the model lost the spatial structure that actually mattered: which pond was connected to which.

Source

The message-passing update I settled on is a gated variant of the standard GNN layer, because plain sum-aggregation over-weighted high-degree nodes (a hub pond feeding six others dominated everything):

Source

During my experimentation, this beat a stacked GCN + separate LSTM by a noticeable margin on multi-step DO forecasting, mostly because it avoided the latency mismatch of running two models in sequence.

Source