← Lectures·L17

AI for Thermohydraulic Models

Tolba · 1 concepts · 7 questions

Concept 1 / 1

Key Concepts to Memorize

Water system components (4 stages):


1. Collection

2. Purification

3. Transmission

4. Distribution


Network topology types:


  • Branching network — tree structure, dead ends (no redundancy)
  • Gridiron network — looped/ring structure (redundancy, more resilient)

Monitoring devices in WDN:


  • Flow and pressure sensors
  • Water quality sensors
  • Noise loggers (for leak detection)
  • Smart meters (consumption measurement)

Pure AI models — limitations:


  • Can produce non-physically correct results (violate conservation laws)
  • Sensitive to noise
  • "Black box" — hard to debug
  • Cannot guarantee physical plausibility

Pure Physical models — limitations:


  • Equations are approximations
  • Sensitive to missing or wrong GIS data
  • Assume constant material properties
  • Difficult to update in real time

Hybrid approach — 2 strategies:


1. Physics for AI (PINN/PIGNN): embed physics equations into the neural network's loss function → model is constrained to follow physical laws during training

2. AI for Physics: use AI to discover missing GIS data, predict discrepancies between model and measurement, or estimate physical properties (e.g., temperature)


Physics Informed Graph Neural Networks (PIGNN):


  • Water networks modelled as graphs: pipes = edges, junctions/nodes = nodes
  • Message Passing (2 steps):

- Aggregate: each node collects information from neighbouring nodes

- Combine: each node updates its own representation using aggregated info

  • Architecture: Encoder → Processor → Decoder
  • Physics constraints (PDE loss) embedded in training loss → ensures physical consistency