IHL Heat Pump Intelligence — Live Monitoring System

A real-time monitoring & alerting layer for LB Energy's Intelligent Heat Link (IHL), built on the research dataset (one climate-controlled room, 4 heat-pump units, heating + cooling weeks).

One undetected fault was burning 37% of the week's heating energy. Our monitor catches it in minutes, prices it in euros, and flags it for F-Gas compliance — then predicts the next failure before it happens.

What it does — predict · detect · visualise

• Detect — a rule-based engine surfaces faults, refrigerant leaks, comfort failures, energy waste and connectivity issues, each with a €/day cost and a recommended action.
• Predict — a physics digital twin of the room (lumped-capacitance model, fit from data, RK4-validated) forecasts heat-up time for predictive preheat and isolates unexplained heat loss (open door / envelope leak).
• Visualise — a Streamlit dashboard with a live-replay cursor that streams the historical week as if it were arriving in real time.

Headline findings (proven on the data)

Finding	Impact
🔴 Device 1 compressor down → running on 43 kW electric backup, undetected	~1,500 kWh excess/week = €29,665/yr avoidable
⚠️ Cooling an empty room (80% of energy while unoccupied)	Hundreds of kWh/week avoidable
🌡️ Occupants below target 68% of occupied time	Comfort failure despite high energy

Run it

pip install -r requirements.txt
streamlit run app.py

Open the EDA: eda_deep_dive.ipynb (deep attribute-by-attribute analysis).

Architecture

DataSource (interface)          <- swap ReplaySource for a live MQTT source
  └─ ReplaySource               to go real-time with ZERO detector/UI changes
data_loader.py   load + occupancy tagging + Modbus alarm decoding
thermal_twin.py  lumped-capacitance room twin: fit C, UA, dynamic COP(T_out);
                 RK4 simulate; one-step-ahead validation (MAE/RMSE/Willmott d)
monitors.py      detection engine -> Alert(severity, €/day, action, compliance tags)
kpis.py          hero KPIs: € avoidable, energy, CO₂, fleet health, comfort
app.py           Streamlit dashboard (live replay)
config.py        tariffs, emissions factors, COP model, thresholds (all tunable)

Methodology note

The digital twin reuses the lumped-capacitance + Runge-Kutta + statistical-validation methodology of Arumugam et al. (2023), Lumped Capacitance Thermal Modelling Approaches for Different Cylindrical Batteries. Their finding that a dynamic internal resistance outperforms a constant one is mirrored here by a dynamic COP(T_out) instead of a fixed COP.

EU compliance hooks

• F-Gas Regulation (EU 2024/573) — refrigerant low-pressure events auto-detected & logged.
• EPBD / EED — continuous technical-system performance monitoring.
• GDPR — occupancy inferred from aggregate CO₂ only; no individual tracking.

Assumptions

Device airflow/rated capacity are not in the dataset; a nominal rating is assumed (documented in config.py) so absolute COP/kW can be shown. All fault detection is fleet-relative, so conclusions do not depend on the exact assumed values.