HAMK Smart City — Particle Swarm HVAC Optimisation for Public Buildings

Design Solution · HVAC & Energy

Design Solution · Dream about it

ML-driven HVAC setpoint optimisation using demand forecasting and particle swarm search to cut energy use in district-heated public buildings.

HAMK Smart City couples GRU/LSTM neural networks (peer-reviewed, 29 kWh MAE accuracy) to forecast district-heating demand with Particle Swarm Optimisation that searches optimal HVAC setpoint schedules. It targets the energy-comfort trade-off in public building operation: reducing heating energy consumption while maintaining occupant comfort bands. The system is implemented in open-source code but has not yet been validated against live in-situ energy savings or deployed in closed-loop operation at scale.

The HAMK approach layers two techniques: GRU/LSTM neural networks forecast district-heating demand (peer-reviewed with a 29 kWh MAE accuracy result) and Particle Swarm Optimisation then searches for setpoint schedules that minimise energy consumption while keeping comfort within defined bands. The architecture separates forecasting from optimisation cleanly, which is a genuine engineering advantage — each layer can be improved without rebuilding the other — and the open-source implementation enables transparency and adaptation. The fundamental gap is that there is no validated in-situ energy savings data from a live deployment; the system runs in open-loop research configuration, not closed-loop building control at scale. That distinction matters enormously for a specifier: predictive accuracy in a paper and autonomous daily HVAC control are different confidence levels. Data quality is the upstream dependency — the model degrades with missing or inconsistent historical weather, demand, and occupancy feeds, which are not uniform across public building portfolios. District-heating system topology varies widely, and the PSO recalculation overhead has not been documented against actual building control response times. The promise is real and the academic foundation is solid; the honest position is that this is a research pipeline approaching deployment, not a product with a commissioning and handover protocol. A building team evaluating it should treat it as a pilot candidate requiring in-situ validation rather than a proven specification.

Strengths

• Reduces HVAC energy consumption by automating setpoint schedules based on predictive demand signals rather than fixed time-clocks.
• Decouples forecasting (peer-reviewed ML model) from optimisation (PSO), allowing each layer to be improved independently.
• Open-source implementation enables transparency, replicability, and adaptation across different district-heating networks.
• Maintains comfort within defined bands during optimisation, avoiding trade-off of occupant wellbeing for energy gain.
• Handles multi-objective search (energy vs. comfort) without requiring hand-tuned rules or manual intervention.

Considerations

• Requires high-quality, continuous historical data (weather, demand, occupancy, comfort feedback) to train and retrain forecasting models; missing or poor data degrades both prediction and optimisation. (high)
• PSO setpoint optimisation introduces computational overhead and latency; real-time or near-real-time recalculation may be impractical in buildings with slow HVAC response or limited edge computing. (moderate)
• District-heating systems vary widely in topology, response time, and control interfaces; the generic approach may not suit all network configurations without re-tuning. (moderate)
• Operator or maintenance staff must understand and trust ML-driven schedules; lack of explainability or intuitive override mechanisms can create resistance and manual circumvention. (moderate)

Risks

• No validated in-situ energy savings data from live deployments (high)
• No closed-loop operational validation at scale (high)
• District heating system dependency limits geographic applicability (moderate)
• Comfort band definition and occupancy forecasting inputs not specified (moderate)
• Particle swarm convergence time and computational overhead not documented (low)

Performance

• Demand forecasting accuracy: 29 kWh MAE
• Primary objective: Heating energy reduction + comfort maintenance

Reality check

Forecasting module has peer-reviewed publication with quantified test error (29 kWh MAE). PSO optimisation layer exists in code but no separate publication, no field trial results, no measured operational energy savings, no evidence of closed-loop live control. GitHub repo provides implementation but lacks deployment logs, control outputs, or pre/post energy bills. Claimed benefit remains algorithmic validation on historical data only—not a proven operational system.

#machine-learning #hvac-optimisation #demand-forecasting #district-heating #setpoint-scheduling #swarm-intelligence #public-buildings #open-source

Source