How Smart Cool Tech Works

Smart Cool Tech uses advanced AI algorithms to optimize chiller plant operations in real-time.

Core Technologies

1. Bayesian Neural Networks (BNN)

Purpose: Predict chiller COP (Coefficient of Performance) with uncertainty quantification

How it works:

Train individual models for each chiller using historical data
Predict COP under various operating conditions
Provide confidence intervals (sigma) for every prediction
Automatically fallback to physics-based models when uncertainty is high

Key Benefits:

Probabilistic predictions - Know when to trust the model
Robust operation - Graceful degradation under novel conditions
No retraining needed - Physics-based fallback ensures continuous operation

2. XGBoost Load Forecasting

Purpose: Predict future cooling load demand

Features used:

Temporal: Hour, day of week, month, holidays
Weather: Temperature, humidity, solar radiation
Building-specific: Historical load patterns, occupancy

Performance:

Typical Accuracy: < 10% MAPE
Forecast Horizon: 1-24 hours ahead
Update Frequency: Hourly with latest weather forecasts

3. Dynamic Programming Optimization

Purpose: Find optimal chiller sequencing that minimizes energy consumption

The Problem: With N chillers and T time steps, there are 2^(N×T) possible sequences.
For 6 chillers over 24 hours: 2^144 ≈ 10^43 combinations!

The Solution: Dynamic Programming exploits problem structure:

Breaks problem into subproblems
Eliminates redundant evaluations
Reduces computation from minutes to seconds

Constraints Handled:

Minimum on/off times (MUT/MDT)
PLR limits (0.3 - 1.0)
Capacity constraints
Safety margins

4. Online Bias Correction

Purpose: Adapt to changing conditions without retraining

Method: EWMA (Exponentially Weighted Moving Average)

bias_t = α × error_t + (1-α) × bias_(t-1)
corrected_prediction = raw_prediction - bias_t

Benefits:

No retraining - Continuous adaptation
Fast response - Adapts within hours
Robust - Handles equipment degradation, fouling, seasonal changes

Optimization Workflow

graph TD
    A[Real-time Data] --> B[Load Forecasting<br/>XGBoost]
    A --> C[BMS Data]
    B --> D[Predicted Load]
    C --> E[Current State]
    
    D --> F[Generate Feasible<br/>Sequences]
    E --> F
    
    F --> G[For each sequence:<br/>Predict Power]
    C --> G
    
    G --> H[BNN COP Models]
    H --> I[Dynamic Programming<br/>Find Minimum]
    
    I --> J[Optimal Sequence]
    J --> K[Dashboard]
    J --> L[Recommendations]

Key Concepts

COP (Coefficient of Performance)

COP = Cooling Load (kW) / Power Consumption (kW)

Higher COP = Better Efficiency
Typical range: 4.0 - 7.0
Varies with load, temperatures, and equipment condition

PLR (Part Load Ratio)

PLR = Current Load / Rated Capacity

Most chillers are most efficient at PLR = 0.5 - 0.8
Operating at very low or very high PLR reduces efficiency
Smart Cool Tech balances load across chillers to maximize system COP

Sigma (σ) - Prediction Uncertainty

Bayesian Neural Networks provide:

Mean prediction - Expected COP
Standard deviation (sigma) - Confidence level

Decision Logic:

if sigma > threshold:
    use physics-based fallback model
else:
    use BNN prediction

This ensures robust operation even when model is uncertain.

Optimization Objective

Minimize:

Total Energy = Σ (Chiller Power + Pump Power + Cooling Tower Power)

Subject to:

Cooling load must be met
All chillers within PLR limits
Minimum on/off times respected
CHWS temperature maintained

Reliability Mechanisms

1. CHWS Temperature Monitoring

Continuous tracking of chilled water supply temperature
Automatic staging if temperature exceeds threshold
Prevents thermal comfort violations

2. Model Monitoring

Track prediction accuracy in real-time
Alert when performance degrades
Automatic fallback to physics-based models

3. Constraint Validation

All recommendations validated against operational constraints
Safety margins included
Human operator always has final control

What Makes Smart Cool Tech Different?

Feature	Smart Cool Tech	Traditional BMS
Optimization	AI-driven, global optimal	Rule-based, local optimal
Adaptability	Online learning, bias correction	Fixed rules
Uncertainty	Quantified with BNN	Unknown
Deployment	Brick Schema, < 1 week	Custom code, months
Performance	10%+ energy savings	0-5% improvement

Technical Papers

For detailed algorithm descriptions, see our publications:

Li, Shuhao, Siqi Li, and Zhe Wang. "Accelerating chiller sequencing using dynamic programming." Energy and Buildings 325 (2024): 115037.
Li, Shuhao, et al. "Field demonstration of model predictive control for chiller sequencing in large-scale commercial buildings." Energy and Buildings (2025): 116021.
Li, Shuhao, et al. "Applying semantic model for easy and fast deployment of chiller sequencing algorithm." Energy and Buildings (2025): 116830.

Core Technologies​

1. Bayesian Neural Networks (BNN)​

2. XGBoost Load Forecasting​

3. Dynamic Programming Optimization​

4. Online Bias Correction​

Optimization Workflow​

Key Concepts​

COP (Coefficient of Performance)​

PLR (Part Load Ratio)​

Sigma (σ) - Prediction Uncertainty​

Optimization Objective​

Reliability Mechanisms​

1. CHWS Temperature Monitoring​

2. Model Monitoring​

3. Constraint Validation​

What Makes Smart Cool Tech Different?​

Technical Papers​