AI for Smart Cities & Building Management

Building Energy Optimization: HVAC, Lighting, and ENERGY STAR Ratings

Commercial buildings consume roughly 35% of total electricity in developed economies — and HVAC alone accounts for 40-60% of building energy use. ASHRAE 90.1 and the IECC set minimum energy performance standards, while ENERGY STAR (Portfolio Manager 1-100 score) and LEED benchmark actual performance. The gap between designed efficiency and operational efficiency is typically 20-40% — the "performance gap" that AI closes.

Open data/building-energy-data.csv in the code panel. Each row is an hourly energy record: building_id, timestamp, zone, hvac_kwh, lighting_kwh, plug_load_kwh, total_kwh, outdoor_temp_c, outdoor_rh_pct, occupancy_count, occupancy_pct, solar_irradiance_w_m2, bms_setpoint_c, actual_zone_temp_c, chiller_cop, ahu_status.

HVAC Optimization with Reinforcement Learning

Traditional BMS operates on fixed schedules and setpoints — 23°C during occupied hours, pre-cooling at 6 AM, night setback. This ignores thermal mass, occupancy patterns, weather forecasts, and time-of-use electricity tariff variation (utility TOU and demand charges). A model predictive control (MPC) approach enhanced with ML:

State variables (per zone):
  zone_temp, zone_rh, co2_ppm
  outdoor_temp (current + 24h forecast from NOAA / national weather service)
  occupancy_prediction (next 4 hours, from access card data + calendar)
  electricity_tariff_current (TOU period + demand charge signal from utility)
  solar_pv_generation (if rooftop solar installed)
  thermal_mass_state (estimated from zone temp trajectory)

Action space:
  chiller_setpoint (6-12°C, 0.5°C increments)
  ahu_supply_air_temp (12-18°C)
  ahu_fan_speed (30-100%)
  zone_setpoint_adjustment (-2 to +2°C from base)
  pre-cool_start_time (4 AM - 7 AM)

Objective: minimize energy_cost subject to:
  zone_temp within comfort band (ASHRAE 55 thermal comfort)
  zone_rh < 65%
  co2 < 1000 ppm

A reinforcement learning agent (Proximal Policy Optimization) trained on 6 months of BMS data from a 50,000 sq ft office achieved:

The primary savings mechanisms: (1) pre-cooling during off-peak tariff hours using thermal mass, (2) occupancy-based setpoint adjustment (raising setpoint by 1°C in zones with <30% occupancy saves 6-8% HVAC energy), and (3) chiller sequencing optimization (running 2 chillers at 70% load is more efficient than 1 chiller at 95% load due to COP curve shape).

Lighting Optimization

Lighting accounts for 15-25% of commercial building energy. ASHRAE 90.1 caps Lighting Power Density (LPD) — roughly 0.9 W/sq ft for open offices. AI-controlled lighting with daylight harvesting and occupancy sensing:

Combined HVAC + lighting optimization moves buildings up 15-20 points on the ENERGY STAR scale and can push a project from LEED Silver toward Gold — representing 15-30% total energy reduction.

Water Supply Network Management: NRW Reduction and Leak Detection

Non-Revenue Water (NRW) in aging urban water systems ranges from 15% to 30% even in developed economies (and higher in legacy networks) — compared to 5-10% in best-managed utilities. AWWA M36 provides the standard water-audit and NRW methodology, and the EPA Safe Drinking Water Act drives infrastructure replacement. The distribution efficiency problem is universal: leaking, aging pipe networks lose treated water before it is billed.

Open data/water-supply-data.csv — columns: zone_id, date, hour, supply_volume_kl, billed_volume_kl, nrw_pct, pressure_bar, pipe_material, pipe_age_years, pipe_diameter_mm, soil_type, supply_hours, consumer_count, reported_leaks, burst_count, dma_id (District Metered Area).

Leak Detection and Localization

The traditional approach to leak detection: wait for a visible leak to surface (which means losing water for weeks or months) or conduct expensive acoustic surveys. AI-enabled DMA (District Metered Area) monitoring changes the economics:

DMA-level anomaly detection:
  Feature: minimum_night_flow (MNF) — flow between 2 AM and 4 AM
  Baseline: historical MNF for each DMA (rolling 30-day average)
  Anomaly: MNF increase > 15% sustained for > 3 days

  Additional features for localization:
    pressure_differential between upstream and downstream sensors
    flow_balance_error (inflow - sum of consumer meters - estimated legitimate losses)
    pipe_burst_probability (function of pipe_material, age, soil_type, pressure)

A gradient boosted model trained on DMA monitoring data from 3 utilities predicts leak probability per DMA per week with AUC of 0.83. The model's top features:

Pressure and Demand Optimization

Even continuously pressurized networks benefit from active pressure and demand management. AI optimizes operation:

Reducing NRW from 30% to 15% in a mid-sized city (500,000 population) saves several million dollars annually in water production costs and deferred infrastructure expansion — the core economic case in any AWWA M36 water audit.

Smart City KPI Monitoring and Benchmarking

Smart-city programs monitor 30+ KPIs across transport, water, sanitation, energy, governance, and citizen services (ISO 37120 provides a standardized indicator set). The challenge: collecting KPIs is not the same as acting on them. Most cities report KPIs quarterly with 2-3 month lag. By the time a declining KPI is noticed, the underlying issue has persisted for 5-6 months.

Open data/smart-city-indicators.json — it contains: city_id, indicator_id, indicator_name, category, value, unit, timestamp, data_source, benchmark_value, national_average.

Real-Time KPI Dashboard with Anomaly Detection

AI transforms KPI monitoring from retrospective reporting to real-time alerting:

For each KPI time series:
  1. Seasonal decomposition (STL — Seasonal-Trend using LOESS)
     - Separate trend, seasonal (weekly + annual), and residual components
  2. Trend change detection (CUSUM or Bayesian Online Change Point Detection)
     - Alert when trend component changes direction or accelerates
  3. Peer benchmarking (percentile rank among similar-size cities)
     - Alert when city drops >10 percentile ranks in 2 consecutive quarters

KPI categories with highest AI value:
  - Water main breaks per 100 miles (real-time from SCADA + work orders)
  - Solid waste diversion % (from weighbridge data at landfill + MRF)
  - Public transport ridership (from AFC / fare systems, not sample surveys)
  - Air quality index (from EPA / EEA continuous monitoring stations)
  - Property tax collection efficiency (from municipal ERP)

Cross-KPI Correlation Analysis

The most valuable insight from multi-KPI monitoring: identifying leading indicators. Examples from smart-city data:

These correlations enable proactive intervention — accelerating water main replacement to prevent service disruptions rather than responding to failures after they occur.

City-Level Energy Benchmarking

For municipal operations (street lighting, water pumping, wastewater treatment, municipal buildings), AI benchmarks energy intensity against peer cities:

Key Takeaways