AI-Powered Air Quality Monitoring System Development

In densely built cities, point sensors give scattered readings. Managers need a concentration map with a 100 m grid and an accurate 48-hour forecast. Deterministic dispersion models require dozens of parameters and often err by 50–70% in variable wind. ML models trained on historical data reduce the error to 15–20%. Over 5 years, we have completed 20+ AI-monitoring projects in cities with different climates and building densities.

System Architecture

Data → Processing → Storage → Analytics → Visualization

Data:
├── Government posts (Rosgidromet, FBU CLM)
├── IoT sensors (own/partner)
├── Satellite (Sentinel-5P, MODIS)
├── Mobile stations (cars, bicycles)
└── NWP weather forecasts (Rosgidromet API, Open-Meteo)

Processing:
├── Low-cost sensor (LCS) calibration
├── Quality assurance/quality control (QA/QC)
├── Spatial interpolation
└── Air quality forecasting

Storage:
└── TimescaleDB (temporal) + PostGIS (spatial)

Analytics:
├── AQI calculation
├── Trend analysis
├── Source attribution
└── Health impact estimation

Visualization:
└── Web portal + mobile app

How ML Interpolation Works

Stations are always fewer than needed. For a 100 m air quality map, interpolation is necessary. Kriging is 30% less accurate than ML interpolation, especially in areas with local sources (factories, roads). We use XGBoost with spatial covariates: distance to highways, NDVI, building density.

def spatial_air_quality_model(station_readings, spatial_covariates):
    """
    Train on station_readings
    Predict for the entire urban 100×100 m grid
    """
    X = pd.merge(station_readings, spatial_covariates, on=['lat', 'lon'])

    # Spatial features
    X['distance_to_highway'] = ...
    X['distance_to_industry'] = ...
    X['ndvi'] = ...  # greening
    X['building_density'] = ...  # building density

    model = XGBRegressor().fit(X, X['pm25'])
    return model

# Predict for the entire city grid
grid = create_city_grid(city_boundary, resolution=100)
grid['predicted_pm25'] = model.predict(grid[feature_cols])

Deep Learning for spatial mapping: U-Net with multispectral satellite imagery + station readings → PM2.5 map at 30–100 m resolution. Training on simultaneous station and satellite data. Sensor maintenance cost savings — up to 40%.

Why LSTM Outperforms Traditional Models

Key forecast factors:

• Meteorology: wind (speed and direction determine transport), atmospheric stability (mixing height), precipitation (PM washout)
• Sources: industrial emissions, transport, heating
• Photochemistry: O3 formation and secondary particles (PM2.5)

LSTM + Weather Attention model:

class AirQualityForecastModel(nn.Module):
    def __init__(self):
        self.pollutant_encoder = LSTM(n_pollutants, 64)
        self.weather_encoder = LSTM(n_weather_vars, 64)
        self.cross_attention = CrossAttention(64, 64)
        self.decoder = nn.Linear(128, n_pollutants * forecast_hours)

Horizon: 24/48/72 hours. Achievable MAPE: < 15% for 48-hour PM2.5 forecast.

Air Quality Index (AQI)

AQI calculation per "Method for Calculating the Atmospheric Pollution Index" (RD 52.04.667):

def calculate_aki(concentrations: dict) -> float:
    """
    AQI = Σ (C_i / MPC_daily_i) for i pollutants
    AQI < 5 — standard air quality
    """
    aqi = 0
    for pollutant, conc in concentrations.items():
        mpc = MPC_DAILY[pollutant]
        aqi += conc / mpc
    return aqi

Color coding:

• Green: AQI < 5 (normal)
• Yellow: 5–7 (slight pollution)
• Orange: 7–14 (moderate)
• Red: > 14 (high, health hazard)

More about Air Quality Index.

Mobile App for Citizens

Features:

• Current AQI at user location
• City air quality map
• AQI forecast for 24/48 hours
• Recommendations: safe to walk/exercise?
• Notifications when thresholds exceed

Personalized recommendations:

• Asthmatics/allergy sufferers: stricter notification threshold
• Cyclists: optimal time/route considering AQI
• Parents with children: playground quality index

Source Attribution

Positive Matrix Factorization (PMF) decomposes PM2.5 chemical composition into sources: industry, transport, residential heating, natural (sea salt, dust). We use EPA PMF 5.0 — the official tool for receptor modeling.

from scipy.optimize import nnls
# G = F × C (observations = sources × contributions)
# PMF minimizes the weighted sum of squared residuals
# under non-negativity of F and C

Result: "30% of city PM2.5 from metallurgy, 40% from transport, 20% from residential heating." This forms the basis for regulatory decisions.

How to Assess ROI of the Monitoring System

Direct savings: reduced fines for exceeding norms (up to 2 million rubles per year in industrial zones), lower laboratory measurement costs (replaced by ML interpolation), optimized filter operation based on actual load. Indirect effect — increased citizen trust and city investment attractiveness.

What's Included in the Work

1. Data audit: assess quality and completeness of existing measurements, choose optimal stack.
2. Architecture design: data flow diagram, model selection, prototyping.
3. Sensor calibration: adjust low-cost sensors against reference stations.
4. ML model development: interpolation, forecasting, source attribution.
5. Web portal and mobile app creation: map, AQI, alerts.
6. Testing and validation: comparison with independent stations, MAPE < 15%.
7. Documentation and training: handover of code, API, instructions for ecologists.
8. 3-month support: monitoring, model retraining, updates.

Timeline: basic IoT network + AQI calculation + map + mobile app — 8–10 weeks. System with ML forecasting, source attribution, and regulatory API — 4–5 months.

Work stages:

1. Data audit — 1 week
2. Design — 1–2 weeks
3. Sensor calibration — 2–3 weeks
4. ML model development — 3–4 weeks
5. Interface development — 3–4 weeks
6. Testing — 2–3 weeks
7. Documentation and training — 1–2 weeks

Contact us — we'll analyze your data in 2 days. Get a consultation on stack selection and work scope estimation. We'll evaluate your project in 2 business days. We guarantee quality thanks to certified engineers and 5+ years of experience in AI monitoring.

When does a time series forecasting model fail in production?

The CFO requests a quarterly sales forecast. An analyst builds SARIMA on three years of data, achieves MAPE 8.3% on the test set, and deploys. Two months later, the metric in production jumps to 23%. The root cause: the model was trained on pre‑COVID data, tested on a stable period, but production hit a promotion and supply chain disruption. Data leakage plus distribution shift—perfect notebook numbers, a broken forecast in reality. We have seen this pattern dozens of times across retail, fintech, and IoT. Our team has delivered more than 50 forecasting projects over 5+ years.

Incorrect cross-validation. Standard train_test_split for time series creates data leakage: the model sees future values during training. The correct approach is TimeSeriesSplit or walk‑forward validation with an expanding window.

Multiple seasonality. Hourly electricity consumption has three seasonalities: daily (24h), weekly (168h), yearly (8760h). SARIMA handles only one. Prophet can handle multiple but scales poorly to thousands of series.

Missing values and anomalies. A missing sensor reading is information (the sensor turned off), not NaN. Linear interpolation destroys this signal. Proper handling depends on the missingness mechanism.

Cold start. A new SKU in a 50,000‑item assortment has no history, yet a forecast is needed. Standard approaches fail; cross‑learning or feature‑based methods are required.

Why is model selection critical for your data?

Prophet (Meta) – a solid start for business data with clear seasonality and holidays. Fast setup, interpretable, built‑in outlier detection. Fails on irregular patterns and does not scale beyond ~10k series without parallelization.

Gradient boosting on features (LightGBM, XGBoost) – often underestimated. Engineer lags (t‑1, t‑7, t‑28), rolling means, day‑of‑week, holidays. The model trains on all series simultaneously, solving cold start via transfer learning. MAPE in retail often beats neural nets with proper feature engineering.

TFT (Temporal Fusion Transformer) – a transformer designed for interpretable forecasting with covariates. Built‑in variable selection, temporal attention, quantile outputs. Available in pytorch‑forecasting. Requires ~10,000+ records per series for stable training.

PatchTST – splits the series into patches (like ViT for images), capturing local patterns better than classic transformers. Excellent for long‑horizon forecasting (96–720 steps ahead).

N‑HiTS, N‑BEATS – attention‑free neural architectures, faster than TFT, competitive accuracy. N‑BEATS won the M4/M5 benchmarks for tasks without covariates.

How do we deploy TFT in production?

A typical pipeline via pytorch‑forecasting:

training = TimeSeriesDataSet(
    data,
    time_idx="time_idx",
    target="sales",
    group_ids=["store", "sku"],
    min_encoder_length=max_encoder_length // 2,
    max_encoder_length=max_encoder_length,  # 120 days
    min_prediction_length=1,
    max_prediction_length=max_prediction_length,  # 28 days
    static_categoricals=["store_type", "category"],
    time_varying_known_reals=["price", "promo_flag"],
    time_varying_unknown_reals=["sales"],
    target_normalizer=GroupNormalizer(groups=["store", "sku"], transformation="softplus"),
)

A common mistake: the default target_normalizer (StandardScaler) breaks predictions for series with zero values (no sales on weekends). GroupNormalizer with transformation="softplus" is the correct choice for count data.

Case study: retail demand forecasting

A chain of 120 stores, 8,000 SKUs, 28‑day forecast horizon. The original system: SARIMA per series, MAPE 18.4%, retraining cycle – 6 hours. We replaced it with TFT on PyTorch + pytorch‑forecasting: a single model for all series, MAPE 11.2%, retraining – 40 minutes on an A10G. Feature importance via variable selection revealed that day_before_holiday influences more than the holiday date itself. Annual savings on inference alone exceeded $50,000.

Step‑by‑step configuration

1. Data collection and preparation. Handle missing values (mark NaN, interpolate only for technical failures), aggregate to required frequency, engineer covariates (holidays, promotions, prices).
2. Create TimeSeriesDataSet. Set group_ids (store + SKU), time index, forecast horizon. Choose target_normalizer based on target distribution.
3. Train a baseline. Prophet or LightGBM first – to understand complexity.
4. Train TFT. Use TemporalFusionTransformer with loss=QuantileLoss(), tune learning rate and hidden layer sizes.
5. Validate and interpret. Walk‑forward test, analyze variable selection, build attention heatmaps.

How to properly evaluate forecast quality?

RMSE alone is misleading – it over‑penalizes large values. Our standard set:

• MAPE – interpretable, unstable near zero.
• sMAPE – symmetric, avoids division by small numbers.
• MASE (Mean Absolute Scaled Error) – normalized relative to a naive seasonal forecast, ideal for comparing series of different scales.
• Pinball loss – for probabilistic forecasting, inventory management.

We guarantee a model card with these metrics on the validation set and walk‑forward results on at least 6 months of history.

What deliverables do you receive?

• Documentation of chosen architecture and hyperparameter rationale.
• Reproducible training and inference pipeline (Docker + CI/CD + Airflow/Prefect).
• Committed code with unit tests for key components.
• Team training: retraining, output interpretation, deployment of new versions.
• 3 months of post‑delivery support (consultations, bug fixes, fine‑tuning).

The model is deployed via FastAPI or Triton Inference Server. Retraining is scheduled (e.g., weekly) via Airflow with drift validation and automatic rollback if metrics deteriorate.

Process and timeline

We start with EDA: visualization, ADF test, STL decomposition, analysis of missing values and outliers. This takes 2–3 days but often reveals systemic data issues that block forecasting. Then we build a baseline (naive seasonal, Prophet), engineer features for LightGBM, and select a neural architecture if needed. Walk‑forward validation with a realistic horizon. Deployment via API with automatic retraining scheduled via Airflow or Prefect.

Timeline: MVP forecast on one data type – 3–6 weeks. Hierarchical forecasting system with automation – 2–5 months. Cost is calculated individually based on data volume, number of series, and required accuracy.

Our team consists of certified ML engineers (AWS ML Specialty, GCP Professional ML Engineer) with 5+ years on the market and over 50 completed forecasting projects. Contact us for a free analysis of your data – we will assess the task and provide initial recommendations within 1–2 days. Request a consultation to ensure your forecasts work in production, not just in a notebook.