1
GSMA Internet of Things Seminar:
MWC Shanghai 2018
Dr. Hau Chen Mike Lee
Far Eastone, Taiwan
AI+IoT Case Study –
Air Quality Forecast Model
and
beyond
2
Air Pollution, an important living quality
issue in Taiwan
• 7 of top 10 causes of death in Taiwan, are
closely related to air pollution.
• Government took many measures but
difficult to measure effectiveness
PM2.5 Taiwan Status
2
Universtiy of Texas. Environ. Sci. Technol. 2015, 49, 8057−8066
Harvard School of Public Health. Environ Health Perspect 2016
CVD & COPD
Polluted Air Can Lead to Cardiovascular and
Respiratory Diseases
Increase 2.1ug/m3 of PM2.5,
eGFR will decrease like the
picture shown below accordingly.
4
Air Quality IoT devices
collected big data
connected devices
to monitor Air
Quality
IoT Real Values to Customers is Information
Information
Sensor Qualities
and Quantities
(many similar competitors)
Qualities of data sets
How often do we have data?
How accurate?
Can we use big data to
forecast air quality?
For how long?
How good?
(Competition on value)
DATA
IoT
Big Data
AI/
Machine Learning
4
Forecast
air quality
5
PoC to Validate Machine Learning
6
Learning from PoC
NO2
Reading
Air Quality
Station Data
Predict Taiwan N02 Value
NO2 _PPB Description NO2 _PPB Index Predicted NO2 Reading
Tempera-
ture
Wind
Direction
Barom-
etric
Pressure
Wind
Speed
Weather
Station Data
Pollution Source
Power Plant data
FET mobile
network data
FET Population
Aggregation
Gradient Boosting
Extra Trees
Random Forest
Data Processing & Model Tuning
Parameter Tuning
Learning Rate,
Max Features,
Max_Depth
Subsamples
GSMA
Model
Result
Grid Search
Filter outlier
Linear
interpolation
Data ProcessingOriginal
NO2 _ Predicted PPB
7
Summary of Model Verification
Best Performance Model
Base 9% 10% 13% 15%
R2 Best Performance
Increase Rate
Predicted Taiwan NO 𝟐
GSMA Machine Learning Model
Weather Data
Interpolation
FET Population
Aggregation
Power Plant
Data
Model Parameter Tuning
(Grid Search)
NO2 Reading = 0
Filtering NO2
FET Carrier
Data
► Best Performance of Gradient Boost R2 : GSMA ML Model + Data
Interpolation + FET Population + Power Plant Data + Model Tuning
Objectives of “Air Quality Forecast Project”
8
PM 2.5 NO2
Build Forecast pipeline for London and Taiwan –
models for urbanization effects
Localize Forecast models for Taiwan by adding
industrialization impact
Expand to PM2.5
Commercialize with local government first
How we Build it
Acquire Taiwan air-
quality & weather
data
Compare London air-
quality model
learning performance
with Taiwanese
models
Adding weather
forecast information
in London and
Taiwan to build air
quality forecast
London/TW AQ &
Weather
Data Acquisition
1
London ML Model
Comparison with
Taiwan Data
2 3
Adding Weather
Forecast Info to build
Air Quality Forecast
9
Localization Taiwanese
Models by adding
Industrial Impacts
Adding industrial Impacts
to Taiwanese Models
4
10
► Higher forecast resolution than area forecast
► Leverage both civilian and government sensors
We can provide reliable Air Quality Forecast
on a per station level
Taipei
Forecast Examples
11
London Taiwan
What we learned
1. Air quality is a key issues in Smart City, Health Care, and Industrial Anti-
pollution
2. Working forecast model need to integrate
► AI machine learning
► Carrier big data + IoT big data platform + weather forecast
► Air Quality IoT Sensors
3. Benchmark is important
4. New tech/new tools are amazing
AQ IoT sensor
open data
FET carrier data
& IoT big data
Health care
Smart City
Industrial
Anti-pollution
AQ forecast
12
Expanding Analytics to other services
With AI + IoT
From smallest tracker
to smartest services
13
PARTNERS CONNECTIVITY AREA
15
Thank you
15
FET is willing to share our development
experiences with carriers

AI+IoT case study - Air Quality forecast for MWC SH 2018

  • 1.
    1 GSMA Internet ofThings Seminar: MWC Shanghai 2018 Dr. Hau Chen Mike Lee Far Eastone, Taiwan AI+IoT Case Study – Air Quality Forecast Model and beyond
  • 2.
    2 Air Pollution, animportant living quality issue in Taiwan • 7 of top 10 causes of death in Taiwan, are closely related to air pollution. • Government took many measures but difficult to measure effectiveness PM2.5 Taiwan Status 2
  • 3.
    Universtiy of Texas.Environ. Sci. Technol. 2015, 49, 8057−8066 Harvard School of Public Health. Environ Health Perspect 2016 CVD & COPD Polluted Air Can Lead to Cardiovascular and Respiratory Diseases Increase 2.1ug/m3 of PM2.5, eGFR will decrease like the picture shown below accordingly.
  • 4.
    4 Air Quality IoTdevices collected big data connected devices to monitor Air Quality IoT Real Values to Customers is Information Information Sensor Qualities and Quantities (many similar competitors) Qualities of data sets How often do we have data? How accurate? Can we use big data to forecast air quality? For how long? How good? (Competition on value) DATA IoT Big Data AI/ Machine Learning 4 Forecast air quality
  • 5.
    5 PoC to ValidateMachine Learning
  • 6.
    6 Learning from PoC NO2 Reading AirQuality Station Data Predict Taiwan N02 Value NO2 _PPB Description NO2 _PPB Index Predicted NO2 Reading Tempera- ture Wind Direction Barom- etric Pressure Wind Speed Weather Station Data Pollution Source Power Plant data FET mobile network data FET Population Aggregation Gradient Boosting Extra Trees Random Forest Data Processing & Model Tuning Parameter Tuning Learning Rate, Max Features, Max_Depth Subsamples GSMA Model Result Grid Search Filter outlier Linear interpolation Data ProcessingOriginal NO2 _ Predicted PPB
  • 7.
    7 Summary of ModelVerification Best Performance Model Base 9% 10% 13% 15% R2 Best Performance Increase Rate Predicted Taiwan NO 𝟐 GSMA Machine Learning Model Weather Data Interpolation FET Population Aggregation Power Plant Data Model Parameter Tuning (Grid Search) NO2 Reading = 0 Filtering NO2 FET Carrier Data ► Best Performance of Gradient Boost R2 : GSMA ML Model + Data Interpolation + FET Population + Power Plant Data + Model Tuning
  • 8.
    Objectives of “AirQuality Forecast Project” 8 PM 2.5 NO2 Build Forecast pipeline for London and Taiwan – models for urbanization effects Localize Forecast models for Taiwan by adding industrialization impact Expand to PM2.5 Commercialize with local government first
  • 9.
    How we Buildit Acquire Taiwan air- quality & weather data Compare London air- quality model learning performance with Taiwanese models Adding weather forecast information in London and Taiwan to build air quality forecast London/TW AQ & Weather Data Acquisition 1 London ML Model Comparison with Taiwan Data 2 3 Adding Weather Forecast Info to build Air Quality Forecast 9 Localization Taiwanese Models by adding Industrial Impacts Adding industrial Impacts to Taiwanese Models 4
  • 10.
    10 ► Higher forecastresolution than area forecast ► Leverage both civilian and government sensors We can provide reliable Air Quality Forecast on a per station level Taipei
  • 11.
  • 12.
    What we learned 1.Air quality is a key issues in Smart City, Health Care, and Industrial Anti- pollution 2. Working forecast model need to integrate ► AI machine learning ► Carrier big data + IoT big data platform + weather forecast ► Air Quality IoT Sensors 3. Benchmark is important 4. New tech/new tools are amazing AQ IoT sensor open data FET carrier data & IoT big data Health care Smart City Industrial Anti-pollution AQ forecast 12
  • 13.
    Expanding Analytics toother services With AI + IoT From smallest tracker to smartest services 13
  • 14.
  • 15.
    15 Thank you 15 FET iswilling to share our development experiences with carriers