machine learning for prediction purposes on pi system...

Machine Learning for Prediction Purposes on PI System Data

By: Ahmad Fattahi, OSIsoft

Michael Christopher, OSIsoft

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 2

Machine Learning

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 3

The Smart Sneezing Colleague

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 4

“We are drowning in information and starving for knowledge”

Extracting important patterns and trends

Understanding what the data says

Applications span many industries

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 5

Relationship with PI System

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 6

PI System as the system of reference for enterprise data

• Collecting and serving historical data efficiently on demand

• Data = Value! This method heavily relies on historical data

• All data in one place

Handling large amounts of valuable data

• Better statistical results

• Improved discovery of correlations

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 7

Prediction

• Load forecasting

• Production forecasting

• Preemptive maintenance

• Enterprise Resource Planning

Patterns and Trends

• Discovering patterns of interest

• Classification and clustering

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 8

Available Tools

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 9

Microsoft SQL Server Analysis Services MS Excel Client

Data

Source

Data mining

add-in

SSAS

PI

DataLink

• Great integration with PI DataLink

• Easy-to-use, wizard-based UI

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 11

tag3 = SIN(2 * tag1 + tag2) + Noise

Build the model and learn

10,000 observations

Make predictions

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 12

MATLAB

• PI SDK, PI OPC, PI ACE, ADO DB

Several ways to integrate with PI System

• Statistics and Neural Network Toolboxes

Powerful machine learning

Can be combined with other advanced analytics

External

Data Source

MATLAB

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 13

Load Prediction

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 14

Using PI System and MATLAB in a Real World Scenario

Predict power demand at OSIsoft headquarters

• Regression (Decision) Tree

MATLAB Statistics Toolbox

• Outdoor temperature

• Day of the week

• Hour of the day

Investigate multiple predictors

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 15

Architecture

HQ Facilities PI Server

MATLAB

Training

Module Regression

Parameters

MATLAB

Forecasting

Module Forecast PI Server

PI XML Interface Weather.gov forecast

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 16

Facilities Historical Data

FAC.OAK.Power-Total_Demand_Calc.PV

151.35

kW

FAC.OAK.Weather-Outside_Temperature-Val .PV

51.957

°F

11/6/2011 12:00:00 AM10/30/2011 12:00:00 AM 7.00 days

100

150

200

250

300

350

400

40

46.667

53.333

60

66.667

73.333

80

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 17

Choosing Predictors

7.3481 7.3481 7.3481 7.3481 7.3481 7.3481 7.3481 7.3481 7.3481

x 105

100

200

300

400

1 - Predictor Model; 60-day Training Period

Actual Power (kW)

Predicted Power (kW)

7.3481 7.3481 7.3481 7.3481 7.3481 7.3481 7.3481 7.3481 7.3481

x 105

100

200

300

400

2 - Predictor Model; 60-day Training Period

Actual Power (kW)

Predicted Power (kW)

7.3481 7.3481 7.3481 7.3481 7.3481 7.3481 7.3481 7.3481 7.3481

x 105

100

200

300

400

3 - Predictor Model; 60-day Training Period

Actual Power (kW)

Predicted Power (kW)

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 18

Forecasts

vCampusDemo.FAC.OAK.Power-Total_Demand_Predict

149.38

forecast.weather.gov/temperature

52.250

F

11/13/2010 12:00:00 AM11/6/2010 12:00:00 AM 7.04 days

100

150

200

250

300

350

400

40

46.667

53.333

60

66.667

73.333

80

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 19

Observations and Forecasts

vCampusDemo.FAC.OAK.Power-Total_Demand_Predict

149.38

forecast.weather.gov/temperature

52.250

F

11/13/2010 12:00:00 AM11/6/2010 12:00:00 AM 7.04 days

100

150

200

250

300

350

400

40

46.667

53.333

60

66.667

73.333

80

11/6/2011 11:47:44.389 AM10/30/2011 12:00:00 AM 7.53 days

100

150

200

250

300

350

400

40

46.667

53.333

60

66.667

73.333

80

* Observations Forecasts

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 20

So What?

• Combine with knowledge of commercial electric rates and internal schedule

Resource Planning

• Alarm on trends off the norm

Preventative Maintenance

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 21

Conclusion

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 22

Introducing Machine Learning

Explaining how it can save money and resources

Showing how it can be done on PI System Data

Demonstrating a real-world example of how it can predict load and detect faults

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C 23

Ahmad Fattahi

Michael Christopher

afattahi@OSIsoft.com

vCampus Team Member

mchristopher@OSIsoft.com

Technical Support Engineer

Thank you

© C o p y r i g h t 2 0 1 1 O S I s o f t , L L C