Forecasting Using

Machine Learning Techniques

2
Acadia Institute for Data Analytics (AIDA)

Foci:
Agriculture,
Oceans,
Environment,
Green Energy,
Health Care

3
Typical Domain: AIS Data
• Automatic Identification System
on vessel transmits:
• time stamp
• vessel position (GPS),
• identity,
• speed and course.
• Received by marine traffic control
and all other AIS equipped
vessels within range.
• Various problems can be posed:
• Type of vessel?
• Current activity?
• Next action?
• Probability of intersection?

4
 Typical Problem and Approaches
Problem can be framed as: • Mix of data
• Structured prediction • Input: temporal, spatial, auto-
• Time series classification regressive, static or independently
• Image recognition varying
• Target: prediction, next state or class
• Temporal variance
Not-fishing
• Convert to common periodic signal
• Include time signal
• Spatial variance
• Normalize for scale, rotation, trans.
• Use convolutional methods
• Sequential data
Fishing
• Convert to sliding window and use
standard ML and Deep NN
• Use recurrent neural network (RNN)

o5
Classifying Fishing Activity
• Structured prediction

under-sampling

• Image recognition

6
 Classifying Fishing Activity
• Structured prediction

Partition-wise Recurrent Neural Networks (pRNNs) to Identify Fishing Activities

7
Solar Energy Forecasting
• Problem: Predict GHI 5-120 min. into the future
• Data:
5 min 10,000 records,60xmin3 sites, 5 min intervals,
inputs: time, weather, satellite, prior GHI
• Performance metric: %RMSE
15 min 90 min
• Also observed value of transfer learning
- no transfer

• BP and Recurrent ANNS - with transfer

30 min 120 min

8
Building Energy Management
• Problem: Predict best HVAC action to better
manage the energy used by a large building
• Data: data at 15 min intervals, inputs:
time, weather, satellite, GHI
• Performance metric: Energy use
• ANNs and Non-linear Optimization

9
Sales Forecasting
• Problem: Predict hourly sales 1 week in advance
• Data: 4 years of sales data from 3 stores, 5 min
intervals, inputs: date/time, special events,
weather, satellite, prior sales
• Performance metric: MAE in $
• Also observed value of transfer learning
• BP ANNs

10
Stream Flow Rate Estimation
• Problem: Predict stream flow rate (m^3/sec)
one day into future using transfer learning
• Data: 30 years of stream flowrate and weather
data (1971-2000), from 4 streams, 1 day
intervals, no auto-regressive terms used
• 17 years train / val, 13 years test
• Performance metric: MAE in m^3/sec
• BP ANNs using a windowing approach

11
Stream Flow Rate Estimation
16

15

MAE (m^3/s)
14

13

12

11
0 1 2 3 4 5 6
Years of Data Transfered

No Transfer Wilmot Sharpe Sharpe & Wilmot Shubenacadie

x = weather data

f(x) = flow rate

Stream flow rate prediction [Lisa Gaudette, 2006]

12
Aquifer Recharge Prediction
• Problem: Predict well levels based on weather
• Data: 30 yrs of weather (1971-2000), stream
flow & well level; 17 yrs train / val, 13 yrs test
• Performance metric: MAE, 5 days in advance
• Interested in value of transfer learning
• Mulitple Task Learning ANNs

13
Cantelope Melon Yield
• Problem: Predict melon yield up to 14 days
• Data: 6 years of melon harvest data (2007-
2013), inputs: date, prior weather, prior
harvest, high resolution weather forecast
• Performance metric: RMSE accum. melons
• BP and Recurrent ANNs

14
Cantelope Melon Yield
Field Cameras

15
Cantelope Melon Yield
Data Preparation

16
Cantelope Melon Yield
• Approach: build a model that predicts tomorrows
accumulated yield
• Repeatly use model 14 times (days) using mix of prior and
forecasted weather and accumulated yield as input

17
Cantelope Melon Yield
Predicted
Actual
Upper error bound
Lower error bound

18
Cantelope Melon Yield
Weather forecast A
Weather forecast B

Actual weather

RMSE of model for

19
UAV Image Classification
• Problem: Classify relevant objects in UAV
field images: bareground, weeds,
investations
• Data: Mosaics of RGB and Near Infrared
(NIRGB) images (up to 1B B pixels)
capture by a drone flying over a blueberry
field
• Performance metric: Sensitivity and
specificity of pixels by object class
• Appoach: Random Forest and
DBN models (semi-supervised)

20
 UAV Image Classification

Use Case:
• Classifying crop
versus
bareground
21
UAV Image Classification

22
UAV Image Classification

Use Case:
• classifying
weeds in NDVI = Normalized
UAV images Difference Vegetation
Index

Goldenrod weed
Classification

23
Estimage – Estimating
Things from Images
• Problem: Count relevant objects in an
image: pints of blueberries, baskets of
grapes, bins of apples, logs,
• Data: Photos from a smart phone
• Performance metric: MSE over pixels,
MAE over count

24
Estimage – Estimating
Things from Images
• Problem: Count relevant objects in an
imgae: pints of blueberries, baskets of
grapes, bins of apples, logs,
• Data: Photos from a smart phone
• Performance metric: MSE over pixels,
MAE over count
• Appoaches:
• Random Forests
• Convolution Neural Networks
• Convolution-Deconvolution Neural Networks

25
Estimage – Estimating
Things from Images
• Problem: Count relevant objects in an
imgae: pints of blueberries, baskets of
grapes, bins of apples, logs,
• Data: Photos from a smart phone
• Performance metric: MSE over pixels,
MAE over count
• Appoaches:
• Random Forests
• Convolution Neural Networks
• Convolution-Deconvolution Neural Networks

26
Estimage – Estimating
Things from Images
• Problem: Count relevant objects in an
imgae: pints of blueberries, baskets of
grapes, bins of apples, logs,
• Data: Photos from a smart phone
• Performance metric: MSE over pixels,
MAE over count
• Appoaches:
• Random Forests
• Convolution Neural Networks
• Convolution-Deconvolution Neural Networks

27
Summary
• There are wealth of problems in the area
spatio-temporal forecasting and classification
• Recent develpoments in deep learning and
recurrent neural networks (LSTM, GRU)
provide exciting possibilities

28
Thank You

• Email: danny.silver@acadiau.ca
• Webpage: aida.acadiau.ca

29