Lossless Compression of Cloud-Cover Forecasts for Low-Overhead Distribution in Solar-Harvesting Sensor Networks

(1)

Lossless Compression of Cloud-Cover

Forecasts for Low-Overhead Distribution in

Solar-Harvesting Sensor Networks

Christian Renner and Phu Anh Tuan Nguyen

ENSsys ’14, Memphis, TN, USA November 6th_{, 2014}

(2)

Motivation

Lossless Compression of Cloud-Cover Forecasts

Making the Most of Harvestable Energy

Objective perpetual operation

Challenge changing weather conditions

Instrument load adaptation

Prerequisite energy harvest prediction

(3)

Motivation

Sources of Harvest Variation

Global Weather conditions

∎ Inter and intra-day variation

∎ Seasonal effects

Local Location-specific pattern

∎ Shades of buildings and trees

∎ Dirt deposits ∎ Hardware aging ∎ Distribution required ∎ Unused ∎ Locally available ∎ Frequently used 2

(4)

Motivation

Advantages of Integrating Cloud Cover Forecasts

0 4 8 12 harvest (mA) 70 71 72 73 74 75 76 77 78 79 80 −2 −1 0 1 2 ME (mA) EWMA Kimball-1 3

(5)

Motivation

Data Distribution

Challenge

∎ One-to-many communication (data distribution)

∎ Reversed data flow↔data collection

∎ Elevated network energy expenditure

Approach

∎ Piggy-back on collection data acknowledgments

∎ Reduces energy overhead

∎ Requires small size↝compression

(6)

1 ₂2 3 4

(7)

Lossless Compression

Compression Checklist

∎ Complete forecasts

differential data problematic in case of packet loss

∎ Lossless compression

coarse-grained input data, no experience with lossy forecasts

∎ Light-weight decoding

execution on resource-constrained sensor nodes, decoding consumes energy

∎ Exploit data properties and patterns

(8)

Cloud-Cover Forecasts

0 6 12 18 24 30 36 42 48 54 60 66 72 0 2 4 6 8 time (h) cloud cover (eighth)

∎ Cloud cover in eighths

∎ Hourly resolution

∎ Multi-day horizon

∎ Hourly updates

(9)

Data Analysis

0 1 2 3 4 5 6 7 0 10 20 30

cloud cover (eighth)

frequency (%) −3 −2 −1 0 1 2 3 0 20 40 60 80 100

cloud cover delta (eighth)

frequency

(%)

∎ difference of adjacent forecast values (deltas) is

∎ zero in 82% of all cases ∎ at most one in 99% of all cases

∎ deltas have lower entropy than absolute values

(10)

Compression Algorithms

Delta Coding (DC)

∎ Huffman encoding of delta values (-8, ..., 8)

∎ First value unencoded (absolute value)

Cloud-cover forecast 5 5 6 5 5 5 8 8 Deltas 5 0 +1 -1 0 0 +3 0 Encoding 0101 0 10 110 0 0 1111110 0 8

(11)

Compression Algorithms

Partial Delta Coding (PDC)

∎ Huffman encoding of delta values -1, 0, 1

∎ Larger deltas encoded as absolute values with prefix

∎ First value unencoded (absolute value)

Cloud-cover forecast 5 5 6 5 5 5 8 8 Deltas 5 0 +1 -1 0 0 +3 0 Encoding 0101 0 10 110 0 0 111 1000 0 9

(12)

Compression Algorithms

Daylight Delta Coding (DDC)

∎ PDC encoding

∎ Omit night slots

∎ Indicate number of values before sunrise / sunset

Cloud-cover forecast 5 5 6 5 5 5 8 8 Deltas 5 0 0 +3 0 Encoding 0101 0 0 111 1000 0 10

(13)

1 2 ₃3 4

(14)

Evaluation

Setup and Metrics

Data set

∎ Cloud-cover forecasts from an online weather forecast service

∎ April 2013 to July 2014

∎ 10 590 forecasts

∎ Hourly resolution, 10-day horizon

∎ Sunrise and sunset times from sunrise equation

Compression

∎ DC, PDC, DDC

∎ Message sizes (including all static information)

(15)

Evaluation

Results

Compression Performance and Comparison DC 1 2 3 4 5 6 7 0 10 20 30 40 forecast horizon (d) encoded length (B) DDC 1 2 3 4 5 6 7 0 10 20 30 40 forecast horizon (d) encoded length (B) 12

(16)

Evaluation

Results

Optimality Study

horizon (d) avg. compression code length (bit/value)

optimal DC PDC DDC 1 1.00 1.50 1.51 1.43 2 0.94 1.41 1.42 1.14 3 0.92 1.37 1.38 1.04 4 0.91 1.35 1.36 0.99 5 0.90 1.34 1.35 0.95 6 0.90 1.32 1.33 0.93 7 0.89 1.31 1.32 0.91 13

(17)

Evaluation

Results

Seasonal Daytime Influence

3-day forecasts: saving of DDC over PDC

Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul −50 0 50 time of year saving (%) 14

(18)

Evaluation

Summary

Results

∎ DDC achieves best results

∎ Compression of up to 76%

∎ Forecasts of up to3d consume at most17byte

∎ Efficient (decoding) implementation

(19)

1 2 3 ₄4

(20)

Conclusion & Future Work

Conclusion

Cloud-cover forecasts

∎ improve (solar) harvest prediction accuracy

∎ are compressible to a few bytes with low computation overhead

∎ can be distributed into the network via acknowledgment

piggy-backing

(21)

Conclusion & Future Work

Current and Next Steps

∎ Implementation for TinyOS

∎ Latency evaluation for piggy-backed data distribution

∎ Field test

∎ Compression / distribution of other weather metrics

e.g., sunshine duration per hour

(22)