Mapping near-real-time power outages from social media

Figures & data

Figure 1. The proposed framework of power outage classification and location detection.

Table 1. Positive and negative power outage tweet examples.

1	1am update to power outages: 13,240 in Brevard County. Numberkeeps climbing. #HurricaneMatthew	positive
2	@ohioedison power outage in Jackson Twp. 44718. Not on map. Any news?	positive
3	I've never lived in an area with so many power outages. I'm sopissed off right now!	positive
4	This power outage isn't cool at all	positive
5	I had a dream Texas state had a huge power outage and my testwas canceled···	negative
6	@realDonaldTrump I would like to c u cut government in half andgive states the power back	negative
7	Did you lose power yesterday? We were lucky and never lost ours	negative
8	Are you prepared in the event of a power outage? https://t.co/GK8TyNdxu4	negative

Table 2. Comparison of machine learning classifiers on power outage tweet classification.

Model	Precision	Recall	F-score	Accuracy
logistic_tf	0.8335	0.8713	0.8485	0.8486
logistic_tfidf	0.8281	0.8760	0.8470	0.8471
logistic_tfidf_ngram	0.8193	0.9146	0.8560	0.8565
mult_nb_tf	0.8148	0.8900	0.8434	0.8437
mult_nb_tfidf	0.8132	0.8790	0.8381	0.8384
mult_nb_tfidf_ngram	0.8224	0.9172	0.8590	0.8594
bern_nb_tf	0.8107	0.9134	0.8493	0.8499
bern_nb_tfidf	0.8107	0.9134	0.8493	0.8499
bern_nb_tfidf_ngram	0.8121	0.9427	0.8613	0.8622
svc_tf	0.8231	0.8433	0.8310	0.8310
svc_tfidf_tf	0.8288	0.8815	0.8495	0.8497
svc_tfidf_ngram	0.8253	0.9236	0.8636	0.8641

Figure 2. Top positive and negative features for power outage classification using the SVC and bag-of-ngram model.

Figure 3. Daily power outage tweet volume (6 September to 8 October 2016).

Figure 4. The network architecture of bi-LSTM-CRF for location detection.

Table 3. Example of sentence encoded with BIO tagging scheme.

A	fire	destroyed	cabins	in	Great	Smoky	Mountains	area	this	morning	.
O	O	O	O	O	B-LOC	I-LOC	I-LOC	O	O	O	O

Table 4. Model comparisons: location detection results on the ALTA 2014 datasets.

Model	Precision	Recall	F-Score
Stanford NER	0.9451	0.3207	0.4788
Stanford NER retrained	0.8668	0.6972	0.7728
ALTA2014 Winner (Malmasi and Dras 2015)	–	–	0.7920
Deep learning method (bi-LSTM-CRF)	0.8481	0.7741	0.8094

Table 5. Datasets used for training, validating, and testing the location detection model.

Data	Train	Validation	Test
Tweets (ALTA 2014)	1851	251	930
Address names (OpenAddresses)	8107	1673	772

Table 6. Location detection using the bidirectional LSTM-CRF model with different training datasets.

Training Data	Precision	Recall	F-Score
Only Tweets (ALTA 2014)	0.8971	0.7438	0.8133
Tweets + OpenAddresses	0.8936	0.8172	0.8537

Table 7. Number of geotagged and geodetected tweets among 40,316 power outage tweets with $p\ge .9$.

Data	No. Tweets	Percentage
Geotagged	1386	3.44%
Geodetected	28,825	71.50%

Table 8. Examples of locations detected from tweets by using bidirectional LSTM-CRF.

1	Hamilton County still experiencing power outages.
2	Another power outage just popped up. This time between Scotts Hill and Hampstead. Over 100 affected.
3	Hey @username or @username any news on a power outage near wynwood?
4	UPDATE — Power restored at #SCState. Downed power line caused by a fallen tree resulted in outage. @username #scnews
5	@username trying to call you and not working. Power outage in Sun Valley Ca.. Valerio St / Cartwright Ave.
6	It appears we've had a power outage here at Smith Stadium. There were some problems with the fire alarm earlier. Mounties on move #arpreps
7	McHenry Outdoor Theater cancels double feature due to power outage - Northwest Herald
8	@username There's been a power outage & now I have no internet. I'm in the red hills road area. When will service be restored?
9	@username power outage in the 90504 area. Any updates?

Figure 5. Maps of selected power outage tweets across the United States ($p\ge 0.998$).

Figure 6. Scatter plot of power outage tweets versus number of customers without power, as reported by National Outage Map, for the period during which Matthew Hurricane affected Florida (6–7 October 2016) at both the US county and state levels. (a) County-level. (b) State-level.

Malmasi S., and M. Dras. 2015. “Location Mention Detection in Tweets and Microblogs.” International Conference of the Pacific Association for Computational Linguistics, 123–134.

 Google Scholar