Caribbean AI Association
Back to the archive
Trinidad and Tobago / Working paper

Screening monthly gas use in Trinidad and Tobago with nearest neighbours

An unsupervised test across four gas-consuming sectors

Lead author · Adrian Dunkley   |   2026-11-20   |   v1.0

In plain English

Two months can use similar amounts of gas overall while factories and power stations receive different shares. We compared each 2024 month with its closest match from 2023, looking at four uses at once.

June stood out whichever single sector we removed. April was a closer call. The distinction tells an analyst where to start checking the records and why the national total alone gives a different list.

Distance from each 2024 month to its closest 2023 match

Standardised distance
Jan0.581
Feb0.775
Mar0.968
Apr1.758
May1.362
Jun4.719
Jul3.824
Aug1.211
Sep1.415
Oct1.391
Nov1.424
Dec1.208
Higher scores indicate less familiar sector combinations. The four-sector reference threshold was 1.703. Scores are unitless. Calculated from [1, 2].

Analysis steps

  1. Scale sector rates
  2. Find closest month
  3. Compare with threshold
  4. Check source records
Read the full paper

We compared monthly gas use in 2024 with twelve reference months from 2023. A nearest-neighbour screen flagged April, June and July. June exceeded the reference threshold even when we omitted each sector in turn. A screen using national gas use alone also flagged August and October but missed April. The results identify records for review. The bulletins contain no incident labels against which to measure detection accuracy [1, 2].

Research question

Does a four-sector screen select the same months for review as a screen based on total national gas use?

Data and methods

From Table 3B of the amended 2023 and 2024 energy bulletins, we extracted gas use for power generation, ammonia, methanol and liquefied natural gas (LNG), plus the national total. Each observation is a monthly average daily rate in millions of standard cubic feet per day [1, 2].

For each pair of months we divided the sector differences by their respective 2023 sample standard deviations, squared them, summed them and took the square root. Each 2024 month received its smallest distance to any 2023 month. This is a one-nearest-neighbour novelty score [3]. The threshold was the largest distance from a 2023 month to its closest other 2023 month. Only scores strictly above it triggered review.

Results

The four-sector threshold was 1.703. April scored 1.758, June 4.719 and July 3.824. June's closest match was August 2023. In June 2024 gas use for ammonia was 294 and methanol 409 million standard cubic feet per day. Both fell below their 2023 monthly minima of 359 and 483 [1, 2].

Repeating the full calculation on the national total alone flagged June, July, August and October. When we removed LNG and recalculated the four-sector test using the three remaining sectors, only June triggered review. June remained above the recalculated threshold in all four single-sector omissions.

Discussion

A reviewer investigating April would see an unusual combination across sectors even though the national total of 2,364 lay within the 2023 range of 2,331–2,665 [1, 2]. The score identifies which records merit comparison. Plant logs and reporting notes would be needed to explain a difference.

Limitations

One reference year cannot represent the full range of seasonal and industrial conditions. The amended bulletins are retrospective. The method has no information about planned shutdowns, reporting errors or confirmed disruptions. A flag estimates neither the probability of an outage nor the cost of a missed event.

Reproduce the analysis

Read the 24 rows in date order. All five volume columns are monthly average daily rates in millions of standard cubic feet per day.

Use the four sector columns to compute 2023 sample standard deviations with denominator 11. Keep these scales fixed when measuring both reference and test distances.

For each reference month, exclude itself before finding its closest match. Use the largest of those twelve distances as the threshold. Compare each 2024 month with all twelve reference months.

Repeat with the national total as the sole feature. For the sensitivity test, omit one sector at a time and recalculate both the threshold and the 2024 scores.

Data table
MonthPowerAmmoniaMethanolLNGNational total
2023-0124746256912642665
2023-0223945456912652638
2023-0325142748812732537
2023-0425340948311902419
2023-0526441951712302529
2023-0627146453411182501
2023-0729341854810942461
2023-0827037752310582331
2023-0929043855911782527
2023-1026137457610822342
2023-1126035959211802488
2023-1224940859010742415
2024-0125641449611582431
2024-0225448055812842686
2024-0326545053611822540
2024-042664485579942364
2024-0529044350610872430
2024-062652944098481905
2024-072844455308002161
2024-0827447857211552586
2024-0927548054410152388
2024-1027244948412612566
2024-1127545248712202536
2024-1224743851211842474

References

  1. Ministry of Energy and Energy Industries, Consolidated Monthly Bulletins, January–December 2023, Table 3B, amended 28 November 2025
  2. Ministry of Energy and Energy Industries, Consolidated Monthly Bulletins, January–December 2024, Table 3B, amended 28 November 2025
  3. S. Ramaswamy, R. Rastogi and K. Shim (2000), Efficient algorithms for mining outliers from large data sets

Data credits · Ministry of Energy and Energy Industries, Table 3B, 2023 and 2024 bulletins, amended 28 November 2025. CAIA calculations.

Download the data and analysis code (ZIP)

Working paper · Internal review · How we check the research