A complete study of seven years of lightning data on the 150 kV Situbondo–Banyuwangi transmission line, and a five-year forward forecast for every one of its 281 towers — written for engineers and decision-makers, not data scientists.
Annual unique strikes ranged from 402 (2025) to 1,718 (2022) — a four-times swing. Using a single long-term average to plan arresters, insulation and maintenance systematically misallocates risk.
The same group of towers (roughly #142–#159, in the Mt. Ijen foothills at 210–245 m elevation) leads every year. Geography wins over weather when you compare towers to each other.
Year-to-year total strike count correlates with ENSO and IOD (the ocean–atmosphere cycles that drive Indonesia's monsoon). El Niño years are drier and quieter; La Niña years are wetter and noisier.
With only seven years of data, our conservative Model C beats a plain average by about 2 %. A new experimental Model D adds recency-weighted learning and forecast-informed climate; section 06 compares the two side-by-side. Use the forecast as a sensitivity guide, not a precise prediction.
Lightning is the single biggest cause of unplanned outages on tropical 150 kV transmission lines. A flash that hits — or induces voltage near — a tower can flash over an insulator and trip the line. On the Situbondo–Banyuwangi corridor we see this most years, in clusters near the mountainous middle section.
The traditional engineering approach uses a single Ground Flash Density (GFD) number from a regional map. This works for new-build design, but it cannot tell you:
This study answers those three questions using the lightning records you have already paid for, plus public climate data.
Figure 1. Annual unique strikes on the line (left axis, bars) and the El Niño/IOD climate index (right axis, lines). When the climate index is negative (La Niña), lightning activity tends to be higher.
Two source files were provided by PLN: a Vaisala-format lightning exposure workbook (2019–2025) and a tower-coordinate sheet. The raw data was messy — 29 sheets, mixed schemas, duplicated columns. We first converted everything into seven clean tidy CSVs so every later step is reproducible.
Every one of the 281 towers, every year from 2019 to 2025, with: total strikes, positive/negative split, mean and peak current (kA), and the local density (flashes per km²).
Strike totals month-by-month for seasonality.
Latitude, longitude, elevation, distance to coast.
Monthly Niño 3.4 (NOAA) and IOD DMI indices, going back to the 1950s — used to test whether wet/dry cycles in the Indian Ocean and Pacific drive year-to-year lightning swings.
Figure 2. Average lightning activity by month across all seven years. The Indonesian wet season (October–April) is when most strikes happen.
We ship two models side-by-side. Model C is the conservative benchmark — equal weight on all seven years, fixed climate scenarios. Model D is an experimental extension that weights recent years more heavily and consumes operational climate outlooks when those exist. Section 06 compares them honestly.
We fit a simple statistical model to the seven historical line totals. Because seven points is very little, we let the model decide (via AICc) whether to use the climate indices (Niño 3.4, IOD DMI) or fall back to a plain average — and we measure how well it would have predicted each historical year if that year had been hidden.
Model D extends this stage by giving
recent years more weight (half-life tuned by cross-validation),
and ingests NOAA / IRI / BoM / BMKG forecast probabilities for
2026 where available. Beyond the operational forecast horizon
(~9 overlapping 3-month seasons), Model D reverts to
documented persistence-decay toward the scenario prior. In
v1.2.1 the count formula is locked to
Niño 3.4 + DMI — AICc still computes the
information criterion as a diagnostic but no longer selects
the candidate; this forces climate to be a functional model
parameter rather than something rejected at n = 7. See caveat F.
Each tower has a stable "share" of the line's total strikes — the high-elevation foothill towers consistently take more than coastal ones. We compute each tower's seven-year average share (with statistical shrinkage to avoid over-fitting one freak storm) and multiply it by Stage 1's annual forecast.
Model D additionally applies a low-degree ridge correction using tower order along the line, elevation, and distance to the Bali Strait — features that the share calculation alone doesn't see.
Deep-learning approaches like LSTMs or graph neural networks were explicitly considered and ruled out. They need hundreds to thousands of training examples — we had seven. The "deep-learning-like" recency behaviour in Model D is implemented transparently as exponential sample weighting, not as a neural network. We chose statistical honesty over technical novelty.
Operational ENSO forecasts cover only the next ~9 overlapping 3-month seasons (about 2¼ years). So we show the line under three explicit climate assumptions (La Niña, Neutral, El Niño) and provide an equal-weighted "marginalised" best-guess. Pick a model below — Model C uses fixed scenario priors; Model D blends in operational outlooks where available and falls back to priors beyond the horizon.
Figure 3. GFD profile along the line — tower by tower (x-axis) for every year. Solid lines are the seven historical years (2019–2025); the peak year 2022 is highlighted in amber. Dashed lines show the median forecast for each year 2026–2030 under the selected climate scenario. Click any year in the legend to show or hide it. Switch the tabs above to compare La Niña, Neutral and El Niño forecasts.
Model D trades cross-validation skill for climate responsiveness. On the seven-year sample, AICc rejects the Niño 3.4 and IOD DMI coefficients as overfit; v1.2.1 overrides that and forces them in. The cost is a ~20 % higher leave-one-year-out RMSE than Model C. The benefit is a ~39 % La Niña-to-El Niño scenario spread that Model C cannot produce. The two models answer different operational questions — we publish both, side by side, so practitioners can pick their assumption transparently.
| Aspect | Model C — benchmark | Model D — experimental |
|---|---|---|
| Climate learning | Limited historical climate relationship | Explicit Niño 3.4 / DMI variables where skillful |
| Future climate input | Fixed La Niña / Neutral / El Niño priors | NOAA / IRI / BoM / BMKG probabilities where available |
| Historical weighting | Equal weight on all 7 years | Recency-weighted, half-life tuned by CV |
| Spatial allocation | Shrunken historical tower share | Shrunken + ridge correction (tower order, elevation, distance-to-coast) |
| Best use | Conservative benchmark, stable hotspot ranking | Rolling planning when forecast outlooks are fresh |
| Risk | May miss recent regime shifts | Can overreact to noisy recent years or uncertain forecasts |
Mean error across all 7 leave-one-year-out folds. Lower is better. A = climatology, C = benchmark, D = forecast-informed.
Figure 4. Mean per-tower GFD across the 281 towers for each forecast year, under the Neutral scenario. Model C and Model D are plotted side by side with their 80% prediction intervals. A gap between the two reflects the recency-weighting effect.
Figure 5. Top-20 towers ranked by predicted 5-year mean GFD under each model. A high Jaccard overlap means both models flag the same hotspots; a low one means the spatial picture changes meaningfully between models. Italicised tower labels marked with* are in only one model's top-20 — both bars are still drawn (showing the value the other model assigned that tower) so the asymmetry is visible by magnitude rather than by missing data.
Figure 6. Per-tower difference (Model D − Model C) in 5-year mean GFD, plotted against elevation. Positive values mean Model D forecasts more lightning than Model C for that tower; negative means less.
The interactive map below shows every tower coloured by its predicted five-year mean GFD under the Neutral scenario. Click a marker for the tower's full forecast and prediction interval. Toggle the scenarios in the layer control, or switch the model below to compare hotspots under Model C vs Model D.
Vertical line is the median forecast; horizontal whiskers show the 80 % prediction interval. The model toggle above swaps between Model C and Model D rankings.
Engineering reports often hide their uncertainty in footnotes. This one puts it in a dedicated section, because using the numbers with the wrong assumptions costs more than not using them at all.
Cross-validation shows Model C lowers prediction error by only ~2 % compared to a plain seven-year average. Model D's CV numbers are reported in section 06 and may improve when more years accrue. With seven years of data, a fancier model would over-fit. Treat both models' climate scenarios as a sensitivity range, not a precise prediction.
When we trained Model C on 2019–2024 and predicted 2025, the tower-by-tower agreement was modest (correlation ≈ 0.24). Per-tower allocation is noisier than we would like — use the rank order of hotspots more than the absolute numbers. Model D's historical CV uses observed climate as a "perfect-forecast" proxy; its real prospective skill will be lower and remains unknown until 2026 is observed.
The per-tower density values in your Vaisala FALLS export are consistently ~5× the line-aggregate density. This is likely because each lightning strike falls inside ~5 adjacent towers' 3.125 km² collection circles. Before comparing our numbers to published GFD climatologies, please ask Vaisala support to confirm whether per-tower counts are overlap-counted (we think so) or unique-attributed.
The forecast assumes the relationship between climate and lightning that held in 2019–2025 will continue through 2030. If a structural change occurs (e.g. sustained warming of the western Pacific), reality may deviate. Re-run this study annually with the newest year of data.
NOAA / IRI / BoM publish ENSO probability forecasts for
~9 overlapping 3-month seasons (~2.25 years).
Beyond that horizon Model D reverts to documented
persistence-decay toward the scenario prior; the per-year
climate input source is stamped in
outputs/forecast_climate_stamps.csv so consumers
can audit which years are forecast-informed vs prior-driven.
Model D must be re-run whenever a fresh outlook is
published — its 2026 numbers should not be treated as
stable for the full 5-year horizon.
Model D's count, mean kA, and max kA formulas
are locked to target ~ Niño 3.4 (JJA) + DMI (JJA)
in v1.2.1. AICc on the v1.2 sample rejected these climate
variables at n_eff ≈ 5.8 and Model D collapsed
to a recency-weighted intercept; v1.2.1 overrides that so
climate becomes a functional model parameter rather than a
candidate that gets silently dropped. The resulting fit is
climatologically sensible: under La Niña
(Niño 3.4 = −1), 2026 mean GFD across the line is
6.16 flashes/km²/yr; under Neutral,
5.47; under El Niño (+1),
4.14. The direction (La Niña > Neutral >
El Niño) matches Indonesia's known wet/dry sensitivity
to ENSO. The magnitude is at the upper limit of what 7
training years can statistically support — coefficient
uncertainty is wide, and Model D may be capturing
sample-specific patterns rather than a stable climate
response. Treat Model D's scenario spread as a
sensitivity dial, not a precise forecast.
Both Model C and Model D place 16 of the same 20 towers (Jaccard 0.67) in their top-20 hotspot lists, and all 16 overlap towers fall in this range (Mt. Ijen foothills, 210–245 m elevation). The cluster is robust across modelling choices and has a stable seven-year history — capex here has the strongest expected return per rupiah.
Model D's forced climate fit puts the 2026 La Niña mean GFD at 6.16 flashes/km²/yr versus 4.14 under El Niño — a 39 % swing along the line. BMKG and BoM publish ENSO outlooks quarterly; treat them as operational signals: in La Niña-leaning years, advance pre-monsoon inspections of insulators and earthing on the foothill section by 1–2 months. In El Niño years, that budget can be redirected to other lines or to capital projects on this one.
Model C is the conservative benchmark — flat weighting across all 7 years, climate as a sensitivity dial. Model D is the forecast-informed extension — climate is a functional model parameter, scenarios are climatologically calibrated. The Hotspots map (section 07) lets you toggle between them. If C and D disagree on a borderline tower, treat it as a tower whose risk depends on which assumption you trust more — and budget accordingly.
The ~5× ratio between per-tower and line-aggregate density is the single biggest interpretation uncertainty in this report — unchanged from v1.0.0. A 30-minute conversation with Vaisala support resolves it permanently for both models.
The entire pipeline is automated (tidy_data.py → lightning_gfd_forecast.ipynb → run_model_d.py → generate_site_data.py). As 2026 lightning data lands, the new run takes minutes. Model D in particular benefits from each new year — its fit weights recent observations more, so a single fresh year measurably shifts the climate coefficients. Re-running closes the feedback loop and lets you compare 2026 actuals against this forecast.
Three sheets per scenario (GFD · Count · kA) plus README with units and caveats.
Download .xlsx ↓Self-contained interactive Leaflet map with three scenario layers and tower popups.
Open .html ↗Highest-GFD towers with 80 % prediction interval, lat/lng, elevation.
Download .csv ↓Per-tower change in GFD per unit Niño 3.4 (ENSO elasticity, year 2028).
Download .csv ↓Cross-validation tables: MAE, RMSE, deviance, CRPS — every metric, every model, every fold.
Open .html ↗The complete reproducible analysis — every cell, every parameter, every figure.
Download .ipynb ↓Source code, tidy data, and this report are open at github.com/fakhrihakim20/stb_bwi_lightning. The analysis can be re-run in under five minutes once Python and the listed packages are installed.