SETU is a forecasting and decision support system for geomagnetically induced currents in the North East Indian power grid. The name stands for Solar Event To Utility, and setu also means bridge, which is what the system does. It carries a measurement made in space all the way through to a switching decision made in a control room.
This repository is a working artefact, so every layer described below is implemented and runs. Nothing here is a mock.
SETU was runner up at the East and North East regional final of the IET India Scholarship Award, run by the Institution of Engineering and Technology. The final was held at the University of Calcutta on 12 September 2026.
A geomagnetic storm does not damage a power grid the way lightning does. The storm changes the magnetic field at the surface of the Earth, that changing field induces an electric field in the crust, and the electric field drives a slow quasi-DC current up the neutrals of grounded transformers. A transformer carrying that current saturates on one half of the cycle, which makes it draw large amounts of reactive power, inject harmonics into the network, and heat its own windings. The grid then fails through voltage collapse and protection misoperation. Hydro-Quebec lost its whole system in ninety two seconds this way in March 1989.
The usual assumption is that India sits at low geomagnetic latitude and is therefore safe. For the North East Region that assumption does not hold, for reasons that have nothing to do with latitude.
- Conductivity contrast. The Shillong Plateau is resistive Precambrian gneiss. It sits directly against the deep conductive sediments of the Brahmaputra valley and the Bengal basin. Induced electric fields concentrate at boundaries like this, so the same magnetic disturbance produces a much larger driving voltage here than it would over uniform ground.
- Network topology. The regional grid is long and radial, and it connects to the rest of India through the narrow Siliguri corridor. Induced voltage on a line grows with its length, and a radial network has little margin to redistribute reactive power when a transformer starts absorbing it.
- The HVDC terminal. The Biswanath Chariali end of the 800 kV link uses single phase converter transformers, which are far more vulnerable to saturation than the three phase three limb units used elsewhere, and it has a ground electrode that couples directly into the earth return path.
- Restoration time. Terrain, seismic risk, and monsoon access make the North East the slowest region in the country to repair. Resilience is about recovery as much as it is about resistance, so the same event costs more here.
The system is five layers, and each one comes from a different discipline.
| Layer | Question it answers | Method |
|---|---|---|
| Sensing | What is the Sun doing right now | Solar wind at the L1 point, live or from the archive, plus ground magnetometers |
| Forecast | How hard will the ground field shake in 30 to 90 minutes | Physics informed dilated convolution network with calibrated quantile heads |
| Induction | What electric field appears in the crust | Plane wave response of a layered Earth model |
| Grid | Which transformers carry how much current | Lehtinen and Pirjola nodal admittance solver, then AC power flow |
| Decision | What should the operator do about it | Scenario based optimisation under a conditional value at risk constraint |
Full tables are in RESULTS.md, which is written by a script from the saved training reports so the numbers cannot drift away from the runs that produced them. The short version follows.
Three storms are held out completely: the Gannon storm of May 2024, the Halloween storm of October 2003, and the October 2024 storm. They never touched the model, the scaler, the alarm threshold, or the calibration map.
| Horizon | Probability of detection | False alarm ratio | Heidke skill score | Brier skill score |
|---|---|---|---|---|
| 30 minutes | 0.83 | 0.37 | 0.67 | 0.34 |
| 45 minutes | 0.78 | 0.38 | 0.63 | 0.28 |
| 60 minutes | 0.78 | 0.43 | 0.60 | 0.28 |
| 90 minutes | 0.71 | 0.44 | 0.57 | 0.28 |
The model has 19,000 parameters, reaches back 315 minutes, and trains in 24 minutes on a laptop with no graphics card.
This is the central design decision and it is worth stating on its own.
The OMNI archive publishes the solar wind against the time it reaches the nose of the bow shock rather than the time it was measured. That is the right convention for studying what the magnetosphere did, and it is exactly the wrong one for building a warning system, because under it a shock front appears in the record at the same instant it strikes the Earth. The first version of this project trained on that convention and had no warning time to forecast into at all.
The archive records the shift it applied, so it can be undone. Each measurement is moved back to the moment the spacecraft actually saw it, using the travel time implied by spacecraft distance and measured speed, which needs nothing a real time system would not already have. Training the same model both ways on the same storms gives the comparison in RESULTS.md. The spacecraft clock wins at every horizon, and the gain in probabilistic skill is much larger than the gain in detection, which is what you would expect from a model that now has real information about the future rather than a better fit to the present.
There is a hard ceiling here that no model can raise, and it is the most useful thing in this project to be able to explain.
A disturbance can only be forecast at a horizon shorter than its own travel time from the spacecraft to the Earth. At any longer horizon it had not yet reached the spacecraft when the forecast had to be issued. During the fast solar wind of 10 May 2024 that travel time was about thirty minutes, so no forecast at forty five minutes or more could have caught that shock, however good the model was. The delay runs from roughly twenty five minutes to eighty depending on speed, which means the available warning is shortest exactly when the storm is fastest.
The system handles this by forecasting at four horizons that straddle the delay, and by giving the network the propagation delay as an input so it can tell which situation it is in.
Through the body of the May 2024 storm, 87 percent of disturbed decision steps carried a standing alarm. The system did not have an alarm standing before the very first disturbed minute, and the reason is the limit above rather than a defect in the model. The shock reached the spacecraft eleven minutes after the forecast for that minute had to be issued. The alarm probability rises from 0.04 to 0.20 within one decision step of the shock becoming visible, which is as fast as the physics allows.
Both numbers are reported side by side, because the first alone flatters the system and the second alone condemns it.
A stated ninetieth percentile should sit above the observation ninety percent of the time, since the decision layer treats these as real probabilities. The predicted quantiles are corrected by a map fitted on part of the validation set and judged on the rest, following Kuleshov, Fenner, and Ermon (2018). Fitting and judging on the same data made the map look almost perfect and it was not, so the check is now on storms the map has never seen.
Calibration remains the weakest part of the system. It improves the whole distribution and it does not fully transfer from the validation storms to the test storms, which is stated in RESULTS.md with the numbers rather than smoothed over.
- The Shillong Plateau produces an induced electric field 7.3 times larger than the Bengal and Tripura basin does, for the same magnetic disturbance at a five minute period.
- On the standard one volt per kilometre benchmark the network is more than three times more exposed to an eastward electric field than to a northward one. The Siliguri corridor and the 361 km direct current route to Biswanath Chariali both run east to west, and the storm time current systems that reach Indian latitudes produce exactly an eastward field. The alignment between the driver and the network is a consequence of the geometry rather than an assumption.
- Earth currents sum to zero across the network to one part in ten to the fourteen, which is the Kirchhoff check on the solver.
Against a hypothetical extreme event, five actions costing 486 lakh rupees remove 4,551 lakh rupees of tail risk, a return of about nine for each rupee. The first action chosen is moving the direct current link off earth return.
Eight neutral blocking devices cut the regional reactive absorption by 75 percent. The third site the search picks is Binaguri, which does not appear anywhere in the top eight of a ranking that scores each site on its own, because it only starts to matter once the two sites above it are blocked.
At the disturbance levels actually observed during the May 2024 storm, the model recommends no action at all. That is the correct answer and it is reported as such.
The archive is what the model learns from and it appears months after the fact. The operational counterpart reads the NOAA real time solar wind feed, which needs no account, and returns the same columns, so the same features and the same trained model run on either one.
python -m setu.cli liveThis prints the current solar wind, how long it will take to arrive, the probabilistic forecast at all four horizons, the resulting current and voltage consequence across the network, and the recommended action. It also checks this project's propagation delay against the operational feed's own propagated product, which agrees to within about half a minute.
Every other result in this repository comes from a storm that had already happened. That is the right way to develop a model and it is not the same as proving one, because the storms were chosen after the fact and the model was tuned while they were on the table. Holding three of them out is real discipline and it still leaves that objection standing.
So the system also runs as a service.
python -m setu.cli nowcastA scheduled job runs that command. The job is scheduled every quarter of an hour, but GitHub starts free scheduled jobs late when its runners are busy, so in practice it runs every three to four hours. Each run issues a forecast for all four horizons from the solar wind as it stands at that minute and writes it into docs/data/ledger.json before anything about the outcome is known. About an hour later, once the valid minute has passed, it fetches what the Indian magnetometer actually recorded at that minute and writes the answer beside the forecast. The persistence baseline is recorded in the same row at the same moment and from the same information, so the comparison between the two cannot drift.
Nobody starts it. The job is in .github/workflows/nowcast.yml and it commits its own results, so the operator console shows a record that grows on its own. That is also why most of the commits in this repository are made by setu-nowcast and say nowcast: in front of them. Filter them out with git log --perl-regexp --author='^(?!setu-nowcast)' to see the engineering history on its own.
Two properties are what make the record worth anything.
A row is written before its outcome exists, so it cannot be selected afterwards. Rows that were replayed through the hours the feed still held when the service first started are marked as backfilled and counted separately, because a forecast made from a record that already exists is a weaker claim than one written down in advance.
Rows older than two days move into docs/data/verification.json, which is a daily count that is added to and never rewritten. Once a day is closed its numbers are final, so a bad week cannot be quietly dropped later.
Whether the record shows skill depends on whether the Sun does anything. Through a quiet stretch there is no detection skill to report and the useful number is the false alarm count, because an alarm system that cries wolf through a quiet week is switched off before the storm arrives.
On 12 September 2026 at 08:57 UTC the record held 61 live forecasts issued since 4 September, scored in 188 checks against the Indian magnetometer, with no false alarms. The ground never reached the 0.1 nT per second alert level in that time, so there is no detection skill to report yet.
Until 12 September the workflow committed only the ledger. The daily record was written on each run and thrown away with the runner, so the closed days were kept only in the history of the ledger. The workflow now commits both files, and scripts/rebuild_rollup.py recovered the closed days from that history through the same fold the service uses.
Every figure in docs/figures is generated by scripts/make_figures.py from the real data and the real model, so none of them can drift away from the result it shows.
Four gaps in this project can only be closed by people outside it. A request for each one was sent on 5 September 2026, to an address read off the holding institution's own website rather than to a guess. The letters themselves are held privately rather than in this repository. One has been answered, and the answer is recorded in the fourth item below.
-
Ground magnetometer data for Shillong and Tirunelveli, held by the Indian Institute of Geomagnetism. This is the largest gap. The model trains on Alibag and Hyderabad, near ten degrees geomagnetic latitude, and is applied to a region near sixteen.
-
Magnetotelluric resistivity profiles across the Shillong Plateau and the Brahmaputra valley, to replace the four ground models assembled from published literature. The geoelectric field scales with the square root of ground resistivity, so this assumption drives the answer more than any parameter in the power network does. The survey that covers this ground was done by the Indian Institute of Geomagnetism rather than by the institute usually assumed, so this request and the one above go in the same letter.
-
Whether any transformer neutral current monitoring exists anywhere in the region. There is no public measurement of neutral current in India, which is why every consequence figure here rests on physics and on measurements made in other countries.
-
The latency of the Aditya-L1 solar wind products. India has its own monitor at the first Lagrange point and a warning system for the Indian grid should be reading it. The data itself is already public, so the question worth asking was how long it takes to arrive, because the whole warning time is the travel time from the spacecraft and any delay comes straight out of it.
This one has been answered. The Aditya-L1 Support Cell replied on 6 September 2026 that it takes at least 48 hours for Aditya-L1 data to become available. The travel time this system forecasts into is 25 to 80 minutes, so a 48 hour archive cannot feed a warning system at all. ASPEX and MAG are science instruments and their archive is built for science, where 48 hours is quick. A low latency operational product is separate engineering and it does not exist yet. Until it does, an Indian warning system for the Indian grid has to read a foreign spacecraft, and that is the finding rather than a complaint.
pip install -r requirements.txt
python -m setu.cli benchmark # physics checks and the network model
python -m setu.cli live # the whole chain on the solar wind right now
python -m setu.cli replay --event 2024-05-10 # walk a held out storm minute by minute
python -m setu.cli placement --budget 8 # where to put blocking devices
python -m setu.cli train --time-base l1 --tag l1The trained model is in the repository, so live and replay work straight after cloning. Data is downloaded on first use and cached under data/raw, so the first run of a command that needs history is slow and later ones are not.
The only dependencies are numpy, scipy, pandas, matplotlib, and requests. The neural network and the power flow are written out by hand rather than pulled from a library, so this installs and runs anywhere.
In scope are the 400 kV and 220 kV transmission levels of the North East Region, quasi-DC geomagnetically induced current effects, the chain from L1 solar wind to transformer consequence, and operator decision support.
Out of scope are distribution networks below 220 kV, three dimensional magnetotelluric inversion, pipeline and railway induction effects, direct control actuation, and security of the data path. These are named as future work rather than treated as solved.
There is no public dataset of measured transformer neutral currents in India, so the consequence model is validated against physics and against published international measurements rather than against Indian ground truth.
The network model is synthetic, because real network parameters are restricted. It is built to be topologically and electrically representative rather than exact, so a conclusion about which parts of the network are exposed will hold while an absolute current in ampere should be read as an estimate.
The forecast model is trained on Alibag and Hyderabad, which sit near ten degrees geomagnetic latitude, and applied to a region near sixteen. Shillong and Tirunelveli are held by the Indian Institute of Geomagnetism, and closing that gap needs a data request to the institute.
The Halloween storm of 2003 loses most of its record on the spacecraft clock, because the particle instruments at the first Lagrange point were saturated by the event itself and the archive carries no usable speed for long stretches of it. That is a real property of that storm rather than a modelling choice, and it is why the October 2024 storm was added as a third test event.
There is no real time Aditya-L1 solar wind product. The Aditya-L1 Support Cell confirmed on 6 September 2026 that the data takes at least 48 hours to become available, which is roughly a thousand times the warning window this system works in. The chain is therefore designed against that interface and demonstrated on the operational NOAA feed and the OMNI archive, and it is ready for an Indian low latency feed on the day one exists.
Development is paused at this stage, as of 12 September 2026. The model, the physics, the grid solver and the decision layer are final as described above, and the numbers in RESULTS.md are the ones the committed weights produce. The nowcast service keeps running on its own, so the standing record will keep growing and will show the first storm that reaches it, whether the forecast goes well or badly.
The next steps are the four requests listed above. The largest is the Shillong magnetometer record, which would let the model be trained at the latitude where it is used. After that comes an ingestion path for a low latency Aditya-L1 feed once one exists, and a check of the network model by somebody who holds the real parameters.
MIT. See LICENSE.