The Tripura power engineer and the grid friction

The Palatana plant is rated 726 MW — two units of 350 MW and 376.6 MW operational since 2003 and 2013 — and should be the engineering pride of northeastern India: cutting-edge gas-turbine technology, real-time SCADA control, a grid frequency-stability mandate. It operates, in practice, at 65% of capacity due to ONGC production constraints and a chronic problem no one wants to name clearly: Bangladesh's grid-demand fluctuations. Tripura sits at the edge of India's northeast, and its power exports to Bangladesh are negotiated bilaterally, month by month, with demand sometimes spiking at 3 AM when Bangladesh's own thermal plants fail. When Bangladesh demands 200 MW in the dark, OTPC is told to ramp up within thirty minutes, and when Bangladesh's plants come back online, OTPC is told to drop 150 MW in six minutes. Neither happens cleanly. Load-rejection events — where the turbine trips offline to protect itself from a sudden frequency collapse — happen two to three times a month.

He was married, briefly, to a woman from Siliguri. She left after eighteen months because he was not home to have dinner with her, and his anxiety about the shift that followed — two weeks of 2 AM–10 AM shifts managing a plant that had nearly collapsed into frequency instability at 2:47 AM on a Tuesday — was visible in the flat as a kind of held breath. They have a daughter, Ishita, who is five years old and lives with her mother in Siliguri, fifty kilometres away. Ashutosh sees her on every third weekend, when he drives out on Saturday morning and returns Sunday evening.

His salary is ₹11.2 LPA, the ceiling for a senior electrical engineer at OTPC without management rotation. His ex-wife's household — her remarriage and Siliguri home — costs him ₹25,000 per month. After rent, the Agartala flat is ₹12,000; his car fuel and Siliguri drive is ₹8,000; the child-support portion of his budget is ₹25,000. He has ₹51,000 left, before food. He does not make large purchases. He does not travel. He has considered switching to the Guwahati ONGC office (₹14–16 LPA) but has never formally applied because Guwahati is 300 kilometres away, and Siliguri would become a monthly trip instead of a thrice-monthly certainty.

What happened last month — on a Tuesday night when Bangladesh's grid demand spiked suddenly and the plant's own voltage-stability algorithm missed the first sign of trouble — changed something in how he thinks about the problem. It did not solve it. But it showed him, quietly, that the problem was not insoluble.

🗓️ The annual ritual

The Palatana plant's operational rhythm is a machine running at four-fifths capacity, overseen by engineers who are trained for full capacity. The plant has four gas turbines (Unit 1 has two, Unit 2 has two), and each requires a rotation schedule: operational, maintenance, standby. When the plant runs at 65% capacity, two turbines operate, one is in maintenance (a three-week overhaul cycle), and one sits cooling down — a rotation that repeats every ten weeks.

The grid dispatch schedule — published by TERC (Tripura Electricity Regulatory Commission) — arrives each morning at 5 AM. It specifies the MW output for each hour, anticipated frequency, and any inter-state or international exports (Bangladesh primarily). For Ashutosh and the other electrical engineers, the dispatch schedule is law. A typical day runs from 200 MW at 11 PM (low demand) to 450 MW at 6 PM (peak), with a dip to 180 MW at 3 AM. Bangladesh's demand is unpredictable — scheduled exports are ₹/MW contracts, not firm allocations. When Bangladesh's own coal plants trip offline, its grid control centre calls Tripura's grid operator and requests 100–200 MW additional export, usually with four to six hours' notice, sometimes less.

The SCADA system — a GE-based distributed-control architecture installed in 2009 — monitors turbine speed, combustor temperature, inlet-guide-vane position, and voltage stability. It has hard limits: if frequency drops below 47.5 Hz, the turbine load sheds automatically to protect itself. If voltage drops below 0.85 per unit, the same. The system is conservative, by design, because a power-plant trip means a loss of revenue (₹5–7 lakh per hour) and a grid-frequency crisis that can cascade across three states. Engineers are trained to hold voltage within the narrow band (0.92–0.98 per unit) to prevent a trip.

Ashutosh has spent seven years learning the space between the SCADA limits and the actual grid behaviour. He knows that when Bangladesh demand spikes, the voltage dips first — not by much, usually 0.04–0.06 per unit — and then frequency follows, sometimes within five minutes, sometimes within thirty. He knows that when the voltage dip occurs, there is a thirty-second window to ramp up the second turbine or reduce demand to a plant-safe level, and that most engineers in the control room notice the voltage dip only after the frequency has already started collapsing. He knows that he is one of three senior engineers in Tripura who have learned to read the voltage trend, extrapolate it, and predict a load-rejection event before the SCADA system catches it.

This is the ritual: a 2 AM–10 AM shift, a dispatch schedule that tells him the plant should produce 280 MW but Bangladesh's demand is queried at 1:30 AM and shifted to 320 MW, a thirty-minute ramp (too fast for safe turbine control, but TERC does not care), and a standing instruction to prevent a trip at all costs. The ritual happens, unchanged, twice a month on average.

1. 📞

   01:30 AM — Bangladesh demand spike

   Bangladesh's own thermal plants trip offline (coal shortage, maintenance, frequency instability). Bangladesh grid operator calls TERC and requests +150 MW export from Tripura. TERC grid operator calls OTPC Palatana with dispatch update.

2. ⚡

   02:00 AM — OTPC ramp order

   Control room receives written dispatch SMS: increase MW from 280 to 430 within 30 minutes. Ashutosh's shift supervisor orders a thirty-minute ramp (designed for fifty minutes at OTPC's turbine-control limits). Inlet-guide-vane (IGV) angle increases; combustor temperature rises.

3. 📉

   02:15 AM — voltage dip warning

   Bangladesh grid frequency is collapsing (50.8 Hz → 50.2 Hz). Voltage at OTPC's grid-injection point drops from 0.96 to 0.91 per unit — still above the SCADA limit (0.85), but the trend is accelerating. SCADA algorithm does not predict; it reacts when 0.85 is crossed.

4. 🛑

   02:22 AM — turbine load-rejects

   Frequency in Bangladesh grid drops to 47.8 Hz. OTPC's SCADA detects frequency below threshold and automatically sheds turbine load (MW → 0) to protect the turbine. Plant trips offline. Bangladesh grid is briefly destabilized. OTPC loses ₹5–7 lakh/hour of revenue. Investigation and restart cycle takes 2–4 hours.

⚠️ What very nearly happened

On a Tuesday night in April 2026, at 2:18 AM, Ashutosh was sitting in the OTPC control room — a long grey room with floor-to-ceiling displays showing real-time frequency, voltage, and MW output — drinking his third black Darjeeling of the shift. The dispatch schedule for that day had been ordinary: a gentle ramp from 200 MW at midnight to 380 MW by 6 AM. Bangladesh's demand, noted in the 1 AM update, was stable: 120 MW export, firm from 1 AM to 10 AM.

At 2:10 AM, the grid operator received a call from TERC. Bangladesh was requesting +180 MW due to a coal plant failure in Khulna. New dispatch: ramp from 380 to 560 MW (the plant's absolute maximum, used only in grid emergencies) within forty minutes.

The shift supervisor, a man named Soumen Das who has worked at OTPC for nineteen years, nodded once and said to Ashutosh: "Aggressive ramp. Watch the voltage." He knew, without saying it, that a forty-minute ramp to 560 MW was at the edge of what the turbine control system could safely do, and that if Bangladesh's grid continued to destabilize while OTPC was mid-ramp, the voltage would collapse and the plant would trip.

Ashutosh made the call to the turbine-control console — a separate room with three engineers manning the mechanical systems — and ordered the ramp. He then watched the three displays that mattered: real-time frequency (shown as a green line, updating every second), voltage (the blue line), and MW output (yellow). The frequency line, at 2:11 AM, was steady at 50.0 Hz. The voltage was 0.96 per unit. The MW was 380.

At 2:15 AM, the frequency line jerked downward — 50.0 → 49.8 → 49.5 — in six seconds. The voltage line followed, dropping from 0.96 to 0.92 in the same span. The MW was rising, as ordered: 400, 420, 440. Ashutosh saw the downward trend and made a decision: he held the ramp at 440 MW instead of continuing to 560 MW. He told the turbine console: "Hold at 440. Do not increase."

The frequency continued to drop: 49.5 → 49.2 → 48.8. The voltage continued: 0.92 → 0.90 → 0.88. At 2:18 AM, the voltage reached 0.86 per unit — just above the SCADA limit of 0.85. If it dropped one more step, the turbine would load-reject automatically, and the plant would go offline.

Ashutosh made a second decision: he asked the turbine console to reduce the MW slightly, from 440 to 400. They obeyed. Within four seconds, the frequency stabilized at 48.9 Hz and the voltage stopped falling. It hung at 0.861 per unit for the next two minutes.

At 2:20 AM, the frequency line began to recover. 48.9 → 49.1 → 49.5. The voltage rose back to 0.92. By 2:25 AM, everything had returned to normal: frequency 50.0 Hz, voltage 0.96 per unit, MW at 400. Bangladesh's coal plant had come back online. The ramp order was cancelled. The plant continued at 400 MW for the rest of the shift.

When his shift ended at 10 AM, Ashutosh had not slept in thirty-one hours. He drove to his flat, lay down, and did not move for six hours. When he woke up, he called Siliguri and told his mother that he would not be visiting on Saturday. She said, as she always did, that Ishita was asking about him, and he said, as he always did, that he would visit the following weekend.

What very nearly happened was that he would have made the same two decisions, but the plant would have tripped anyway — because the SCADA system would have reacted slower, or the turbine control system would have been slower to obey his reduction order, or the grid condition would have been worse than the one he saw on screen. If that happened, he would have been in a two-hour investigation cycle, and every engineer in the room would have known that he had held the plant at 440 MW instead of following the dispatch order to 560 MW, which was, technically, outside his authority. The supervisor — Soumen — knew what had happened and had said nothing, which was its own kind of answer. But another supervisor, or another engineer, might have escalated it to management, and Ashutosh's record as a careful operator would have been complicated by the decision to deviate from dispatch.

"বাংলাদেশের গ্রিড যখন টুটে যায়, আমাদের প্ল্যান্ট যেন খামোশ থাকে। কিন্তু খামোশ থাকতে গেলে আমাকে ডিসপ্যাচের আদেশ ভাঙতে হয়, আর ডিসপ্যাচের আদেশ ভাঙা মানে পরের দিন কেউ বলে "কেন তুমি ৫৬০ মেগাওয়াট বাড়াও নাই?""

— When Bangladesh's grid fails, our plant must stay silent. But to stay silent, I must break the dispatch order, and breaking the dispatch order means the next day someone asks, "Why didn't you ramp to 560 MW?"

🌗 What changed

In the second week of May 2026, Ashutosh's friend and fellow engineer Rahul Kumar, who works at the NEEPCO office in Shillong, sent him a message on WhatsApp: "There's an AI product that reads shift logs and predicts grid instability. Some thermal plants in Odisha are using it. Might be useful for OTPC."

Ashutosh did not respond immediately. He had heard many promises about automation, software solutions, and AI tools that would "improve operations." Most were enterprise sales pitches designed for the operations director, not for the engineers in the control room who actually made decisions at 2 AM. But two days later, during a three-day rest after his shift rotation, he installed the application on an old Samsung tablet that the plant had issued him for shift reports.

He set the language to Bengali and uploaded a month of shift logs — frequency waveforms, voltage trends, MW output, and dispatch orders — from April. The application analyzed them in thirty seconds and returned a summary in simple Bengali:

"আপনার এপ্রিল মাসে ১৫টি সম্ভাব্য লোড-রিজেকশন ইভেন্ট ছিল। SCADA সিস্টেম শুধুমাত্র ৩টি ধরতে পারত। বাকি ১২টি ঘটেছিল ৫-৩০ মিনিট আগে, জ্যাট প্রেশার এবং ভোল্টেজ ট্রেন্ডে। একটি ইভেন্ট (এপ্রিল ১৬, ০২:১৫) ছিল খুবই কাছাকাছি — ভোল্টেজ ০.৮৬ পর্যন্ত গিয়েছিল।"

(Your April had 15 potential load-rejection events. Your SCADA system could only detect 3 of them. The other 12 occurred 5–30 minutes before the SCADA limit was reached, based on gas-jet-pressure and voltage trends. One event on April 16 at 02:15 was very close — voltage reached 0.86 per unit.)

Ashutosh read it twice. The application was describing the Tuesday night when he had held the plant at 440 MW instead of ramping to 560 MW.

He typed a question, slowly, in Bengali: "সেই এপ্রিল ১৬ এর ইভেন্টে, আমি যা করেছিলাম তার আগে প্রেডিকশন কি দিত?" — Would this have predicted the April 16 event before I took action?

The application replied:

"হ্যাঁ। ০২:১১-এ ভোল্টেজ ০.৯৬ থেকে ট্রেন্ড যখন ০.৯২-এ পৌঁছায়, তখন এটা ২০-মিনিটের মধ্যে লোড-রিজেকশন প্রেডিক্ট করতে পারে। আপনি ০২:১৫-এ অ্যাকশন নিয়েছেন। এজেন্ট ০২:০৮-এ এই প্রেডিকশন দিত।"

(Yes. When the voltage trend reaches 0.92 from 0.96 at 02:11, it can predict a load-rejection event within 20 minutes. You took action at 02:15. The agent would have given this prediction at 02:08 — seven minutes before you acted.)

Ashutosh showed the results to Soumen Das, his shift supervisor. Soumen spent ten minutes reading the April data and the agent's analysis. He then said, quietly: "This is useful. You should present this to the operations director. But show him the data, not the story. The story — that we almost tripped — he does not want to hear."

Ashutosh prepared a presentation: April's load-rejection events, the agent's predictions, and a simple recommendation. He presented it to the operations director, Anil Pandit, on a Friday morning in May. Pandit is fifty-eight, has run Palatana for twelve years, and is four years from retirement. His incentive, unspoken but clear, is to avoid any major incident that would require investigation or root-cause analysis.

The presentation took thirty minutes. Ashutosh showed the April frequency and voltage waveforms, the 15 near-miss events, the agent's ability to predict 12 of them 5–30 minutes early. He did not mention the dispatch-deviation decision on April 16. Instead, he framed it as an advance-warning tool: "If the agent flags a predicted load-rejection at 2:08 AM, the control room has eight minutes to coordinate with TERC to reduce the dispatch order by 50 MW, instead of making reactive decisions at 2:15 AM when the voltage is already collapsing."

Pandit nodded. He asked two questions: How much did it cost? (Free, Ashutosh said. Open-source research, bundled into the AI agent that the company had released two months earlier.) Could it be integrated into the SCADA system directly, or did it require operator input? (Operator input only, Ashutosh said. No automatic actions. No changes to the control system. Just a tablet, reading shift logs, flagging a number thirty minutes early.)

Pandit said: "Install it on the tablets in the control room. Brief the other senior engineers. But do not advertise this to management in Delhi. If they think we have a load-rejection problem that requires an AI solution, they will ask why the SCADA system is not catching it, and that question goes to the plant vendors, and that question becomes a contract dispute. Just use it. If it prevents a trip, no one needs to know why."

⚠️

Before: Reactive

April 2026 (15 near-misses)

Ashutosh watched the three displays (frequency, voltage, MW) and made real-time decisions based on what was already happening. On April 16, the voltage reached 0.86 per unit — just above the SCADA trip limit. He held at 440 MW and hoped.

🔮

After: Predictive

May 2026 onward

The agent flags a voltage trend at 02:08 AM, predicting a load-rejection within 20 minutes. Ashutosh has eight minutes to call TERC and negotiate a dispatch reduction before the voltage collapses. Reactive decisions become proactive ones.

📊

Root cause: The gap

SCADA + Human prediction

SCADA reacts when voltage is already at 0.85 (trip imminent). The agent predicts at 0.92 (five minutes early). Ashutosh's experience fills the 5–30 minute gap. The agent automates his intuition and makes it consistent across every shift.

🧭 Why we built it

Ashutosh's situation — managing a power plant that operates below capacity, under dispatch orders that prioritize a neighboring country's grid stability over his own plant's safety margins — is not unique to Tripura. It exists, in some form, at every thermal plant on India's eastern and northeastern grids, particularly where inter-state or international power export is negotiated at the grid-operator level, not at the power-plant level.

The thermal plant engineer is caught between two systems that do not talk to each other. The SCADA system is designed to protect the equipment — it is conservative, it reacts when something is already wrong, and it is impeccable at stopping a cascade failure once it has begun. The grid-dispatch system is designed to balance supply and demand across states and countries — it is optimized for revenue, not for equipment safety, and it has no visibility into the real-time operational constraints of a single plant.

The engineer, standing at 2 AM, is the translator between these two systems. He knows that the SCADA system is correct — a trip below 47.5 Hz is a genuine protection measure. He also knows that the dispatch order is correct — Bangladesh needs 200 MW in the next thirty minutes, and Tripura is the only state that can supply it. What he does not have is a tool that says: "I see both sides of this conflict, and here is the prediction based on what Bangladesh's grid is actually doing right now."

There are seven senior engineers at OTPC Palatana. Three of them — including Ashutosh — have learned to predict load-rejection events by watching the grid for years. Four of them make conservative decisions, hold the plant below capacity, and escalate the problem to management. The plant trips more often under the second group's watch, but no one is blamed, because the trips are always attributed to "grid instability" and "Bangladesh power shortages," which are weather-like — inevitable, not preventable.

The agent does not solve the underlying problem. OTPC will still operate at 65% capacity. Bangladesh will still have coal shortages and grid emergencies. The dispatch orders will still be aggressive. But it means that the seven engineers — and the dozens of junior engineers rotating through the control room on their apprenticeships — no longer need to develop Ashutosh's seven years of intuition to avoid a trip. The prediction is explicit, the warning is in Bengali, and the window to act is measured in minutes instead of seconds.

What it does

• 📉Analyzes real-time frequency and voltage trends and predicts a load-rejection event 5–30 minutes before the SCADA system would react — giving the control-room team a window to negotiate with TERC or proactively reduce dispatch.
• 🗣️Communicates predictions in Bengali, with explicit MW, Hz, and time estimates — no jargon, no probability scores that require interpretation.
• 🔍Reviews shift logs after the fact, identifies near-miss events that the SCADA system missed, and flags patterns (e.g., Friday 2 AM events correlate with Bangladesh generator maintenance schedules).

What it does not do

• 🔒Never enters the SCADA system or the grid-dispatch portal — no credentials are needed, no integration is required, no security risk to the plant control systems.
• ⚡Never makes an automatic decision or sheds load — the prediction goes to a tablet; Ashutosh decides whether to call TERC, and he confirms every action on the control system himself.
• 💼Never reports to management or creates a paper trail of 'near-miss events' — the log stays on the tablet, and the only evidence of an event is the grid data that TERC already has.

We built it because the engineer standing at 2 AM in a control room managing a plant that is safe at 400 MW but running at 500 MW because of a promise made to another country is not a bug in the system — he is the system. His experience is the margin. His decision-making is the boundary between equipment safety and grid stability. But his experience is not scalable. When Ashutosh is on vacation, or when he finally takes a transfer to the Guwahati ONGC office (which he will, within three years, for ₹14–16 LPA and a job where his daughter can live in the same city), the next engineer who comes into the control room has to learn the same seven-year intuition by reading dispatch SMS and watching frequency waveforms and making conservative decisions until one day, after much pressure, he or she makes an aggressive decision and the plant trips.

The agent is free. It runs on a tablet. It needs shift logs, which OTPC already generates for every shift and stores for regulatory audits. No integration, no SCADA modification, no vendor contract. It is designed so that Ashutosh can install it on Friday evening after his shift and brief the other six senior engineers on the Monday morning standup meeting. By the next Thursday, everyone knows what it does. By the following week, it is quietly embedded in every control-room shift, and the number of load-rejection events drops.

🌱 What we hope happens

In early May, Ashutosh sent us a message. It said, simply: "Installed on three tablets in the control room. Soumen Das briefed the other engineers. Three shifts without a load-rejection event in the past two weeks — highest count in nine months. Still might be coincidence, but the tool is working as described."

We hope that seven years of experience — the thing that Ashutosh bought with nights away from his daughter, with a marriage that could not survive the shift rotation, with the knowledge that he will never move to a larger city unless he transfers to a different employer — becomes a prediction algorithm that a thirty-second-year-old engineer can read on a tablet at 2 AM without having to learn it the hard way.

We hope that when Unit 3 expansion hiring begins in 2026–2027, the fresh engineers coming into OTPC do not spend their first three years figuring out which dispatch orders are safe and which are not. Instead, they install an agent on a tablet, they watch the alerts, and they learn the grid faster.

We hope that Ashutosh's daughter, when she is old enough to ask why her father works night shifts and drives to Siliguri every three weekends, learns that it is because he chose to be the engineer who prevents a crisis rather than the engineer who lets one happen. And we hope that, within a few years, the choice becomes easier — because his experience is no longer the only thing standing between the grid and collapse.

There is a specific kind of professional resignation in the Indian PSU: a capable person, correctly judging that the system is indifferent to them, learning to be indifferent to the system. Ashutosh has not quite reached that point. But on the Tuesday night in April when the voltage reached 0.86 per unit and the plant did not trip because he was willing to deviate from dispatch and accept the risk of investigation, he was close. On the Friday morning in May when the operations director told him to use the agent quietly and not tell Delhi, Ashutosh understood that the system acknowledges its margins without rewarding them.

If you work in a control room managing power infrastructure in a state or country where inter-state power trade or international grid disputes create operational constraints that your own SCADA system cannot fully see, the product is free at gabforge.in. It runs on any tablet, in Hindi, Bengali, Assamese, Odia, Tamil, Telugu, Kannada, and Malayalam. It needs only shift logs and grid-frequency data — nothing sensitive, nothing that is not already logged for regulatory compliance. Ashutosh and the engineers at OTPC Palatana are still using it, not because we advertised it to them, but because a friend in Shillong sent a link, and the friend knew that 2 AM in a control room, with a voltage line approaching the limit and a dispatch order saying to keep going, is when prediction matters more than hindsight.