Why Our Transformer Agent Remembers Every Oil Test (Built with Hindsight)

The power sector is changing fast. More people need electricity, industries keep growing, and renewable energy is being added to the grid every year. All of this puts new pressure on how we operate and maintain power systems.

At the heart of this infrastructure are transformers.

They step voltage up and down so electricity can travel long distances and reach homes, businesses, hospitals and factories safely. When a transformer fails, the impact is immediate: outages, costly repairs and days of disruption.

Keeping transformers healthy over many years needs regular monitoring and timely maintenance. Inspections, oil testing and engineering experience are still the foundation. But we wanted to see how AI could help engineers use the information they already have, better.

That's why we built Transformer Health Historian: a DGA trend agent with memory.

Why Smarter Transformer Maintenance Is Needed

A transformer can run reliably for years while its inner condition changes slowly.

Heat, electrical stress, insulation ageing and moisture all take a toll. The challenge is that these changes are not always obvious right away.

One test report gives you a snapshot. A developing problem often only becomes clear across several reports over time.

Now multiply that by hundreds of transformers. Engineers have to look through lab reports, maintenance documents, loading records and inspection notes. When those records sit in different places, understanding the full story of a single transformer can take hours.

The Problem: Every Report Is Judged Alone

Most monitoring compares each measurement against fixed reference limits. That's an important part of condition assessment, but it misses something.

Consider a transformer whose acetylene went from 0 to 4 ppm.

A threshold-based check may say the reading is still within limits. But an experienced engineer would immediately ask:

What did this transformer look like before?
Was there any maintenance work just before this test?
Has a similar transformer shown this pattern before it failed?

Our agent is built to ask exactly those questions, automatically.

How Hindsight Gives the Agent a Memory

The key piece of our project is memory. We used Hindsight, an open-source agent memory system by Vectorize.

It works in two simple steps:

1. Retain. Every event (a new DGA result, a repair, an overload, a failure, an engineer's verdict) is saved into a memory bank for the whole fleet, with its date.

2. Recall. When a new reading arrives, the agent first searches its memory for that transformer's history and lessons from other transformers. Then the AI reasons over what it found.

In our code, saving a memory looks like this:

def remember(self, content, context, timestamp=None): if timestamp: content = f"[{timestamp[:10]}] {content}" # keep the date visible to the AI  self.hs.retain(bank_id=self.bank_id, content=content, context=context)

If you want to understand the idea behind this, the Hindsight documentation and Vectorize's guide on what agent memory is explain it well.

What the Agent Does for Engineers

📁 Organised history: keeps DGA results, loading events and maintenance records for every transformer.

📈 Trend awareness: compares new readings with that unit's own past values, not just generic limits.

🛠 Maintenance context: links unusual readings to events that happened around them.

💬 Plain-language answers: engineers can simply ask questions and get clear summaries.

🔁 Pattern matching across the fleet: spots when one transformer is repeating what another did before it failed.

🧠 Learning from outcomes: when an engineer confirms a fault or a false alarm, that verdict is remembered for next time.

A Real Example: Same Reading, Two Answers

A new oil test comes in for transformer T-3, showing acetylene at 4 ppm.

Without memory, the agent gave a generic answer: medium risk, probably normal ageing, check again in a month.

With Hindsight memory, the engineer asks: "Is there anything unusual in T-3's latest DGA result?" The agent replies:

"T-3's acetylene increased from 0 to 4 ppm following last month's tap-changer maintenance. A similar pattern was recorded on T-1 before its documented failure in 2024. This is not confirmation of a fault. Review the DGA results and maintenance records, and consider arranging an inspection and follow-up test."

It also works the other way. When transformer T-5 showed a similar jump, the agent remembered that the same thing in 2025 was caused by a leaking gasket and confirmed as a false alarm, so it advised fixing the leak instead of an emergency shutdown.

The AI model was the same in both cases. The only difference was memory.

Who Can Benefit

⚡ Distribution companies (DISCOMs): one organised history for transformers across many substations.

🔧 Substation O&M contractors: faster access to test results, inspections and past repairs.

🏭 Large industries with their own substations: better maintenance planning for critical equipment.

For all of them, the value is the same: making existing information easier to find and use.

Humans Stay in Charge

AI should never make independent decisions about electrical equipment.

A change in gas levels can have several explanations, and engineers must weigh all the evidence before deciding on an inspection, extra testing or maintenance.

That's why our agent is designed as a decision-support system. It shows its evidence (every answer lists the exact memories it used), explains the change and flags possible concerns.

Final decisions always stay with qualified engineers. Reliability and safety come before automation.

What Was Challenging

Building this taught us a lot:

Data quality: incomplete or wrong records lead to wrong comparisons.
Dates matter: at first, the AI mixed up the order of events because it couldn't see when each one happened. Putting the date inside every memory fixed it.
Keeping assets separate: the agent must never confuse one transformer's records with another's.
Explainability: engineers need to see why something was flagged, so we always show the memories behind each answer.
Validation: before any real-world use, a tool like this must be tested against real cases and reviewed by domain experts.

What's Next

Automatic report reading: extracting values directly from lab DGA reports instead of typing them.
Trend charts: visualising how each gas changes over time.
Smart notifications: alerts when a trend or pattern needs attention.
Integration: connecting with maintenance management and substation monitoring systems.

Why We Chose This Project

As electrical engineering students, we wanted a problem that connects core power engineering with new AI technology.

Transformer maintenance is central to the power sector, and DGA is real engineering data with real consequences. At the same time, AI agents with memory are becoming one of the most exciting areas in software.

Instead of building a chatbot that answers general questions, we explored how an agent can use structured history to give engineers meaningful, evidence-backed answers.