Most electricity distribution companies (DISCOMs) in India lose 15–20% of the power they buy before it reaches homes and businesses. Andhra Pradesh’s Eastern Power Distribution Company Limited (APEPDCL) has already performed much better, reducing losses to 5–6%, comparable to the technical losses of leading global utilities.
Now it faces a tougher challenge: finding and reducing the last remaining losses, which are much harder to detect.
To address this, APEPDCL has approved a pilot with Stanford-affiliated AI startup, Pravāh, to pinpoint where power is being lost in its network. The project is being funded by the Ministry of Power under its Powerton program.
Prudhvitej Immadi, the Chairman and Managing Director of the DISCOM says, “The project enables a shift from reactive to predictive DISCOM operations through AI-driven grid intelligence and load flow analysis.”
Why the last 5% is the hardest
APEPDCL serves around 8 million consumers across coastal Andhra Pradesh, including the industrial hub of Visakhapatnam. By Indian standards, it is a strong-performing utility, with a reliable power supply, growing digital systems, and losses that compare well even internationally.
However, even a 5% loss means a large amount of electricity is wasted every year, costing hundreds of crores of rupees. More importantly, these losses are not evenly distributed. Some areas show losses of 10–12%, while others operate at just 3–4%, and it is difficult to clearly explain why.
Although the utility has invested in GIS mapping, smart meters, and feeder monitoring, these systems operate in silos. They do not yet provide a unified, physical picture of what is happening across wires, transformers, and feeders.
What this means for Andhra residents and solar growth
Reducing electricity losses can directly lower costs for consumers, since wasted power is ultimately built into tariffs. Even a 1% reduction could save hundreds of crores each year. Areas with higher losses also tend to experience poorer power quality, including low voltage during peak hours.
“When we identify high losses due to undersized conductors or overloaded transformers, fixing that directly improves the quality of power reaching homes and businesses,” explained Dhruv Suri, CTO of Pravāh and recent Stanford PhD.
The pilot also supports Andhra Pradesh’s rapid expansion of rooftop and agricultural solar, with over 1,000 MW approved under PM-KUSUM. As solar feeds power back into networks designed for one-way flow, voltage fluctuations and inverter tripping can occur. The physics-based model helps APEPDCL identify which feeders need upgrades before new solar is connected.
According to Prudhvitej Immadi, the DISCOM’s Chairman, “This project will support higher renewable energy integration without proportional infrastructure expansion.”
How AI sees what humans can’t
Pravāh builds a physics-based digital model of the electricity network that simulates how power flows through every wire. The model accounts for factors such as resistance, heat, voltage drop, and loading.
When the model’s predictions match real meter data, engineers know the network is well understood. When they do not, it signals that something is wrong and needs investigation.
“We can tell APEPDCL that losses on a specific feeder spike between 2 PM and 6 PM because of phase imbalance, or that a transformer is bleeding energy because it’s oversized for its load,” explained Mohak Mangal, Pravāh’s CEO.
What success looks like
The three-month pilot will focus on substations where losses remain unusually high. If successful, APEPDCL plans to expand the approach across its entire network. The Ministry of Power is also watching closely, with the goal of applying similar methods to other high-performing utilities.
The pilot begins in the coming weeks, with results expected by May 2026. In an industry where 5% losses are already considered excellent, APEPDCL is aiming to set a new benchmark, using AI to deliver real financial and service improvements for millions of consumers.
Disclaimer
Views expressed above are the author’s own.
END OF ARTICLE
