Forecast or Pay: Why AI Load Forecasting Is Now a DISCOM Balance-Sheet Decision

Executive Summary

India met an all-time peak demand of 270.8 GW on 21 May 2026. Only a month earlier, the record had been 256.1 GW. The system held. But the evening told a different story: thermal plants ramped to nearly 185 GW in non-solar hours, and a 2.57 GW shortage was recorded in the evening period.

That gap between the solar afternoon and the dark evening is where DISCOMs now make or lose money. Every megawatt we under-forecast is bought late, in a tighter market. Every megawatt we over-forecast is paid for and never sold.

My view is simple. Load forecasting is no longer a spreadsheet task in the power purchase cell. It is a balance-sheet decision. And the only way to forecast a grid shaped by heatwaves, rooftop solar and prepaid smart meters is with AI trained on our own meter data.

The Evening Is the New Peak

For years, we planned around a single afternoon peak. That model is gone.

On 25 April 2026, when demand touched 256.1 GW at 15:38, solar supplied 56.2 GW, about 21.5% of generation. Solar carries the afternoon. It does not carry the evening. When the sun drops, demand stays high because cooling loads stay on. The load that solar was serving must shift back to coal, hydro, storage and the market within a short window.

This creates a steep ramp and a narrow, expensive evening block. A forecasting model that treats the day as one curve with one peak will miss it. A model that learns the shape of each feeder's evening, by temperature, by humidity, by day of week, will not.

In North Bihar, I see this at feeder level. Rural feeders with new pump and cooler loads behave nothing like urban commercial feeders. Averaging them into one district curve hides exactly the risk we need to price.

Why Forecast Error Is Now a Finance Problem

The market is getting more active and more expensive. In May 2026, the average Day-Ahead Market price on IEX was ₹4.88 per unit, up 18.3% year on year. Real-Time Market volume grew 15.9%.

Those are averages. The evening blocks on hot days are where the pain concentrates. When a DISCOM under-forecasts, it buys the gap in the real-time market or deviates from schedule. When it over-forecasts, it either sells surplus at a loss or carries fixed costs for capacity it did not need.

Here is how I frame the exposure for a CFO:

A one-percentage-point improvement in forecast accuracy is not a data science metric. It is a line item. When I present forecasting to leadership, I present it in rupees per day, not in MAPE.

Smart Meters Are the Training Data We Already Paid For

India's smart metering push under RDSS has given DISCOMs something they never had before: interval data at the consumer and distribution transformer level. With 6M+ endpoints in my own experience, I can say this plainly. Most of that data is used for billing and collection. Very little of it is used for forecasting.

That is a missed return on a very large investment.

A good AI forecasting stack for a DISCOM combines four layers of data:

1. Meter interval data from consumers and DT meters, cleaned for gaps and communication failures from the head-end system.
2. Weather data at a fine grid, especially temperature, humidity and cloud cover, because cooling load drives the evening.
3. Calendar signals such as festivals, harvest seasons, school holidays and exam weeks, which matter enormously in Bihar.
4. Behind-the-meter solar and prepaid behaviour, because rooftop solar and prepaid recharge patterns change net demand in ways a top-down model cannot see.

Gradient-boosted models and sequence models trained on these layers can produce forecasts at feeder level, then roll them up. Bottom-up forecasts expose where the risk sits. Top-down forecasts only tell you the total.

A Framework I Use: The Forecast Maturity Ladder

Most DISCOMs I work with sit on the first or second rung. The value sits on the fourth.

The jump from Level 3 to Level 4 matters most. A point forecast says "expect 1,200 MW at 20:00." A probabilistic forecast says "there is a 10% chance it exceeds 1,300 MW." That second number tells the CFO how much cover to buy. It turns forecasting into risk management.

The Next Load Wave Is Already Scheduled

Summer is not the only pressure. The Ministry of Power told Parliament that AI-driven data centres could add 26.3 GW of demand by 2031-32, nearly double the earlier estimate of 13.56 GW.

These are large, flat, always-on loads. They will be concentrated in specific states and specific substations. A DISCOM that cannot forecast its base load and evening ramp today will struggle when a hyperscale campus asks for a firm connection tomorrow.

There is an irony here I enjoy. AI will add load to the grid. AI is also the best tool we have to plan for it.

What This Means for Leaders

1. Put forecasting on the CFO's agenda. Report forecast error in rupees per day and tie it to power purchase cost. Make the scheduling desk and finance share one number.
2. Use the smart meter data you already own. Ask your AMISP for clean, time-aligned interval data feeds into an analytics layer, not just billing extracts. Write this into SLAs.
3. Forecast bottom-up, at feeder and DT level. Start with 20–30 high-risk feeders. Prove the gain there before scaling across the circle.
4. Move from point to probabilistic forecasts. Give procurement a range with confidence levels, so cover decisions are based on risk, not gut feel.
5. Build a forward-deployed team. Pair data scientists with SDOs and JEs who know why a feeder behaves the way it does. Models fail when they never meet the field.
6. Re-forecast intra-day. Link forecasts to real-time market bidding so the desk can correct before the evening block, not after.

Key Takeaways

• India's peak hit 270.8 GW in May 2026, and the evening, not the afternoon, is now the hardest hour.
• Forecast error lands directly on DISCOM finances through market purchases, deviation and stranded capacity.
• Smart meter interval data is the most under-used asset for forecasting in Indian DISCOMs.
• Feeder-level, probabilistic AI forecasts turn a technical task into a risk-management tool.
• Data centre growth will make accurate forecasting even more valuable over the next five years.

Randhir Verma is a senior enterprise leader in digital transformation and smart metering, working with DISCOMs, AMISPs and analytics teams to turn meter data into better decisions.

Sources: PIB – India meets all-time highest peak power demand of ~256 GW (28 Apr 2026) · Down To Earth – India meets all-time high demand of 270.8 GW (22 May 2026) · Business Standard – IEX trade volume rises 18.6%, DAM price up 18.3% in May (3 Jun 2026) · Energetica India – AI data centre boom set to add 26.3 GW load (28 Jul 2026)