Probabilistic wind and solar power forecasting for bidding into aFRR

The shift to shorter aFRR capacity blocks is underway across Europe, with markets moving from 24-hour blocks to 4-hour, 1-hour, even 15-minute blocks. For a renewable power trader, committing capacity in a 4-hour block instead of a 24-hour one means making that call more often, and a point forecast alone doesn’t say how much margin it needs.

This article explains how probabilistic power forecasting can support that sizing decision, using an approach some Power Forecasting customers already apply in practice.

From point forecasts to probability distributions

The default wind or solar power generation forecast is a point forecast: the model’s best estimate of where production will land. This deterministic approach places all probability mass on one outcome.

The probabilistic forecast answers the question of how confident you should be in that outcome by providing quantiles alongside the point estimate: the P5, P10, P25, P50, P75, P90, P95, and so on. Each quantile defines a threshold with a specific probability. The P5 forecast, for example, represents the level below which production is expected to fall in only 5 out of 100 scenarios. The P95 represents the level below which production is expected to fall in 95 out of 100 scenarios.

The interval between these quantiles is itself informative. A narrow P5–P95 range signals high confidence: the model sees limited spread in likely outcomes. A wide range signals the opposite: conditions are uncertain, and the production outcome could vary significantly. This distinction matters for any decision that depends not just on expected output, but on the reliability of that expectation.

Why aFRR capacity changes the calculus

Germany and Belgium have procured aFRR capacity in 4-hour blocks for years. France operates on 1-hour blocks. In November 2025, Dutch TSO TenneT introduced 4-hour contracts for aFRR capacity in the Netherlands, replacing the traditional 24-hour blocks. From late April 2026, the share of 4-hour products has been increasing by 2% per week, targeting 40–45% by July 2026 and 100% by the end of the year.

While 24-hour contracts favored predictability and rewarded simple availability – primarily gas plants – 4-hour products lower the participation threshold and better align with intraday dynamics. Batteries, solar, and wind become increasingly competitive as portfolios are pushed to actively steer between markets.

For solar power traders, this creates a specific challenge: committing aFRR capacity in a 4-hour block requires confidence about what you can deliver within that window.

A conservative commitment floor

Probabilistic forecasting becomes useful with shorter aFRR capacity blocks. Rather than bidding based on the P50 – the expected output – a trader can, for example, use the P25 as a conservative lower bound (the level the asset is expected to reach or exceed in 75% of scenarios).

If you sell a block of aFRR upward capacity, you need to be confident you can steer production in the committed direction when called upon. Using the P25 as your commitment ceiling means you are building in a margin of safety against forecast error. You can then optimize the remaining block between your committed volume and the P50 in the day-ahead, intraday, or imbalance market.

How conservative to go is a choice. A P25 floor accepts a roughly 1-in-4 chance of falling short of your commitment; a P10 floor lowers that to roughly 1 in 10, at the cost of a smaller committed volume. The decision depends on what a missed delivery actually costs you.

Sizing a bid in practice

A single day

The figure below shows a single day from a solar aFRR upward capacity book: 12 May 2026 in the Netherlands, a 10 MW park bidding into 4-hour blocks. It plots the point forecast, the P5–P95 distribution, and what the park actually delivered.

For the 12:00–16:00 block, we show two sizing approaches. A haircut approach bids a fixed percentage of the point forecast as a safety margin – here, 0.8 of it – and takes the lowest value in the block as committed capacity. A probabilistic approach takes the P10 forecast and does the same, more conservative than the P25 discussed above and reflecting this trader’s own appetite for risk.

Both hold for most of the window, but between roughly 15:15 and 15:45, realized availability drops below the haircut line, meaning the asset-backed trader has sold capacity it can’t deliver. The P10 line stays below realized output for the full four hours.

Eight months of 4-hour blocks

Below, we look at the first months of four-hour aFRR capacity blocks since TenneT introduced them in the Netherlands in November 2025, restricted to blocks where the aFRR capacity price cleared above €50/MW/h. This filter keeps the comparison focused on blocks where the sizing decision actually has commercial stakes.

We compare six ways of sizing the aFRR bid: four fixed haircuts of the point forecast (0.4x–0.7x) and two probabilistic floors (P5, P10) for a 10MW solar portfolio in the Netherlands.

The clearest pattern is that more aggressive sizing doesn’t reliably pay off. The two most conservative rules, 0.4x and 0.5x, never breach across the 20 blocks, but they also leave revenue on the table: €54,974 and €61,539 respectively. Pushing the haircuts higher doesn’t fix that: 0.6x and 0.7x fail on 4 and 5 of 20 blocks (20% and 25%) and still earn less than P10.

P10 fails on 2 of 20 blocks – a 10% rate that lines up closely with what a P10 commitment implies by definition – and produces the highest net revenue of any rule tested: €65,552. It beats every fixed haircut in the comparison, including the more conservative ones that fail less often and the more aggressive ones that fail more often.

The revenue case for P10

Whether you base your aFRR capacity bid on a haircut of the point forecast or a probabilistic floor like P1 ultimately boils down to risk appetite. P10 still carries some chance of falling short when called, but even accounting for that, it earned more than every fixed haircut we tested, including ones both more and less conservative. A haircut can get you to zero failures, but it gets there by giving up more revenue than P10 loses to its occasional misses.

Read more

Probabilistic forecasting involves several layers of methodology, ranging from quantile regression and conformal prediction to evaluation metrics such as CRPS and quantile loss. For a deeper understanding, please see our white paper, “Probabilistic forecasting for short-term power trading.”