Forecasting wind and solar for short-term power trading: Three shifts and ways to adapt

We asked many short-term power traders what defines a good day. The answer was rarely about the hours that go to plan, but about how quickly they can respond when they don’t. When a storm shifts, a unit trips, or prices bounce without warning, the difference between a loss and a gain often comes down to seconds. The reliability of their power forecast plays a big role during these moments.

In this environment, a high-performing model alone won’t be enough to deliver a reliable forecast. The best results come from a clear understanding of market behavior, data quality, and the ability to respond to edge cases as they happen.

Now, halfway through the year and deep into a volatile summer, we explore three market shifts that are impacting the requirements for wind and solar power forecasting.

Market context: Volatility and rising balancing costs

While the acute shocks of the energy crisis have passed, a new pattern of volatility is emerging. The imbalance spread – defined here as spot price minus the imbalance price – continues to widen. The box plot below displays the 25th and 75th percentiles, the minimum and maximum values, and the ±10% range. Notice that the maxima, but especially the minima – typically caused by large imbalance price peaks – are still climbing.

A striking example occurred recently in the Netherlands, where activation prices exceeded €3,990/MWh during a heatwave – a period marked by high temperatures, low wind, and soaring demand.

This growing spread is a strong proxy for rising balancing costs and highlights a key risk: even in relatively stable day-ahead markets, the cost of getting it wrong in real time is increasing.

To navigate this trend, forecasts need to do more than predict averages. They need to capture risk, respond to market shifts, and reflect real-world constraints. In our experience, three trends are reshaping what short-term traders and analysts require from wind and solar forecasts.

Traded volumes shifting to the intraday market

Across Europe, trading activity is shifting away from the day-ahead market and accelerating on the intraday level. This chart of transactions over time shows a near-exponential increase in activity on the Single Intraday Coupling (SIDC).

As volumes move toward intraday, the need for high-resolution, short-term forecasts grows more urgent. The analogy we often use is hiking: the day before, you check tomorrow’s forecast. But if you’re about to leave now, you open a weather app with real-time radar imagery and optical flow technology to track incoming rain. Or: you just look out the window.

Forecasting in power markets works the same way. Close to delivery, combining numerical weather prediction (NWP) models with real-time observations leads to better accuracy. But further out, say six or more hours ahead, just looking out the window (or relying only on real-time data) won’t be enough. You still need a robust weather model to predict how the situation will evolve.

This is reflected in our own benchmark analysis for a wind portfolio. The chart below compares two approaches over a short forecast horizon (exact time frame not shown for confidentiality reasons). A forecast based purely on numerical weather prediction (NWP) holds an average Normalized Mean Absolute Error (NMAE) of just over 11%. However, when we integrate online asset data, such as real-time production, into the ensemble, performance improves significantly, especially close to delivery. Our approach to incorporating this data ensures that the current measured production from the park is filtered, interpreted, and integrated intelligently.

Increasing tail risk from rare weather events

Extreme or rapidly changing weather events are playing an increasingly significant role in trading risk. These “tail risks” are difficult to predict, but their financial impact can be severe, potentially leading to losses of millions in a single day. Examples include:

Dense morning fog or stratoclouds, reducing irradiance and delaying solar generation

• Sahara dust, coating panels and lowering solar output
• Wind turbine icing, causing widespread shutdowns
• Snow or ice accumulation on PV panels, delaying generation until melt-off
• Thunderstorms, major storms, or changing wind directions, disrupting wind production

We’ve illustrated this ‘pain’ using a graph of balancing costs plotted against imbalance volume and price – a format inspired by a discussion we had with a short-term power trader. The x-axis shows whether the forecast was too high or too low. The y-axis shows whether the imbalance price was higher or lower than the day-ahead price.

The red zones mark the most damaging combinations: underrforecasting when prices are negative, or overforecasting when imbalance prices spike. These ‘wrong-wrong’ scenarios – where both forecast direction and the market price move against you – account for a disproportionately large share of total losses. The lines across the plot, which we could call isocost lines, intuitively show that the combination of both factors brings you into the expensive territory.

At Dexter, we’ve built several features into our forecasting model to help traders address these rare but impactful events:

• We benchmark every relevant weather model to evaluate which performs best in specific conditions.
• Rather than relying on a single best estimate, we provide a range of possible outcomes (probabilistic forecasts). This helps traders understand the uncertainty surrounding extreme events and make more informed decisions.
• Our models detect and correct for expected performance losses, such as reduced output due to icing, directly in the forecast.
• We notify users when a tail risk is on the horizon. These alerts are designed to prompt vigilance, support internal checks, and allow traders to adapt strategies before the impact occurs.

Persistent bias from poor data quality

Apart from weather and market dynamics, there’s a quieter but equally powerful source of forecast error: persistent bias due to poor data quality.

Power forecasting curves based on wind speed or solar irradiance often reflect lower-than-expected outputs due to grid or commercial curtailments, redispatch constraints, or maintenance. These factors create data pollution. If not filtered out, models trained on this data tend to under- or over-forecast.

As renewable penetration grows, curtailment becomes both more common and more consequential. According to reports compiled by the Financial Times, in 2024, nearly 10% of Britain’s planned wind output was curtailed, along with almost 30% in Northern Ireland; Germany saw about 5% of its renewable generation curtailed. Moreover, in the first five months of 2025, Poland reported a 36% increase in solar curtailment. And, in the U.S., Nat Bullard’s annual decarbonization deck highlighted the growing scale of curtailment in California, which has more than tripled since 2019.

To counter the impact of poor-quality data, we detect and exclude affected data points before they influence the training pipeline. Our cleaning process is:

• Monitored manually by our forecasting team (a ‘human-in-the-loop’ approach)
• Supported by automated detection tools
• Fully integrated into our algorithmic pipeline

In addition, we deliver our forecasts through a modern API that allows customers to retrieve their latest outputs and dynamically update their portfolios. When an asset is added or removed, the very next forecast already reflects the change.

Get the most out of your power forecast

To conclude, effective forecasting for day-ahead and intraday trading currently depends on three things:

• Near-time data, such as online asset data, satellite imagery, and weather station observations, to improve near-term accuracy
• Active management of tail risk during rare or extreme weather events
• High data quality, through automated cleaning or expert review

At Dexter, we cover all three aspects. If you’d like to learn more about how we support short-term power trading with accurate, reliable forecasts, get in touch.