The intraday value of WeatherNext 3’s better PV forecasts.

Executive Summary:

WeatherNext 3 (WN3) is more accurate than ECMWF for day-ahead solar forecasting. In a six-month backtest of a 100 MWp PV portfolio across Belgium and Germany, WN3’s initial day-ahead nominations left 12–14% less absolute imbalance than ECMWF’s, and the advantage held after intraday updates. Its residual imbalance costs were 29–37% lower in Belgium and Germany, respectively, excluding 6 April. Total revenue with intraday participation was also higher, by roughly 0.9% in Belgium and 0.3% in Germany. WN3's main value lies in a better initial position on the day-ahead market, with intraday updates providing further control. Financial gains are not guaranteed, since imbalance prices are volatile and can outweigh forecast improvements, but lower imbalance volume reduces a BRP’s exposure to that price risk.

How we set up the PV comparison with ECMWF IFS MRC

In our previous WeatherNext 3 analysis, we examined how forecast accuracy translates into imbalance costs when a renewable asset holds its day-ahead nomination until delivery. Intraday trading gives operators further opportunities to update that position as new forecasts arrive.

We now test those opportunities for exemplary 100 MWp Belgian and German PV sites. We compare the median (p50) of the WeatherNext 3 (WN3) ensemble forecast with ECMWF’s Integrated Forecasting System (IFS) Medium-Range Control forecast (MRC), following each model through the day-ahead market and three intraday auctions. At every auction, the strategy uses the latest forecast already available when bidding closes. The backtest covers 1 February through 31 July 2026. We represent a 100 MWp PV installation using a simple irradiance-to-power conversion: output scales linearly with global horizontal irradiance (GHI), with a performance ratio of 0.85. At 1,000 W/m², the modeled site produces 85 MW. We apply the same conversion to both forecasts and the ground-truth observed irradiance.

We compare both models at weather-station locations, averaged across sites, so measured irradiance is the ground truth, instead of having to resort to reanalysis datasets such as ERA5. We use GHI measured across 13 Royal Meteorological Institute (KMI) weather stations in Belgium and 62 German Weather Service (DWD) weather stations in Germany, highlighted in Figure 1. This gives us a feasible, local observational reference for evaluating solar forecasts without needing production telemetry from a particular PV installation. ERA5, on the other hand, used in our previous blog post, reconstructs weather through a numerical model informed by observations. For this site-level experiment, we chose that direct measurement as the preferred ground truth. The resulting PV generation is still an idealized proxy; it does not reproduce the panel geometry, equipment, or operating constraints of a real 100 MWp plant.

Fig. 1: Stations used for ground-truth GHI measurements.

We retrieve ten-minute GHI measurements through KMI’s and DWD’s open-data services and average them into hourly values, following their documented hourly aggregation. Each hour uses the six ten-minute intervals ending at that hour and requires at least five valid samples to be considered during settlement. Both forecasts are requested at the points nearest the 13 KMI and 62 DWD weather stations. WN3’s forecasts are served on a 0.1° grid. ECMWF uses a different grid: an octahedral reduced Gaussian grid. The grids don’t align perfectly, so we extract the nearest source cell for each weather station. The selected model points lie around 0.6 to 3.9 km from the stations but are not identical. Comparing grid-scale forecasts with point measurements therefore includes differences in local cloud conditions that the grids cannot fully resolve.

The strategies share an average station generation profile of 67,490.9 MWh across 16,960 settled quarter-hours in Belgium, and 68,900.6 MWh across 17,048 settled quarter-hours in Germany. We evaluate each model using a day-ahead-only strategy and one that updates its commercial position at each intraday auction: IDA1, IDA2, and IDA3. We update only the quarter-hour intervals covered by the auction with a cleared price. Physical generation stays fixed; this experiment tests nomination changes rather than optimized curtailment or continuous intraday trading.

Which forecasts can be used at each auction?

An intraday auction collects buy and sell orders and clears them together. For a PV operator, it provides another opportunity to adjust an earlier sale. If the generation forecast falls, the operator can buy back some of that sale. If it rises, the operator can sell additional volume. Our strategy updates our nominations to be equal to the current forecasted generation of the PV site, without forecasting prices or picking trades based on an expected market spread. The goal of this strategy is thus to diminish our imbalance volume at delivery as much as possible, not to pick the most lucrative trades. The three auctions follow the Single Intraday Coupling (SIDC) intraday auction schedule. IDA1 closes at 15:00, while IDA2 closes at 22:00. Both cover the full following delivery day D. IDA3 closes at 10:00 on delivery day D but covers only quarter hours from 12:00 onward. All auction gate closures and delivery times are referenced in local Belgian time (CE(S)T), with D denoting the delivery day.

To determine which forecast applies at an auction, we need to distinguish its initialization time, publication time, and lead time. The nominal initialization time is the starting point from which forecast lead time is counted and is referenced in UTC. Lead time measures the interval from that starting point to the hour of electricity delivery. The forecast becomes available to the operator at its publication time. For example, a forecast initialized at 00:00 UTC with a six-hour publication delay is available at 06:00 UTC (07:00 CET/08:00 CEST); any auction deadline after that time can use it. In this backtest, we assume a publication delay of seven hours for WN3 and six hours for ECMWF. At each auction deadline, we select the most recently initialized run that has already been published. This ensures that the simulated strategy only uses forecasts that would have been available when the trading decision was made.

The update frequency determines how old that latest available run is. WN3 produces a new forecast every hour. Because the auction deadlines also fall on whole hours, its latest available run is always seven hours old. Shifting the same local deadline from CET to CEST moves WN3’s selected UTC initialization one hour earlier, so the local time of that initialization stays the same. ECMWF produces four runs a day, at 00:00, 06:00, 12:00 and 18:00 UTC. Each becomes available six hours later, and the operator then waits up to six hours for the next publication. The latest available ECMWF run is therefore between six and eleven hours old at a whole-hour deadline. At these four auctions, ECMWF uses the same UTC initialization in both seasons, so that run is one hour older under CET than under CEST.

That age is part of the lead time. For example, at IDA3’s 10:00 deadline the first tradable hour is 12:00, two hours later. WN3’s run started at 02:00 UTC (03:00 CET) or 01:00 UTC (03:00 CEST), and its 12:00 forecast has a nine-hour lead in both seasons. ECMWF’s run was initialized at 00:00 UTC (01:00 CET/02:00 CEST), so its forecast for 12:00 has an eleven-hour lead in CET and a ten-hour lead in CEST.

Figure 2 shows how these publication delays and update schedules meet the auction deadlines. WN3’s lead ranges are the same in both seasons: 19–42 hours at day-ahead, 16–39 hours at IDA1, 9–32 hours at IDA2 and 9–20 hours at IDA3. ECMWF’s summer ranges are 22–45, 16–39, 10–33 and 10–21 hours, and each range is one hour longer in winter. WN3 holds the more recent initialization at these gates, except at IDA1 in summer, when both models use the 06:00 UTC run. Its hourly cycle provides that fresher initialization even though each run takes one hour longer to publish.

Fig. 2: Lead times of the WN3 and ECMWF forecasts used at each day-ahead and intraday auction.

To compare the performance of both models at each auction’s gate closure time, we compare their mean absolute errors (MAE) for each forecasted hour. We will only consider hours with an average measured GHI of at least 20 W/m² for this, the hours that actually matter for PV generation. Looking at Figure 3, our biggest takeaway is that WN3’s day-ahead forecast is already more accurate during these daylight hours than ECMWF’s final auction forecast at IDA3, with a 0.55 MW gap at 15:00 in Belgium and a 0.41 MW gap at 15:00 in Germany.

The most volatile, and therefore most important, hours for PV generation are the daylight hours between 12:00 and 16:00, highlighted in Figure 3. At IDA3, the first tradable quarter hour is 12:00, two hours after bidding closes. IDA3 thus plays an important role in correcting our nominations for these volatile hours with the freshest information available. We can indeed see WN3’s mean absolute error fall from 6.62 MW at day-ahead to 6.16 MW at IDA3 in Belgium, and from 7.52 MW to 7.00 MW in Germany. ECMWF improved from 7.69 MW to 7.08 MW in Belgium, and from 8.65 MW to 7.87 MW in Germany. These last-minute corrections thus only carry marginal improvements for our portfolio. WN3 shines in medium-range forecasts; for short-term corrections like in IDA3, we should use other forecasting solutions, such as a solar nowcast.

Fig. 3: MAE at each delivery hour of the used forecasts.

Intraday corrections reduce imbalance volume

To compare the performance of both weather forecasting models when informing day-ahead and intraday positions, we first measure the physical mismatch at delivery. This mismatch, or imbalance, indicates how precise each PV forecast was, since a perfect PV forecast would result in no imbalance. For each quarter-hour, absolute imbalance is the absolute difference between nominated and modeled actual power. In Belgium and Germany, each out-of-balance position leads to a financial outcome in the imbalance settlement procedure, regardless of sign. We therefore plot the summed absolute values in Figure 4 to prevent overproduction and underproduction in different quarter hours from canceling each other out.

Fig. 4: Cumulative absolute imbalance energy at delivery; lower is better.

With total physical generation output of 67,490.9 MWh in Belgium and 68,900.6 MWh in Germany, Table 1 shows the absolute imbalance volume across both day-ahead and intraday auction strategies. ECMWF has more errors to correct, and thus removes slightly more volume (up to 952.1 MWh or -6.61% in Germany). WN3 nevertheless maintains a smaller total mismatch, before and after the intraday auctions, retaining an over 11% volume advantage in both countries. The main advantage is already present when WN3 sets the initial day-ahead position.

Tab. 1: Absolute imbalance volumes after day-ahead nominations and intraday updates.

How that volume translates into imbalance costs

Each MWh of imbalance has a different financial effect depending on its direction and the current imbalance price. To illustrate the direct financial impact of a more precise PV generation forecast, we look at residual imbalance costs after following each model's generation forecast. We calculate the residual imbalance invoice as nominated power minus actual power, multiplied by the quarter-hour duration and imbalance price. Positive values are costs; negative values are revenue. Since our strategy does not use price forecasts to guide trading decisions, the financial results reflect both forecast performance in the imbalance volume and the market conditions observed during the backtest. They should therefore be interpreted cautiously and do not, on their own, establish a repeatable trading advantage.

Extreme price events, such as imbalance prices reaching €-15,000/MWh in Belgium on Easter Monday, April the 6th, 2026, and the German imbalance price (reBAP) reaching €-6,642.76/MWh in Germany, can significantly affect individual station results. In the KMI 13-station mean in Belgium, WN3 lost €11,885 on Easter Monday intraday corrections, while ECMWF gained €17,447. In Germany, WN3 gained €3,013, while ECMWF lost €13,776. These outliers are plotted in Figure 5. To evaluate overall performance, we analyze the window both with and without Easter Monday.

Fig. 5: Residual imbalance costs after each intraday auction for both models; Easter Monday is a clear outlier.

The cumulative curves in Figure 6 consider only imbalance costs resulting from the imbalance settlement. Cash paid or received in the day-ahead market and intraday auctions will be considered later.

Fig. 6: Cumulative imbalance costs, settled on the imbalance price, following each strategy and each model; lower is better.

Table 2 shows the financial impact of intraday updates on residual imbalance settlement invoices, listing figures for both the full window and the period excluding the extreme outlier event on Easter Monday. It is striking that ECMWF’s intraday updates reduce imbalance volume more, but this isn’t reflected everywhere in greater imbalance cost savings. WN3 achieves up to €43,407 in savings in Belgium (excl. Easter Monday) compared to ECMWF's €25,547. This financial outcome also depends on the timing and direction of the remaining imbalances and the prices at which they are settled. The sharp April step also further illustrates why we report the sample without 6 April, given the size of its impact. In addition, we can see that both models perform very similarly for most months in our sample in both Belgium and Germany. Only from May onwards does ECMWF's performance diverge from WN3. Indeed, higher summer irradiance creates more megawatt-hours of exposure per hour of forecast error. WN3’s benefit really starts accumulating during these months.

Tab. 2: Imbalance cost deltas after intraday updates compared to only day-ahead nominations. Negative values are invoice savings. 

Total profit and loss also includes the day-ahead and intraday trades

To evaluate the final financial outcome, we combine day-ahead sales, cash from all three intraday auctions, and residual imbalance settlement. A smaller imbalance invoice can require expensive auction purchases. Conversely, an auction trade can earn cash without substantially reducing the final invoice.

Fig. 7: Impact of intraday updates on total profit and loss and imbalance volume when following each model, without Easter Monday.

All arrows in Figure 7 point up and to the left: intraday participation following increasingly precise forecasts reduces mismatch and increases cash over this window. Table 3 outlines total overall revenue, intraday financial gains, and their percentage share of overall revenue across regions and models. Overall, WN3 finished €33,874 ahead of ECMWF with intraday updates in Belgium (€3,690,900 vs €3,657,026), and €12,205 ahead in Germany. Relative to overall revenue, these incremental intraday gains remain modest across both models (ranging from 0.15% to 0.31%).

Tab. 3: Total profit and loss and impact of intraday updates.

In this sample, ECMWF’s intraday updates begin to close the gap with WN3’s DA-only position. The question is whether this improvement is repeatable. We resample whole delivery days 10,000 times to estimate a 95% bootstrap interval for the cumulative gain, plotted in Figure 8. The interval runs from −€10,834 to +€25,572 for WN3 in Belgium and −€3,356 to +€13,896 in Germany; and from −€15,889 to +€39,675 for ECMWF in Belgium and −€3,259 to +€18,157 in Germany. ECMWF’s wider interval reflects greater variability in its daily financial gains. Noisier forecast updates can drive this variability by prompting more frequent, larger intraday trades, while volatile imbalance prices amplify the financial consequences of remaining forecast errors. The interval alone, however, cannot distinguish these contributions. Both intervals include zero, so this backtest does not establish a repeatable cash advantage from generation-following nominations using either model’s intraday updates. The large jump needed for ECMWF to match WN3’s results illustrates the added risk Independent Power Producers (IPPs) take on when steering their assets based on ECMWF’s forecast rather than WN3's.

Fig. 8: Bootstrapped 95% confidence intervals for the cumulative value of the intraday corrections following each model, without Easter Monday.

What does this mean for setting PV nominations?

For renewable portfolio managers and BRPs in Belgium and Germany, WeatherNext 3's primary value lies in risk mitigation and baseline portfolio protection, not in chasing speculative intraday trading margins. By delivering a day-ahead forecast already more accurate than ECMWF’s final IDA3 update, WN3 lets market participants secure cleaner initial market positions with longer lead times. This significantly reduces structural exposure of IPPs to volatile imbalance settlement mechanisms.

WN3’s day-ahead nominations achieve 13.9% less absolute imbalance than ECMWF in Belgium and 11.9% in Germany. Furthermore, intraday auction updates add precision, reducing WN3's absolute imbalance by an additional 5.21% (588.2 MWh) in Belgium and 5.66% (717.8 MWh) in Germany over the full sample. These volume reductions directly translate into significant risk reductions for IPPs and imbalance cost savings, lowering residual imbalance invoices in both Belgium and Germany.

Although pure generation-following nominations capture modest, incremental revenue gains (+0.15% to +0.31%), WN3’s primary commercial advantage stems from its superior baseline accuracy and systematic imbalance reduction. To fully use WN3’s forecasts, market participants should combine its robust medium-range forecasts with real-time solar nowcasting and price-aware trading signals. Integrating these tools allows BRPs to maintain strong baseline portfolio protection while optimizing intraday position updates when market conditions are most favorable.