Managing forecast accuracy expectations

Overview

Forecasting is the process of predicting solar irradiation, PV or wind power output, and meteorological parameters across different forecast horizons. The result is time-series data — a collection of timestamps with a requested resolution and corresponding forecast values.

Despite high expectations, forecasting is a challenging and complex domain. Forecast accuracy is determined by a combination of factors, including seasonal weather variability, location-specific climatic and geographical characteristics, the forecast horizon, and the inherent limitations of Numerical Weather Prediction (NWP) models. Understanding these factors is essential for setting realistic accuracy expectations and for meaningful communication between forecast data providers and PV power plant operators.

Note: A prediction model that delivers highly accurate forecasts every day does not exist. Forecast accuracy should always be evaluated over a sufficiently long period — ideally a full year — to account for all seasonal patterns and provide statistically reliable results.

How seasonality affects forecast accuracy

Seasons significantly affect the accuracy of PV power output forecasts. Seasonal weather patterns vary in cloud cover, humidity, temperature, and other atmospheric conditions that directly impact solar radiation levels. For example, summer typically brings clear skies and consistent sunlight, whereas autumn and winter bring more frequent cloud cover, rainfall, and fog.

NWP models generally perform better during stable weather conditions and worse during rapidly changing ones. The following examples illustrate changes in day-ahead forecast accuracy across individual months of 2023 for four different locations.

Reunion Island

Weather patterns on Reunion Island do not change significantly throughout the year. The year divides into two seasons: a warmer and more humid period from November to April, and a fresher and drier period from May to October. There are no major differences in temperature, humidity, cloud cover, or sunlight intensity between the seasons. As a result, forecast accuracy expectations can be relatively consistent across most months.

Figure 1: Seasonality of forecast deviations — Reunion Island (PV power output, 2023). Mean absolute deviation [kWh] by month.

Central Mexico

Central Mexico has a tropical climate. The rainy season typically runs from late spring (May–June) to early autumn (September–October). The remainder of the year is hotter with more sunny days. Forecast accuracy is expected to be worse during the rainy season due to more frequent and unpredictable weather pattern fluctuations.

Figure 2: Seasonality of forecast deviations — central Mexico (PV power output, 2023). Mean absolute deviation [kWh] by month.

Southern France

Southern France has four distinct seasons. Summer (June to August) is characterized by long daylight hours, clear skies, and minimal rainfall — conditions that support high forecast accuracy. Autumn and winter (September to February) bring increased rainfall and less stable weather, raising forecast uncertainty. In spring (March to May), rainfall decreases and weather stability improves, which positively influences forecast performance.

Figure 3: Seasonality of forecast deviations — southern France (PV power output, 2023). Mean absolute deviation [kWh] by month.

Central Vietnam

In central Vietnam, the rainy season — which includes occasional typhoons and tropical storms — runs from September to December. The dry season, with more stable weather, lasts from January to August. October and November show the highest forecast uncertainty, driven by rapidly changing weather patterns where clear skies can give way to intense downpours within minutes. Such conditions represent the most significant challenge for weather forecasting models.

Figure 4: Seasonality of forecast deviations — central Vietnam (PV power output, 2023). Mean absolute deviation [kWh] by month.

Challenge of forecasting

One of the biggest challenges for weather forecasting models is accurately predicting the exact moment a weather pattern changes. In everyday contexts, a one-hour timing error in cloud arrival may go unnoticed. In PV power forecasting, however, the same error can have significant financial consequences. PV power plant operators who provide generation forecasts to Distribution System Operators (DSOs) or Transmission System Operators (TSOs) may face financial penalties for inaccurate forecasts.

Predicting the precise timing of a weather change is difficult for all existing prediction models, particularly at longer forecast horizons. The following examples show day-ahead forecasts with prediction horizons from 24 to 48 hours ahead.

Example: Correct magnitude, shifted timing

In this case, an NWP model correctly predicted a sudden drop in PV power output and even its magnitude. However, the drop occurred approximately one hour later than predicted, resulting in a large forecast deviation of 100 MW at 11:30. The forecast was issued at 6:00 the day before — meaning the model was required to anticipate the event approximately 30 hours in advance.

Figure 5: NWP day-ahead forecast vs. reference — PV power output. A correctly predicted power drop occurred approximately one hour later than forecast, causing a deviation of ~100 MW at 11:30.

Example: Correct timing, steeper-than-forecast drop

Here, the NWP model correctly predicted a drop in PV power output, but in reality the drop started one hour later and was steeper than forecast. Although overall accuracy for the day is acceptable, the forecast deviation at 14:30 exceeded 100%. The forecast horizon at that point was nearly 33 hours ahead.

Figure 6: NWP day-ahead forecast vs. reference — PV power output. The actual power drop was steeper than forecast and began approximately one hour later, resulting in a deviation exceeding 100% at 14:30.

Example: Variable weather with time-shifted deviations

During this day, clear and overcast conditions alternated, creating an oscillating reference PV power output profile. The NWP model correctly predicted the general up-and-down pattern, but time differences between the prediction and reality produced both positive and negative forecast deviations. The average absolute deviation between 6:00 and 18:00 was approximately 25 MW, with a forecast horizon ranging from 24 to 36 hours ahead.

Figure 7: NWP day-ahead forecast vs. reference — PV power output. Variable weather caused alternating positive and negative forecast deviations, with an average absolute magnitude of 25 MW between 6:00 and 18:00.

Forecast horizon and uncertainty

The forecast horizon directly influences the accuracy of PV power output predictions. Shorter forecast horizons generally yield higher accuracy because there is less time for weather conditions to change unpredictably. Over shorter timeframes, atmospheric conditions are more stable and thus more predictable.

For longer forecast horizons, accuracy decreases because the atmosphere is a chaotic system — small errors in initial conditions can grow exponentially over time. Even the most advanced NWP models have limitations in their ability to predict weather reliably beyond a few days.

The following tables show PV power output forecast accuracy across six forecast horizons (leadtimes) ranging from intra-hour (H0) to two days ahead (D2), for three example locations. Data represent averages for 2022 and 2023.

Note: For all example locations, forecast accuracy deteriorates as the forecast horizon extends.

Table 1 — Bratislava, Slovakia, Europe

Table 2 — Johannesburg, South Africa, Africa

Table 3 — Singapore, Asia

Location specifics

The location of a PV power plant significantly influences forecast accuracy. Climatic, geographical, and environmental factors all play a role in how predictable local solar irradiance is.

Climatic conditions

Locations with stable, predictable weather patterns — such as deserts with consistently clear skies — tend to yield more accurate PV power output forecasts. Lower variability in solar irradiance simplifies the forecasting process. Regions with highly variable weather, such as those experiencing frequent cloud cover changes, storms, or rapid atmospheric shifts, pose greater challenges. Coastal areas or regions with monsoon seasons can experience rapid and unpredictable changes that directly affect solar irradiance.

Geographical attributes

Topography is an important factor. Mountainous regions create complex weather patterns through orographic effects, where rising and cooling air leads to cloud formation and more frequent, sudden changes in cloud cover. These conditions reduce forecast accuracy. Flat terrain, such as plains, typically has more stable and predictable weather patterns, which supports more accurate forecasts.

Environmental factors

Proximity to large bodies of water can produce specific phenomena — such as lake-effect clouds or coastal fog — that are difficult to forecast accurately and can cause sudden changes in solar irradiance. Dense forests influence local humidity and temperature, contributing to localized cloud formation and weather patterns that are harder to predict. Conversely, arid regions with sparse vegetation tend to have clearer skies, making solar irradiance more predictable and improving forecast accuracy.

The following example compares forecast accuracy for two contrasting locations: Singapore (variable weather with abundant rainfall and frequent thunderstorms) and Dubai (relatively stable weather with minimal rainfall, especially in summer). Data represent averages for 2022 and 2023.

Table 4 — Singapore, Asia

Table 5 — Dubai, United Arab Emirates

Comparing Bias, MAD, and RMSD percentage values across the two locations confirms that forecast accuracy for Dubai is significantly better than for Singapore. This difference reflects the greater weather stability at the Dubai site and demonstrates the strong influence of local climate on forecast performance.

Acceptable forecast accuracy for a variable weather day

"Acceptable forecast" is a relative term that depends on weather variability, forecast horizon, and the perspective of the evaluator. On a day with sudden and frequent weather pattern changes — visible as sharp ups and downs in the reference PV power output — an NWP model is required to predict the exact moment of each change with high precision. This is inherently challenging.

In practice, NWP models often produce a smoothed forecast on such days, generating positive and negative deviations around the actual output. However, if the daily sums of predicted and actual PV power output are approximately equal, many forecast domain experts would consider the result acceptable — particularly for day-ahead forecasts with a horizon from 24 to 48 hours.

Figure 8: NWP day-ahead forecast vs. reference — PV power output on a variable weather day. The model provides a smoothed forecast while the actual output fluctuates sharply. Daily totals for forecast and reference are approximately equal.

Forecast inaccuracy during consecutive days

It is common for PV power plant operators to react strongly when forecasts underperform over a short sequence of consecutive days. When financial penalties are tied to forecast accuracy, even one or two days of poor performance can prompt complaints. However, from a forecast domain perspective, drawing conclusions from such short periods is not statistically meaningful.

PV power output is highly dependent on weather conditions, which are inherently variable and can be unpredictable — especially over longer forecast horizons. Sudden weather events, such as unexpected cloud cover or storms, can significantly affect accuracy for isolated days without reflecting overall model quality. Forecast performance should be assessed over a much longer period to obtain a reliable picture. Even the best models will occasionally produce less accurate predictions given the complex nature of atmospheric dynamics.

Important: Evaluating forecast performance based on one or two days is not statistically significant and can lead to unrealistic accuracy expectations. A minimum evaluation period of one full year is recommended.

Forecast accuracy evaluation

Setting realistic forecast accuracy expectations requires thorough analysis of historical forecast and reference data. At a minimum, the following factors must be considered:

• Seasonality and the time of year.

• Time of day.

• Forecast horizon (leadtime).

• Location-specific climate and geography.

• Length of the evaluation period.

Ideally, a full year of the most recent available data should be used. Shorter periods — such as individual months — may not capture all seasonal weather patterns. Very short periods — days or weeks — can produce misleading conclusions due to an insufficient sample size.

The forecast horizon is a crucial consideration. Accuracy generally deteriorates as the horizon extends, so expectations must be calibrated accordingly. For example, accuracy expectations for a two hours-ahead (H2) forecast must differ from those for a two days-ahead (D2) forecast. Properly calibrated expectations prevent misunderstandings between forecast data providers and PV power plant operators.

What is behind the forecast product

The final forecast product looks simple: a time series with a requested resolution and corresponding forecast values. However, the process behind that output is complex. Before a forecast time series is delivered, significant volumes of data must be downloaded, pre-processed, transformed, calculated, weighted, and post-processed.

The final product is a combination of multiple prediction models — including Cloud Motion Vector (CMV) models and several Numerical Weather Prediction (NWP) models. These combinations are not static; they vary by location, season, forecast horizon, and data resolution. All forecast data sources are continuously evolving, and the supporting infrastructure must accommodate rapid development in the field.

Forecast accuracy continues to improve, but all prediction models have inherent limitations that must be considered when setting accuracy expectations. Forecasting is a complex domain, and the primary goal of forecast data providers is to deliver outputs that are as accurate as possible given the current state of the science.

Further reading

• Operational and historical forecast

• Origin of the data for Monitor and Forecast

• Validation of Forecast data

• Evaluating forecast accuracy - best practices