The role of traditional and AI weather models in the future of renewable energy forecasting

Accurate weather forecasts are fundamental in making wind and solar energy production more predictable, ultimately accelerating the integration of renewables into the grid. Unexpected shifts in weather patterns can lead to volatile energy prices and costly imbalances for power producers or traders.

Fortunately, advancements in weather forecasting are helping us tackle these challenges. A diverse array of weather models – from global deterministic systems to breakthrough AI ensembles – is now available. For those trading renewables on short-term power markets, the question arises: Which weather forecast is most reliable for forecasting the production of a particular wind or solar park and the power price in a specific market?

This article reviews all types of weather models, exploring their strengths and weaknesses as well as recent trends. Drawing on our extensive historical weather forecast data lake and experience in power and price forecasting, we argue that the most accurate wind and solar power generation and price forecasts result from a careful combination of multiple ‘old’ and ‘new’ models. If you’re looking for ways to assess the accuracy of a forecast, we also recommend our recent white paper exploring how to evaluate power generation forecasts for optimal decision-making.

Deterministic global NWP models

Let’s start with the ‘traditional’ weather models, which we typically refer to when talking about numerical weather prediction (NWP). These models simulate the atmosphere and oceans by solving physical equations (i.e., the flow of fluids and physical processes) and return predictions for hundreds of meteorological variables, ranging from the ocean surface to the upper atmosphere.

Some of the most well-known (at least by acronym) and widely used global deterministic NWP models are:

• European Centre for Medium-Range Weather Forecasts (ECMWF’s) Integrated Forecasting System (IFS)
• The Global Forecast System (GFS), run by the United States National Weather Service (NOAA)
• The Icosahedral Nonhydrostatic (ICON) model, created by Germany’s national weather service, Deutscher Wetterdienst (DWD)
• The Global Environmental Multiscale Model (GEM) from the Meteorological Service of Canada
• The UKV-Global model, developed by the UK Met Office

There is a consensus that ECMWF’s IFS, visualised below, is the most accurate forecast for trading power in the short-term markets, day-ahead to a few hours ahead. Nonetheless, combining outputs from multiple models makes it possible to reduce forecasting errors further, achieving better overall accuracy than relying on any single model. Global NWP models are invaluable because, as we will see, they serve as the foundation for all local NWP models.

ECMWF IFS model 2m temperature and 100m wind speed, forecast of 2024-09-26T00 valid for timestep +12h, based on 0.4deg open data. ©ECMWF Source: www.ecmwf.int License: CC-BY-4.0

High-resolution local NWP models

As their name indicates, high-resolution (HRES) local NWP models are more granular and region-specific. They are initialized on and use boundary conditions of global models and can have multiple levels of nesting.

Their advantage lies in their ability to zoom in on smaller geographic areas, offering a much finer level of detail, as shown in the image below. As a result, they are mainly valuable for regions with complex terrain. For instance, they can more accurately simulate phenomena like wind flows over a mountain range.

A comparison between the level of detail of a high-resolution local NWP model (on the left) versus a deterministic global NWP model (on the right). The forecast’s variable is 10m wind speed with initialization time 2024-09-26T00 and all 48 timesteps. Source: Dexter Energy

However, the increased resolution comes with a trade-off. Local NWPs introduce more detail into the forecast – precisely forecasted phenomena that might not occur in that exact space or time. In energy forecasting, we must interpret this ‘noise’ carefully to avoid incorrect predictions, which can lead to costly trading errors.

Deterministic AI models

After 2023 turned out to be a notable year for developments in AI weather models, ECMWF recently announced its new Artificial Intelligence/Integrated Forecasting System (AIFS), a data-driven ’emulator’ of its IFS weather forecast model. These launches signal a new stage in weather forecasting, where AI is beginning to play a central role. Nonetheless, while we acknowledge that deterministic AI models represent a significant advancement in the field, it is important to approach the surrounding hype with caution. As with many AI innovations, the potential benefits are sometimes overstated.

In essence, AI models emulate the global weather models. They are trained on extensive datasets spanning decades of global climate data, and their learning is based on atmospheric reanalysis, such as EMCWF’s Reanalysis v5, commonly known as ERA5.

One notable characteristic is their ability to ‘blur‘ specific details, especially at longer lead times. This is possible because they are not bound by the same physical constraints as traditional models. For instance, while traditional models might attempt to pinpoint the exact location of a rain shower, an AI model will predict the broader area of possible impact, optimizing the overall forecast accuracy, especially when looking at forecasts further out.

Comparison of IFS and AIFS precipitation for the forecast of 2024-08-29T00 between +90h and +96h, based on 0.4deg open data. ©ECMWF Source: www.ecmwf.int License: CC-BY-4.0

This might be problematic for meteorologists interpreting the forecast; a weather front won’t be as easily recognizable, and weather patterns will vary depending on the forecast’s lead time. Yet, the effect on power forecasting accuracy can be positive or negative depending on the forecast’s properties and error metric used.

Additionally, these models seem to maintain consistency and show good scores for forecasts beyond two days, already outperforming IFS HRES at longer lead times in terms of error scores. Moreover, they offer the advantage of being more cost-effective to run.

That said, a key limitation of many deterministic AI models is their six-hourly resolution. This is inadequate for energy forecasting, where hourly timesteps are required. Furthermore, their effective resolution is quite low, and the current generation cannot forecast the butterfly effect.  Although not directly a problem for operational forecasting, this makes them unsuitable for climate research.

Ensemble models

Another significant milestone in weather forecasting is the development of ensemble models, marking the shift from deterministic to probabilistic methods.

If you are unfamiliar with the probabilistic concept or could use a refresher, my colleagues in data science have written a detailed article on probabilistic price forecasting in the context of power trading. Although not strictly necessary, ensemble weather forecasts can contribute to the overall reliability of forecasts.

Ensembles work by running multiple simulations – known as ‘members’ – of a forecast, each with slightly varied initial conditions and model settings. Rather than returning a single forecast like deterministic models, ensembles generate a range of future weather scenarios, offering a more comprehensive view of uncertainty. If we again consider the rain shower scenario, an ensemble would indicate the broader area where rain is likely and the probability of it occurring at various locations within that area.

ECMWF Ensemble (ENS) runs 50 such members of perturbed forecasts, with each simulation equal in computational power to the high-resolution model – an incredible feat. Furthermore, ECMWF’s plans to stop seeing the high-resolution forecast as different from the ensemble clearly indicate the organization’s recognition that ensembles are the future of weather forecasting.

An advantage in this case is that the mean of the ensemble often filters out unnecessary detail present in both global and local high-resolution deterministic models, which can lead to improved error scores. Additionally, the range (or spread) of these 50 members helps predict extremes more accurately, making probabilistic forecasts better calibrated and more reliable.

The downside: the volume of data generated is enormous, and maintaining an archive spanning several years is complex. We’ve written a separate blog on the monumental challenge of weather data storage and the paradigm shift we employed to solve it.

AI ensemble models

Given the value of ensemble weather forecasting is becoming increasingly clear, a new generation of models is emerging: AI ensembles. They hold immense promise for better predicting extremes while retaining the established benefits of AI weather models.

Recent work indicates that AI ensembles might perform as well as the EC ENS product in terms of error scores. The result could be a combination of the fast run time of AI weather models with the probabilistic nature of ensembles – a highly exciting avenue.

AI models assimilating their own data

The next frontier for AI models – already in development – is the ability to assimilate their own data, separate from traditional NWP models.

Currently, all AI weather models in production start running from the 0th timestep – the ‘initial conditions’ – of a traditional NWP model. The NWP model constructs these conditions by taking its previous forecast of this timestep and using all the data it has to turn it into an ‘analysis.’ This is necessary because we don’t fully observe our atmosphere. We have sparse observations from weather stations, aircraft, satellites, and many more sources; however, we don’t measure, for example, the temperature at a height of 500m everywhere on Earth.

Initial strides in assimilating a model’s own data have been made with precipitation nowcasting, a spectator-friendly target we’ve been following with MetNet-3. Progress has accelerated over the past year, and we expect production systems to emerge soon.

The advantages here are two-fold. Firstly, forecasts can be available sooner than traditional models, with every minute making the forecast more valuable as it becomes less ‘stale.’ Secondly, AI weather models might be able to assimilate more data that is typically challenging to assimilate into traditional models, such as satellite imagery.

This advancement has the potential to blur the line between today’s rudimentary nowcasting methods and full weather models, making it particularly relevant for intraday energy markets.

Mixing models for precision forecasting

At Dexter Energy, we harness the strengths of various weather data sources to achieve the highest possible precision in our power and price forecasts. By combining traditional NWP and AI-driven models, we can maximize their unique advantages to create a more robust and accurate forecasting system.

In this process, we collect millions of data points from all relevant weather models. Moreover, we continuously benchmark models to identify ‘edges’ – specific scenarios and moments where certain inputs deliver the best performance.

We’ve achieved this mix because our teams have diverse skills in data science and advanced domain knowledge. This combination allows for a deep understanding of both statistics and the physical processes behind the data, from meteorological phenomena to wind turbine engineering.

Never a dull moment in weather forecasting

Weather models have come a long way, and we have good reasons to expect even more fascinating developments. Just last month, we saw the launch of a new high-resolution NWP ensemble model for a large part of Europe, which is a valuable new input for our forecasting pipelines. Also, the expected move towards foundation AI weather models has started with ECMWF announcing its WeatherGenerator project.

We conclude this overview with three key takeaways:

• The first wave of production AI weather models is already adding value by slightly improving error scores.
• Ensemble models, offering a probabilistic rather than deterministic approach, are the future and will become more critical in energy forecasting and markets.
• The true game-changer for AI weather models will be their ability to assimilate more real-time data faster than traditional NWP models.

At Dexter Energy, we’ll continue to efficiently incorporate the most accurate weather models into our forecasting mix and be ready to promptly leverage all future breakthroughs in this field.