min read

The future of power forecasting: how AI will help you staying ahead

Renewable energy power forecasting represents one of Jungle’s solutions that accelerate our mission to develop the world’s most effective tools to resolve underperformance stemming from operational complexity.

Problem Formulation

Over the past few months, we have been working on enhancing our power forecasting solution, based on state-of-the-art deep learning models, to reach the highest possible accuracy. In this article, we are excited to present to you what we have accomplished so far.

To better understand the problem of power forecasting, we will start by explaining the data we rely on to operate. Our models use Numerical Weather Predictions (NWP) as the main data input to predict the power generated by a renewable energy farm over a period of time. Additionally, there's the option to incorporate the farm's online power production data as input (i.e., power produced over the last hours), which can further improve the accuracy of our predictions. The only information we need to train the models is the farm’s location and the historical power data.

Figure 1: End-to-end model architecture. The input data of various weather features, in conjunction with online power production are transformed into power predictions, versatile for diverse applications.

The Importance of Multiple Forecasters

A challenge in power forecasting is understanding which NWP provider is most suitable. Which one proves most accurate? Or the most reliable? Maybe the one with the higher grid resolution? It is not trivial to choose the best NWP provider for a particular farm, and the answer may even vary over time, location, or application (i.e. solar or wind).

Our solution is simple: we do not choose. Instead, we feed the model with data from multiple NWP providers and let it decide which weather features warrant more attention. Despite its simplicity, this is only possible because our models are built based on state-of-the-art machine learning (ML) architectures that power today’s most powerful artificial intelligence (AI) applications. This enables them with the capability to discover complex patterns in the data that even domain experts might overlook, giving more attention to the most relevant features, while ignoring the noisy ones.

Another advantage of this approach is that we become much more resilient to missing data. We are now less vulnerable to possible accessibility issues, such as delays on the delivery of the forecasts, or even days of total unavailability. Furthermore, by using multiple forecasters, we can harness more weather features exclusive to some providers.

Figure 2: Our model is data resilient. Even if one entire NWP provider is missing, the model is capable of running and providing accurate power forecasts.

In terms of impact on performance, Figure 3 shows an example of the improvements over the addition of multiple weather forecasters. We can see that by adding one, and two more NWPs as input to our models, there is a 22.83% and 25.78% accuracy increase, compared to only using one NWP.

Leveraging the Farm’s Online Power

Consider a scenario where the farm has some kind of underperformance that is not weather related. How do our models adjust the predictions over an unexpected behaviour (e.g. downtimes, curtailments)? This is where the online power data comes in. The model learns how different behaviours impact the overall performance of the farm, by leveraging the information of the power that was produced over the last hours. That way, the output can be calibrated for unexpected events. Figure 3 illustrates that a model with the knowledge about the power produced over the last hours has a 3.56% accuracy improvement compared to the one without it.

Figure 3: Enhanced accuracy seen when using 2, 3 forecasters, and 3 forecasters plus the online power. Improvement is reported against Jungle’s predictions using only 1 forecaster.

As you can see from Figure 4, it becomes quite clear how impactful this improvement is during times of underperformance. This snippet shows the power forecasted 24 hours ahead, during a time where there was a period of under production. We can analyse that the forecasted values are higher without the help of online power, since it has no information about the current underproduction of the farm.

Figure 4: Snippet of the time-series from our solar case study. The blue line corresponds to the ground truth, and the red and green lines correspond to the model’s forecasts with and without the use of online power, respectively.

Tailored Models

We wanted to ensure that our forecasting solution could be tailored to best fit each customer needs and specific requirements:

  • Resolution: We can forecast power at various time intervals (e.g., 10 minutes, 30 minutes, 1 hour, etc.)
  • Forecast Horizon: Some customers may need the power forecast for the current day. Others may need to have the predictions for the next few days. Our product can be customised to predict any preferred horizon.
  • Output Flexibility: We enable the capability of providing our customers with point-wise predictions, as well as probabilistic bands (we will cover this in a future blog post, stay tuned!).
  • Performance Metrics: Different energy markets may penalise wrong power forecasts under different schemes. Our forecasting solution is able to optimise to any metric, MAE or NMAE for instance, that best describe the penalisation scheme.
  • Missing Data: Are there missing values in the historical power data? With the flexibility of our new ML models, we can feed the data as is, without any significant accuracy impact.


Jungle is actively developing leading-edge deep learning models to tackle power forecasting challenges in both wind and solar farms. Our solution stands out by using state-of-the-art architectures that allows us to get the best possible results, while focusing on having a modular product that can be adjusted to a broad spectrum of needs. The only requirements are the farm’s location and the historical production data; we handle the rest for you!

Manuel Santos

Machine Learning Engineer

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.

Back to blogs