Aphelion Wind: A Decision-Support Tool for Wind Farm Operators

1. Introduction

Aphelion Wind is a decision-support tool tailored for wind farm operators to leverage their decisions based on weather and energy forecasts obtained using artificial intelligence based forecasting models. This abstract describes the main features of this tool. Further insights are provided in terms of model validation as well as the main challenges faced during the development of this tool.

2. Aphelion Wind

From an user perspective, Aphelion Wind is a web-based application used for asset management and for visualizing both weather and energy forecasts, as shown in Fig. 1. Confidence intervals are built to provide a range of values where the forecast will fall in a certain time step in the future.

The main business cases are automated using an in-house tool which has three key components: a) regionalization of numerical weather prediction (NWP) data and feature extraction, b) automated model training, and c) model deployment. The whole work cycle of a ML model applied for energy forecasting is automated with this tool (Fig. 2).

The system is triggered when historical data from a target (e.g., power production data) is inserted. First, features in the form of weather forecasts are required to train an AI/ML model in order to find relationships between those weather forecasts and the target. NWP models provide those weather forecasts by solving a set of equations based on mathematical models of the atmosphere. Thus, relevant features from any available NWP model are extracted for any particular asset in the time period where the historical target data is available. This will allow to train the AI/ML models available in the system and subsequently evaluate them in the testing period. Based on the performance evaluation metrics, the model or set of models (depending on the requirements of the client) with best scores will be deployed to provide forecasts operatively. The tool allows to update these forecasts by extracting the features of every new run of the available NWP models.

3. Model Validation.

Currently, the automated tool includes a set of AI/ML models which have been developed for every main target to be forecast, such as wind power, wind speed, and wind direction.

One of the most important aspects to ensure great product quality is to find the most suitable models for a given asset, as well as providing metrics with realistic scores to meet the expectations of the service. To achieve that, a thorough performance evaluation process is carried out following the good practices developed by the IEA Wind Task 36 (Möhrlen et al., 2022). A critical point to validate the models is the amount of data available. Ideally, at least 2-3 years of the target data should be available for training and validation, using at least 1 year for testing, so any existing seasonal effects can be captured in the evaluation process.

In terms of performance evaluation metrics, we use Normalized Mean Absolute Error (NMAE) as a metric where the scores are normalized by the installed capacity of the wind farm (González-Sopeña et al., 2021). In other instances, such as wind speed forecasts, we use MAE to evaluate model performance. For example, Fig. 3 shows the validation considering forecasts up to 24 hours considering the AI/ML models implemented in our automated tool. In the upper subfigure, potential refers to power production where effects such as curtailment are not considered.

4. Challenges

While AI based decision-support applications become more and more refined and are able to provide a great benefit to wind farm operators, there are still many challenges yet to be addressed.

One challenge we have identified is the lack of awareness on how such weather and energy forecasts can improve the decision-making process faced by wind farm operators. A clear example is near-future maintenance operation scheduling for wind turbines. When wind speed forecasts (ideally refined with a ML model to reduce forecast errors) are available up to 1 week ahead, we can optimize the scheduling of certain O&M activities within that week, thus leading to reducing O&M costs of the asset.

From a technical point of view, there are at least two main challenges to be taken into consideration when developing a weather forecast and decision-support tool with a strong focus on AI. First, the lack of open-source wind farm data to validate any AI/ML forecasting model. It is difficult to persuade a potential client we can provide a state-of-the-art solution without any previous validation of our AI/ML models. This is gradually improving and more and more data are available for model development (Effenberger and Ludwig, 2022), but there is still a long way to go before being able to validate energy forecasting models. The second challenge, often overlooked in the forecasting community, is the importance of having a well-oiled machine that covers the whole lifecycle of ML models. In general, authors try to reduce the forecast error by developing very complex models, which might not be as useful from an industrial point of view as their deployment can become cumbersome. On another note, we might disregard how difficult can become the automation and operationalization of ML products without using Machine Learning Operations (MLOps) (Kreuzberger et al., 2023), that is, the best practices set up in the software engineering community applied to ML to cover adequately the lifecycle of a ML based product. MLOps cover aspects such as versioning for reproducibility and traceability, continuous ML training and evaluation (periodic retraining with new feature data), and continuous monitoring not only of data, models, and code but infrastructure and model deployment performance to detect any errors or changes which could hamper model performance.

5. Conclusions

This abstract presents the main features of Aphelion Wind, a decision-support tool tailored for wind farm operators to leverage their decision-making process using AI/ML models able to provide weather and energy forecasts. We emphasize the importance of correctly validating this models to ensure the product quality, and we point out some social and technical challenges we have come across during the development of Aphelion Wind.