The development of artificial intelligence models to forecast renewable energy generation faces a less visible but critical challenge: validating input data and ensuring successful model training. In many cases, plant operators and asset managers provide production data with errors, gaps, time shifts, or corrupted records. Data quality and filtering are fundamental to achieving accurate forecasts.

The Silent Challenge: Reliable Data Validation and Model Training

Currently, this validation process is performed manually. Analysts examine internal correlations, relationships with meteorological variables, and other indicators to assess whether the data is suitable for model training. This process is time-consuming and requires deep technical expertise.

Let’s take an example. A solar plant submits its historical production data to train a predictive model. Upon reviewing the data, the following issues are found:

• For several days, production appears as zero despite sunny conditions (indicating a measurement error or export issue).

• At other times, there are unrealistic production spikes far exceeding the plant’s capacity.

• Some dates are out of order or duplicated.

Today, such issues are detected by a technician manually reviewing graphs, calculating correlations with forecasted solar irradiance, and assessing data validity. With AI and an automated workflow, a model could learn to recognize these errors and label files as valid or invalid within seconds.

Even after validation, training deep learning models can still fail. Factors such as poor parameter initialization, low data quality, or imbalanced variables may prevent the model from learning correctly, requiring the process to be repeated.

This double bottleneck, data validation and training supervision, limits the scalability of the services provided by Ravenwits. Since each plant requires a model tailored to its specific characteristics, it becomes urgent to develop automatic ways to detect and fix issues before they reach the end user.

The Challenge of Efficient Large-Scale Operations

Scalability is a common challenge for many tech startups. To maintain efficiency in delivering services, it is essential to decouple the growth of clients and assets from the growth of the human team.

In Spain alone, there are around 1,400 wind farms and an even greater number of solar installations. At the European and global scale, these figures multiply by hundreds. Naturally, these numbers continue to rise as the energy transition progresses. Each installation has unique conditions, data, and requires an adapted model.

Providing services to thousands of assets without relying on a massive human workforce requires robust systems for automatic data and model quality control. These systems not only improve efficiency but also reduce errors, standardize the service, and free up resources for innovation.

AI for AI: Automating Validation and Training

To tackle this challenge, Ravenwits is developing, together with the Community of Madrid and its 2024 Subsidy Program for AI Applications in Industry, a computing platform designed to automate the entire workflow, from data ingestion to validation and model training.

This platform will include:

• A storage system for meteorological and production data.

• An AI module based on decision trees and multilayer perceptrons (MLPs) to automatically validate the quality of input data.

• A module to train predictive models (CNNs, GNNs, etc.).

• A second AI module that evaluates whether the model has been trained correctly, estimating the likelihood of failure.

• A simple interface where operators can visualize key indicators and make informed decisions.

Initially, the system will be trained using real-world cases previously handled by the team, with the goal of functioning as an intelligent assistant. In the medium term, it will retrain itself automatically to become a fully autonomous tool.

Renewable energy forecasting requires more than just strong models, it demands a robust, scalable, and reliable ecosystem for data validation and model training. Automating these critical tasks will enable service providers to scale without sacrificing control or quality.