Client: Covolt
Industry: Solar energy
Process: Asset management
Covolt B.V. is a company in the renewable energy sector, where they employ their intelligent energy management system. This system optimizes the production of solar parks and automatically offers the energy to the market. This results in a substantial improvement in both efficiency and reliability of these parks.
As part of their energy management solution, Covolt aims to predict future energy production for one day-ahead and intraday periods. This predictive capability would enable them to refine their supply bids on the market, as the effectiveness of energy trading heavily relies on accurate predictions of the production. Within this framework, the primary challenges lie in training forecasting models with limited historical data (± 2 years) and generating predictions for entirely new locations that lack any historical data.
Bright Cape has trained multiple machine learning models to forecast energy production, utilizing both internal and external data sources. Since weather data was indicated as the primary forecasting driver, a comparison study of various external weather sources was conducted to select the most accurate weather APIs. Subsequently, Bright Cape assisted in creating the foundational data infrastructure to retrieve and integrate these diverse data sources.
Next, this infrastructure was extended by setting up dedicated training, validation, and forecasting pipelines to ensure continuous operability on the Microsoft Azure cloud platform. These pipelines were designed with modularity in mind, facilitating the simultaneous development, evaluation, and deployment of both one day-ahead and intraday forecasting models. Furthermore, a roll-out strategy was implemented to enable the scalability of the models across locations throughout the Netherlands, including entirely new sites without sufficient available data.
Feature importance techniques were used to assess the significance of each variable and to determine the most significant predictors, including factors like irradiation and time of the year. These techniques provide valuable insights for data scientists and business users into what the model has learned from the data and improves the model’s transparency.
Bright Cape assisted in the design, development, and deployment of the forecasting pipelines and the underlying data infrastructure. The trained models significantly outperformed both the baseline market prediction strategies for the pilot locations. Forecasting errors were cut in halve, where hourly day-ahead forecasting errors were reduced from 3.3% to 1.5%, and 15-minute intraday forecasting errors improving from 4.7% to 2.3%. Consequently, this led to more effective energy trading on the market and increased profit margins. Additionally, the project yielded a return on investment in under a year as it enabled the acquisition of new clientele.
as the project enabled the acquisition of new clientele
reduced hourly day-ahead baseline forecasting error
reduced 15-min intra-day baseline forecasting error
We are here to help you.
