Improving Near Real-Time Predictions of Renewable Electricity Production at Substation Level
Authors/Creators
- 1. Institute for Molecules and Materials, Radboud University Nijmegen, The Netherlands
- 2. Institute for Computing and Information Sciences, Radboud University Nijmegen, The Netherlands
- 3. Alliander N.V., P.O. Box 50, 6920 AB Duiven, the Netherlands
- 4. Department of Environmental Science, Radboud Institute for Biological and Environmental Sciences, Radboud University, Nijmegen, The Netherlands
Description
With the recent urgency of the transition to a low-carbon economy, the monitoring and prediction of regional renewable energy generation over time have become increasingly important. The difficulties of renewable energy data transfer between multiple stakeholders have also caused elevated the need for more trustworthy data analytics within the energy grid. However, only a few voltage transformation facilities in the grid (i.e. substations) contain complete information about the renewable energy generated within the region. A large number of incomplete-information substations with fully-missing renewable energy data limits the analysis and policy-making related to renewable energy. This work studies the potential to transfer information from perfect-information substations to incomplete-information substations with fully-missing renewable energy data. To preserve the practicality of renewable energy prediction, a domain adaptation for zero-shot learning in sequence (DAZLS) strategy is proposed for fully-missing renewable energy prediction in substations. DAZLS is a model agnostic technique which can utilize any base model within its framework for the prediction of renewable energy generation. Using total measured power and weather information (solar irradiation and wind speed) in information-complete substations in 10 locations within the Netherlands, we developed a model to predict solar and wind power from energy producers associated with information incomplete substations via additional real-time weather data, metadata information (e.g. geospatial position, existence and capacity of renewable facilities). Using DAZLS, the average root-mean-squared error for prediction (RMSEP) is 0.07, while that of a default transfer learning model is 0.70. This meant that renewable energy sources in information-incomplete substations could be reliably monitored using weather data, meta-data and physical data, resulting in lesser investment in power meters. This approach was demonstrated for the highest frequency prediction possible in the grid, with a near-real-time frequency of 15 minutes.