Reduction of Wind Speed Forecast Error in Costa Rica Tejona Wind Farm with Artificial Intelligence

Abstract

The energy sector relies on numerical model output forecasts for operational purposes on a short-term scale, up to 10 days ahead. Reducing model errors is crucial, particularly given that coarse resolution models often fail to account for complex topography, such as that found in Costa Rica. Local circulations affect wind conditions at the level of wind turbines, thereby impacting wind energy production. This work addresses a specific need of the Costa Rican Institute of Electricity (ICE) as a public service provider for the energy sector. The developed methodology and implemented product in this study serves as a proof of concept that could be replicated by WMO members. It demonstrates a product for wind speed forecasting at wind power plants by employing a novel strategy for model input selection based on large-scale indicators leveraging artificial intelligence-based forecasting methods. The product is developed and implemented based on the full-value chain framework for weather, water, and climate services for the energy sector introduced by the WMO. The results indicate a reduction in the wind forecast RMSE by approximately 55% compared to the GFS grid values. The conclusion is that combining coarse model outputs with regional climatological knowledge through AI-based downscaling models is an effective approach for obtaining reliable local short-term wind forecasts up to 10 days ahead.

Keywords

• wind forecasts
• model error reduction
• artificial intelligence