6th International Conference Energy & Meteorology: Abstract Submission

Development and implementation of wind and solar generation forecasts for system operations in Colombia (774)

Sebastián Ortega 1 , Maria Camila Angel 1 , Lina Marcela Ramirez Arbelaez 1 , Juan Carlos Morales 1
  1. XM, Medellin, ANTIOQUIA, Colombia

Objective and Background

The Colombian power system is characterized by a large hydrological component, more than 60% of its generation is due to large hydropower plants which regulate their inflows and follow generation schedules by managing their reservoirs. In turn, this ability translates into an operation of the power system with a relatively low uncertainty in horizons from hours to weeks.

In the near future, a large integration of renewable energies is expected in the Colombian power system. According to UPME, Colombia┬┤s energy and mining planning unit, close to 35% of the installed capacity of the country will be due to wind and solar plants by 2023. This integration will translate into a power system that at times might reach 100% of its generation due to renewables, but this integration is also expected to translate into a higher uncertainty for its operation.

As the operators of the Colombian power system, we have been preparing in XM for this integration. Since 2017 we have been building our internal capabilities in operating variable renewable energy and have put forward regulatory proposals to best integrate them into the system. As a fundamental component of these efforts, we have developed a prototype of a system capable of forecasting solar and wind generation, a system which continues to be developed during 2019, and which will be used for planning and operating the system in the presence of renewables.

Methods

During the prototype, we developed our own forecasting methods and received forecasts from different providers. We developed and validated statistical and dynamical models to find the more suitable ones for each plant. The statistical models included linear regressions, neural networks and support vector machines. The dynamical model included statistical post-processing techniques (MOS) of the Global Forecast System (GFS). Additionally, we developed multi-model combination approaches such as linear regressions, Bayesian averaging and machine learning techniques to improve upon these forecasts.

Results, conclusion and discussion

The prototype prepared the operation of the Colombian power system to receive and generate the renewable energy forecasts required to maintain a safe and reliable operation of its power system. The experiences gained during the project allowed for a clear picture of the advantages and limitations of renewable forecasts, as well as the requirements to implement them in an operative manner.

Regarding the deviations expected from the forecasts, results showed that multi-model ensembles had a higher skill for wind forecast. The minimum mean absolute errors (MAE) for the 3-hour ahead wind forecasts were 1.5 MWh (7.5% of installed capacity), while for the week ahead horizon were 1.8 MWh (9% of capacity). For solar generation forecasts, dynamical GFS-based models showed better results than other approaches. The minimum 3-hour ahead MAE was 1 MWh (10% of installed capacity), while for day to week ahead horizons the MAE increased to 1.3 MWh (13% of capacity).