Self-adaptive Wind Speed Forecasting Method for New Wind Farm

Introduction

As a green and renewable energy resource, the utilization of wind energy has been growing rapidly all over the world. However, the stochastic and uncontrollable characteristics of wind can reflect on wind power, which would affect the safety and stability of power grid. Therefore, the generated power of wind farm has to be accurately forecasted. Especially, accurate short-term wind power forecasts with a prediction horizon from one hour to several days, which is short-term prediction, are crucial for optimizing the scheduling of wind farm maintenance and electricity reserves.
Many approaches have been proposed for short-term wind power forecasting, which can be categorized as statistical methods/physical methods. However, most of the existing methods need several months’ historical data for model training, which is a major obstacle for newly built wind farms: there is no or little historical operational data collected yet. To solve this problem, a self-adaptive method with sequential learning ability based on Gaussian Processes is proposed. This kind of method could operate with very limited or no historical information at the beginning of a new built wind farm.


Approach

The whole self-adaptive wind speed forecasting method for new wind farms based on on-line learning Gaussian process, named as O-GP model, includes several steps: 1) Appropriate method for choosing dataset from other long-term run wind farm is designed, to provide proper data for initial model training; 2) Gaussian process is adopted to build wind speed forecasting model – Gaussian process is a stable regression method, which needs small training set; 3) New operational data of real wind speed, wind power and NWP data from new built farm is obtained each day, and the training set is updated by new data to update the forecasting model accordingly.

Main body of abstract

In this paper, firstly the whole modeling process of wind speed forecasting for new wind farm is described, including the 3 steps as mentioned above. Secondly the mathematical methods are introduced, which are 1) dataset chosen method based on distance calculation to maximum the effectiveness of existing data; 2) the on line learning method to adjust Gaussian processes for better and faster fitting the characteristics of target wind farm as time goes. Thirdly, details on experiments are given, covering the dataset information, case study, experimental results and the corresponding analysis, etc.
Two years real-world data from a wind farm in China is used to validate the model performance, and the experimental results proved that the proposed method can achieve higher accuracy for wind speed forecasting, with about 20% improvement comparing to the classical persistent method, 15% improvement comparing to normal GP method.


Conclusion

Accurate wind speed forecasting can still be achieved with absence of local historical data, therefore ensures the stability of grid, even for integration of newly built wind farms.


Learning objectives
Accurate wind forecasting is possible using less historical wind data, therefore can decrease the dependency of forecasting methods on measurement condition.