Wind power forecasting based on WRF and ANN enhanced with Statistical Capabilities

Introduction

The significance of wind power is at its peak with ever increasing energy demand and environmental concerns associated with fossil fuel based power generation. Integrating the available wind power capacity to the grid is a challenging task because of its intermittent nature. Wind power penetration to the grid can be improved and grid stability can be attained by enforcing accurate wind power forecasting. Forecasting gives an estimate of wind power that can be generated at any given instance, with which power generation at conventional power stations can be properly planned and scheduled. Among different forecasting techniques available, hybrid forecasting model, combing different forecasting methods, outperforms any individual model based on Numerical Weather Prediction (NWP) methods or statistical approaches or Artificial Neural Network (ANN) techniques.

Approach

Numerical Weather Prediction based Weather Research and Forecasting (WRF) model, statistical method based Auto Regressive Integrated Moving Average (ARIMA) model and Computational Fluid Dynamics (CFD) along with Artificial Neural Network (ANN) technique are employed to develop a hybrid model to overcome the limitations of individual models.

Main body of abstract

Physical conditions of the atmosphere can be clearly represented with the help of WRF. Global Forecast data from different global models are used to specify the boundary conditions for a limited area of interest using nesting. Site specific data from turbine SCADA and weather monitoring stations are incorporated in WRF to specify the local atmospheric and physical conditions with greater accuracy. Forecasted wind speed output of WRF is fed into CFD based ANN solution and mesoscale wind speed output of WRF is interpolated to microscale level using its built-in down-scaling technique. Forecast tool, using its ANN feature, train the model with the actual wind speed obtained at site and forecasted wind speed obtained from WRF and calibrates the model accordingly to adjust itself to the error between forecasted and actual wind speed. CFD along with forecast tool is used to forecast the wind power at site level. Univariate statistical model based ARIMA is used in the final stage to improve the wind power accuracy by training the model with historical data to follow the trend or pattern associated with wind power generation. Performance of the developed wind power forecasting system is tested at three different sites in India, each having different levels of terrain complexities. Absolute Percentage Error (APE) is used as evaluation criteria to validate the result with actual data. Validation is done by calculating APE for forecasted wind power with respect to the available capacity. For a period of six months with 15 minutes time stamp, the validation result shows, in two out of three sites, 86% and 84% of time the results were within the forecasting error band of 15% and in third site it was around 70%. On investigation, curtailment (de-rating) due to grid constraints was found to be the reason for a low 70% here. When the curtailment affected time stamps were removed from calculation, the percentage of time the model giving results less than 15% error got improved to 81% from 70%.

Conclusion

Developed system proves to perform better in all the three test sites. Thus this hybrid system outperforms the individual model as physical conditions are well represented with WRF in first stage, phase and bias error associated with mesoscale modeling is reduced and accuracy of forecast at site level is improved with CFD based ANN solution in second stage and the performance of forecasting system is tuned in accordance with the historical trend using statistical ARIMA in final stage.


Learning objectives
This work exhibits how hybrid model improves the accuracy of wind power forecasting over individual models and also examines the suitability of single hybrid forecast system for different terrain conditions.