Improved Wind Speed Prediction Results by Artificial Neural Network Method

Abstract

Accurate and reliable wind prediction is vital for sustainable wind power system. Especially in the atmospheric boundary layer, the difficulties of short-term wind forecasts affect the reliability of the model results. The forecast ability of the numerical weather models may be improved through artificial neural network (ANN), principle component analysis (PCA), genetic algorithm (GA), and other similar methods. In this study, the evaluation forecasts were made with the Weather Research and Forecasting/Advanced Research (WRF/ARW) model run with six different planetary boundary layer (PBL) parameterizations. The site of test station is located in the northern part of Istanbul with coordinates 41° 30′ N and 28° 66′ E at 51 m over sea level; it was found by Wind Atlas Analysis and Application Program (WASP) (Fig. 1 and Table 1). The performance of WRF/ARW for wind forecasting is assessed with measured wind variables at different hub heights at test station. The observed wind profiles are compared with WRF/ARW forecast, which uses the BL schemes based on turbulence kinetic energy. All the simulated schemes tend to underestimate or overestimate the wind at hub height during day and night. The diurnal evolution and the expected transitions of wind speed, temperature, and the alpha-parameter are evaluated by all the schemes.

In the proposed study, the weather research and forecasting model (WRF) is first run with six different physical conditions to find the appropriate variables of actual atmospheric condition during the experiment time period. Next, the study explored artificial neural network (ANN) methods to forecast the wind speed in Terkos (Durusu), Istanbul. To reduce the noises of the numerical weather prediction model, ANN is applied to forecast the short-term wind speed and is approved at the designed wind turbine farm of Terkos. Finally, the performance of the proposed approach is evaluated using observed data. The forecasting performance was improved by the ANN method.