Short-Term Wind Speed Forecasting Based On Non-stationary Time Series Analysis

Wind speed forecasting is very important in the planning and designing of the wind electric farm. But wind speed time series is non-stationary, and has time-varying variance. Therefore, wind speed is often considered as one of the most difficult meteorological parameters to forecast. Wind speed in itself has a timing sequence and auto-correlation, so ARIMA model can be set up to predict it. But it exists the problem of delay and low multi-step prediction accuracy. To solve this problem, this paper uses a hybrid algorithm based on wavelet transform and ARIMA model. To some extent, it achieves good results. However, it’s still not ideal.In order to further improve the prediction accuracy, the paper considers the characteristics of the wind speed series. Then, a hybrid model combined wavelet transform with ARIMA model, as well as ARCH model, is given. First, wind speed series is decomposed and reconstructed into approximate series and detailed series by wavelet analysis. Then use ARIMA model to analyze each part, simultaneously considering the heteroscedasticity effect of the residual series, the corresponding ARIMA-ARCH model is set up. The final forecasting wind speed values are the sum of the predicted approximate and detailed values.The traditional ARIMA model, wavelet ARIMA model, ARIMA-ARCH model and wavelet ARIMA-ARCH model are respectively established to the sample wind speed series. And the analysis of the prediction errors of each model shows that the wavelet ARIMA-ARCH model which takes the heteroscedasticity effect of the wind speed sequence into account can effectively improve the accuracy of wind speed forecasting.