Artificial neural network and kalman filter approaches based on arima for daily wind speed forecasting

• Main Author: Shukur, Osamah Basheer
• Format: Thesis
• Language: English
• Published: 2015
• Subjects: QA Mathematics
• Online Access: http://eprints.utm.my/id/eprint/54833/1/OsamahBasheerShukurPFS2015.pdf; http://eprints.utm.my/id/eprint/54833/
• Institution: Universiti Teknologi Malaysia

The wind speed forecasting is important to observe the wind behaviour in the future and control the harms caused by high or slow speeds. Daily wind speed is more consistent and reliable than other time scales by providing vast monitoring and effective planning. Although a linear autoregressive integrated moving average (ARIMA) model has been used for wind speed forecasting in many recent studies, but the model is unable to identify the nonlinear pattern of wind speed data. ARIMA modelling process causes the stochastic uncertainty as a second reason of inaccurate forecasting results. Wind speed data collection process faces several problems such as the failure of data observing devices or other casual problems that lead losing parts of data. Therefore, wind speed data naturally contains missing values. In this study, an ARIMA-artificial neural network (ANN) and ARIMA-Kalman filter (KF) methods are proposed to improve wind speed forecasting by handling the nonlinearity and the uncertainty respectively. A new hybrid KF-ANN method based on the ARIMA model improves the accuracy of wind speed forecasting by rectifying both nonlinearity and uncertainty jointly. These proposed methods are compared with others such as AR-ANN, AR-KF, and Zhang’s method. AR-ANN method is also used to impute the missing values. It is capable to overcome the missing values problem in wind speed data with nonlinear characteristic. It is compared with linear, nearest neighbour, and state space methods. Two different daily wind speed data from Iraq and Malaysia have been used as case studies. The forecasting results of the ARIMA-ANN, ARIMA-KF and the new hybrid KF-ANN methods have shown in better forecasting than other compared methods, while AR-KF and AR-ANN methods provided acceptable forecasts compared to ARIMA model. The ARIMAANN and the new hybrid KF-ANN methods outperformed all other methods. The comparison of missing values imputation methods has shown that AR-ANN outperformed the others. In conclusion, the ARIMA-ANN and the new hybrid KFANN can be used to forecast wind speed data with nonlinearity and uncertainty characteristics more accurately. The imputation method AR-ANN can be used to impute the missing values accurately in wind speed data with nonlinear characteristic.