Wind power data analytics - Simulation, prediction, and statistics

[Abstract] As the renewable and green energy, wind power has been popular and recently developed countries like Europe and North American have to focus on the development of the wind power in order to relieve the stress from increasing demanding in fossil fuels. Nonetheless, the wind power as one of the natural resources instinctively set barriers towards its prosperity in the developments. The wind power, if developed, will confront the challenges mainly from three points. Firstly, wind power as energy is non-dispatchable. Secondly, unlike traditional fuels, wind power highly depends on geographic and meteorological condition. Thirdly, the wind power down ramp, if happens, will adversely influence reliable wind power generation. Among these challenges, the third one may bring the most hazards to reliable wind power industries. Therefore, my research is around on how to supply reliable wind power generation. As to the research on the supplies of reliable wind power generation, it is necessary to figure out why wind power down ramp causes unreliable wind power generation. The reason is that the wind power down ramp disrupts the backup the wind power generation prepares as the reserve to compensate small variation in the other wind power generation. Since it is clear that the wind power down ramp causes unreliable wind power generation, the next research must be conducted on the cause of wind power down ramp. Instinctively, the wind power down ramp is caused by the induced front or drastic wind change. Therefore, the research should be a focus on how to mitigate hazards the wind power down ramp brings. In order to effectively mitigate hazards, the research should be conducted on the precaution of wind power down ramp. In other word, the question of when the wind power down ramp occurs is critical. Hence, the entire of my doctoral researches is around on critical techniques to when the wind power down ramp occurs. As to the question of when the wind power down ramp occurs, I convert the question to the prediction of the wind power down ramp. In order to effectively predict the wind power down ramp, I have taken considerations from a different perspective. Those perspectives are mainly from the meteorology, statistics and machine learning in order to provide the detection of the wind power down ramp more geographic and meteorological information. First of all, as the wind power down ramp highly depends on geographic and meteorological conditions, with the purposes of making more informative in the case study of wind ramp on the date of May.28th, 2015, I did the simulation by the WRF model (Weather Research Forecasting). The WRF model is based on the theorem of the NWP (Numerical Weather Prediction). The NWP takes advantage of physical law as governing equations to govern the meteorological factors like atmospheric pressure, radiation, temperature, and air flow. In addition, the NWP utilizes mathematics as methodologies to compare different pattern governed by physical laws, so-called physical schemes. Based on the NWP, The WRF model implements weather forecasting, not only global wide but also regional size. According to the case study of wind ramp on May.28th, 2015, I used parameterization to implement the WRF simulations. The implementation contains: drawing the WRF model’s domains mapped with West Texas where the case study of wind ramp happened, parameterizing both chronical and geographical data on the WRF model, and parameterizing different physical schemes like YSU and MYJ. Associated with the case study of the wind ramp, the WRF simulations have proven that the YSU outperforms in the global meteorological study and that the MYJ performs well in regional meteorological researches. Therefore, the conclusions are drawn in my research on the wind ramp that MYJ as a physical scheme in the WRF model might be better in the wind ramp research and that high resolution should be adopted in the WRF simulation of future wind ramp research, due to its mathematical property. Secondly, for the research on the detection of the wind ramp, according to the case study of wind ramp on May.28th, 2015, I utilized observation data from West Texas Mesonet and programmed the observation data for the visualization of observation tracking. In the visualization of the observation data, a dynamic procession of the wind ramp data serves to grab geographic tendency of wind ramp events. By the implementation, the core design is to map a palette table with a different color to wind speed at a site with different intensities. As results, it turns out that the observation tracking with the case study of May.28th’s wind ramp enables to offer the geographic tendency of the case study in the wind ramp, from northwest to southeast orientation in the map of U.S. and the conclusion as research results bring significance to the future research on regional wind ramp detection. Thirdly, since there are geographical tendencies from northwest to southeast orientation in West Texas when the wind ramp happened on May.28th, 2015, it is easier to make the hypothesis that the wind speeds at sites on West Texas Mesonet have correlation during the wind ramp. In order to verify the hypothesis, the research is conducted and ten sites’ observation data are taken samples, due to the case study of wind ramp on May.28th. 2015. For implementation, the correlation coefficient is calculated among any two sites of wind speeds during the date of May.28th, 2015 when wind ramp happened and the correlation coefficient among any two sites are plotted. By the plots of there are linear among any two of ten sites in the wind speeds and thus liner relation indicates high correlation among sites. Therefore, the conclusions are drawn that wind speeds of sites in West Texas Mesonet have geographic correlation when the wind ramp happens and the conclusion as research results bring significance to the future research on regional wind ramp detection. Fourthly, I proposed adapted supervised logistic classifier as a predictive model for the detection of the wind ramp. For optimization, compare with different methodologies in the art of status in Computer Science and Statistics, the supervised classifier with the trait as supervised outperforms to the tools in traditional statistics for the detection of the wind ramp. Furthermore, among different measurements in Machine Learning, supervised logistic classifier is optimized as well because the supervised logistic classifier has characteristics of cross-entropy as loss log function and gradient descent to find the minimum distance between the model and the factual data. Whereas, the use of supervised logistic classifier has a challenge. The challenges are that input data with the time serials should keep it unchanged for the predictive output and that the logistic classifier requires the input data to be changed as assorted for computational efficiency. Faced with challenges, I overcame the limitation that the use of logistic classifier brings to ask the input data as assorted which destroyed the input data with the feature as time serials and proposed adapted supervised logistic classifier for the detection of the wind ramp. Due to the case study of the wind ramp on the date of May.28th, 2015, I utilized the observation data at sites. After trained and tested models, the plots indicate a common rise up fitting lines in the middle between May,28th, and May,29th, 2015. According to ERCoT’s records, as there is wind ramp event in the early morning of May.29th, 2015, the rise up fitting lines illustrated as drastic changes of wind speed at a site, have proven as the wind ramp event. The conclusion is drawn that adapted supervised logistic classifier is capable to detect wind ramp. Fifthly, as my research turned from the wind ramp detection to the large wind power ramp detection, the research has been statistically considered the wind power down ramp as rare events. From this, extreme value theory is applied and thus Pareto distribution has been considered for the detection of wind power down ramp. Because the Pareto distribution has heavy tail distribution, empirical data by the excess meaning plot are used to draw a line and, if one terminal of such a line exhibits as the heavy tail, the slope and interception of the line are useful to estimate the location, shape, and scale parameters in the Pareto distribution. As three parameters in the Pareto distribution are decided, the Pareto distribution is determined to predict the wind power down ramp. By the Pareto distribution, the quantile value at risk could be decided. After verification, the hypothesis has proven. The conclusion is drawn that the extreme value theory is useful for the detection of the wind power down ramp. Sixthly, space dependency and Markov Random field are proposed for the detection of the wind power down ramp. In terms of power generation in 2011, more than 200 power generations are tracked and found that, if those 200 wind power generations are trained into the model, the model cannot detect the wind power down ramp. Even though Frechet distribution is capable of the wind power down ramp, the likelihood for the wind power down ramp is still very low. From this, space dependency is come up with and the related indication is that neighbor turbines in a wind farm most likely have space dependency. In order to verify the space dependency among neighbor turbines, the Markov Random Field are applied. Moreover, if there is the space dependency among the neighbor turbines, the Bayesian posterior distribution is used in order to build the relation between space dependency among the neighbor turbines and the indicators of the detection of the wind power down ramp. Furthermore, the linear classification is used to partition three regions: up-stream, ramping, and downstream regions. Those three regions are used for performance evaluations. The conclusions are drawn that the Markov Random Field and space dependency among the neighbor turbines in a wind farm enables to detect large wind power down ramp and that only up-stream is useful as next-times input data for the detection of the wind power ramp after compared with other measurements like AR (Auto Regression).