CHAPTER 6 CONCLUSION AND FUTURE SCOPE 146 CHAPTER 6 CONCLUSION AND FUTURE SCOPE 6.1 SUMMARY The first chapter of the thesis highlighted the need of accurate wind forecasting models in order to transform a wind farm into a wind power plant. A detailed review of the forecasting models developed till now for wind speed, power curve and wind power was presented in the second chapter. The challenges that are yet to be addressed in the area of wind power forecasting were identified and three major objectives for this research work were defined. Linear and non-linear models have been developed for forecasting of wind speed and power. For all the non-linear models developed in this research work, the performance of seven or eight data mining algorithms was tested for prediction. The best performing three and four algorithms have been considered for tabulation and comparison. In this research work, the best performing models or algorithms are ascertained based on the MAE and RMSE error metrics, when compared to the other models or algorithms chosen. Development of Wind Speed Forecasting Models The first objective of the research work was to develop accurate multistep wind speed forecasting models, for different time horizons, incorporating the impacts of meteorological variables and annual trends. This has been carried out and the methodology and results are presented in Chapter 3. Two different models
CHAPTER 6 CONCLUSION AND FUTURE SCOPE 147 were developed for wind speed forecasting: Models for hourly forecasting and 10-min ahead forecasting. Performance of Hourly Forecasting Wind Speed Models The hourly forecasting model is realized using meteorological data and time-series data. The meteorological data included atmospheric parameters like wind direction, air temperature, relative humidity, solar radiation and atmospheric pressure. The model was developed on one year hourly data of Sulur region and tested for the next year s data. The time series wind speed data was used to develop NAR models with and without exogenous variables. Four different data mining algorithms were used to develop the meteorological model, namely Bagging, RBFNetwork, MLP and M5P. Among the models based on meteorological data, model developed using bagging algorithm performed the better in terms of MAE and RMSE (MAE = 3.4184, RMSE=4.7165). Among the models developed using time-series data, NARX1 model solved by M5P algorithm registers the better performance (MAE = 2.2514, RMSE=3.2207), which is followed by the NARX1 model solved by bagging algorithm (MAE = 2.2581; RMSE =3.2338). On comparison, it can be observed that the multivariate time-series model developed with temperature, wind direction and relative humidity as exogenous variables outperform all models. Bagging and M5P algorithm perform better when operating on large datasets. Panov and Dzeroski (2007) state that the greatest strength of Bagging algorithm is that it uses random subsamples of the training data and hence larger the dataset, better is the accuracy. The M5P algorithm uses a decision tree algorithm to build a tree and hence has the benefit of randomizing the algorithm for learning base-level classifiers. The M5P combines a conventional decision tree with linear
CHAPTER 6 CONCLUSION AND FUTURE SCOPE 148 regression functions at the nodes. Hence M5P generates models that are compact and relatively comprehensible. Performance of 10-min Ahead Wind Speed Forecasting Models Wind speed forecasting models for 10-min ahead forecasting were developed using linear and nonlinear ARMAX models. The exogenous variables used include wind direction, wind speed at different heights, solar radiation, temperature and wind speed in the same place in the previous two years. The models have been developed on two different datasets. The non-linear models were developed based on three data mining algorithms SMOreg, bagging and M5R. The performance was determined based on multi-step prediction for the next six consecutive steps. For dataset 2, according to both MAE and RMSE measure, SMOreg algorithm based ARMAX1 (2,3,2) model performs better and is followed by SMOreg algorithm based ARMAX3 (5,1,1). The standard deviation values of the MAE and RMSE over the six consecutive time steps for the first model are 0.187 and 0.249 against 0.188 and 0.249 of the second model. ARMAX models based on wind direction as the only exogenous variable perform better than all the other ARMAX models. For dataset 3, according to MAE measure, M5R algorithm based ARMA2 (2,2) performs better followed by M5R algorithm based ARMAX4 (2,1,2). The standard deviation values of the MAE over the six consecutive time steps for the first model is 0.491 against 0.512 of the second model. According to the RMSE measure, SMOreg algorithm based ARMAX4 (2,1,2) performs better and is closely followed by M5R algorithm based ARMA2 (2,2). The standard deviation values of the RMSE over the six consecutive time steps for the first model is 0.61 against 0.635 of the second model.
CHAPTER 6 CONCLUSION AND FUTURE SCOPE 149 The SMOreg algorithm performed extremely well for 10-min Ahead forecasting. The SVM has been a useful tool for solving pattern recognition and regression problems. Smola and Scholkopf (1998) proposed SMO, an iterative algorithm for solving regression problem using SVM. Shevade et al (2000) pointed out an important source of inefficiency in SMO algorithm proposed by Smola and Scholkopf and suggested further improvements by using two threshold parameters. These modifications made SMOreg better and faster. Weka 3.6.4 implements the modified algorithm and hence gives better results. SMOreg performs better especially when the number of input data is lesser and when the variation in the input parameters is less. Development of Wind Turbine Power Curve Models The second objective to develop a wind turbine power curve model with least error has been done in Chapter 4. The parametric and non-parametric models were identified and validated for five different datasets, one statistically generated and the remaining four real-time datasets. The parametric models were developed based on four and five parameter logistic expression whose parameters were solved using GA, EP, PSO and DE. A model based on the conventional linearized segmental model was also developed for comparison. The nonparametric models were developed using neural networks, fuzzy c-means clustering and data mining algorithms. The algorithms were ranked based on the error measures and the average rank for the five datasets was calculated. The five parameter logistic expression based wind turbine model developed using DE gave better results for both the error measures. When the RMSE measure alone was considered, NN was ranked 2, and the power curve model developed using five parameter logistic function with the parameters solved using PSO was ranked 3. If the MAE measure is considered, the linearized segmented model is ranked 2 and NN is ranked 3. When the statistically generated dataset was ignored, it was
CHAPTER 6 CONCLUSION AND FUTURE SCOPE 150 observed that the parametric models totally outperformed all the non-parametric models. In general, when independent datasets are considered five parameter logistic function solved by PSO or DE can give comparable results. According to Kachitvichyanukul (2012), in DE, the mechanism to generate new solutions is based of floating point arithmetic, similar to PSO. Though the exploration ability of DE might be comparable to PSO, the diversification of DE is better because the best solution in the population does not exert any influence on other solutions in the population. Since the mutant vector is not from the original population, the crossover operation in DE is always between a solution from the population and a newly generated one. PSO and DE share the advantages of having a dense continuous search space and high ability to reach good solution without local search. The influence of population size on solution time is linear in both PSO and DE. PSO has the disadvantage of having a higher tendency for premature convergence and the possibility of average fitness getting worse. Both these possibilities are ruled out in DE. Development of Wind Power Forecasting Models The third objective to develop an accurate multi-step wind power forecasting model has been realized in Chapter 5. The performance of the combined model of wind power prediction is compared with the direct prediction model. Performance of Direct Models The Direct prediction model was developed using the time-series wind power data with and without exogenous variables. Two different non-linear AR direct prediction models were developed: NAR and NARX4. Four different
CHAPTER 6 CONCLUSION AND FUTURE SCOPE 151 data mining algorithms were used namely SMOreg, MLP, Bagging and M5R. The models developed using SMOreg gave excellent results for one-step prediction for both the models (NAR MAE=8.043; RMSE=14.0117 and NARX4 MAE = 8.0402; RMSE = 14.0155). According to MAE measure, the NARX model developed using wind speed as exogenous variable performed better than the NAR model and the RMSE value of NAR was better than NARX model. Performance of Combined Models The Combined Model was developed based on the wind speed prediction model and wind turbine power curve that have least error metrics. The wind speed prediction models were developed using time-series wind speed data with and without exogenous variables. The wind direction and the wind speeds in the previous two years were used as the exogenous variables. The performance of these models for multi-step prediction was also ascertained. The wind turbine power curve model was developed using four and five parametric logistic expression and the parameters solved using PSO and DE. Non-parametric models for wind turbine power curve were developed based on four different data mining algorithms namely MLP, bagging, M5R and M5P. The wind speed predicted by the prediction model with least error metrics was given as input to the wind turbine power curve model. The performance of the Direct and Combined models was tested for multistep prediction. Though the performance of the combined prediction model was not better than the direct model for one-step prediction, it outperformed the direct model in the multi-step prediction. The combination of non-linear autoregressive wind speed model without any external variables (NAR1) developed using M5R algorithm and wind turbine power curve model developed using four parameter logistic expression solved by DE algorithm registers lowest mean and standard deviation when the RMSE is considered. Though the mean of its MAE errors are
CHAPTER 6 CONCLUSION AND FUTURE SCOPE 152 higher than the other models, its standard deviation is the lowest. The standard deviation values of the MAE and RMSE over the six consecutive time steps for the Combined Model (NAR1(M5R)+DE(4P)) are 2.547 and 3.198 against 3.333 and 4.498 of the Direct Model (NAR2(SMOreg)). Wind Resource Assessment of Sulur, a town near Coimbatore, Tamil Nadu, India has been carried out as an application of the developed Combined Model for wind power forecasting. Wind speed data of one year was used to predict the wind speed of the next year. The predicted wind speed was extrapolated to three different heights 50m, 80m and 100m. The wind turbine power was calculated based on a model power curve (Dataset 4). The annual energy produced was estimated for the next year at three different heights. 6.2 FUTURE SCOPE OF WORK The future scope of this work is presented below: Hybrid Models for Wind Farm Power Forecasting Recent research on wind power forecasting has involved the combination of the physical and statistical modeling. Models for Wind Power Ramp Forecasting The variability in wind power can present a substantial challenge to the grid, when the penetration of wind power is high. The ramp event is a very critical issue and is characterized by a sudden large change (increment or decrement) in wind power. Accurate models for ramp event detection and forecasting is important for maintaining the stability of electrical grid.
CHAPTER 6 CONCLUSION AND FUTURE SCOPE 153 Estimation of Forecast Uncertainty Uncertainty analysis of the wind forecasts made plays a key role in grid integration and other power system operations. Use of Wind Power Forecasting in Power System Operations. Accurate models for wind power forecasting aids in several power system operations like operating reserve requirements, unit commitment, dispatch formulations etc. Wind Power Trading Under Uncertainty in Electricity Markets Accurate models for wind power forecasting can transform a wind farm into a wind power plant. It will help the wind farm operators to trade the electricity for higher revenues, competitively with the other conventional sources. Peik- Herfeh et al (2013) developed a decision making model of a virtual power plant under uncertainties for bidding in a day-ahead market using point estimate method. Condition Monitoring of Wind Farms Good performing forecasting models for wind power and wind turbine power curve models can also be used as performance indicator for the health of a wind turbine. Wind Forecasting and Offshore Wind Farms Due to limited land area in several regions, many countries are already into offshore wind farms. These offshore wind farms have opened up huge prospects of research in operation, maintenance, control of wind farms and wind resource assessment to identify potential sites etc.