Wind Power Forecasting & Its Role In Wind Integrated Power System Operations

Abstract

Wind power forecasting plays a critical role in wind integrated power system operations such as unit commitment, economic dispatch, reserve allocation, storage management and trading of energy. Accurate assessment of wind data forecasting is vital for wind power dispatch. This is required by transmission system operators, utility energy service providers, wind farm managers and energy traders. A critical analysis of literature on short-term forecasting of wind data shows that most of models developed suffer with limited accuracy. This is due to unstructured data, varying scenario and selection criterion employed in models. Amongst combinational models, impact of all possible techniques is not considered. In addition, correlation of time varying error distribution with forecast results is not investigated and thereby results in significant error in forecasted values. In view of this, the thesis contributes to solution by developing a novel combinational methodology with a mixture of Statistical and Artificial Intelligence (AI) techniques. In continuation, a novel mathematical derivation is proposed for combined error minimization model.

Highlight of this work is that, the proposed models are tested using ten years historical wind speed data from 2007 to 2017. The data is measured from 50 m wind mast installed at Basaveshwar Engineering College (Autonomous), Bagalkot, Karnataka state, India. The system was installed under R&D grant from Technical Education Quality Improvement Program (TEQIP) of World Bank in 2007. Wind data is recorded using NRG Symphonie Data logger at 10 min interval. Further, to find site dependent characteristics of wind data and forecast models, 7 years actual measured wind data from 2009 to 2017 at 50 m level is collected from Regional Agricultural Research Station (RARS) Vijayapur and State Load Dispatch Centre (SLDC) KPTCL Bengaluru.

In this thesis, Autocorrelation Function (ACF), Partial Autocorrelation Function (PACF), spectrum, cumulative frequency distribution and box plots are used to analyze behavior of wind speed time series data. Detailed analysis from statistics and graphs indicate underlying seasonal pattern in wind speed time series. ACF and PACF plots revealed that, each site has a distinct characteristics and patterns associated with them. Raw wind speed data are differenced and smoothed to attain stationarity before using them in development of forecast models.

Mathematical modeling of six types of statistical and time series techniques are critically investigated in this research work. Least Mean Square (LMS) model is considered as bench mark forecasting model to compare with other models. Holt Winters exponential smoothing model is developed to identify effect of smoothing on forecast accuracy. To avoid misspecification, Akaike Information Criteria (AIC), Bayesian Information Criteria (BIC) and Final Prediction Error (FPE) are employed to optimally select appropriate Autoregressive Integrated Moving Average (ARIMA) model. Analytical and iterative methods are investigated for identification of ARIMA models. Further Transfer Function-ARIMA, GARCH and Wavelet-ARIMA models are developed.

A case study is carried out to investigate effect of quality, volume of data and forecast horizon on wind speed forecasting accuracy. Statistical tests are conducted on forecasted wind speed and error values of these six methods. An accuracy of 83.35% is achieved with Wavelet-ARIMA model as compared to 72.37% from benchmark LMS model. Results show that, as uncertainty increases forecast accuracy decreases. Further, forecast accuracy increases with increase in volume of training data up to 3 years. Statistical tests reveal that, Wavelet-ARIMA model has higher score in t-test, h-test and regression tests.

Four AI Techniques namely Neural Network (NN), Fuzzy Logic (FL), Adaptive Neuro Fuzzy Inference System (ANFIS) and Support Vector Machine (SVM) models are developed and tested. Levenberg-Marquardt (LM), Bayesian Regularization (BR) and Scaled Conjugate Gradient (SCG) Algorithm are used to train NN model and tested for three sites. It is found that SCG algorithm performed better than remaining two algorithms for both 24 hour and 168 hour prediction. A Fuzzy Logic model is developed for wind speed forecasting model using rules framed from input data training pattern. ANFIS model is developed for wind speed forecasting using Generalized bell, Triangular, Trapezoidal and Gaussian type membership functions. A new Support Vector Machine model is developed using Radial Basis Function training algorithm. Forecasting results of SVM indicates a remarkable improvement over NN, FL and ANFIS model. Forecast accuracy of 88.90% is obtained with SVM model as compared to 86.63% with NN model. Further these models are tested with data of 18 patterns at three sites. Results revealed that SVM model has outperformed remaining individual Statistical and AI models.

A novel approach, for optimally combining results of 10 different forecasting methods is presented to utilize merits and features of individual models. Mathematical models are developed for Statistical and AI models. These mathematical models are used in an algorithm for optimally combining model equations to obtain a global optimal solution for short term wind speed forecast. Results shown that, SVM model combined with any other model has outperformed remaining combinations. Models are validated by testing with 12 types of patterns for 24 hour and 168 hour wind speed prediction. An accuracy of 92.87% for 24 hour and 90.50% for 168 hour prediction is achieved with 6-pair model. The t-test, F-test and regression tests are conducted on forecasted wind speed values of paired models.

A novel approach for short term wind speed forecasting using combined error minimization technique for optimally selecting combinational models is developed. A mathematical expression is derived for minimizing forecast error based on error output of each combinational model. Optimization algorithm is used to minimize forecast MAPE and maximize regression coefficient. Time varying weights are assigned to each model depending on forecast error at training time so as to capture random and symmetrical error distribution at all intervals. Proposed model is tested using ten year historical wind speed data. It is observed that there is an improvement in R2 of 0.725 and forecast accuracy of 93.3% with Combined Error Minimization Model (CEMM) as compared to 0.705 and 89.8% in absence of CEMM. Further there is significant improvement in average error from 0.99 to 0.80 with incorporation of CEMM.

It is concluded that, proposed novel approach is a powerful tool to mitigate effect of uncertainty of wind power system operation and to increase value of wind energy.