Comparative Study Of Artificial Neural Network And Support Vector Machine For Load Forecasting

Electricity load forecasting problem has been investigated in this thesis.Forecasting of electricity demand has become one of the major research fields in electrical engineering.The electric utilities in industry require forecast with lead time which ranges from short term to long term according to their requirements.Load forecasting is a complex multi variable and multi-dimensional estimation problem.With inclusion of large number of different distributed energy resources and large amount of data collection due to smart meters and sensors is further increasing its complexity.Thus classical methods to predict accurate load forecasts are answerless,as they fail to track the seemingly random trends accurately,in which machine learning algorithms are much better.Therefore,machine learning techniques have grabbed a lot of attention from researchers due to their flexibility and precise results in data modeling.This thesis aims to address the problem of improving load forecasting using machine learning methods.Support Vector Machine(SVM)and Artificial Neural Network(ANN),two most common machine learning algorithms for load forecasting are used in this study.The work mainly focusses on utilizing kernel trick property of Support Vector machine which has the ability to model complex nonlinear relationships and propose the SVM model with best kernel amongst the three famous kernels namely Polynomial Kernel,Radio basis function(RBF)kernel and Pearson function(PUK)kernel.By this way,we can propose a SVM model which have a universal kernel which can transform any type of complex nonlinear input data into linear data.The main idea in this thesis is to select the best kernel amongst the three available kernels.To further endorse our model,the result of best kernel function model is compared with another popular machine learning model,ANN.Methodology is validated by testing the reliability of forecasting model on real time data.The performance of proposed approach is verified with simulation software WEKA.Optimization of SVM parameters and kernels along with parameters of ANN model has also been proposed.To achieve the task,historical data has been collected from distribution grid in Pakistan for a period of six weeks.This data was then preprocessed and missing values and outliers are removed.Then data become normalized to achieve better performance.The resultant dataset is then used to optimize the design of SVM model and ANN model.Error Metrics have been used to compare the performance among the kernels and machine learning models.Our research finding suggest that PUK based SVM model performs better than ANN model and can improve overall system prediction accuracy.