Please use this identifier to cite or link to this item:
Title: Self - Organizing Maps Based Hybrid Approaches to Short Term Load Forecasting
Keywords: self organizing maps, short term load forecasting, hybrid linear neural model
Issue Date: 25-May-2010
Citation: VINEET YADAV (2010-05-25). Self - Organizing Maps Based Hybrid Approaches to Short Term Load Forecasting. ScholarBank@NUS Repository.
Abstract: Accurate short term load forecasting has become an essential task for effectively managing the modern power systems. Important operating decisions such as scheduling of power generation, scheduling of fuel purchase, maintenance scheduling, and planning for energy transactions depend upon the electricity load forecasts. An improvement in accuracy of load forecasting results in substantial savings in operating costs, and an increased safety and reliability of the electric system. For short term load forecasting, two approaches coming from two different scientific disciplines and using different ideas for solving the same problem coexist. On one side, there are the conventional load forecasting techniques which are rooted in classical statistics. These include methods such as multiple regression and Box-Jenkins time series models, which study the relationship between load behavior and its influencing factors in a linear fashion. While these methods are widely used in the industry for forecasting load profiles on normal days, they have severe limitations in terms of accuracy of results, especially on weekends, holidays and special days. On the other side, there are the soft computing methods such as artificial neural networks (ANN) which are fully capable of learning the non-linear relationship between the load and its influencing factors due to their excellent learning and generalization capabilities. But they are associated with problems as well, the more inherent ones being no assurance of convergence, network falling into local minimum and a need for frequent re-training due to changing seasonal conditions. In this thesis, two hybrid models are proposed which combine elements from both the above mentioned approaches ? statistical methods and artificial neural network methods. The hybridization approach helps achieve two objectives. In the first model, the forecasting accuracy of the resulting hybrid model is significantly improved over either of the individual component models. In the second model, hybridization helps to tackle one of the major challenges faced while training a neural network ? its weight initialization. In the first model, a two-stage hybrid network is proposed with weighted self-organizing maps (SOM) and autoregressive (AR) model. In the first stage, a weighted SOM network is applied to split the past dynamics into several clusters in an unsupervised manner. Then in the second stage, a local linear AR model is associated with each cluster to fit its training data in a supervised way. Though this method can be used for forecasting any time series, it is best suited for processes which are non-linear and non-stationary and show cluster effects, such as the electricity load time series. The second hybrid model describes a linear model with time varying coefficients which are the outputs of a single hidden layer feedforward neural network. The hidden layer is responsible for partitioning the input space into multiple sub-spaces through multivariate thresholds and smooth transition between the sub-spaces. A new method is proposed to smartly initialize the weights of the hidden layer before the network training. Applied to the electricity markets, the proposed method is better able to model the smooth transitions between the different regimes which are present in the load demand series because of market effects and season effects.
Appears in Collections:Master's Theses (Open)

Show full item record
Files in This Item:
File Description SizeFormatAccess SettingsVersion 
header.pdf40.41 kBAdobe PDF


intro.pdf132.21 kBAdobe PDF


thesis.pdf1.55 MBAdobe PDF



Page view(s)

checked on Apr 21, 2019


checked on Apr 21, 2019

Google ScholarTM


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.