Please use this identifier to cite or link to this item: https://scholarbank.nus.edu.sg/handle/10635/182794
Title: ADAPTIVE SHORT-TERM ELECTRICAL LOAD FORECASTER USING ARTIFICIAL NEURAL NETWORKS
Authors: TAN SWEE SIEN
Issue Date: 1998
Citation: TAN SWEE SIEN (1998). ADAPTIVE SHORT-TERM ELECTRICAL LOAD FORECASTER USING ARTIFICIAL NEURAL NETWORKS. ScholarBank@NUS Repository.
Abstract: For many years, load forecasting has played a major role in power utilities around the world. The reason behind its importance lies in the fact that a study of the load forecast is essential before appropriate economic operations and planning activities can be successfully performed. There are basically three types of load forecast, namely long-term load forecast, medium-term load forecast and short-term load forecast. This thesis focuses on short term load forecasting with particular emphasis on one-day ahead forecast. However, a short section on a one-hour ahead load forecaster that was developed by the author has been added. An analysis of the literature based on short-term load forecast reveals that tremendous amount of research has gone into this field. A wide variety of short-term load forecasters (STLF) have been developed in a bid to obtain one that is capable of producing results with the highest accuracy for all operating conditions. In the past, methodologies have been based on traditional statistical methods. However, in recent years, there has been a gradual shift towards employing artificial intelligence in STLF. This can be attributed to the discovery that far better results can be derived with models based on AI techniques. An analysis on the various artificial intelligence (AI) based STLF techniques will be among the topics covered in this thesis. In addition, this thesis also provides an insight into some AI-based load forecasting models developed by the author. The first is a pure neural network model for one-day ahead load forecast [ 16]. The second predicts the one-hour ahead forecast and uses only Kohonen's neural networks. The final and undoubtedly, the best-performing model in this thesis involves a parallel neural-fuzzy expert system [27] that is used to derive an enhanced one-day ahead forecast. It has been put through various testing phases and has proven its versatility by producing very good results at all times. Implementations and simulations of the respective models have been carried out on a Sun Spare 10 Workstation using an AI-based software ECANSE (Environment for Computer-Aided Neural Software Engineering) provided by Siemens Pte Ltd. A comparison of the results with other load forecasters currently in use reveals that the hybrid fuzzy-neural model is able to produce predictions with a higher level of accuracy, but at a much lesser time and fuss. The pure neural network model, on the other hand, is less versatile when faced with changes in conditions such as unusual weather conditions and proximity to holidays. Furthermore, a longer training time is required for the two sets of Kohonen's neural networks. Overall, the greatest improvement in accuracy can be observed in the holiday load forecast, which dips from 2.71 % down to 0.84% when the fuzzy expert system is incorporated into the model. The fuzzy-neural load forecasting model presented in this thesis has been extensively tested with actual load data provided by PUB (now know as PowerGrid Ltd). It has been proven to produce very accurate results. In addition, it is fully automated and requires little human interference. The electric load forecaster is a major component of any EMS (Energy Management System). Integration of this model into the modem EMS used by PowerGrid Ltd can be easily performed and it will play an essential role in forecasting the hourly load for the next day.
URI: https://scholarbank.nus.edu.sg/handle/10635/182794
Appears in Collections:Master's Theses (Restricted)

Show full item record
Files in This Item:
File Description SizeFormatAccess SettingsVersion 
b20650140.pdf5.77 MBAdobe PDF

RESTRICTED

NoneLog In

Google ScholarTM

Check


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