Hybrid artificial neural network for short-term electricity load forecasting

Name: Abeyrathna K., Darshana

keyword: Short-term electricity demand forecasting

LCSH: Electric power consumption -- Thailand -- Forecasting

; Artificial neural networks

LCSH: Neural networks (Computer science)

; Genetic algorithm

; Particle swarm pptimization

; Backpropagation

Abstract: Numerous research have been being conducted by many researchers in the field of electricity forecasting with the purpose of matching the forecasted outcomes with the actual consumption for the next day. This is important since both the under and over forecasting lead to revenue reductions. Even though the existing forecasting techniques can forecast the electricity consumption up to a considerable level, small errors can reduce the revenue in large amounts. Therefore, the electricity forecasting field is still active and strategies and improvements for the existing techniques for reducing the forecasting errors are always appreciated. Thus, the objective of this research is to improve the electricity forecasting outcomes or further reducing the electricity forecasting errors for the case of Thailand. Based on the literatures, Artificial Neural Network (ANN) is selected as the technique to forecast the daily electricity consumption. ANN has the ability of recognizing and learning non-linear patterns in data. However, ANN has many internal and external parameters and training methods which have to be arranged in the optimum way. These optimally arranged parameters and training algorithms can significantly improve the forecasting outcomes. Nevertheless, this research mainly focuses on finding the best training algorithm to train the ANN for electricity demand forecasting. The main experiment is supported by two other experiments where they help to find the optimum number of inputs and outputs to the ANN, the amount of data needed to train the ANN, and the optimum number of hidden units of ANN for electricity demand forecasting. Due to the limitations of using backpropagation training algorithm to train the ANN, two stochastic optimization techniques are introduced to optimize the weights and bias of the ANN: Genetic Algorithm (GA) and Particle Swarm Optimization (PSO). Using the positive characteristic of both the algorithms, a combined algorithm is created to train the ANN. An experiment is arranged to find the best training algorithm out of these four algorithms. A sample data set is selected from the data gathered by Electricity Generating Authority of Thailand (EGAT). The sample set consist with data from 1st January, 2012 to 31st December, 2013 where one year data is used for training the ANN and one year data is used for testing. The experiment is started with cleaning the selected data set using the time window based data cleaning technique. ANN with four inputs (Lt(d-7), Lt(d-1), Tt(d-1), Tt(d)) and one output (Ft(d)) is created to forecast each time period t of each day d. The ANN consist with one hidden layer and four hidden neurons. Backpropagation training algorithm, GA, PSO, and hybridized training algorithm are used to train the created ANN with one year data, respectively. Forecasting outcomes from each training algorithm are evaluated in term of the percentage error (Mean Absolute Percentage Error). According to the yearly average MAPE, the hybridized training algorithm performs well compared to the other three training algorithms. Genetic algorithm also gives competitive results compared to the hybridized training algorithm. However, Particle Swarm Optimization and backpropagation training algorithms are fairly good compared to the other two training algorithms. Regardless of the training algorithm, days in December are hard to forecast as their electricity consumption is significantly low compared to the other months. Therefore, forecasted outcomes are always higher than the actual consumption. Findings of this research will add a great value to the electricity forecasting filed and the users of ANN. The research will help to define the inputs, outputs, and required data for electricity forecasting and most importantly for selecting the best training algorithm for the ANNs

Thammasat University. Thammasat University Library

Address: BANGKOK

Email: preserv@tu.ac.th

Name: Chawalit Jeenanunta

Role: advisor

Created: 2016

Modified: 2021-08-13

Issued: 2021-08-13

วิทยานิพนธ์/Thesis

application/pdf

DOI: 10.14457/TU.the.2016.1110

eng

DegreeName: Master of Science

Level: Master's Degree

Descipline: Engineering and Technology

Grantor: Thammasat University

©copyrights Thammasat University

ลำดับที่.	ชื่อแฟ้มข้อมูล	ขนาดแฟ้มข้อมูล	จำนวนเข้าถึง	วัน-เวลาเข้าถึงล่าสุด
1	3684abeyrathna.pdf	4.03 MB	3	2021-10-25 15:17:58

ใช้เวลา

0.01962 วินาที