Viet - Prediction of Natural Gas Consumption With Feed-Forward and Fuzzy Neural Networks

8/6/2019 Viet - Prediction of Natural Gas Consumption With Feed-Forward and Fuzzy Neural Networks

http://slidepdf.com/reader/full/viet-prediction-of-natural-gas-consumption-with-feed-forward-and-fuzzy-neural 1/17

Prediction of Natural GasConsumption with Feed-forward andFuzzy Neural Networks

N.H. Viet Institute of Fundamental Tech. Research

Polish Academy of Sciences Poland,

J. MadziukFaculty of Mathematics and Information Science

Warsaw University of Technology Poland.



N.H. Viet, J. Madziuk: Prediction of Natural Gas Consumption with Neural Networks

Presentations schedule

Introduction

Feed-forward neural networks

Fuzzy neural networks

Experimental results and conclusions




Introduction

Prediction of gas consumption is an important element in business planning.

The challenges:

the volatility of consumer profile, the strong dependency on weather conditions,

the lack of historical data.

The purpose of this work: an application to gas

load prediction using various neural networkmodels.




Introduction

Three types of prediction: One day (short-term) prediction,

One week (mid-term) prediction,

Four week (long-term) prediction.




Introduction

The data contains the daily gas loads and the averagedaily temperatures.

Seasonality

Strongdependency ontemperature.

An overview of the data:




Introduction

The inputs: Historical daily gas loads,

Average daily temperatures

Time factor (the season inputs) For the n-day period: [t + 1, t + n], two values were used:

Where:

One additional bit indicating the work day/weekend day inthe case of daily prediction.




Feed-forward model

General network architecture:

Previous daily loads

Previous daily temperatures

Time encoding




Feed-forward model

Chosen configurations: One day prediction: 9(3+3+3)-8(3+3+2)-3-1

One week prediction: 12(5+5+2)-10(4+4+2)-4-1

Four week prediction: 16(7+7+2)-10(4+4+2)-4-1




Fuzzy neural model

Why to use the fuzzy neural model?: Impreciseness of data (only average daily

temperature is available),

Fuzzy neural networks generally have a better

performance,




Fuzzy neural model

Fuzzy neural network architecture:

Membership layer Defuzzification layer




Fuzzy neural model

Fuzzy neural network dynamics: Gaussian membership function:

Output:




Fuzzy neural model

FNN can be trained using the gradient-basedtechnique.

An equivalent rule sets:





Training data: from Jan. 01, 2000 to Dec. 31,2001.

Testing data: from Jan. 01 2002 to Jul. 31, 2002.

The moving window technique was used togenerate the training and the testing samples.

The following experiments were performed: Single feed-forward network (SingleN)

Single fuzzy network (FuzzyN) 3 feed-forward networks (3AvgN)

3 temperature context networks (3TempN)





Temperature context networks: Divide the training set into 3 overlapping subsets

(denoted by Low, Medium and High ) using theaverage temperature,

Train 3 types of networks with these setsindependently,

Combine 3 networks into one module while testing.

Remark: training the networks within a particular

context should be easier than in the entire input space.









An example of one week and four week prediction:



Thank you for your Thank you for your attentionattention

Viet - Prediction of Natural Gas Consumption With Feed-Forward and Fuzzy Neural Networks

Documents