Short-Term Load Forecasting Using System-Type Neural Network Architecture

cracklegulleyΤεχνίτη Νοημοσύνη και Ρομποτική

19 Οκτ 2013 (πριν από 3 χρόνια και 7 μήνες)

68 εμφανίσεις

Short
-
Term Load Forecasting Using
System
-
Type Neural Network Architecture

Shu Du, Graduate Student

Mentor: Kwang Y. Lee, Professor and Chair

Department of Electrical and Computer Engineering

Baylor University


Outline


Introduction and Background


Objectives


Load Forecasting Categories


Load Forecasting Methods


Proposed Approach


Regression and Rearrangement


System
-
Type Neural Network Method


Learning Algorithm of System
-
Type Neural Network


Extrapolation and Interpolation


Simulation Results


Rearrangement


Output of Semigroup Channel


Extrapolation


Conclusions

Introduction and Background


Objective


Electric power generation, transmission, distribution, security


Increase or decrease output of generators


Interchange power with neighboring systems


Prevent overloading and reduce occurrences of equipment failures


Electric power market


Price settings


Schedule spinning reserve allocation properly


Load Forecasting Categories


Short
-
term load forecasting


One hour ~ One week


Control and schedule power system in everyday operations


Medium
-
term and Long
-
term load forecasting


One week ~ longer than one year


Determine capacity of generation, transmission, distribution systems,
type of facilities required in transmission expansion planning,
development of power system infrastructure, etc.

Introduction and Background

Introduction and Background


Load Forecasting Methods


Parametric methods


Regression method


Time series

Autoregressive Moving Average (ARMA)

Spectral expansion technique (Fourier Series)

State equations


Artificial intelligence methods


Artificial neural networks

Feedforward network

Recurrent network


Fuzzy logic


Expert systems

Proposed Approach


Regression and Rearrangement


Regression


Objective

Represent given load with respect to two major variables

time

and
temperature


Load Form

-----
Base load component (time factor)

-----
Weather sensitive load component (weather factor)

-----
Load component (other factors)

Proposed Approach


Regression and Rearrangement


Rearrangement


Objective

Minimize the fluctuation caused by hourly temperature

Obtain the smoothness of the given load data

Hour

1

24

2

Rearrangement

Temperature

Day

Hour

1

24

2

Load before Rearrangement

Load after Rearrangement


Implementation

Align given load based upon
magnitudes of hourly temperatures

Proposed Approach


System
-
Type Neural Network Method


Algebraic Decomposition


Objective

Form an approximation load data to



Implementation


Reorganize given load into a parameterized set



Select elements and orthonormalize them to a basis set by Gram
-
Schmidt
process



Determine the linear combination of basis set for each element



Combine the coefficient vector and the basis set to achieve an approximation



Proposed Approach


System
-
Type Neural Network Method

Function Channel

(NN1)

Semigroup Channel

(NN2)


Function Channel


Structure


RBF networks


Each

network

implements

one

of


orthonormal

basis

functions


Semigroup Channel


Structure

Simple Recurrent Network


Smoothen the coefficient vector and


Realize semigroup property

Proposed Approach


Learning Algorithm of System
-
Type Neural Network


Function Channel


RBF network can be designed rather than trained



RBF networks emulate selected basis functions


Semigroup Channel


Primary Objective


Replicate and smoothen the vector with a vector which has
the semigroup property


Secondary Objective


Acquire a semigroup property in the weight space which is the basis for
extrapolation


The entire trajectory is sliced into a nested sequence of trajectories

Proposed Approach


Extrapolation and Interpolation


Extrapolation


Extrapolation is needed only when temperature forecast at a given hour exceeds the
historical bounds at the same time


Interpolation


Interpolation is needed when temperature forecast at a given hour falls into the
historical temperature range at the same time

Load after Rearrangement

Extrapolation of Coefficient

Hour

1

24

2

Decompose &
Smoothen

4

Extrapolated
Coefficient

4

Temperature

3

1

2

5

Hour

1

24

2

Decompose &
Smoothen

3

4

Interpolated
Coefficient

4

Temperature

3

1

2

Load after Rearrangement

Interpolation of Coefficient

Simulation Results


Forecasting Procedure


Data Source


New England Independent System Operator


Historical Data


Load


load for the year
2002


Temperature


weighted average hourly temperature of 8 stations in


the New England area


Pattern


Weekday pattern (Mon ~ Fri) and Weekend pattern (Sat, Sun)


Next Day Forecasting


Previous loads and temperatures in the length of four weeks

Simulation Results


Simulation of Forecasting A Weekday Load


Rearrangement

Rearrange

Simulation Results


Simulation of Forecasting A Weekday Load


Output of Semigroup Channel

Simulation Results


Simulation of Forecasting A Weekday Load


Extrapolation

Simulation Results


Regression Load Forecasting Results

Conclusions


Next Day Load Forecasting based upon Weather
Forecast


A mathematical approach referred to as algebraic decomposition is
investigated


The system
-
type neural network architecture combining Radial Basis
Function Networks and a Simple Recurrent Network is proposed


A new training algorithm in the SRN is proposed


Regression and Rearrangement are performed to guarantee smoothness of
coefficient vector


Interpolation and Extrapolation are implemented based on temperatures


Much better results with respect to actual load and removal of regression
are expected if load and temperature are highly correlated to each other





THANK YOU