International Core Journal of Engineering 2020-26 | Page 166
2019 International Conference on Artificial Intelligence and Advanced Manufacturing (AIAM)
Distribution Network Load Forecasting Based on
Improved BP Neural Network
Zheng Gong Zhenghua Zhang* MengSheng Wu Hang Ni
School of Information
Engineering
Yangzhou University
Yangzhou, China
[email protected] School of Information
Engineering
Yangzhou University
Yangzhou, China
[email protected]* State Grid Yangzhou
Power Supply Company
Yangzhou, China Yangzhou Xinyang
Switchgear Co., Ltd.
Yangzhou, China
II. D ISTRIBUTION N ETWORK L OAD
Abstract—In order to overcome the inherent defects of
traditional neural networks, and further improve the short-
term power load prediction accuracy of the station, the genetic
algorithm is introduced to optimize the classical neural network
algorithm, the initial value and threshold of the neural network
are recalculated, and the prediction model is established. The
predicted results are obtained through simulation training and
compared. Finally, the effectiveness of the algorithm is proved
by an example of a city.
The load is usually divided into four types: industrial
power load, agricultural power load, transportation power
load, municipal life and lighting power. The municipal life
and lighting power that is closer to us include utilities,
residents, businesses, schools, institutions, and troops.
Among them, the utility load is relatively stable, and the
residential and commercial power load is affected by the
season, and it is extremely obvious. Therefore, in general,
this type of power load fluctuates greatly, sometimes directly
affecting the seasonal variation of the peak load of the
distribution network.
Keywords—Power system; genetic algorithm; neural network;
short-term load forecasting;
I. I NTRODUCTION
The load characteristics of the distribution network reflect
the inherent nature and variation of the distribution network
load, such as amplitude, power factor, fluctuation, seasonal
variation, concentration, and environmental impact. Studying
and mastering the load characteristics of the distribution
network is conducive to rationally configuring the capacity
and form of the equipment, and can continuously improve the
economic benefits of the distribution network and the safety
and reliability of the power supply.
With the further development of modern technology, 澳
based on the physical grid, the smart grid [2] , which is formed
by the integration of advanced sensing technology [1] ,
communication technology, information technology,
computer technology and control technology with the
physical grid, has entered people's lives. 澳 At the same time,
the requirements for smart grids are also increasing. One of
them is that the power consumption mode can be adjusted
according to the fluctuation of real-time electricity price,
which requires accurate monitoring of the load of the power
grid [3] and pre-judgment [4] . 澳 The magnitude of the load is
directly related to the stability of the grid, so effective
monitoring of the load is necessary. By prediction of the
power and electricity consumption of the load in a certain
period of time in the future, not only can 澳 provide the data
needed for the development of the work plan and
development plan for the distribution network operation
department and the planning and design department but also
through the smart home [6] , intelligent control equipment and
other terminals, use the applications installed on the mobile
phone can remotely control water heaters, air conditioners,
refrigerators and other electrical appliances [7] , to achieve the
Internet of Things. 澳 This is not only in line with the needs of
the ‘Internet +’, but also has great practical significance for
ensuring stable and efficient operation of the power grid,
ensuring smooth social production security, and energy-
saving and emission-reduction peak power [8] .
A. Load Forecasting
According to the forecast period, it can be divided into
short-term, near-term (5-year), medium-term (10-year), and
long-term (20-year) forecasts. The predictions taken in this
paper are based on short-term load forecasting [14] . Short-term
forecasting is generally used for the dispatching operation of
the distribution network, with the predicted power as the
main, in order to formulate the power production plan,
determine the power distribution operation mode and arrange
equipment maintenance and replacement procedures. 澳
According to the needs of short-term load forecasting,
continuous forecasting can be made on a monthly, weekly,
daily, hourly, or even minute basis during the year. At present,
the more conventional forecasting method mainly includes
four steps: 澳 澳
1) Determine the amount (power or electricity
consumption) to be predicted and the time interval between
the forecast quantities, and ask for enough historical data.
This paper proposes to optimize on the basis of neural
network [10] ,introduce genetic algorithm [11] to get better
network initial weight and threshold [12] , improve the
accuracy of load forecasting [13] , through the urban power
load The prediction contrast proves the superiority of the
algorithm.
978-1-7281-4691-1/19/$31.00 ©2019 IEEE
DOI 10.1109/AIAM48774.2019.00036
2) Analyze the historical data of the load. Establish a
model that is applied to the process of surface ripe load
change, and identify it and estimate the parameters roughly.
3) Perform accurate estimation of parameters and
scientific testing of models.
144