Consumer load prediction based on NARX for electricity theft detection
A range of load prediction techniques has largely been used for energy management at various levels. However, the data used for the prediction are cumulative energy data, which reveal the activities of consumers and not individual consumer, on the distribution power network. Individual consumer...
| Main Authors: | , , |
|---|---|
| Format: | Conference or Workshop Item |
| Language: | English English |
| Published: |
IEEE
2016
|
| Subjects: | |
| Online Access: | http://irep.iium.edu.my/55565/ http://irep.iium.edu.my/55565/ http://irep.iium.edu.my/55565/ http://irep.iium.edu.my/55565/2/55565.pdf http://irep.iium.edu.my/55565/8/55556_Consumer%20load%20prediction%20based_Scopus.pdf |
| Summary: | A range of load prediction techniques has largely
been used for energy management at various levels.
However, the data used for the prediction are cumulative
energy data, which reveal the activities of consumers and not
individual consumer, on the distribution power network.
Individual consumer data is essential for real time
prediction, monitoring and detect of electricity theft. A new
approach of monitoring individual consumer based on
consumer load prediction using nonlinear autoregressive
with eXogenous input (NARX) network is considered in this
study. One month average energy consumption data
acquired from consumer load prototype developed was used.
Consequently, 5-minute step ahead load prediction was
achieved. The NARX architecture was based on nine hidden
neurons and two tapped delay and the network trained using
Bayesian regulation backpropagation technique. The data
set contains a total of 8928 data points representing energy
consumed at five minute interval for one month. The data
was divided into two sets at ratio 70:30 for training and
validation, respectively. The training data equals 6206 while
the validation data is 2722. MATLAB environment was used
for the processing of the data. The training and validation
MSE is 0.0225 and 0.0533 respectively, while the total time
for the training is 0.016s. |
|---|