Hindawi

Mathematical Problems in Engineering

Volume 2018, Article ID 4276176, 16 pages
https://doi.org/10.1155/2018/4276176

Research Article
A Simple Method of Residential Electricity Load Forecasting by
Improved Bayesian Neural Networks

Shubin Zheng, Qianwen Zhong , Lele Peng, and Xiaodong Chai

School of Urban Railway Transportation, Shanghai University of Engineering Science, China

Correspondence should be addressed to Qianwen Zhong; datouzqw@aliyun.com

Received 4 April 2018; Revised 5 July 2018; Accepted 16 August 2018; Published 13 September 2018

Academic Editor: Gaetano Zizzo

Copyright © 2018 Shubin Zheng et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Electricity load forecasting is becoming one of the key issues to solve energy crisis problem, and time-series Bayesian Neural
Network is one popular method used in load forecast models. However, it has long running time and relatively strong dependence on
time and weather factors at a residential level. To solve these problems, this article presents an improved Bayesian Neural Networks
(IBNN) forecast model by augmenting historical load data as inputs based on simple feedforward structure. From the load time
delays correlations and impact factors analysis, containing different inputs, number of hidden neurons, historic period of data,
forecasting time range, and range requirement of sample data, some advices are given on how to better choose these factors. To
validate the performance of improved Bayesian Neural Networks model, several residential sample datasets of one whole year from
Ausgrid have been selected to build the improved Bayesian Neural Networks model. The results compared with the time-series load
forecast model show that the improved Bayesian Neural Networks model can significantly reduce calculating time by more than
30 times and even when the time or meteorological factors are missing, it can still predict the load with a high accuracy. Compared
with other widely used prediction methods, the IBNN also performs a better accuracy and relatively shorter computing time. This
improved Bayesian Neural Networks forecasting method can be applied in residential energy management.

1. Introduction regression trees algorithm, which allows performing closed-

loop control for demand response (DR) strategy synthesis
In recent years, how to continuously meet people's energy for large commercial buildings. Chen et al. [6] designed a
needs has become a hot research topic of global concern. Support Vector Regression (SVR) forecasting model with
In addition to improving energy efficiency and developing the ambient temperature of two hours before DR event as
new energy sources, rational management of energy is also input variables. Cabrera et al. [7] proposed a methodology to
a useful way to solve the energy crisis. The premise of obtain probabilistic forecasts of electricity load that is based
effective energy management is more accurate energy load on functional data analysis of generalized quantile curves.
forecasting. Therefore, researchers have done a lot of research Marszal-Pomianowska et al. [8] presented a high-resolution
work on the prediction of electricity demand. Yu et al. [1] model of household electricity use developed based upon a
studied the use of sparse coding for modeling and forecasting combination of measured and statistical data. Liang et al.
individual household electricity loads. Qiu et al. [2] presented [9, 10] proposed two electricity demand forecasting methods.
an ensemble method composed of Empirical Mode Decom- One applied a new weight determination method with
position (EMD) algorithm and deep learning approach. forecasting accuracy as induced variables based on extreme
Hsiao [3] proposed an approach to model the very short- learning machine and multiple regression models. The other
term load of individual households based on context infor- was from a carbon emission view to give a hybrid model
mation and daily schedule pattern analysis. Aman et al. [4] based on wavelet transform and least squares support vector
analyzed dynamic demand response (D2R) of prediction machine optimized by an improved cuckoo search.
models, especially deeply discussed the scenario of small With the vigorous development of artificial intelligence,
customers. Behl et al. [5] provided a model based control with the neural network method has also been applied by many
2 Mathematical Problems in Engineering

Hidden Hidden Hidden Output

Input Layer Layer Layer Layer

x a1 LW 2 a2 an
IW1 LWn LWo ao-L

b1 b2 bn bo

Figure 1: Structure of FFNN.

WO (x) N

VI aH
WH WO
ao -L 
L
Er Y
bH
bO
WH (x) B(x) N

Figure 2: Structure diagram of BNN training algorithm.

researchers in the field of electricity load forecasting [11– comparison also is made between the IBNN method and sev-
15]. Neural network methods already have some mature eral common applied machine learning regression methods.
training algorithms and network structures. Among various The results show that the IBNN model performs relatively
neural network algorithms that have been widely used, like better in all evaluation indicators. Finally, conclusions are
Scaled Conjugate Gradient, Levenberg Marquardt, Bayesian, summarized and the future work is briefly mentioned.
etc., the performance of Bayesian Neural Networks (BNN)
is validated as one of the most effective way to build the 2. Method of Improved Bayesian Neural
electricity load prediction model [16–22]. And related to the Networks (IBNN)
structures, Time-Series Neural Networks (TSNN) structure
is more reasonable and effective than the simple Feedforward This section firstly provides the basic BNN model structure,
Neural Networks (FFNN) structure since the electricity load then the improved BNN model is presented in this paper.
has obvious time-cycle characteristics [23–25]. However, as The selected evaluations of IBNN prediction model are briefly
the traditional continuous time delay feedback increases, the introduced. Finally, the impacts of inputs and relative factors
efficiency of the prediction model is significantly reduced are discussed.
under TSNN structure.
To improve the above problem, an improved BNN 2.1. Basic BNN Model Structure. Bayesian feedforward neural
(IBNN) model of residential short-term electricity demand network structure (BFFNN) model is selected as the basic
is proposed by a relatively simple method with both high neural networks in our study for residential load forecasting.
performance and high efficiency. This method is based on The principle of Bayesian approach is described in [16, 26].
the basic BNN method and the simple FFNN structure The FFNN structure is shown in Figure 1.
and considers the characteristics of electricity demand cyclic To obtain high accuracy of the electricity demand predic-
changing over time. Through analyzing the correlation results tion, the authors also apply the most popular structure of NN
between historical electricity demand data and current elec- which is called multilayer perceptron (MLP). This structure
tricity demand data at different time ranges, the historical mostly has an input layer, one or several hidden layers, and an
demand data with stronger correlations are selected as the output layer. Every layer obtains its weight and bias matrixes
predictive vectors to construct the prediction model instead through Bayesian training algorithm.
of continuous feedback in the TSNN. It can be found from the process of BNN in Figure 2
The remaining of this paper is organized as follows. that the NN forecasting model is built through finding a
Firstly, an IBNN based on FFNN is built by adding histor- minimum error between the predict values and the actual
ical demand data as inputs through correlation analysis of observed values, which is adjusted by some defined rules,
electricity consumption at different delayed time scales. 𝐵(𝑥), 𝑊𝑂(𝑥), and 𝑊𝐻(𝑥) in the software until it satisfies a
Moreover, inputs selection of the IBNN forecast model is defined error rule (< Er), as shown in the following:
discussed and further with the effect analysis of relative
𝐿 (𝑉𝐼 ) = 𝐿 (𝑉𝐼 )
factors on the forecast performance. Then, the results of the
time-series BNN model and the IBNN model are compared + min (𝑓𝑂 (𝑓𝐻 (𝑉𝐼 ⋅ 𝑊𝐻 + 𝑏𝐻 ) ⋅ 𝑊𝑂 + 𝑏𝑂) (1)
and discussed, especially on the program running time and
the dependence on time or meteorology factors. Further the − 𝐿 (𝑉𝐼 ))
Mathematical Problems in Engineering 3

Input Hidden Output

L n-1
Layers Layer
L n-2
VITS T
L n-3 aH WOTS ao -L
D WHTS
L n-4 L
bHTS bOTS
L n-t d
T
D WHTS_TDL
L

Figure 3: Structure of TSNN.

Input Hidden Output

Layer Layer

Hidden
Time Neurons
WH1,1
WH1,2 Forecasting
WH1,n
L n-1 WH2,1 Load
Day Type WH2,2
L n-2
WH2,n
WH3,1 WO1,1
L n-3 WO2,1
WH3,2
L n-4 Temperature
WH3,n
L n-24/t i WH4,1
WH4,2
WH5,2 WOn,1
L n-2×24/t i WH5,1
L n-7×24/t i Humidity WH4,n

L n-7×2×24/ti
WH5,n
Historical
Load

Figure 4: Structure and process of IBNN model.

Here in (1), 𝐿(V𝐼 ) is the real historical value, W𝐻 is the increasing number of time delay values, the computing time
weight matrix of hidden layer, b𝐻 is the bias vector of hidden is significantly growing. Due to the fact that there are only
layer, and W𝑂𝑇𝑆 and b𝑂𝑇𝑆 express the weight matrix and several input vectors, such as time of everyday, day-type,
bias vector of output layer, respectively. 𝑓𝑂 and 𝑓𝐻 are the ambient temperature, and relative humidity, many of them
functions of output layer and hidden layer, respectively, as are meteorological data, and another problem may be that the
shown in model is greatly affected by the meteorological data.
2
𝑓𝐻 (𝑥) = 𝑡𝑎𝑛𝑠𝑖𝑔 (𝑥) = −1 (2)
1 + exp (−2 ∗ 𝑥) 2.2. Improved Bayesian Neural Networks (IBNN) Model. The
structure of IBNN is illustrated in Figure 4, besides the time
𝑓𝑂 (𝑥) = 𝑝𝑢𝑟𝑒𝑙𝑖𝑛 (𝑥) = 𝑥 (3) inputs like time and day-type and meteorology inputs like
After obtaining the forecast model, the forecast load can temperature and humidity, the model also uses the historical
be given by (4). Here in the following equation, IV is a vector load data as inputs.
contains the latest values of inputs which form this forecast Different from traditional time-series with continuous
model, and there may be one or several past historical load historical values or fixed interval historical values feedback,
̂ 𝐹𝐵 (IV) is the forecast load value by the built IBNN
values. 𝐿 an improved model with highly correlation historical values
model. of prediction target as inputs is designed to obtain relatively
higher accuracy but shorter computing time. The chosen
̂ 𝐹𝐵 (IV) = 𝑓𝑂 (𝑓𝐻 (IV ⋅ W𝐻 + b𝐻) ⋅ W𝑂 + b𝑂 )
𝐿 (4) historical data may be only one vector which is close to the
From a time-series view, a dynamic BNN method is forecast interval or several vectors from same time intervals
widely applied in prediction models. The past data are seen and close intervals.
as feedback in the model. In this paper, the time-series Pearson correlation coefficient (𝑃) and Spearman corre-
forecasting problem is defined as Nonlinear Autoregressive lation coefficient (𝑆) are used to measure their correlations.
with External (Exogenous) Input (NARX), with feedback If each variable has 𝑁 scalar observations, the Pearson
connections enclosing several layers of the network [27, 28]. correlation coefficient, 𝑃, is defined as [29–31]
The Bayesian time-series neural network structure can be
simplified as Figure 3, where TDL means tapped delay line.
However, the performance of the time-series BNN model 1 𝑁 𝑀1𝑖 − 𝜇𝑀1 𝑀2𝑖 − 𝜇𝑀2
𝑃 (𝑀1 , 𝑀2 ) = ∑( )( ) (5)
is affected by the setup of time delay line, and with the 𝑁 − 1 𝑖=1 𝜎𝑀1 𝜎𝑀2
4 Mathematical Problems in Engineering

1 1

Spearman Correlation
Pearson Correlation

Coefficient
Coefficient

0.5
0.5

0
0

0 7 14 21 28 0 7 14 21 28
Past Time (day) Past Time (day)
(a) (b)

Figure 5: Correlation coefficients of past time load vectors of No. 11.

1 1

Spearman Correlation
Pearson Correlation

0.8
Coefficient

Coefficient
0.6
0.5
0.4
0.2
0 0
−0.2
0 7 14 21 28 0 7 14 21 28
Past Time (day) Past Time (day)
(a) (b)

Figure 6: Correlation coefficients of past time load vectors of No. 17.

1 1
Pearson Correlation

Spearman Correlation

0.8
Coefficient

0.6
Coefficient

0.4 0.5
0.2
0
0
−0.2
0 7 14 21 28 0 7 14 21 28
Past Time (day) Past Time (day)
(a) (b)

Figure 7: Correlation coefficients of past time load vectors of No. 50.

where 𝜇𝑀1 and 𝜎𝑀1 are the mean and standard deviation line is the line connecting the highest value point of the
of 𝑀1 , respectively, and 𝜇𝑀2 and 𝜎𝑀2 are the mean and correlation coefficients in the delays of the daily cycle. It can
standard deviation of 𝑀2 . The above equation can also be be clearly seen from the dotted line and the calculated values
described as correlation coefficient based on the covariance that the highest value of the correlation coefficients in the
of 𝑀1 and 𝑀2 , daily cycle is obtained by delaying an integral multiple of
The Spearman correlation coefficient (𝑆) [32, 33], 𝑆, can 24 hours, that is, the same time period of the past day. In
be computed by the following equation: addition, the calculation results of No. 11 and No. 50 can
also be seen that the correlation coefficients calculated by
∑𝑖 (𝑀1𝑖 − 𝑀1 ) (𝑀2𝑖 − 𝑀2 ) the delay of weekly cycle may be slightly higher than other
𝑆 (𝑀1 , 𝑀2 ) = (6) daily calculation values. This is because most households
2 2
√ ∑𝑖 (𝑀1𝑖 − 𝑀1 ) ∑𝑖 (𝑀2𝑖 − 𝑀2 ) have significant differences between working days and rest
days and individual households have a special electricity
Figures 5–7 are the correlation coefficients results of three consumption cycle mode in one week, which leads the
residences from a sample of 300 homes supplied by Ausgrid current forecast value have a high correlation values with the
from July 2010 to June 2011, which are No. 11, No. 17, and No. same time delays on the same day-type of past weeks. In order
50, respectively. to further numerically compare the correlation coefficients,
As can be seen from the figures, the solid line indicates the partial operation results are listed in Table 1.
the correlation coefficients calculated by different time delay It can be seen from the table that the correlations between
vectors and the current predicted value, and the dotted the historical electricity consumption over the past two hours
Mathematical Problems in Engineering 5

Table 1: Correlation coefficients of past time load vectors of three residences.

Past No. 11 No. 17 No. 50
Load/number of Pearson Spearman Pearson Spearman Pearson Spearman
home Correlation Correlation Correlation Correlation Correlation Correlation
L𝑛−1 0.82 0.83 0.85 0.79 0.21 0.53
L𝑛−2 0.64 0.73 0.71 0.67 0.10 0.39
L𝑛−3 0.54 0.64 0.59 0.55 0.07 0.33
L𝑛−4 0.48 0.56 0.50 0.44 0.04 0.21
L𝑛−5 0.42 0.50 0.42 0.39 0.01 0.13
L𝑛−6 0.36 0.42 0.35 0.31 -0.02 0.07
L𝑛−24/t𝑖 0.56 0.63 0.56 0.56 0.14 0.30
L𝑛−24/t𝑖 ⋅2 0.51 0.54 0.50 0.53 0.14 0.31
L𝑛−24/t𝑖 ⋅7 0.55 0.60 0.40 0.47 0.17 0.35
L𝑛−24/t𝑖 ⋅7⋅2 0.50 0.52 0.35 0.40 0.16 0.32

and the current forecast period are significantly lower and are only the MAPE is considered, large percentage errors will be
even generally lower than the past delay time of day cycles obtained in small power consumption periods, so that the
to the current forecast period. For families with obvious MAPE throughout the day will become very large. The low
weekly cycles, as No. 11 and No. 50, the weekly same time power consumption is often only tens to hundreds of watts
period correlations are even higher than the closest daily per hour, but at peak times, electricity is often used more
delay correlation. than kW per hour. In other words, it is more important to
It can also be obtained from the above analysis that there accurately predict the electricity consumption during peak
are differences in the electricity consumption patterns of hours, so the mean absolute error (MAE) is also adopted to
different households, and different input historical data make evaluate the prediction accuracy. The MAE can be calculated
the forecast model different and the performance different. as
Specific analysis of specific targets is required, so that it is 󵄨 󵄨
∑𝑁 󵄨󵄨𝑦 − 𝑦̂𝑖 󵄨󵄨󵄨
possible to select predictive variables with higher prediction 𝑀𝐴𝐸 = 𝑖=1 󵄨 𝑖 (9)
accuracy. To find how to select the historical data as inputs 𝑁
can better improve the performance of forecast model and Another statistic metric, regression coefficient, 𝑅, is also
how the model related factors impact the performance, applied to indicate the amount of variance explained by the
several compared forecast models are designed in the next model. 𝑅 is defined as (10), 𝑦𝑖 is the mean of the observed data
section. value, 𝑦̂𝑖 is the mean of the predicted values, ∑𝑁 ̂𝑖 )2
𝑖=1 (𝑦𝑖 − 𝑦

2.3. Performance Evaluation. The following four indexes are is the residual sum of squares, and ∑𝑁 2
𝑖=1 (𝑦𝑖 − 𝑦𝑖 ) is the
selected to evaluate the forecast model performance. MSE is explained sum of squares. 𝑅 can take on any value between
the mean square error or the residual mean square. Equation 0 and 1, with a value closer to 1 indicating that a greater
(7) shows how to calculate the MSE value. 𝑦̂𝑖 is a vector proportion of variance is accounted for by the model [35].
of 𝑁 predictions, and 𝑦𝑖 is the vector of observed values ∑𝑁 𝑦𝑖 − 𝑦̂𝑖 ) (𝑦𝑖 − 𝑦𝑖 )
𝑖=1 (̂
corresponding to the inputs to the function which generates 𝑅= (10)
the predictions. 𝑒𝑖 stands for the square of the errors. An MSE √ ∑𝑁
2 2
𝑦𝑖 − 𝑦̂𝑖 ) √∑𝑁
𝑖=1 (̂ 𝑖=1 (𝑦𝑖 − 𝑦𝑖 )
value closer to 0 indicates a fitting that is more useful for
prediction when the model is not overfitted.
3. Inputs Selection and Relative
1 𝑁 2 1 𝑁 2
Factors Analysis
𝑀𝑆𝐸 = ∑ (𝑒𝑖 ) = ∑ (𝑦̂𝑖 − 𝑦𝑖 ) (7)
𝑁 𝑖=1 𝑁 𝑖=1
In this part, aiming at giving a better chosen on input vectors
MAPE is the mean absolute percentage error, which is of IBNN method load forecasting model, authors define
accuracy evaluation and comparison of a forecasting method several models of BNN to analyze different factors.
in statistics and especially a widely used metric in energy
[32, 34]. The definition of MAPE is as follows: 3.1. Basic Inputs Selection of IBNN Model. At first, an IBNN
model is built with 16 load related inputs, named as BNN 16.
1 󵄨󵄨󵄨 𝑦 − 𝑦̂𝑖 󵄨󵄨󵄨󵄨
𝑁
𝑀𝐴𝑃𝐸 = ∑ 󵄨󵄨󵄨 𝑖 󵄨 × 100% (8) The vectors of inputs are shown in
𝑁 𝑖=1 󵄨󵄨 𝑦𝑖 󵄨󵄨󵄨
V𝐼 = [t, dt , T, T𝑛−1 , T𝑛−2 , RH, RH𝑛−1 , RH𝑛−2 , L𝑛−1 , L𝑛−2 ,
However, for small household electricity consumption, (11)
the daily electricity consumption varies greatly with time. If L𝑛−3 , L𝑛−4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 ]
6 Mathematical Problems in Engineering

1.5

Load (kWh) 0.5

0 Sunday Monday
10/10/2010 11/10/2010 12/10/2010
Time

Real BNN_2 months

BNN_2 weeks BNN_0.5 hours
BNN_1 month
Figure 8: Comparison of real data and forecasting values under BNN 0.5hour, BNN 2 weeks, BNN 1 month, and BNN 2 months of 8 hidden
neurons.

Here in the above equation, the input factors which are and discuss the built forecasting models as it has the largest
mostly considered in the existed models would be the time average daily electricity demand. The average daily electricity
category and the meteorological category. In this article, consumption is 36.83 kWh and the average interval electricity
time of everyday, t, day-type (which is defined as integers consumption is 0.77 kWh. As the data is observed every half
from 1 to 7 to express Monday to Sunday and 8 to express hour, the total sample number is 17520 in one whole year
special holidays), dt , ambient temperature, T, and relative with 365 days. The training set, validation set, and test set
humidity, RH, are firstly considered as the inputs. 𝑛 is used is created by a fixed partition algorithm called ‘divideind’ in
to represent the series order number of historical sample Matlab with 60%, 20%, and 20% respectively.
data’s intervals, and t𝑖 is the observed interval time. For
instance, T𝑛−1 means the temperature vector observed from 3.2. Models with Different Periods of Historical Data. As the
the past one interval. Due to the fact that there is no record electricity demand is testified to vary with changes of the
before the historical first interval, here use 𝑇1 as the initial weather factors, the period of historical data, to great extent,
value to complement the vector. If other vectors lack some decides the performance of obtained forecasting model. In
items, use the same complement method. According to [36], Table 2, number of input vectors is increased with extending
for marine climate or inshore areas, relative humidity also the range of historical data.
may affect the consumption on electricity. As in this model, Due to the demand usually with a weekly cycle, here only
there is no time delay in the NN structure and there is a add vectors of real load data at the same time in past weeks.
study suggesting that human’s perception of temperature and The model is studied with real load data in the next part.
relative humidity is delay with some time [37]. Taking this Because the time interval of the data applied in this paper
study as a reference, then adding historical environmental is 0.5 hour, to clearly illustrate the different time period, the
data as input vectors also very likely can improve the BNN compared models are named with real time period in inputs
prediction model’s performance. With similar meaning of selection.
subscript, the eight historical data vectors in the back are According to the models built in Table 2, the following
actual load in the first past interval, L𝑛−1 , second past interval, values are obtained from the IBNN program. When increas-
L𝑛−2 , third past interval, L𝑛−3 , fourth past interval, L𝑛−4 ,same ing the number of inputs with historic data, in Table 3, the
interval of yesterday, L𝑛−24/t𝑖 , same interval of the day before results show that performance of training set will be slightly
yesterday, L𝑛−24/t𝑖 ⋅2 , same interval in last week of the same better with the raising number of inputs. However, the test
day-type, L𝑛−24/t𝑖 ⋅7 , and same interval in week before last week set is obtained with random algorithm, and from the table,
of the same day-type, L𝑛−24/t𝑖 ⋅7⋅2 , respectively. These eight the performance of test set does not keep improving with
inputs of the forecast model basically cover the most relevant the increasing inputs number. This suggests if the real load
historical load values within past two weeks. The target vector observations for training are enough, there is no need to add
is defined as L𝑛 , which is the time series of last observation. too many historic inputs vectors.
The actual load data used in this article are from a sample The comparison between real data and the IBNN model
of 300 homes supplied by Ausgrid from July 2010 to June 2011. with different numbers of historic inputs is illustrated in
The related weather information is from Australian Bureau Figure 8. Obviously, the prediction model BNN 0.5 hour is
of Meteorology. No. 17 of 300 homes is selected to validate significantly less effective than the other three. The graph also
 Mathematical Problems in Engineering

Table 2: Inputs of IBNN with different period of historic data.

Model Inputs vectors (number of input vectors)
BNN noL [t, dt , T, T𝑛−1 , T𝑛−2 , RH, RH𝑛−1 , RH𝑛−2 ] (8)
BNN 0.5hours [t, dt , T, T𝑛−1 , T𝑛−2 , RH, RH𝑛−1 , RH𝑛−2 , L𝑛−1 ] (9)
BNN 1hour [t, dt , T, T𝑛−1 , T𝑛−2 , RH, RH𝑛−1 , RH𝑛−2 , L𝑛−1 , L𝑛−2 ] (10)
BNN 2hours [t, dt , T, T𝑛−1 , T𝑛−2 , RH, RH𝑛−1 , RH𝑛−2 , L𝑛−1 , L𝑛−2 , L𝑛−3 , L𝑛−4 ] (12)
BNN 2days [t, dt , T, T𝑛−1 , T𝑛−2 , RH, RH𝑛−1 , RH𝑛−2 , L𝑛−1 , L𝑛−2 , L𝑛−3 , L𝑛−4 , L𝑛−24/t𝑖 ] (13)
BNN 3days [t, dt , T, T𝑛−1 , T𝑛−2 , RH, RH𝑛−1 , RH𝑛−2 , L𝑛−1 , L𝑛−2 , L𝑛−3 , L𝑛−4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 ] (14)
BNN 1week [t, dt , T, T𝑛−1 , T𝑛−2 , RH, RH𝑛−1 , RH𝑛−2 , L𝑛−1 , L𝑛−2 , L𝑛−3 , L𝑛−4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 ] (15)
BNN 2weeks (BNN 16) [t, dt , T, T𝑛−1 , T𝑛−2 , RH, RH𝑛−1 , RH𝑛−2 , L𝑛−1 , L𝑛−2 , L𝑛−3 , L𝑛−4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 ] (16)
BNN 1month [t, dt , T, T𝑛−1 , T𝑛−2 , RH, RH𝑛−1 , RH𝑛−2 , L𝑛−1 , L𝑛−2 , L𝑛−3 , L𝑛−4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 , L𝑛−24/t𝑖 ⋅7⋅3 and L𝑛−24/t𝑖 ⋅7⋅4 ] (18)
BNN 2months [t, dt , T, T𝑛−1 , T𝑛−2 , RH, RH𝑛−1 , RH𝑛−2 , L𝑛−1 , L𝑛−2 , L𝑛−3 , L𝑛−4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 , L𝑛−24/t𝑖 ⋅7⋅3 , L𝑛−24/t𝑖 ⋅7⋅4 , L𝑛−24/t𝑖 ⋅7⋅5 , L𝑛−24/t𝑖 ⋅7⋅6 , L𝑛−24/t𝑖 ⋅7⋅7 and L𝑛−24/t𝑖 ⋅7⋅8 ] (22)
7
8 Mathematical Problems in Engineering

Table 3: Results of IBNN models with different range of history data.

Model IBNN (Inputs Training set (60%) Test set (20%) Computing time
number/Hidden neurons) MSE R MAPE (%) MSE R MAPE (%) (s)
BNN 0.5hours (9/8) 9.83e-2 8.71e-1 11.58 9.43e-2 8.71e-1 11.66 9
BNN 1hour (10/8) 9.63e-2 8.72e-1 11.30 9.54e-2 8.76e-1 11.88 10
BNN 2hours (12/8) 9.46e-2 8.77e-1 11.45 1.00e-1 8.58e-1 12.32 14
BNN 2days (13/8) 8.83e-2 8.85e-1 10.80 8.62e-2 8.84e-1 10.36 14
BNN 3days (14/8) 8.66e-2 8.86e-1 10.65 8.45e-2 8.91e-1 10.60 13
BNN 1week (15/8) 8.37e-2 8.92e-1 10.64 8.97e-2 8.76e-1 10.68 16
BNN 2weeks (16/8) 8.43e-2 8.89e-1 10.63 8.57e-2 8.89e-1 10.54 18
BNN 1month (18/8) 8.30e-2 8.92e-1 10.81 7.96e-2 8.92e-1 10.35 34
BNN 2months (22/8) 7.44e-2 9.05e-1 10.67 8.30e-2 8.85e-1 11.44 37

1.5

1
Load (kWh)

0.5

0 Sunday Monday
10/10/2010 11/10/2010 12/10/2010
Time

Real BNN_12h
BNN_0.5h BNN_24h
Figure 9: Comparison of real data and forecasting values under BNN 0.5h, BNN 12h, and BNN 24h of 8 hidden neurons.

shows that when historical data reaches a certain time range, To predict the electricity load of several hours in advance,
using more input vectors with longer history data can no’t five forecasting models which applied the same sample data
apparently improve the accuracy of the forecasting model and are built to examine the effectiveness of the IBNN model in
only increase the computing time. the above. Table 5 gives the results of the model with different
forecasting time ahead.
3.3. Models with Different Prediction Time Range. To test the From the table it is clearly that the IBNN forecasting
accuracies of the IBNN models under different prediction model can give a relatively high accuracy in very short-term
time, compared models with 16 inputs are designed, which forecasting. With the prediction time range extending, the
are listed in Table 4. MSE and R-square values become lower. However, when the
For better upgrading the forecast learning model, in range is more than 4 hours, there is not evident changing
our design the chosen historical inputs should be the latest trend from the results. That means when existing sample data
observed actual load data. With this consideration in Table 4, are used to forecast the load in a very short time in advance,
it can be seen that the load historical input vectors are the accuracy is very high, but when to forecast the load in the
changed according to the prediction time range. First is a following few hours or a day, the accuracy seems to be around
model to forecast load half hour ahead, second is to forecast a lower boundary.
load 1 hour ahead, until the last is to forecast load 24 As the discussion above, Figure 9 shows two days’ com-
hours ahead. The models are distinguished by the subscripts parison between the real data and the forecasting data with
which represent the prediction time range. The prediction of different forecasting times. From the graph, it is apparently
electricity demand is usually used for energy management that the dashed line, which is results from the model of
and electricity devices control, so with different optimization forecasting 0.5 hour ahead, is very close to the black line,
methods, the forecasting model needs to have reasonable which is the actual load observation. The other two lines are
prediction time. results from forecasting models of 12 hours and 24 hours
 Mathematical Problems in Engineering

Table 4: Inputs of IBNN under different prediction time.

Model Inputs vectors (number of input vectors)
BNN 0.5h V𝐼𝑝𝑡1 = [t, dt , T𝑛−1 , T𝑛−2 , T𝑛−3 , RH𝑛−1 , RH𝑛−2 , RH𝑛−3 , L𝑛−2 , L𝑛−3 , L𝑛−4 , L𝑛−5 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 ]
BNN 1h V𝐼𝑝𝑡2 = [t, dt , T𝑛−1/t𝑖 , T𝑛−1/t𝑖 −1 , T𝑛−1/t𝑖 −2 , RH𝑛−1/t𝑖 , RH𝑛−1/t𝑖 −1 , RH𝑛−1/t𝑖 −2 , L𝑛−−1/t𝑖 −1 , L𝑛−1/t𝑖 −2 , L𝑛−1/t𝑖 −3 , L𝑛−1/t𝑖 −4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 ]
BNN 4h V𝐼𝑝𝑡3 = [t, dt , T𝑛−4/t𝑖 , T𝑛−4/t𝑖 −1 , T𝑛−4/t𝑖 −2 , RH𝑛−4/t𝑖 , RH𝑛−4/t𝑖 −1 , RH𝑛−4/t𝑖 −2 , L𝑛−−4/t𝑖 −1 , L𝑛−4/t𝑖 −2 , L𝑛−4/t𝑖 −3 , L𝑛−4/t𝑖 −4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 ]
BNN 12h V𝐼𝑝𝑡4 = [t, dt , T𝑛−12/t𝑖 , T𝑛−12/t𝑖 −1 , T𝑛−12/t𝑖 −2 , RH𝑛−12/t𝑖 , RH𝑛−12/t𝑖 −1 , RH𝑛−12/t𝑖 −2 , L𝑛−−12/t𝑖 −1 , L𝑛−12/t𝑖 −2 , L𝑛−12/t𝑖 −3 , L𝑛−12/t𝑖 −4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 ]
BNN 24h V𝐼𝑝𝑡5 = [t, dt , T𝑛−24/t𝑖 , T𝑛−24/t𝑖 −1 , T𝑛−24/t𝑖 −2 , RH𝑛−24/t𝑖 , RH𝑛−24/t𝑖 −1 , RH𝑛−24/t𝑖 −2 , L𝑛−−24/t𝑖 −1 , L𝑛−24/t𝑖 −2 , L𝑛−24/t𝑖 −3 , L𝑛−24/t𝑖 −4 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 ]
9
10 Mathematical Problems in Engineering

Table 5: Results of BNN 0.5h, BNN 12h, and BNN 24h.

Training set (60%) Test set (20%)
Model MAPE MAPE Computing time (s)
MSE R MSE R
(%) (%)
BNN 0.5h 8.35e-2 8.90e-1 10.62 8.71e-2 8.86e-1 10.65 7
BNN 1h 1.32e-1 8.20e-1 12.26 1.37e-1 8.15e-1 13.00 12
BNN 4h 1.80e-1 7.43e-1 14.41 1.91e-1 7.31e-1 14.47 26
BNN 12h 1.84e-1 7.39e-1 14.45 1.83e-1 7.32e-1 14.77 8
BNN 24h 1.82e-1 7.43e-1 14.41 1.84e-1 7.30e-1 14.56 12

Table 6: Results of BNN 16 model with different periods of historical data.

Training set (60%) Test set (20%)

Historic period MAPE MAPE Computing time (s)
MSE R MSE R
(%) (%)
1 month 6.15e-2 9.61e-1 9.20 1.32e-1 9.14e-1 9.36 4
3 months 9.94e-2 9.00e-1 11.73 1.54e-1 8.47e-1 11.92 4
Half year 7.61e-2 8.91e-1 10.88 8.88e-2 8.76e-1 11.05 8
9 months 8.26e-2 8.91e-1 10.76 9.05e-2 8.83e-1 10.87 11
1 year 8.28e-2 8.92e-1 10.63 8.89e-2 8.83e-1 10.59 12

in advance respectively, it cannot easily define which one is den neurons on the TS BNN forecasting model. From the
better. results it can be found that TS BNN with the increasing of
the closed-loop input time delays and the performance of pre-
3.4. Range Requirement of Load Sample Data. Evidently that diction model is evidently improved. However, when the time
the forecasting model will have higher accuracy with longer delays are set to 50 intervals, 25 hours, which is one more
period of historic electricity data. However, it is more useful hour than one day, the running time reaches half past three
to give a forecasting model which can predict the electricity minutes.
load even with a relative short load recorded period, such as From Table 7, it shows that, with the increasing of hidden
one year or just several months. To discuss such problem in neurons number, the MSE of the training set becomes lower
this article, the authors try to assess the IBNN forecasting and R-square values become higher; these performances
model (BNN 16) with five different periods of actual sample seem facially better. However, the MSE values of test set do
data under 8 hidden neurons, which are 1 month, 3 months, not decrease and the R-square values do not increase com-
half year, 9 months, and a whole year, respectively. pletely with the increasing of hidden neurons. Above com-
The obtained values in Table 6 show that, with longer puting results show that prediction model can not obtain
history period, the MSE becomes higher and the R-square completely better performance with the number of hidden
becomes lower. However, the performance of the test set neurons. In Table 7, when the hidden neurons of BNN 16
gets better. As the case studied here is one home in Sydney, are 8, the MSE and R-square of test set have reached close
seasonal weather factors must be taken into consideration. to the optimal value. When the hidden neurons are more
From the table, it can be concluded that, to obtain higher than 8, the computing results are in repeated fluctuations.
accuracy of prediction, at least half year load record should Correspondingly, it can be seen that, with the increase of
be used for building the forecasting model. hidden neurons, the operation time of BNN prediction model
is obviously prolonged. Therefore, it is very important to
4. Results and Discussion select appropriate hidden neurons according to the model
to improve the performance and efficiency of BNN model
4.1. Comparison between IBNN and TSNN. The same data set and reduce the computation time. In general, the numbers of
as above is chosen for the following validation and analysis. hidden neurons are within the range of 3∼8. It also should
The authors use a toolbox named Neural Net in Matlab to to be noted that the best performance of TS BNN model
run the designed forecasting model. The time-series Bayesian is under the condition with 50 time delays and 2 hidden
Neural Networks (TS BNN) method has been described in neurons, with more than three minutes of running time. The
the introduction part. The inputs of time-series BNN are closest performance of BNN 16 is under the condition of 8
defined as follows: hidden neurons with only six seconds. Obviously BNN 16
forecast model significantly reduces the time.
V𝐼 𝑇𝑆 = [t, dt , T, RH] (12) The randomly selected comparison between actual re-
Table 7 lists the computing results of TS BNN model and corded data and the IBNN model can be seen in Figure 10,
BNN 16 under different hidden neurons or time delays. Here which is from 8th Oct 2010 to 15th Oct 2010. Here the
the authors firstly analyze effect of the time delays and hid- illustration of one week’s forecasting load values and real load
Mathematical Problems in Engineering 11

Table 7: Results of time-series BNN and BNN 16 model.

Model Training set (60%) Test set (20%)
MAPE MAPE Computing time (s)
Time delay Hidden neurons MSE R MSE R
(%) (%)
1 2 1.07e-1 8.58e-1 11.75 1.05e-1 8.57e-1 11.86 6
2 2 1.06e-2 8.58e-1 11.79 1.04e-1 8.63e-1 12.25 5
TS BNN 50 2 8.79e-2 8.83e-1 10.74 8.76e-2 8.87e-1 11.34 192
1 6 1.01e-2 8.65e-1 11.46 9.80e-2 8.74e-1 11.67 27
1 8 9.84e-2 8.68e-1 11.73 1.05e-1 8.66e-1 11.50 29
- 2 9.27e-2 8.78e-1 11.38 9.41e-2 8.75e-1 11.40 2
- 3 8.90e-2 8.84e-1 11.41 9.65e-2 8.68e-1 11.55 2
- 5 8.56e-2 8.88e-1 10.93 9.46e-2 8.75e-1 10.72 3
- 8 8.26e-2 8.92e-1 10.64 8.56e-2 8.87e-1 10.52 6
- 10 8.16e-2 8.93e-1 10.54 8.77e-2 8.88e-1 10.92 7
BNN 16 - 15 7.78e-2 8.99e-1 10.55 8.97e-2 8.80e-1 10.42 18
- 18 7.72e-2 8.99e-1 10.36 8.95e-2 8.83e-1 10.87 139
- 20 7.62e-2 9.01e-1 10.43 8.80e-2 8.83e-1 10.94 148
- 30 7.19e-2 9.07e-1 10.13 9.29e-2 8.78e-1 10.98 355
- 50 6.47e-2 9.17e-1 9.62 9.41e-2 8.73e-1 11.59 732
- 100 5.81e-2 9.26e-1 9.04 1.07e-1 8.59e-1 12.50 1856

Table 8: Results of time-series BNN under different input vectors.

TS-BNN (Hidden Training set (60%) Test set (20%) Computing

neurons 4) Input vectors MSE R MAPE (%) MSE R MAPE (%) time (s)
[t, dt , T, RH] 9.97e-2 8.61e-1 11.67 1.02e-1 8.55e-1 11.78 15
[dt , T, RH] 1.07e-1 8.55e-1 11.82 1.11e-1 8.58e-1 12.49 7
[t, dt , RH] 1.08e-1 8.55e-1 11.60 1.05e-2 8.61e-1 11.45 9
[t, dt , T] 1.04e-1 8.60e-1 11.76 1.14e-1 8.52e-1 11.13 15
[t, T, RH] 1.01e-1 8.66e-1 11.91 1.09e-1 8.53e-1 11.44 5

2 values comparison is just randomly selected. From the figure,

1.8 it is obvious that the forecasting model performs well with
the trend varying of the real data. However, when the actual
1.6 data changed severely, the deviation between actual data and
1.4 forecast data increased to a certain extent.
Table 8 discussed the influence of input vectors under
1.2
Load (kWh)

TS BNN model. Due to the time delay of historical load data,

1 the TS BNN model only uses four related factors as input
0.8 vectors, which are time, temperature, relative humidity, and
day-type. The approach used here is to eliminate one of the
0.6 relevant factors to observe the results of the computation.
0.4 From the above results, although it is difficult to say
what kind of factors related has more impact on electricity
0.2
consumption of the consumer, by comparing the situations
0 Fri. Sat. Sun. Mon. Thue. Wed. Thur. of eliminating a factor and using total four input vectors, the
0 9/10 10/10 11/10 12/10 13/10 14/10 15/10
results show whatever reducing anyone of the four inputs; the
Time (2010)
performance of the prediction model became lower than the
IBNN model with total four inputs. It means that all the four factors
Real should have influence on the consumer’s electricity load.
To analyze the impacts of different considered input
Figure 10: Comparison of real data and forecasting values under factors on IBNN model, here the authors define 13 models
IBNN of 8 hidden neurons. with different numbers of inputs vectors (listed in Table 9).
 12

Table 9: Comparison models with the basic IBNN with 16 inputs.

Model No. Inputs vectors describe (number of input vectors) Inputs selection
1 BNN 16 (16) [t, dt , T, T𝑛−1 , T𝑛−2 , RH, RH𝑛−1 , RH𝑛−2 , L𝑛−1 , L𝑛−2 , L𝑛−3 , L𝑛−4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 ]
2 No historical Load (8) [t, dt , T, T𝑛−1 , T𝑛−2 , RH, RH𝑛−1 , RH𝑛−2 ]
3 Less T (14) [t, dt , T, RH, RH𝑛−1 , RH𝑛−2 , L𝑛−1 , L𝑛−2 , L𝑛−3 , L𝑛−4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 ]
4 Less RH (14) [t, dt , T, T𝑛−1 , T𝑛−2 , RH, L𝑛−1 , L𝑛−2 , L𝑛−3 , L𝑛−4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 ]
5 No T (13) [t, dt , RH, RH𝑛−1 , RH𝑛−2 , L𝑛−1 , L𝑛−2 , L𝑛−3 , L𝑛−4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 ]
6 No RH (13) [t, dt , T, T𝑛−1 , T𝑛−2 , L𝑛−1 , L𝑛−2 , L𝑛−3 , L𝑛−4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 ]
7 No time (15) [dt , T, T𝑛−1 , T𝑛−2 , RH, RH𝑛−1 , RH𝑛−2 , L𝑛−1 , L𝑛−2 , L𝑛−3 , L𝑛−4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 ]
8 No day-type (15) [t, T, T𝑛−1 , T𝑛−2 , RH, RH𝑛−1 , RH𝑛−2 , L𝑛−1 , L𝑛−2 , L𝑛−3 , L𝑛−4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 ]
9 Only Load (8) [L𝑛−1 , L𝑛−2 , L𝑛−3 , L𝑛−4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 ]
10 Only T and Load (9) [T, L𝑛−1 , L𝑛−2 , L𝑛−3 , L𝑛−4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 ]
11 Time, T and Load (10) [t, T, L𝑛−1 , L𝑛−2 , L𝑛−3 , L𝑛−4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 ]
12 day-type, T and Load (10) [dt , T, L𝑛−1 , L𝑛−2 , L𝑛−3 , L𝑛−4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 ]
13 T, RH and Load (10) [T, RH, L𝑛−1 , L𝑛−2 , L𝑛−3 , L𝑛−4 , L𝑛−24/t𝑖 , L𝑛−24/t𝑖 ⋅2 , L𝑛−24/t𝑖 ⋅7 , L𝑛−24/t𝑖 ⋅7⋅2 ]
T-Temperature RH-Relative Humidity.
Mathematical Problems in Engineering
Mathematical Problems in Engineering 13

Table 10: Results on analyzing impacts of related input factors.

3 Hidden neurons 8 Hidden neurons 15 Hidden neurons

Training set (60%) Training set (60%) Training set (60%)
Inputs impacts (number of input vectors)
MAPE MAPE MAPE
MSE R MSE R MSE R
(%) (%) (%)
1 BNN 16 (16) 9.02e-2 8.82e-1 11.24 8.71e-2 8.86e-1 10.64 8.48e-2 8.85e-1 10.35
2 No Load (8) 2.77e-1 5.84e-1 18.06 2.63e-1 5.86e-1 17.84 2.32e-1 6.25e-1 17.87
3 Less T (14) 9.07e-2 8.82e-1 11.29 9.02e-2 8.82e-1 10.51 8.13e-2 8.95e-1 10.41
4 Less RH (14) 9.11e-2 8.78e-1 11.12 8.23e-2 8.85e-1 10.63 8.35e-2 8.89e-1 10.48
5 No T (13) 9.21e-2 8.74e-1 11.18 9.10e-2 8.74e-1 10.66 9.07e-2 8.79e-1 10.41
6 No RH (13) 9.26e-2 8.76e-1 11.14 8.81e-2 8.83e-1 10.67 8.43e-2 8.83e-1 10.38
7 No time (15) 8.96e-2 8.85e-1 11.11 9.05e-2 8.79e-1 11.10 8.39e-2 8.91e-1 10.78
8 No day-type (15) 9.64e-2 8.79e-1 11.42 9.15e-2 8.86e-1 10.76 9.04e-2 8.89e-1 10.35
9 Only Load (8) 9.63e-2 8.70e-1 11.29 9.38e-2 8.76e-1 11.05 9.22e-2 8.83e-1 10.94
10 Only T and Load (9) 8.78e-2 8.88e-1 11.19 8.44e-2 8.90e-1 10.89 8.85e-2 8.83e-1 10.80
11 Time, T and Load (10) 9.14e-2 8.66e-1 11.17 8.65e-2 8.91e-1 10.86 8.11e-2 8.91e-1 10.52
12 day-type, T and Load (10) 9.33e-2 8.77e-1 11.53 8.70e-2 8.81e-1 11.14 8.21e-2 8.90e-1 10.93
13 T, RH and Load (10) 9.31e-2 8.71e-1 11.23 8.39e-2 8.91e-1 10.90 8.63e-2 8.86e-1 10.65

Besides discussing the impact of each factor, here the the robustness of the prediction model. In other words,
authors also want to find if increasing the number of historic it may also be noted that the increasing of inputs vectors
temperature or humidity inputs can improve the forecasting on historical load data may improve the stability of the
performance. Under this consideration, 2 models called Less prediction model, and the influence of other factors on the
T and Less RH are defined, unlike applying another two performance is relatively reduced.
past historical input vectors of temperature and humidity, In order to further validate the effectiveness of the model,
applying only one temperature input and one humidity input. Table 11 uses the same BNN 16 structure to train the predic-
As mentioned above, the effects of different input factors tion models for 15 households which are randomly selected
on the model are analyzed. The MSE and R-square of training from the same Ausgrid yearly data set.
set are calculated with 3, 8, and 15 hidden neurons, respec- As can be found from Table 11, the performances of TSNN
tively, to find similar results and exclude the impact on system and IBNN training are very close, but the computing time is
randomly selecting data. reduced by an average of 31 times. The shortest reduction time
From the calculated values in Table 10, there are the is from No. 28 family, which is 8 times shorter, and the longest
following results. First, the comparison of the ‘Less T’ and reduction time is from No. 55 family which is shortened by
‘Less RH’ with the normal 16 inputs model (BNN 16) shows more than 83 times.
that increasing the input number of historic weather data
cannot improve the model’s forecasting performance and 4.2. Comparison with Other Prediction Methods. The com-
surprisingly slightly lower the performance. However, the parison is between the proposed IBNN method and the
results in the table only show that this unit is not very sensitive methods from the MATLAB Statistics and Machine Learning
with the temperature and humidity or there is a deduction Toolbox. The parameters of machine learning methods are
that adding historical load vectors as inputs enhances the defined as defaults. Table 12 lists all the computing results
robustness of the prediction model. Second, deleting one with the inputs the same as BNN 16 of No. 17 household
related factor in the input vectors, there are ‘no T’, ‘no RH’, ‘no dataset.
time’, and ‘no day-type’, four compared items with the results As can be seen from Table 12, the MSE and R values
of BNN 16. The temperature has the largest impact on the of the IBNN method show the best results compared to
model’s accuracy as the MSE of no T becomes the biggest and other machine learning methods. Although the MAE value
R-square gets the lowest value. There is not noticeable effect of of IBNN is slightly inferior to the Bagged Trees method,
the other three factors from the obtained values. However, it it is greatly reduced in computation time by nearly five
cannot just from the simple calculated values define that these times. In order to more intuitively compare various methods,
three factors have positive effects on improving the model. the authors have defined the corresponding number as the
Time and day-type to the basic load forecasting knowledge horizontal coordinate to draw the four evaluation indicators,
will significantly affect some users’ consumption mode, so the respectively, as shown in the Figure 11.
above results only can show that, under this IBNN model, the
effect of these three factors is not obvious. 5. Conclusion
From Tables 9 and 10, the comparison of results
by TS BNN model and BNN 16 model validates the first Traditional time-series BNN load forecast model has some
deduction that the input historical load vectors enhance problems when applied in residential load forecast area, such
14 Mathematical Problems in Engineering

Table 11: Results on different datasets.

IBNN (BNN 16) TSNN (TS BNN, time delays = 25)
Home number\Methods
MSE R MAE Time MSE R MAE Time
11 2.47e-2 0.87 0.10 17.48 2.38e-2 0.89 0.10 476.87
13 3.43e-2 0.85 0.09 12.05 3.24e-2 0.88 0.10 350.22
16 2.86e-2 0.87 0.10 7.79 2.83e-2 0.88 0.10 477.66
22 7.76e-2 0.72 0.16 9.73 7.24e-2 0.75 0.16 480.80
25 8.60e-2 0.90 0.15 11.94 8.34e-2 0.89 0.15 264.94
28 2.32e-2 0.87 0.07 21.89 2.68e-2 0.85 0.09 199.21
31 4.30e-2 0.83 0.09 10.56 4.78e-2 0.86 0.11 207.40
34 4.03e-2 0.82 0.11 12.84 4.25e-2 0.84 0.12 482.75
37 2.43e-2 0.89 0.09 12.86 1.81e-2 0.91 0.09 330.83
40 1.11e-2 0.91 0.06 14.64 1.12e-2 0.91 0.06 199.69
44 1.98e-2 0.79 0.07 11.09 2.11e-2 0.84 0.08 263.15
47 2.34e-2 0.90 0.09 9.64 2.36e-2 0.89 0.09 143.94
50 4.67e-2 0.51 0.13 7.49 4.58e-2 0.54 0.13 222.25
55 4.63e-2 0.89 0.16 5.63 4.83e-2 0.88 0.16 475.61
58 7.52e-2 0.84 0.17 11.77 8.09e-2 0.83 0.19 315.73

Table 12: Performance results of prediction methods.

Method Corresponding number MSE R MAE Time (s)

Linear 1 9.90e-02 0.75 0.20 2.10
Interactions Linear 2 9.39e-02 0.77 0.20 4.79
Robust Linear 3 1.05e-01 0.74 0.19 3.74
Stepwise Linear 4 9.40e-02 0.77 0.20 1256.50
Fine Tree 5 1.30e-01 0.68 0.22 7.09
Medium Tree 6 1.09e-01 0.73 0.20 3.53
Coarse Tree 7 1.01e-01 0.75 0.19 2.73
Linear SVM 8 1.03e-01 0.75 0.19 96.97
Quadratic SVM 9 9.78e-02 0.76 0.19 214.74
Cubic SVM 10 9.25e-02 0.77 0.18 828.17
Fine Gaussian SVM 11 2.07e-01 0.49 0.29 159.63
Medium Gaussian SVM 12 9.05e-02 0.78 0.18 112.76
Coarse Gaussian SVM 13 9.88e-02 0.76 0.19 78.12
Boosted Trees 14 9.39e-02 0.77 0.19 23.57
Bagged Trees 15 8.45e-02 0.79 0.17 54.57
BNN 16 16 8.26e-02 0.89 0.18 9.33

as long running time and relatively strong dependence on machine learning methods show that the IBNN model can
time and weather factors. To solve these problems, based significantly reduce calculating time and even when the time
on basic BNN training method and simple FFNN structure, or meteorological factors are missing; it can still predict
an improved BNN forecast model is built by augmenting the electricity demand with a high accuracy. Future work
historical load data as inputs through correlation analysis will focus on the application of IBNN forecasting model in
of electricity consumption at different delayed time scales. renewable residential energy management, especially for PV-
Further from the impact factors analysis, containing different storage system.
inputs, number of hidden neurons, historic period of data,
forecasting time range, and range requirement of sample data, Data Availability
some advices are given on how to better choose these factors.
To validate the effectiveness of IBNN model, several residen- The data used in this article are provided by a power company
tial sample datasets of a whole year from Ausgrid have been named ‘Ausgrid’, which can be found by the following link.
selected to build the IBNN models. The results compared https://www.ausgrid.com.au/Industry/Innovation-and-re-
with the time-series prediction model and common applied search/Data-to-share/Solar-home-electricity-data.
Mathematical Problems in Engineering 15

0.25 1

0.2
0.8
MSE

0.15

R
0.6
0.1

0.05 0.4
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Method corresponding number Method corresponding number

30

1000
25

Time (s)
MAE

20 500

15 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Method corresponding number Method corresponding number

Figure 11: Performance comparison of prediction methods.

Conflicts of Interest calculate the demand response baseline for office buildings,”
Applied Energy, vol. 195, pp. 659–670, 2017.
The authors declare that they have no conflicts of interest. [7] B. L. P. Cabrera and F. Schulz, “Forecasting generalized quantiles
of electricity demand: a functional data approach,” Journal of
Acknowledgments the American Statistical Association, vol. 112, no. 517, pp. 127–136,
2017.
The authors would like to thank Ausgrid and Bureau of [8] A. Marszal-Pomianowska, P. Heiselberg, and O. Kalyanova
Meteorology in Australia for the data used in this article. Larsen, “Household electricity demand profiles - A high-
This work was supported by the National Natural Science resolution load model to facilitate modelling of energy flexible
buildings,” Energy, vol. 103, pp. 487–501, 2016.
Foundation of China (Grants nos. 51478258 and 51405287)
and Shanghai Committee of Science and Technology (Grant [9] Y. Liang, D. Niu, M. Ye, and W.-C. Hong, “Short-term load fore-
casting based on wavelet transform and least squares support
no. 18030501300).
vector machine optimized by improved cuckoo search,” Ener-
gies, vol. 9, no. 12, 2016.
References [10] Y. Liang, D. Niu, Y. Cao, and W.-C. Hong, “Analysis and mod-
eling for China’s electricity demand forecasting using a hybrid
[1] C.-N. Yu, P. Mirowski, and T. K. Ho, “A Sparse Coding Approach method based on multiple regression and extreme learning
to Household Electricity Demand Forecasting in Smart Grids,” machine: A view from carbon emission,” Energies, vol. 9, no. 11,
IEEE Transactions on Smart Grid, vol. 8, no. 2, pp. 738–748, 2017. 2016.
[2] X. Qiu, Y. Ren, P. N. Suganthan, and G. A. J. Amaratunga, [11] A. S. Ahmad, M. Y. Hassan, M. P. Abdullah et al., “A review on
“Empirical Mode Decomposition based ensemble deep learning applications of ANN and SVM for building electrical energy
for load demand time series forecasting,” Applied Soft Comput- consumption forecasting,” Renewable & Sustainable Energy
ing, vol. 54, pp. 246–255, 2017. Reviews, vol. 33, pp. 102–109, 2014.
[3] Y.-H. Hsiao, “Household electricity demand forecast based on [12] Z. Hu, Y. Bao, T. Xiong, and R. Chiong, “Hybrid filter–wrapper
context information and user daily schedule analysis from feature selection for short-term load forecasting,” Engineering
meter data,” IEEE Transactions on Industrial Informatics, vol. 11, Applications of Artificial Intelligence, vol. 40, pp. 17–27, 2015.
no. 1, pp. 33–43, 2015. [13] H. Shayeghi, A. Ghasemi, M. Moradzadeh, and M. Nooshyar,
[4] S. Aman, M. Frincu, C. Chelmis, M. Noor, Y. Simmhan, and V. “Simultaneous day-ahead forecasting of electricity price and
K. Prasanna, “Prediction models for dynamic demand response: load in smart grids,” Energy Conversion and Management, vol.
Requirements, challenges, and insights,” in Proceedings of the 95, pp. 371–384, 2015.
IEEE International Conference on Smart Grid Communications, [14] L. Xiao, J. Wang, R. Hou, and J. Wu, “A combined model based
2016. on data pre-analysis and weight coefficients optimization for
[5] M. Behl, F. Smarra, and R. Mangharam, “DR-Advisor: A data- electrical load forecasting,” Energy, vol. 82, pp. 524–549, 2015.
driven demand response recommender system,” Applied [15] L. Y. Xiao, J. Z. Wang, X. S. Yang, and L. Y. Xiao, “A hybrid model
Energy, vol. 170, pp. 30–46, 2016. based on data preprocessing for electrical power forecasting,”
[6] Y. Chen, P. Xu, Y. Chu et al., “Short-term electrical load fore- International Journal of Electrical Power & Energy Systems, vol.
casting using the Support Vector Regression (SVR) model to 64, pp. 311–327, 2015.
16 Mathematical Problems in Engineering

[16] P. Lauret, E. Fock, R. N. Randrianarivony, and J.-F. Manicom- [35] X. Chai, S. Zheng, S. Geng, and L. Zhang, “The prediction of
Ramsamy, “Bayesian neural network approach to short time railway vehicle vibration based on neural network,” Journal of
load forecasting,” Energy Conversion and Management, vol. 49, Information and Computational Science, vol. 12, no. 16, pp. 5889–
no. 5, pp. 1156–1166, 2008. 5899, 2015.
[17] H. S. Hippert and J. W. Taylor, “An evaluation of Bayesian [36] J. Hong and W. S. Kim, “Weather impacts on electric power load:
techniques for controlling model complexity and selecting Partial phase synchronization analysis,” Meteorological Applica-
inputs in a neural network for short-term load forecasting,” tions, vol. 22, no. 4, pp. 811–816, 2015.
Neural Networks, vol. 23, no. 3, pp. 386–395, 2010. [37] D. Wey, A. Bohn, and L. Menna-Barreto, “Daily rhythms of
[18] L. Hernandez, C. J. M. Baladron, B. Aguiar et al., “A Survey on native Brazilians in summer and winter,” Physiology & Behavior,
Electric Power Demand Forecasting: Future Trends in Smart vol. 105, no. 3, pp. 613–620, 2012.
Grids, Microgrids and Smart Buildings,” Ieee Communications
Surveys and Tutorials, vol. 16, no. 3, pp. 1460–1495, 2014.
[19] J. G. Jetcheva, M. Majidpour, and W.-P. Chen, “Neural network
model ensembles for building-level electricity load forecasts,”
Energy and Buildings, vol. 84, pp. 214–223, 2014.
[20] M. Ghayekhloo, M. B. Menhaj, and M. Ghofrani, “A hybrid
short-term load forecasting with a new data preprocessing
framework,” Electric Power Systems Research, vol. 119, pp. 138–
148, 2015.
[21] M. Ghofrani, M. Ghayekhloo, A. Arabali, and A. Ghayekhloo,
“A hybrid short-term load forecasting with a new input selection
framework,” Energy, vol. 81, pp. 777–786, 2015.
[22] S. Hassan, A. Khosravi, and J. Jaafar, “Examining performance
of aggregation algorithms for neural network-based electricity
demand forecasting,” International Journal of Electrical Power &
Energy Systems, vol. 64, pp. 1098–1105, 2015.
[23] P. Bento, J. Pombo, M. Calado, and S. Mariano, “A bat optimized
neural network and wavelet transform approach for short-term
price forecasting,” Applied Energy, vol. 210, pp. 88–97, 2018.
[24] Y. Wang and J. M. Bielicki, “Acclimation and the response of
hourly electricity loads to meteorological variables,” Energy, vol.
142, pp. 473–485, 2018.
[25] X. Zhang and J. Wang, “A novel decomposition-ensemble
model for forecasting short-term load-time series with multiple
seasonal patterns,” Applied Soft Computing, vol. 65, pp. 478–494,
2018.
[26] D. J. MacKay, “Bayesian Interpolation,” Neural Computation,
vol. 4, no. 3, pp. 415–447, 1992.
[27] B. P. Hayes, J. K. Gruber, and M. Prodanovic, “A Closed-
Loop State Estimation Tool for MV Network Monitoring and
Operation,” IEEE Transactions on Smart Grid, vol. 6, no. 4, pp.
2116–2125, 2015.
[28] K. M. Powell, A. Sriprasad, W. J. Cole, and T. F. Edgar, “Heating,
cooling, and electrical load forecasting for a large-scale district
energy system,” Energy, vol. 74, pp. 877–885, 2014.
[29] R. A. Fisher, Statistical methods for research workers, 1958.
[30] M. G. Kendall, The Advanced Theory of Statistics, Charles Grif-
fin, 1976.
[31] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery,
Numerical Recipes in C. Art of Scientific Computing, vol. 10, 1992.
[32] C. Tofallis, “Erratum: A better measure of relative prediction
accuracy for model selection and model estimation,” Journal of
the Operational Research Society, vol. 66, no. 3, pp. 524-524, 2015.
[33] D. J. Best and D. E. Roberts, “Algorithm AS 89: The Upper Tail
Probabilities of Spearman’s Rho,” Journal of Applied Statistics,
vol. 24, no. 3, pp. 377–379, 1975.
[34] C. Bergmeir, R. J. Hyndman, and B. Koo, “A note on the validity
of cross-validation for evaluating autoregressive time series
prediction,” Monash Econometrics & Business Statistics Working
Papers, 2015.
Advances in Advances in Journal of The Scientific Journal of
Operations Research
Hindawi
Decision Sciences
Hindawi
Applied Mathematics
Hindawi
World Journal
Hindawi Publishing Corporation
Probability and Statistics
Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 http://www.hindawi.com
www.hindawi.com Volume 2018
2013 www.hindawi.com Volume 2018

International
Journal of
Mathematics and
Mathematical
Sciences

Journal of

Hindawi
Optimization
Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

Submit your manuscripts at

www.hindawi.com

International Journal of
Engineering International Journal of
Mathematics
Hindawi
Analysis
Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

Journal of Advances in Mathematical Problems International Journal of Discrete Dynamics in

Complex Analysis
Hindawi
Numerical Analysis
Hindawi
in Engineering
Hindawi
Differential Equations
Hindawi
Nature and Society
Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

International Journal of Journal of Journal of Abstract and Advances in

Stochastic Analysis
Hindawi
Mathematics
Hindawi
Function Spaces
Hindawi
Applied Analysis
Hindawi
Mathematical Physics
Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018