PI-controlled ANN-based Energy Consumption Forecasting

for Smart Grids

Gulsum Gezer

, Gurkan Tuna

, Dimitris Kogias

, Kayhan Gulez

and V. Cagri Gungor

Department of Control and Automation Engineering, Yildiz Technical University, Istanbul, Turkey

Department of Computer Programming, Trakya University, Edirne, Turkey

Department of Electronics Engineering, Piraeus University of Applied Sciences, Aigaleo, Greece

Department of Computer Engineering, Abdullah Gul University, Kayseri, Turkey

Keywords: Smart Grid, Demand Forecasting, Artificial Neural Network, Optimization.

Abstract: Although Smart Grid (SG) transformation brings many advantages to electric utilities, the longstanding

challenge for all them is to supply electricity at the lowest cost. In addition, currently, the electric utilities

must comply with new expectations for their operations, and address new challenges such as energy

efficiency regulations and guidelines, possibility of economic recessions, volatility of fuel prices, new user

profiles and demands of regulators. In order to meet all these emerging economic and regulatory realities,

the electric utilities operating SGs must be able to determine and meet load, implement new technologies

that can effect energy sales and interact with their customers for their purchases of electricity. In this

respect, load forecasting which has traditionally been done mostly at city or country level can address such

issues vital to the electric utilities. In this paper, an artificial neural network based energy consumption

forecasting system is proposed and the efficiency of the proposed system is shown with the results of a set

of simulation studies. The proposed system can provide valuable inputs to smart grid applications.

1 INTRODUCTION

The traditional power grid in many countries suffers,

amongst others, from huge maintenance costs

because of the system's age, from scalability issues

because of the globally increasing demand for more

power along with the expense of building new

power stations and from lack of efficient system

monitoring that could increase the overall

performance by acting proactively in preventing

damages. Therefore, the Smart Grid (SG) solution

has been presented as an evolutionary system for

power generation and distribution. A SG is a

modernized power transmission and distribution

network which uses robust two-way data

communications, distributed computing technologies

and smart sensors to improve reliability, safety and

efficiency of power delivery and use (Gungor et al,

2010; Gungor et al, 2011). SGs use renewable

energy production, smart meters and modern

communication technologies for effective system

monitoring, thus succeeding in addressing many of

the requirements of a modern power grid system

while significantly increase its performance.

Using a sophisticated information processing and

communication technology infrastructure, the SG is

able to fully use and benefit from its distributed

power generation system, while maximizing the

whole system's energy efficiency. Consequently, the

SG is also considered as a data communication

network which, by supporting many power

management devices, achieves seamless and flexible

inter-operational abilities among different advanced

system components that leads to an efficient

performance. The power system infrastructure is

comprised of all the devices found on existing

electrical grids, as described above, with the

additions of smart meters and sensors deployed

throughout the grid to detect outages and measure

critical performance metrics that should be

forwarded to the backhaul system's detection and

decision data centres. This infrastructure generally

includes the power generation, transmission and

distribution system and the customer premises.

At the core of a SG, Advanced Metering

Infrastructure (AMI) lies. Basically, AMI is a two-

way communications network between customer

premises and the backhaul of the SG and consists of

smart meters, advanced sensors and monitoring

110

Gezer G., Tuna G., Kogias D., Gulez K. and Gungor V..

PI-controlled ANN-based Energy Consumption Forecasting for Smart Grids.

DOI: 10.5220/0005516801100116

In Proceedings of the 12th International Conference on Informatics in Control, Automation and Robotics (ICINCO-2015), pages 110-116

ISBN: 978-989-758-122-9

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

systems that collect and distribute information

between the connected devices in order to enable the

gathering and transfer of energy usage information

in near real-time. Since the amount of collected data

is huge and the collected data is important, the utility

communication infrastructure is expected to be

scalable and provide high bandwidth capabilities and

low latency (Sood et al, 2009; Tuna et al, 2013).

Although SGs can address their current challenges

with their smart metering and AMI, forecasting tools

can help them further optimize their operations.

Because the complexity of challenges has been

growing and a SG is a data-rich environment with

many data inputs from several applications.

Basically, forecasting is a data-intensive numeric

discipline and used by utilities for various planning,

investment and decision-making purposes. Using

forecasting tools which manage large quantities of

usage data from new inputs from smart meters, the

utilities can determine how their customers will use

energy and try to understand when their customers

will change their behaviour in response to personal

economic conditions, demand response programs,

government initiatives, and whether and climate

change concerns. In this way, the utilities can plan

their operations. Finally, among many important

questions that can be answered by forecasting tools,

from the return on investment view, the key question

is "what types of pricing programs are likely to

produce the largest benefits?".

Several modelling techniques have been

proposed for demand forecasting. They can be

classified into nine main categories (Alfares and

Nazeeruddin, 2002; Gajowniczek and Zabkowski,

2014; Chan et al, 2012), namely multiple regression,

exponential smoothing, iterative reweighted least-

squares, adaptive load forecasting, stochastic time

series, ARMAX models based on genetic

algorithms, fuzzy logic, artificial neural networks,

and expert systems. In comparison with the other

techniques, artificial neural networks (ANNs) are

better at solving forecasting problems due to their

hidden layers and ability to learn (Gajowniczek and

Zabkowski, 2014). They are able to identify hidden

trends thereby finding the trends in time series and

use them to produce more accurate results.

Therefore ANNs are very popular and attractive for

practical applications.

There are many studies in the literature that deal

with the models/techniques for demand forecasting.

In (Javed et al, 2012), autonomic demand side

management is presented as a paradigm to provide

demand side management and demand response in

micro-grids. Liu et al. in (Liu et al, 2012) mainly

focus on a hybrid model with parameter

optimization for load forecasting of micro-grids. A

design problem in a setting where several agents can

generate estimates of independent future demands at

a cost is investigated in (Egri and Vancza, 2013). In

the proposed approach, an aggregator agent elicits

the forecasts and based on this information,

optimises a procurement decision. Schachter and

Mancarella define the functionality required for

developing a short-term load forecasting module for

demand response applications and propose an

algorithm to provide high forecasting performance

(Schachter and Mancarella, 2014).

In this study, a forecasting approach for SGs is

proposed. The approach is based on the use of the

back propagation neural network algorithm which is

a multilayer feed-forward network trained according

to error back propagation algorithm. Using the

proposed approach, SG operators can estimate future

demand from the past and respond with their

purchases. This also allows them to offer electricity

generated from renewable sources. The rest of the

paper is organized as follows. Section 2 presents the

details of the artificial neural network based

forecasting approach for smart grid and focuses on

its implementation. Section 3 presents the evaluation

results. Finally, the paper is concluded in Section 4.

2 ARTIFICIAL NEURAL

NETWORKS FOR DEMAND

FORECASTING

ANN is basically a parallel distributed processor

consist of simple processing units that has an

inherent trend for storing experiential knowledge

and making it available for use (Haykin, 2008).

Multilayer feed-forward neural networks are

generally used based on minimization of an error

function and back propagation (BP) learning is the

common training method used in these networks. BP

learning uses the gradient descent procedure to train

the connection weights (Fazayeli et al, 2008).

Multilayer feed-forward neural networks consist of a

layer of input units, one or more layers of hidden

units, and one output layer of units. Every layer has

neurons and is connected with next layers. Each

connection between nodes has a weight associated

with it. The trend of on-line learning for the

supervised training of multilayer perceptrons has

been further raised by the development of the back

propagation algorithm. BP algorithm can be

described using the following equations.

PI-controlledANN-basedEnergyConsumptionForecastingforSmartGrids

111



+===

jijijjj

ownetthenxfnetfo

)()(

(1)



∈

−=

)(

outj

pjpjp

otE

(2)

()

pj pj pj

pji

θε

=−





∂



Δ=−



∂





∂



Δ=−



∂



(3)

where j is the layer number and i is neuron number,

is neuron output,

net

is weighted sum,

bias,

is weight,

is learning rate,

represents error value in layer j,

is target output

and

is actual output. Equation (2) is used to

estimate the entire error in the output layer for the

pth sample pattern, the error that is eventually

minimized by variating the weights and biases using

gradient descent (3).

Levenberg-Marquardt algorithm, one of the most

important optimization method (Roweis, 2009), was

used for the development the proposed NN. This

technique, due to Levenberg (Levenberg, 1944) and

Marquardt (Marquardt, 1963), is a sequence of the

following two methods (Ranganathan, 2004):

the Gauss-Newton's method approaches

rapidly near to a global or local minimum but

sometimes may deviate;

II.

the Gradient descent method certainly

approaches through a proper selection of step

size parameter but does slowly.

The block diagrams and the architecture of the

developed ANN are shown in Figure 1 and Figure 2,

respectively.

Figure 1: The structure of the proposed multilayer NN

model prepared using Neural Network Toolbox of

MATLAB.

Proportional-Integral-Derivative (PID) controller

and its alternatives P, PI and PD are the one of the

best popular controllers used in nearly all control

Figure 2: Architecture of the system for energy consumption load forecast.

ICINCO2015-12thInternationalConferenceonInformaticsinControl,AutomationandRobotics

112

applications. Base of the popularity is from its

simplicity. The tuning method of Ziegler-Nichols

(Ziegler and Nichols, 1942) is implemented widely

in the industrial process area. The reproduction of

the PID gains relies on the presence of a process

time delay. The control system proposed in this

paper is based on neural networks. To improve the

performance of the system, PI controller is added to

the system. Figure 3 illustrates the block diagram of

the closed loop system.

Figure 3: Block diagram of the closed loop system.

The block diagram shows that a PI controller is

linked with the neural network controller system in

closed loop. Then, the PI controller becomes:

() () ( )

tet ed

ττ



(4)

where e(t) is the closed loop error function. The

basic action of the proportional gain is to determine

the system response time or bandwidth. The integral

gain improves steady-state and low frequency

trajectory following by adding stiffness against

disturbances and steady state errors.

3 PERFORMANCE

EVALUATION

In recent years, different load forecasting methods

have been developed and implemented. Artificial

neural network based methods are popular in smart

grid applications. This study focuses on comparing a

house's neural network based energy consumption

load forecast and PI+Neural Network energy

consumption load forecast. There are 7 inputs in this

house. These inputs include refrigerator, washing

machine, dishwasher, television, lighting, heating/air

condition and computer. The dataset is collected

hour by hour for 1 week. Table 1 lists the partial

dataset. The simulated results were observed using

MATLAB Neural Network Toolbox (MathWorks,

2014) and they were compared with their actual

datasets. The improved NN model is a 1-input layer,

2-hidden layers (including 16 and 15 nodes,

respectively) and 1-output layer. As already

explained, the Levenberg-Marquardt optimization

algorithm was used since it presents certain

advantages over BP (Michailidis et al, 2014).

Table 1: Partial data set.

Inputs Outputs

RG WM DW TV Light. H/AC PC RG WM DW TV

ight. H/AC PC

1 00:00

50,0 0,0 0,0 0,0 50,0 200,0 0,0 50,0 0,0 0,0 0,0 51,0 200,0 0,0

2 01:00

50,0 0,0 0,0 0,0 50,0 200,0 0,0 49,0 0,0 0,0 0,0 50,0 196,0 0,0

3 02:00

50,0 0,0 0,0 0,0 0,0 200,0 0,0 51,0 0,0 0,0 0,0 0,0 203,0 0,0

4 03:00

50,0 0,0 0,0 0,0 0,0 200,0 0,0 50,0 0,0 0,0 0,0 0,0 201,0 0,0

5 04:00

50,0 0,0 0,0 0,0 0,0 200,0 0,0 50,5 0,0 0,0 0,0 0,0 198,0 0,0

6 05:00

50,0 0,0 0,0 0,0 0,0 200,0 0,0 50,0 0,0 0,0 0,0 0,0 199,0 0,0

7 06:00 50,0 0,0 0,0 0,0 0,0 200,0 0,0 51,0 0,0 0,0 0,0 0,0 197,0 0,0

8 07:00 50,0 0,0 0,0 0,0 0,0 200,0 0,0 49,5 0,0 0,0 0,0 0,0 202,0 0,0

9 08:00 50,0 0,0 0,0 0,0 0,0 200,0 0,0 49,0 0,0 0,0 0,0 0,0 200,0 0,0

10 09:00 50,0 0,0 0,0 0,0 0,0 200,0 0,0 50,0 0,0 0,0 0,0 0,0 204,0 0,0

11 10:00 50,0 0,0 0,0 0,0 0,0 200,0 0,0 49,0 0,0 0,0 0,0 0,0 199,0 0,0

12 11:00 50,0 0,0 0,0 0,0 0,0 200,0 0,0 48,5 0,0 0,0 0,0 0,0 200,0 0,0

13 12:00 50,0 0,0 0,0 0,0 0,0 200,0 0,0 51,5 0,0 0,0 0,0 0,0 201,0 0,0

14 13:00 50,0 0,0 0,0 0,0 0,0 200,0 0,0 50,0 0,0 0,0 0,0 0,0 205,0 0,0

15 14:00 50,0 0,0 0,0 0,0 0,0 200,0 0,0 49,5 0,0 0,0 0,0 0,0 200,0 0,0

16 15:00 50,0 0,0 0,0 0,0 0,0 200,0 0,0 50,0 0,0 0,0 0,0 0,0 199,0 0,0

17 16:00 50,0 0,0 0,0 0,0 0,0 200,0 0,0 51,0 0,0 0,0 0,0 0,0 202,0 0,0

18 17:00 50,0 0,0 0,0 0,0 0,0 200,0 0,0 50,5 0,0 0,0 0,0 0,0 201,0 0,0

19 18:00 50,0 0,0 0,0 0,0 50,0 200,0 20,0 49,0 0,0 0,0 0,0 48,0 200,0 19,0

20 19:00 50,0 50,0 0,0 150,0 50,0 200,0 20,0 50,0 49,0 0,0 149,0 49,0 203,0 21,0

21 20:00 50,0 50,0 0,0 150,0 50,0 200,0 20,0 49,0 50,0 0,0 151,0 51,0 199,0 20,0

22 21:00 50,0 0,0 0,0 150,0 50,0 200,0 20,0 48,5 0,0 0,0 150,0 50,0 197,0 18,0

23 22:00 50,0 0,0 0,0 150,0 50,0 200,0 20,0 51,0 0,0 0,0 152,0 51,0 198,0 21,0

24 23:00 50,0 0,0 0,0 0,0 50,0 200,0 0,0 50,0 0,0 0,0 0,0 48,0 200,0 0,0

PI-controlledANN-basedEnergyConsumptionForecastingforSmartGrids

113

The comparison of the training dataset with the

corresponding ANN result and PI+ANN result for

Refrigerator (RG), Washing Machine (WM),

Dishwasher (DW), Television (TV), Lighting

(Light.), Heating/Air Conditioning (H/AC) and

Computer (PC) are given in Figures 4-10,

respectively. Before being entered into the ANN

model, the input dataset was normalized for more

reliable results and then rescaled to the original

dataset. The activation functions used for the

neurons of the proposed ANN model were tangent-

sigmoid functions for the two hidden layers and

linear transfer functions for the output layer. The

performance of the NN model and PI+NN model

and regression between target and NN output and

between target and PI+NN output are shown in

Figure 11 and Figure 12, respectively.

Figure 4: a) ANN result for Refrigerator, (b) PI+ANN

result for Refrigerator.

Figure 5: (a) ANN result for Washing Machine, (b)

PI+ANN result for Washing Machine.

Figure 6: (a) ANN result for Dishwasher, (b) PI+ANN

result for Dishwasher.

Figure 7: (a) ANN result for Television, (b) PI+ANN

result for Television.

Figure 8: (a) ANN result for Lighting, (b) PI+ANN result

for Lighting.

ICINCO2015-12thInternationalConferenceonInformaticsinControl,AutomationandRobotics

114

Figure 9: (a) ANN result for Heating/Air Condition, (b)

PI+ANN result for Heating/Air Condition.

Figure 10: (a) ANN result for PC, (b) PI+ANN result for

PC.

Figure 11: (a) ANN performance result, (b) PI+ANN

performance result.

Figure 12: (a) ANN regression, (b) PI+ANN regression.

4 CONCLUSIONS

To be able to meet emerging economic and

regulatory realities, the electric utilities must be able

to determine and meet demand, interact with their

customers and implement new hardware and

software technologies. Although the transformation

from the traditional power grid to the smart grid

brings many advantages and helps them realize

these, they still need to use forecasting tools to

further optimize their operations from the return on

investment view. For this purpose, in this paper, an

energy consumption forecasting approach for smart

grids has been presented. Based on the data input

received from smart meters, the proposed approach

enables the electric utilities to forecast electricity

demand. In addition, it provides a valuable input to

many smart grid applications.

REFERENCES

Gungor, V. C., Bin, L., Hancke, G. P. 2010. Opportunities

and challenges of wireless sensor networks in smart

grid. IEEE Trans. Industrial Electronics, 57(10),

3557-3564.

Gungor, V. C., Sahin, D., Kocak, T., Ergut, S., Buccella,

C., Cecati, C., Hancke, G. P. 2011. Smart Grid

Technologies: Communication Technologies and

Standards. IEEE Transactions on Industrial

Informatics, 7(4), 529-539.

Sood, V. K., Fischer, D., Eklund, J. M., Brown, T. 2009.

Developing a communication infrastructure for the

smart grid. IEEE Electrical Power and Energy

Conference (EPEC 2009), 1-7.

Tuna, G., Gungor, V. C., Gulez, K. 2013. Wireless Sensor

Networks for Smart Grid Applications: A Case Study

on Link Reliability and Node Lifetime Evaluations in

Power Distribution Systems. International Journal of

Distributed Sensor Networks, 2013(2013), Article ID

796248, Available from: doi: 10.1155/2013/796248

[Accessed 3rd March 2015].

PI-controlledANN-basedEnergyConsumptionForecastingforSmartGrids

115

Alfares, H. K., Nazeeruddin, M. 2002. Electric load

forecasting: literature survey and classification of

methods. International Journal of Systems Science,

33(1), 3-34.

Gajowniczek, K., Zabkowski, T. 2014. Short term

electricity forecasting using individual smart meter

data. Procedia Computer Science, 35, 589-597.

Chan, S.C., Tsui, K.M., Wu, H.C., Hou, Y., Wu, Y.-C.,

Wu, F. F. 2012. Load/Price Forecasting and

Management of Demand Response for Smart Grids:

Methodologies and Challenges. IEEE Signal

Processing Magazine, 29(5), 68-85.

Javed, F., Arshad, N. Wallin, F., Vassileva, I., Dahlquist,

E. 2012. Forecasting for demand response in smart

grids: An analysis on use of anthropologic and

structural data and short term multiple loads

forecasting. Applied Energy, 96, 150-160.

Liu, N., Tang, Q., Zhang, J., Fan, W., Liu, J. 2014. A

hybrid forecasting model with parameter optimization

for short-term load forecasting of micro-grids. Applied

Energy, 129, 336-345.

Egri, P., Vancza, J. 2013. Efficient Mechanism for

Aggregate Demand Prediction in the Smart Grid.

LNAI, 8076, 250-263.

Schachter, J., Mancarella, P. 2014. A Short-term Load

Forecasting Model for Demand Response

Applications. Proc. 11th International Conference on

the European Energy Market (EEM).

Haykin. S. O. 2008. Neural Networks and Learning

Machines, Prentice Hall. New York, 3

edition.

Fazayeli, F., Wang, L., Liu, W. 2008. Back-Propagation

with Chaos. IEEE Int. Conference Neural Networks &

Signal Processing, 5-8.

Roweis, S. 2009. Levenberg-Marquardt optimization.

[Online] Available from: http://www.cs.nyu.edu/

?roweis/notes/lm.pdf [Accessed 3rd March 2015].

Levenberg, K. 1944. A method for the solution of certain

non-linear problems in least squares. Quarterly of

Applied Mathematics, 2, 164-168.

Marquardt, D. 1963. An algorithm for least-squares

estimation of nonlinear parameters. SIAM Journal on

Applied Mathematics, 11(2), 431-441.

Ranganathan, A. 2004. The Levenberg-Marquardt

Algorithm. Honda Research Institute, USA.

Ziegler J.G., Nichols B. 1942. Optimal setting for

automatic controllers. Trans ASME, 64, 759-768.

MathWorks. 2014. Neural Network Toolbox, [Online]

Available from: http://www.mathworks.com/products/

neural-network/ [Accessed 2nd March 2015]

Michailidis, E. T., Tuna, G., Gezer, G., Potirakis, S. M.,

Gulez, K. 2014. ANN-Based Control of a Multiboat

Group for the Deployment of an Underwater Sensor

Network. International Journal of Distributed Sensor

Networks, 2014 (2014), Article ID 786154, Available

from: doi: 10.1155/2014/786154 [Accessed 3rd March

2015].

ICINCO2015-12thInternationalConferenceonInformaticsinControl,AutomationandRobotics

116