Combination of Kohonen SOM and Multi- layer Perceptron for
modeling and forecasting the energy needs of a Solar House
D.P. Iracleous U.H .
G. Kathreptis Technological Educational Institute of Piraeus
gkathr@gdias.teipir.gr Tel 210 5381304
T.I. Xanthopoulos U.H. University of Hertfordshire
xanthopoulos@ist.edu.gr Tel. 210 5381304
Abstract
Neural Networks are used with success to model non-linear systems, as this of a Solar
House. Multi- layer Perceptron is used in the majority of papers for forecasting the energy that
is required in a Solar House. In this paper the energy behavior of a Solar House is examined,
to perform on-line peak power forecast. ΝΝ Kohonen is used to order the energy and
meteorological facts, classify them in similar complexes and after that, feed them in an MLP,
to achieve faster and more precise forecast of energy demands for a Solar House.
1. Introduction
Artificial Neural Networks (ANN) has been widely used for a range of applications in the area
of energy modelling. The literature has demonstrate their superior capability overconventional methods, such as times series or regression, their main advantage being the
high potential to model non-linear processes, such as utility loads or individual buildings
energy consumption.
The application of the ANN methodology to the problem of short term load forecasting for
electric power utilities has received much of the attention and excellent results have been
achieved in real applications [1–5].
A combined approach based on Kohonen and Multi layer neural networks for the electric
energy demand forecasting of a Solar House with a prediction time of 24 h is illustrated.
1.1 Kohonen Neural Networks
The Kohonen network is formed by a single layered neural network. The data are presented
to the input and the output neurons are organized with a simple neighbourhood structure.
Each neuron is associated with a reference vector (the weight vector), and each data point is
“mapped” to the neuron with the “closest” (in the sense of the Euclidean distance) reference
vector. In the process of running the algorithm, each data point acts as a training sample
which directs the movement of the reference vectors towards the value of the data of this
sample. The vectors associated with neurons, called weights, change during the learning
process and tend to the characteristic values of the distribution of the input data. The
characteristic value of one cluster can be intuitively understood as the typical value of the
data in the cluster and will be defined more precisely in the next subsections. At the end of
the process the set of input data is partitioned into disjoint sets (the clusters) and the weight
associated with each neuron is the characteristic value of the cluster associated with the
neuron in the one-dimensional case, which is the case of interest to us. We limit our analysis
to this case because the condition of convergence of the algorithm is easier to check, the
cluster of the partitions are easier to visualize and it is not difficult to compare the behaviour of
the genes in the clusters corresponding to the different biological conditions. Each neuron
individuates one cluster, and the physical or biological entities with measure values belonging
to the same cluster are considered to be involved in the same cellular process. Thus the
genes with expressions belonging to the same cluster might be functionally related.
The following points show the main properties which make the Kohonen network useful for
clustering:
i) Low dimension of the network and its simple structure.
ii) Simple representation of clusters by means of vectors associated with each neuron.
iii) Topology of the input data set is somehow mapped in the topology of the weights of the
network.
iv) Learning is unsupervised.
v) Self-organized property.
Points i and ii are simple to understand. Point iii, means that neighbouring neurons have
weight vectors not very different from each other. Point iv, means that the reasoned to have
an external constraint to drive the weights towards their right values beside the input to the
network and that the learning process finds by itself the right topology and the right values.
This holds only if the learning process with which the network is constructed converges.
The self-organization is formulated in the current literature referring to some universality of the
structure of the network for a given data set. It is connected to point iii and is also a
consequence of a.e. convergence or of the convergence of the average of the weights over
many different learning processes.
Figure 1
Kohonen Self – Organising Map (SOM)
1.2. Feed Forward Neural networks as non-linear regression models
Feed forward artificial neural networks, is a class of flexible and widely applied models, used
to find the relation between the input and the output variables. The problem can be stated as
finding a function f (F: Rd !R) such as to obtain an estimate of the output values y from the
input values x. The neural network approach of performing prediction is to induce this function
in a standard multilayer perceptron (MLP).
Figure 2
Standard Multi Layer Perceptron
Many studies have pointed out the overwhelming sensitivity of electricity consumption to
weather variables, in particular focusing attention on the forecast limited to 24 h ahead.
Recent research activities have also focused on the impacts of climatic changes both on
supply (studies about the potential impact of climate change on renewable energy resources,
such as wind power and hydroelectric power) and demand (studies about electricity demand
and natural gas demand correlation with weather variability) for energy. Weather sensitivity
has also been examined in order to correlate electricity consumptions to the increases in
market saturation of air conditioning induced by long term climatic changes.
Furthermore, it should be noticed that in order to model energy use, environmental variables
such as ambient temperature, solar radiation or wind speed are often used, as weather
strongly affects the energy consumption of buildings.
However, for predictive control applications where the input data at the target time are
needed, i.e. to get a prediction 24 h ahead, a pre-process is required before the prediction
can be made and usually is based on first predicting the weather data for the next day, either
using a predicting methodology, or by directly linked to weather reports, say via the Internet.
2. Our proposal
The combination of Kohonen SOM and Multi Layer Perceptron is our proposal in order to
manage the energy use in a Solar House.
In our scenario, a family of four persons is resident of a 70m2 house. The location of the
house is a very important matter that will not be discussed in this paper.
For standard house electrical equipment, that covers the needs of the residents (refrigerator,
cooker, electric bulbs and entertainment electrical appliances that are allowed to consume
pre- determined monthly amount of electrical energy every month)
θ out
Η out
θ in
H in
Day
Month
Sun Flux
Kohonen
SOM
Multi layer
Perceptron
Solar House
Figure 3
θ out external temperature
H out external humidity
θ in internal temperature
Day
Month
Sun flux measurement outside the Solar House
Kohonen SOM will use these parameters to extract and classify the information needed to
feed the MLP in order to achieve thermal stability inside the Solar House.
It is our belief that using Kohonen SOM in combination with the MLP will increase the
response and the accuracy of the system that controls the energy use in the Solar House.
References
Gale TM, Laws KR
Category-specificity can emerge from bottom-up visual characteristics: Evidence from a
modular neural network
BRAIN COGNITION 61 (3): 269-279 AUG 2006
Chen CH, Khoo LP, Yan W
An investigation into affective design using sorting technique and Kohonen self-organising
map
ADV ENG SOFTW 37 (5): 334-349 MAY 2006
Malone J, McGarry K, Wermter S, et al.
Data mining using rule extraction from Kohonen self-organising maps
NEURAL COMPUT APPL 15 (1): 9-17 MAR 2006
Godin N, Huguet S, Gaertner R
Influence of hydrolytic ageing on the acoustic emission signatures of damage mechanisms
occurring during tensile tests on a polyester composite: Application of a Kohonen's map
COMPOS STRUCT 72 (1): 79-85 JAN 2006
Bryan BA
Synergistic techniques for better understanding and classifying the environmental structure of
landscapes
ENVIRON MANAGE 37 (1): 126-140 JAN 2006
S. Kalogirou, Artificial neural networks in renewable energy systems applications: a review,
Renewable &Sustainable Energy Reviews 5 (2001) 373–401.
C. Cetiner, F. Halici, H. Cacur, et al., Generating hot water by solar energy and application of
neural network, Applied Thermal Engineering 25 (2005) 8–9.
A. Argiriou, I. Bellas-Velidis, C. Balaras, Development of a neural network heating controller
for solar buildings, Neural Networks 13 (2000) 811–820.
A. Ben-Nakhi, M. Mahmoud, Energy conservation in buildings through efficient A/C control
using neural networks, Applied Energy 73 (2002) 5–23.
S. Wang, Y. Chen, Fault-tolerant control for outdoor ventilation air flow rate in buildings based
on neural network, Building and Environment 37 (2002) 691–794.
M. Mahmoud, A. Ben-Nakhi, Architecture and performance of neural networks for efficient
A/C control in buildings, Energy Conversion and Management 44 (2003) 3207–3226.
A. Argiriou, I. Bellas-Velidis, M. Kummert, P. Andre, A neural network controller for hydronic
heating systems of solar buildings, Neural Networks 17 (2004) 427–440.
I. Yang, K. Kim, Prediction of the time of room air temperature descending for heating
systems in buildings, Building and Environment 39 (2004) 19–29.
A. Ben-Nakhi, M. Mahmoud, Cooling load prediction for buildings using general regression
neural networks, Energy Conversion & Management 45 (2004) 2127–2141.
G. Mikhlakakou, M. Santamouris, A. Tsangrassoulis.
On the energy consumption in residential buildings,
Energy and Buildings 34 (2002)727–736.
K. Funahashi,
On the approximate realization of continuous mappings by neural networks,
Neural Networks 2 (1989) 183–192.
E.M. Crispim, M.D. Martins, A.E. Ruano, C.M. Fonseca
Remote data acquisition system of environmental data, in: Proceedings of the Sixth
Portuguese Conference of Automatic Control (Controlo 2004),
University of Algarve, June, 2004.
S. Haykin
Learning strategies in Neural Networks: A Comprehensive Foundation,
Prentice Hall, 1998.
P.M. Ferreira, E.A. Faria, A.E. Ruano
Neural network models in greenhouse air temperature prediction
Neurocomputing 1 (43) (2002) 51–75.
C.M. Fonseca, P.J. Fleming
Multiobjective optimization and multiple constraint handling with evolutionary algorithms: a
unified formulation
IEEE Transactions on Systems, Man and Cybernetics—Part A: System
and Humans 1 (28) (1998) 26–37.
P.M. Ferreira, A.E. Ruano, C.M. Fonseca
Genetic assisted selection of model structures for greenhouse inside air temperature
prediction
Proceedings of the 2003 IEEE Conference on Control Applications (CCA
2003), 2003, pp. 576–581.
S. Billings, Q. Zhu
Nonlinear Model Validation using Correlation Tests
Department of Automatic Control and Systems Engineering, University of Sheffield, Sheffield
S14DU, UK, 1993 (Research Report 463).
P.M. Ferreira, A.E. Ruano, C.M. Fonseca, Evolutionary multi-objective design of radial basis
function networks for greenhouse environmental control
16th IFAC World Congress, Prague, July, 2005.
694 A.E. Ruano et al. / Energy and Buildings 38 (2006) 682–694