predicting home service demands from appliance ...

GRENOBLE INSTITUTE OF TECHNOLOGY

PREDICTING HOME SERVICE DEMANDS FROM APPLIANCE USAGE DATA

by Kaustav Basu

A thesis submitted in partial fulfillment for the degree of Master of Science in Informatics

in the M2-GVR Master of Science in Informatics at Grenoble (MoSIG)

JUNE 2011


Résumé M2-GVR Master of Science in Informatics at Grenoble (MoSIG) Master of Science in Informatics par Kaustav Basu

La gestion énergétique dans les bâtiments, que ce soit des maisons individuelles ou des bureaux, a besoin d’utiliser la prédiction de consommation pour palier le manque d’informations fournies par l’utilisateur. Le caractère aléatoire de l’utilisation des appareils électriques fait que prédire ` a partir de données de consommations le moment o` u ils se mettront en marche est relativement compliqué. Un modèle général pour réaliser cette prédiction n’existe pas ` a l’heure actuelle. Dans ce rapport, nous proposons d’enrichir les algorithmes d’apprentissage avec des connaissances. Nous proposons un modèle géneral utilisant la connaissance pour prédire si un équipement particulier se mettra en route ou pas ` a une heure donnée. Cette approche utilise à la fois les connaissances et les données. Le système global de gestion énergétique de la maison a besoin de connaˆıtre les prédictions de consommation pour les 24 prochaines heures. Le modèle proposé est ensuite testé sur des données et on compare les résultats avec ceux fournis par des prédicteurs standards.


Abstract M2-GVR Master of Science in Informatics at Grenoble (MoSIG) Master of Science in Informatics by Kaustav Basu

Power management in homes and offices requires appliance usage prediction when the future user requests are not available. The randomness and uncertainties associated with an appliance usage make the prediction of appliance usage from energy consumption data a non-trivial task. A general model for prediction at the appliance level is still lacking. In this work, we propose to enrich learning algorithms with expert knowledge and propose a general model using a knowledge driven approach to forecast if a particular appliance will start at a given hour or not. The approach is both a knowledge driven and data driven one. The Overall energy management for a house requires that the prediction is done for the next 24 hours in the future. The proposed model is tested over an Irise data set and the results are compared with some trivial knowledge driven predictors.

Keywords : Appliance Usage Prediction, Enriched Learning Algorithm, Energy Management in Homes, Data Mining.

Acknowledgements First and foremost I offer my sincerest gratitude to my supervisors, Prof Stephane Ploix and Prof James Crowley, who have supported me thoughout my thesis with their patience and knowledge while allowing me the room to work in my own way. I attribute the level of my Masters degree to their encouragement and effort and without there help this thesis, too, would not have been completed or written. One simply could not wish for a better or friendlier supervisors. I will take this oppurtunity to thank Mathieu Guillame-bert for his co-supervision of the project and his kind words of encouragement and fellow student and researchers in the G-SCOP laboratory. Last but not the least I will like to thank my Friends and Family in India and also Prof Ujjawal Maulik for there support and encouragement.

iii

Table des mati` eres

Abstract

ii

Acknowledgements

iii

List of Figures

vi

List of Tables

vii

Abbreviations

viii

1 Introduction 1.1 Introduction to Home Automation System . . 1.2 Requirement of Prediction and Approach . . 1.3 Problem Statement . . . . . . . . . . . . . . . 1.4 Summary and Research Report Organization 2 Review of Past Literature 2.1 Past Approaches in Appliance Prediction 2.2 Energy Management in Smart Homes . . 2.3 Load Forecasting Algorithms . . . . . . . 2.3.1 Similar-day approach . . . . . . . . 2.3.2 Regressive methods . . . . . . . . 2.3.3 Time series . . . . . . . . . . . . . 2.3.4 Expert system . . . . . . . . . . . 2.3.5 Neural network . . . . . . . . . . . 2.3.6 Support vector machines . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . .

. . . . . . . . .

. . . .

. . . . . . . . .

. . . .

. . . . . . . . .

. . . .

. . . . . . . . .

. . . .

. . . . . . . . .

. . . .

. . . . . . . . .

. . . .

. . . . . . . . .

. . . .

. . . . . . . . .

. . . .

. . . . . . . . .

. . . .

. . . . . . . . .

3 Proposed Approach 3.1 Detailed problem statement . . . . . . . . . . . . . . . . . . . . 3.2 Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 Proposed Model . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.1 An expert system that generates knowledge : the oracle 3.3.2 Past consumption history . . . . . . . . . . . . . . . . . 3.3.3 Hour of the day . . . . . . . . . . . . . . . . . . . . . . . 3.3.4 Day of the week . . . . . . . . . . . . . . . . . . . . . . iv

. . . .

. . . . . . . . .

. . . . . . .

. . . .

. . . . . . . . .

. . . . . . .

. . . .

. . . . . . . . .

. . . . . . .

. . . .

. . . . . . . . .

. . . . . . .

. . . .

. . . . . . . . .

. . . . . . .

. . . .

1 1 3 4 4

. . . . . . . . .

6 6 7 10 10 11 11 11 11 12

. . . . . . .

14 14 14 15 16 17 17 18

Contents

3.4 3.5 3.6

3.3.5 Season of the year . . . 3.3.6 Previous day same hour 3.3.7 Oracle output . . . . . . Data Selector . . . . . . . . . . Predictor . . . . . . . . . . . . Predictor Controller . . . . . .

v . . . . . .

. . . . . .

4 Experimental Results 4.1 Introduction . . . . . . . . . . . . . 4.2 Oracle knowledge results . . . . . . 4.3 Choice of classifier, parameters and 4.4 Overall Result . . . . . . . . . . . . 4.5 Discussion . . . . . . . . . . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . . . . . . . . . . . . . . . . implementation . . . . . . . . . . . . . . . . . . . . .

. . . . . .

. . . . .

. . . . . .

. . . . .

. . . . . .

. . . . .

. . . . . .

. . . . .

. . . . . .

. . . . .

. . . . . .

. . . . .

. . . . . .

. . . . .

. . . . . .

. . . . .

. . . . . .

. . . . .

. . . . . .

. . . . .

. . . . . .

. . . . .

. . . . . .

19 20 20 20 21 21

. . . . .

22 22 22 28 30 30

5 Conclusion and Perspective

32

Bibliographie

34

Table des figures 1.1 1.2

Smart home concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Goal of the prediction system . . . . . . . . . . . . . . . . . . . . . . . . .

2.1 2.2

Architecture of a power manager . . . . . . . . . . . . . . . . . . . . . . . 8 STLF predictor input configuration . . . . . . . . . . . . . . . . . . . . . . 12

3.1 3.2

Screenshot of the data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 Schematic representation of the prediction system . . . . . . . . . . . . . . 16

vi

2 3

Liste des tableaux 3.1

Overall oracle output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 4.10 4.11 4.12 4.13 4.14 4.15 4.16 4.17 4.18 4.19 4.20 4.21

Oracle Result : House- 900 ; Appliance-Lamp . . . . . . . Oracle Result : House- 900 ; Appliance- Water Heater . . Oracle Result : House- 911 ; Appliance- Television . . . . Oracle Result : House- 911 ; Appliance- Lamp. . . . . . . Oracle Result : House- 925 ; Appliance- Lamp . . . . . . . Oracle Result : House- 925 ; Appliance- Oven . . . . . . . Oracle Result : House- 932 ; Appliance- Lamp . . . . . . . Oracle Result : House- 932 ; Appliance- Oven . . . . . . . Oracle Result : House- 939 ; Appliance- Washing Machine Oracle Result : House- 939 ; Appliance- Washing Machine Oracle Result : House- 951 ; Appliance- Cooker . . . . . . Oracle Result : House- 983 ; Appliance- Electric Heater . . Oracle Result : House- 951 ; Appliance- Lamp . . . . . . . Oracle Result : House- 986 ; Appliance- Television . . . . Classifier Comparison . . . . . . . . . . . . . . . . . . . . Neural Network parameter . . . . . . . . . . . . . . . . . . Training Algorithm . . . . . . . . . . . . . . . . . . . . . . Neural Network Architecture . . . . . . . . . . . . . . . . Neural Network parameters . . . . . . . . . . . . . . . . . Overall Result 1 . . . . . . . . . . . . . . . . . . . . . . . Overall Result 2 . . . . . . . . . . . . . . . . . . . . . . .

vii

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

23 23 23 24 24 25 25 25 26 26 26 27 27 28 28 29 29 29 29 31 31

Abbreviations STLF

Short Term Load Forecasting

TOD

Time Of Day

DOW

Day Of Week

SOY

Season Of Year

SVM

Support Vector Machines

NN

Nearest Neighbours

RBF

Radial Basis Function

viii

To my friends and family. . .

ix

Chapitre 1

Introduction

1.1

Introduction to Home Automation System

A home automation system basically consists of household appliances connected by an energy network and by a communication allowing the interaction between appliances. Home and building automation is traditionally used to increase comfort, to enable remote access to buildings and to increase the efficiency of buildings. These systems may aim at determining the best energy assignment plan and a good compromise between energy production and energy consumption [1]. In this study, energy is restricted to the electricity consumption and production. [1, 2] present a three-layers (anticipative layer, reactive layer and device layer) household energy control system. This system is both able to satisfy the maximum available electrical power constraint and to maximize a ratio between user satisfaction and cost. The objective of the anticipative layer explained in [3] is to compute plans for production and consumption of services. This approach is based on the available predictions. Figure 1.1 also describes the concept of energy management in homes and offices. Housing with the appliances aims at providing comfort to inhabitants thanks to services. Services can be decomposed into three kinds : the end-user services which produce directly comfort to inhabitants, the intermediate services which manage every storage and the support services which produce electrical power to intermediate and end-user services. Generally, when the home automation system is able to modify the behavior of a service, this service is qualified as modifiable by the system, for example, the modification of the starting time of a cooking service or the interruption of a washing service, etc. A service is qualified as permanent if its energetic consumption/production/storage covers

1

Chapter 1. Introduction

2

reactive and anticipative solver

satisfaction models

cost

physical variables (temperatures, ending times,...)

expectations

comfort

predictor (weather, occupancy, energy costs, usages)

Figure 1.1: Smart home concept

the whole time range of the energy assignment plan, otherwise, the service is named temporary service. In a home automation system, the user is not supposed to give the system his expectations (requested services). When the user’s demand is not known during a given period, the system must take into account this uncertainty by anticipating the energy needed for services. This helps the system to avoid some problems like peak consumption in this period. Therefore, the behavior of the inhabitant has to be modeled and integrated into the home automation system. In order to keep under control the total amount of consumed energy every hour, and then avoid peak consumptions and minimize the energy cost, the Home Automation System has to schedule as much as possible the energy consumptions in the most appropriate time periods. For example, the washing machine could be planned before or after the oven in a low energy cost period as far as such a plan satisfies the predicted user’s request. The efficiency of the anticipated plan is as good as the prediction of the user’s request. Indeed if the actual user’s behavior is far from the predicted one, then the reactive layer has to stop an appliance in order to satisfy the availability energy constraint for example, and schedule this appliance later without any energy cost optimization.


3

knowledge

observations, known data

local database : household history

UI

predictor

learning time time adaptative time unsupervised services

supervised services

predictions for all services

Figure 1.2: Goal of the prediction system

1.2

Requirement of Prediction and Approach

To anticipate the energy needed for a service in a home automation system, the system must take into account the uncertainty which can be provided by the user. In this context, a proper prediction of energy demand in housing sector is very important. A bottom-up approach can be used : first, the energy consumption prediction is done for each appliance in a home, then the forecast will be made for the total energy consumed in a home and, finally, a prediction can be made regarding the households supplied by a certain energy provider. Even if it is easier to predict overall consumption, it is important to be able to predict the consumption of each appliance because, regarding dynamic demand side management, it is also important to evaluate how much energy can be saved thanks to request to customers like unbalancing requests or energy price variations. The energy savings depend on appliances : some can be unbalanced, some can be postponed and some cannot be changed. The overall goal of the prediction is described in Figure 1.2. The overall goal also includes an User Interface where the user may provide his plans for the future. Our proposed approach is restricted to the prediction of appliance usage using only appliance consumption data and time of the event.

The objective of this work is to statistically predict the user energetic service request for the next 24 hours using a enrich learning algorithms with expert knowledge. We sample


4

the time space into 24 hour and wish to predict the user appliance usage requirement for a particular hour. At each point of time the Prediction system will predict for the following 24 hours and then shift to the next hour and predict the following 24 hours. The novelty of our approach lies in the proposed general model which is still lacking in the domain of appliance usage prediction for home automation system. It is a difficult problem because prediction at appliance level does not benefit from any profusion. It is very difficult to tackle these prediction problem with usual prediction approaches. The model is tested in the IRISE data set and the results indicate that the model is well suited for the application.

1.3

Problem Statement

Supervised services require predictions of home service demands. The goal of this project is to determine how to generate such predictions from patterns of appliance usages. The G-SCOP lab possesses one year of recorded activity data with a resolution of 10 minutes recorded for each electric appliances of a hundred of homes. The problem is to determine a appropriate model for learning to predict home service demands from such data. In the course of this report, we discuss the requirement of prediction for home automation system, the difficulty associated with prediction at appliance level, the proposed general model and its validation on the IRISE data set.

1.4

Summary and Research Report Organization

The work is divided into 6 chapters and we give a brief introduction to them in this section. Chapter 1 provides the introduction to the problem. Chapter 2 reviews the past literature, looking into details the past approaches commonly used in home automation system. Looking closely also to the past forecasting algorithms used at the grid level and the approaches which are applicable for the prediction of appliance in houses. Chapter 3 provides the details about the proposed approach and how we formalize the knowledge provided by expert. The details of the functions and statement which formalize the knowledge is discussed along with a step by step analysis of the proposed model.


5

Chapter 4 gives us the experimental results of our approach is a stepwise manner and also compares with other trivial knowledge driven predictor. Firstly, the results for the choice of the predictor is shown, then results obtained after the integration of expert knowledge into the prediction is shown. Lastly we give the results of the prediction of the 24 hours and also propose a weighted scoring which is more applicable in home automation system. Concluding remarks and scope of future work are presented in chapter 5.

Chapitre 2

Review of Past Literature 2.1

Past Approaches in Appliance Prediction

The problem of appliance usage prediction through consumption data is new. [4] deals with the problem of the user behavior prediction in a home automation system using a Bayesian network for a single appliance but a general model for appliance prediction is still lacking. Short term energy prediction at the grid level has been there for some time now and is discussed in section 2.3 but at the appliance level, these techniques are yet to be tested. In our effort we try to build a general model using a suitable predictor which can be applied to different Houses. We take into account the approaches in different domains and built a model which is general for appliance usage prediction. [4] and other subsequent work in the G-SCOP laboratory shows that regressive model is not very useful in appliance usage prediction. This is due to the randomness associated with the use of appliances. To get over this problem, we divide the day into 24 hour samples and try to predict if the appliance is consuming or not in the time slots. This sampling of the continuous time space in 24 discrete samples makes the prediction more realistic. The details of the proposed scheme is provided in chapter 3.

The literature is terms of methodology is still lacking in field of appliance prediction. Therefore firstly, in section 2.2 we review the Energy Management in smart homes which is the overall domain of the problem. Secondly, in section 2.3 we look in the domain of energy load forecasting at the grid level and review the approaches taken in this domain.

6

Chapter 2. Review of Past Literature

2.2

7

Energy Management in Smart Homes

Different authors studied control systems dedicated to homes for tracking purpose. [5] and [6] have proposed optimal control strategies for HVAC (Home Ventilation and Air Conditioning) system taking into account the natural thermal storage capacity of buildings that shift the HVAC consumption from peak-period to off-peak period. [6] has shown that this control strategy can save up to 10% of the electricity cost of a building. However, these approaches do not take into account the energy resource constraints, which generally depend on the autonomy needs of off-grid systems [7] or on the total power production limits of the suppliers in grid connected systems. [8] propose a temperature tracking control using dynamic matrix control (DMC), a variant of model predictive control (MPC). But, optimal tracking does not maximize energy usage efficiency. A home automation system basically consists of household appliances linked via a communication network allowing interactions for control purposes [10]. Thanks to this network, a load management mechanism can be carried out : it is called distributed control in [11]. The notion of building energy management and control system is described in [12]. This system consists of a set of appliances fitted with micro-controllers able to communicate via standard protocols. Some authors considered in particular the management of local production means and storage systems [13, 15]. Energy load management has been presented in [16, 17]. [18] used a dynamic programming approach. Other researchers used a multi-agent approach [20]. But, general approach of the energy management in living places yield new issues : – solving energy management problems where uncertainties are predominant. A three layer architecture has been proposed in [21, 22]. It is both able to satisfy the maximum available electrical power constraint and to maximize user satisfaction criteria. In order to adapt the consumption to the available energy. The home automation system controls the appliances by determining the starting time of services and also by computing the temperature set points of HVAC systems. This problem has been formulated as a multi-objective constraint satisfaction problem and has been solved by a dynamic Tabu Search and more efficiently by mixed linear programming. [23] proposes a way to take uncertainties into account during the optimization step. – solving large dimension optimization problems. It has been tackled using a mixed integer linear programing approach that can manage thousands of binary and continuous variables in [24]. Ways of transforming an energy management problem into a MILP, which is a regular problem, have been shown. – solving singular problems. Multi-agent approaches have been used to manage services that can only be modeled by nonlinear equations [26, 27].


8 prediction system

GUI GUI

energy cost and limitation service request forecast component predictor

weather forecast component anticipative solving system

smart plug

activation order & set points measurements

activation order smart appliance

current state control order

smart appliance

load level

anticipative solver service proxy service proxy

service parameter learning system reactive solver

service proxy supervision system

contextual problem data anticipative plan (set-points, starting times)

tranversal problem generator appliance andservice service service type service service problem problem problem problem problem generator generator generator

projected problem solver solution

MILP optimization solver

MILP with heuristics optimization solver

generator generator other solver...

generator repository service with user appliances satisfaction model model

environment models

power cost models

related to a given type of services

Figure 2.1: Architecture of a power manager

– generating dynamically the energy management problems to solve them because each living place is unique and evolving. Dynamic optimization problem generation has been studied in [32]. Software architecture and solving process have been depicted in [33]. Figure 2.1 illustrates the proposed solution. Technical aspects related to communication means within the smart house have been detailed in [34]. Hardware and software architectures to predict and optimally manage energy have been presented in [35]. The PlaceLab is a real home where the routine activities and interactions of everyday home life can be observed, recorded for later analysis, and experimentally manipulated. Volunteer research participants individually live in the PlaceLab for days or weeks, treating it as a temporary home. Meanwhile, a detailed description of their activities is recorded by sensing devices integrated into the fabric of the architecture [38, 39]. Three key challenges must be overcome from such work : – Need for comprehensive sensing ; – Need for labeled training datasets ; – Need for complex, naturalistic environments, to evaluate how typically users will react to a prototype technology in a representative setting


9

In [40], they are developing technologies and design strategies that use context-aware sensing to empower people with information that helps them to make decisions, but they do not control their environment. There are technical and human-computer interface advantages of creating systems that attempt to empower users with information at “teachable moments” rather than automating much decision-macking using “smart” or “intelligent” control. Uniqueness of housing systems involves a set of new issues in control system science : it is necessary to develop new tools and algorithms for globally optimized power management of the home appliances, able to anticipate difficult situations but also able to take into account the actual housing system state and the occupant expectations. This global control approach leads to the concept of energy smart home, which is more ambitious, that home automation. It should help to keep the balance between consumption and electricity production on the home scale but also at building, neighborhood and grid scales. Smart homes should be able to take into account external signals, like energy prices or unbalancing orders, and to modify the home appliance behaviors to compromise between occupants’ expectations and external actor wishes.

Three tools for acquiring data about people, their behavior, and their use of technology in natural settings are described [42] : – a context aware experience sampling tool, offers a variety of options for acquiring self-report data from users or subjects in experiments (software) – a ubiquitous sensing system that detects environmental changes can collects data via measurement of objects in the environment and compliments the self report data collected by the context-aware experience sampling device (software + hardware) – an image-based experience sampling system. this tools combines scene-based sensing and sampling techniques Moreover, uniqueness also requires cheap installation and maintenance costs because economy of scale is not possible. It means that the new tools and algorithms will have to be easy to install thanks to auto-discovering and auto-learning capabilities, easy to reconfigure and easy to repair. These issues involve sensing capabilities and intuitive human machine interfaces.

In [43] a system for recognizing activities in the home setting using a set of small and simple state-change sensors is introduced. The sensors can be quickly and ubiquitously installed in home environments. This system presents an alternative to sensors that are sometimes perceived as invasive, such as cameras and microphones.


10

Anticipating problematic situations require also prediction capabilities. Weather forecasts have to be fit to real housing environments, taking into account building aspects and masks involving shadings. Occupant behavior should also be predicted in order to avoid interrogating inhabitants about their intended activities. But predicting the use of an oven is a quite difficult problem with no apparent regularity. New prediction algorithms have to be found where data comes both from history and from the expression of intentions by occupants using proper HMIs.

2.3

Load Forecasting Algorithms

We need look into the domain of Load Forecasting at the grid level and review the approaches used in this domain. In the subsequent sections we will look into details and different approaches taken in this domain and select the approach which is relevant to our problem with modification. Load forecasts can be divided into three categories : – Short term forecasts : From one hour to one week. – Medium forecasts : From a week to a year. – Long-term forecasts : Longer than a year. The problem of appliance usage prediction have similarity with the short-term load forecasting (STLF) and we list some of the common approaches used in the later. Though STLF uses regressive approaches whereas our approach is based on classification but the strategies used in the domain of energy load prediction guides us in our choice of input to the predictor. [44] does a study on the approaches used in load prediction and we look into them in the following sub-sections.

2.3.1

Similar-day approach

This approach is based on searching historical data for days within one, two, or three years with similar characteristics to the forecast day. Similar characteristics include weather, day of the week, and the date. The load of a similar day is taken as a forecast [44].


2.3.2

11

Regressive methods

For electric load forecasting regression methods are usually used to model the relationship of loads consumption and other factors such as weather, day type, and customer class. [45] presented several regression models for the next day peak forecasting. But for appliance prediction [4] and other subsequent work in G-SCOP laboratory indicate that this model is not well suited.

2.3.3

Time series

Based on the assumption that the data have an internal structure, such as autocorrelation, trend, or seasonal variation. In particular ARMA, ARIMA, ARIMAX are the most often used classical time series methods. The idea of the time series approach is based on the understanding that a load pattern is nothing more than a time series signal with seasonal, weekly, and daily periodicities. Generally, techniques in time series approaches work well unless there is an abrupt change in the environmental or sociological variables which are believed to affect load pattern. Our work tries to take into account the time series approaches using a Neural Network predictor but it must be emphasized unlike in time series based energy forecasting our approach is classification based where the energy values are discretized in time as well as consumption of energy.

2.3.4

Expert system

Expert systems, incorporates rules and procedures used by human experts in the field of interest into software that is then able to automatically make forecasts without human assistance. [46] proposed a knowledge-based expert system for the short-term load forecasting of the Taiwan power system. Operators knowledge and the hourly observations of system loads along with weather parameters were taken into consideration. Our model also tries to make a general model which takes the expert knowledge into account. The detailed approach regarding how to formalize expert knowledge is discussed in 3.

2.3.5

Neural network

[47, 48, 49, 50] gives us the details of implementation of neural networks in the domain of energy load forecasting. These studies gives us an idea of the architecture and parameters that are most commonly used for energy load forecasting. Neural networks are essentially non-linear circuits that have demonstrated the capability to do non-linear curve fitting. The outputs of an artificial neural network are some linear or non-linear


12

Figure 2.2: STLF predictor input configuration

mathematical function of its inputs. The inputs may be the output of other network elements are arranged in a relatively small number of connected layers of elements between network inputs and outputs. For load forecasting the commonly used architecture is Hopfield, Back propagation, Boltzmann machine. In our work we used the Back propagation architecture. The following figure shows the input configuration of an implementation of neural networks in Short term load forecasting.

2.3.6

Support vector machines

Support vector machines perform a nonlinear mapping (by kernel functions) af the data into a higher dimensional feature space. Then support vector machines use a simple linear functions to create linear decision boundaries in the new space. [51] proposed a SVM model to predict daily load demand for a month. The problem of choosing an


13

architecture for a neural network is replaced here by the problem of choosing a suitable kernel function for support vector machine.

Chapitre 3

Proposed Approach 3.1

Detailed problem statement

The objective of this work is to build an enriched learning algorithm which takes knowledge into account and formalize it to statistically predict the user energetic service request for the next 24 hours. For the prediction of appliances from consumption data we first reduce the problem to a two class classification problem, i.e if an appliance is consuming at a particular hour or not. We discussed in chapter 2.1 regressive based model are not well suited for the application. The model for the prediction is for each appliance in each house. So we sample the time space into 24 hour and wish to predict the user appliance usage requirement for a particular hour. At each point of time the Prediction system will predict for the following 24 hours and then shift to the next hour and predict the following 24 hours. Our approach is that an expert proposes certain knowledge based on his domain expertise and then to formalize and represent this knowledge. We considered the knowledge representation in an incremental manner and at every stage validate the knowledge in terms of accuracy of predictor. The model proposed is a general one and all aspects of it are still not implemented in an automated way. The details of the proposed approach is discussed in section 3.3. Before we go into our proposed model, in section 3.2 we give a brief description of the database to understand the representation of the raw data.

3.2

Database

A database is obtained from Residential Monitoring to Decrease Energy Use and Carbon Emissions in Europe (REMODECE) which is a European database on residential 14

Chapter 3. Proposed approach

15

Figure 3.1: Screenshot of the data

consumption, including Central and Eastern European Countries, as well as new European Countries (Bulgaria and Romania). This database stores the characterization of residential electricity consumption by end-user and by country. The IRISE project has been chosen from REMODECE which deals only with houses in France. Each database concerns one house. In such a database, information is recorded every 10 minutes for each appliance in house and over one year. This information represents the consumed energy by each service, its data and its time. An example of this data is given in 3.1. Moreover, it is possible to know the number of people who live in each house. However, this data is not directly available. Let us notice that appliances are just involved in services : they are not central from the inhabitant point of view. Consequently, they are not explicitly modeled. The presence of the user is important but it is not predictable at the moment.

3.3

Proposed Model

The proposed model consists of enriched learning algorithm which proposes a general way to take expert knowledge into account. The proposed model divides the task into modules each of which has its own purpose. The general model is shown in figure 3.2 and in the following sub-sections we describe each of the processing modules in details. – Raw data contain – Energy consumption for an appliance . – Contextual information (Time, Date, Weather). – Oracle is composed of statements leading to entities (factors) that may be taken into account. – Data Selector is a non-temporal Matrix which stores and transforms the data in order to present them to the classifiers. – Predictor.


16

erreur

{{U(t); t ∈ τi }; i ∈ {0, −1, . . . }}

{Pi (t0 )}

prediction oracle

{Ci (t0 )}

switch

{D(t0 )}

memory

data selector

{{Y (t); t ∈ τi }; i ∈ {0, −1; . . . }}

predictor1

prediction selector

final prediction

erreur

{{U ! (t); t ∈ τi! }; i ∈ {0, −1, . . . }}

switch prediction predictor2

predictor controller

t0 is the current time τi is a given set of times that does not exceed ti Figure 3.2: Schematic representation of the prediction system

– Predictor Controller.

3.3.1

An expert system that generates knowledge : the oracle

We define Oracle knowledge as Statements leading to entities (or factors) that need to be taken into account. The oracle receives the raw data from the database giving the consumption at an particular hour and the date, time and weather information at that hour. The oracle proposes knowledge and then gives the necessary function which represents the data in a form interpretable by the Prediction system. The knowledge which are relevant for a particular appliance in a house might not be relevant for another house using the same appliance. So all knowledge proposed by the Oracle have to be validated and knowledge which doesn’t increase or reduces the accuracy of prediction for a particular appliance have to be rejected. The part of validating and structuring of the Oracle output is done in the subsequent processing module as seen in figure 3.3. Statements and the functional representation inside the oracle : – Immediate past history of consumption is meaningful to appliance usage prediction. – Hour of the day is meaningful to appliance usage prediction. – Day of the week is meaningful to appliance usage prediction. – Season of the year is meaningful to appliance usage prediction. – Previous days same hour is meaningful to appliance usage prediction.


17

In the following sub-sections we look at each of the proposed knowledge by the Oracle and their representation in details.

3.3.2

Past consumption history

The Oracle proposes that the past sequence of energy consumption prior to an event is meaningful in appliance usage prediction. We first try to describe the function mathematically and then give a detailed insight in what it is trying to represent.

Mathematically, it is formalized by the following predicate function Predicate function inputs : Consumption(H -1) ; Consumption(H -2) ;... ; Consumption(H -n) ; where, n = Is the size of the past time history H = Current hour Output : {0, 1}n Here the output is a thresholded binary vector of size n which signify if there is consumption at the hours prior to the event on not. Illustration : Suppose an appliance is used every three hours. To take this knowledge into account we need to know the consumptions just prior to that event. In this case we will find that for a event that the appliance is switched ON it was also switched ON three hour prior to that event.

3.3.3

Hour of the day

The Oracle proposes that the “time of the day” is meaningful to appliance usage prediction. In practice, by this knowledge the time space is discretized into 24 hour slots and the “actual time” the event occurs assumes priority. We first try to represent it mathematically and then provide some illustrations to better understand the representation.

Predicate function HOD inputs : Current Hour in the day, X X ∈ {(0 − 1), (1 − 2), (3 − 4), (5 − 6), ..., (23 − 24)} Hours Output : {0, 1}24 This is an orthogonal representation of an hour rather than a numeric value.


18

Illustration : If the Time of the day is 6.00 am, instead of using the numeric value 6 we use “0, 0, 0, 0, 0, 1, 0,...,0” to represent the same.

So for a day example of the representation will be :

Hour 0

—

(1,0,0,0,0,...,0)

Hour 1

—

(0,1,0,0,0,...,0)

.

—

(0,0,1,0,0,...,0)

.

—

(0,0,0,1,0,...,0)

Hour 23

—

(0,0,0,0,0,...,1)

Here the Oracle proposes that the time of the day is important in appliance prediction. Therefore it takes into account the time at which the event took place. Example : Let us assume, for a particular house the inhabitants sleep at 11.00 pm and wake up at 7.00 am. If we take time of the day into account then this knowledge can be learnt by the predictor as it will most of the time find the appliance between 11.00 pm to 7.00 am switched off.

3.3.4

Day of the week

The Oracle proposes that “Day of the week” is meaningful in appliance usage prediction. Similar to the way in 3.3.3 by taking this knowledge into account we want to discretized the whole week into 7 days. Instead of representing this with a numeric value, we use the orthogonal representation as in 3.3.3. Predicate function DOW inputs : Current day of the week, X X ∈ {Sunday, M onday, ..., Saturday} Output : {0, 1}7

So for a week example of the representation will be :

Chapter 3. Proposed approach Day (Sunday)

—

(1,0,0,0,0,0,0)

Day (Monday)

—

(0,1,0,0,0,0,0)

Day (Tuesday)

—

(0,0,1,0,0,0,0)

Day (Wednesday)

—

(0,0,0,1,0,0,0)

Day (Thursday)

—

(0,0,0,0,1,0,0)

Day (Friday)

—

(0,0,0,0,0,1,0)

Day (Saturday)

—

(0,0,0,0,0,0,1)

19

Here the Oracle proposes that the time of the day is important in appliance prediction. Therefore we sample the week in terms of 7 days. Example : Let us assume, for a particular house there is distinctive different behavior in weekdays and weekends. This knowledge will be taken into account if we sample the week into 7 days as illustrated.

3.3.5

Season of the year

Similar to the prior sub-sections the oracle proposes that season of the year is meaningful in appliance usage prediction. There are appliances in houses which show distinctive different behavior depending on the season of the year. As like the prior representations here we choose an orthogonal representation over a numeric one. Predicate Function SOY inputs : Current season of the year, X where X ∈ {Spring, Summer, Autumn, W inter} Output : {0, 1}4 So the season oracle output will be :

Season (Spring)

—

(1,0,0,0,)

Season (Summer)

—

(0,1,0,0)

Season (Autumn)

—

(0,0,1,0)

Season (Winter)

—

(0,0,0,1)

Here the Oracle proposes that the season of the year is important in appliance prediction. Therefore we sample the year into four seasons. Example : Let us assume that the electric heater is used primarily in the winter. This knowledge will be taken into account if we sample the year in terms of seasons.


3.3.6

20

Previous day same hour

Here the oracle proposes that what happens on previous days on the same hour is important in appliance prediction. So we look if there is consumption or not in the previous days for the same hour. The output is a vector of thresholded binary values of previous days at the same hour. Predicate function input : Consumption(H -24) ; Consumption(H -48) ;.. ; Consumption(H -n) ; where n = Is typically taken as 168 H = Current Day Output : {0, 1}7

Example : If we assume, in a particular house that people eat every day at a particular hour and also every week they do sport at a particular hour. These rules will be taken into account if we use this knowledge proposed by the Oracle.

3.3.7

Oracle output

The overall oracle output after the representation of all the knowledges proposed by the oracle is shown in table 3.3.7 where each row represent the proposed knowledge at a particular time in a incremental manner. The table is obtained by the incremental addition of knowledge proposed by the oracle, so the knowledge proposed in section 3.3.2 to 3.3.6 are added incrementally in order. The oracle has an available memory so we do it for the the hours available in the past history.

3.4

Data Selector

We define a data selector as an non-temporal matrix processor which stores, selects and structures the data for the predictor. This matrix is the input to the predictor. The data selector may choose the whole or the subset of the output from the oracle and this is controlled by the predictor controller discussed in section 3.6. It should be noted that all the knowledge proposed by the oracle might not be useful for a particular appliance in a house and there are different possible structuring of the different knowledge proposed by the oracle. All the outputs of the oracle is stored in the data selector, but only those which are validated by the predictor are selected for the overall prediction. This is controlled by the predictor controller. In our work we

Chapter 3. Proposed approach Knowledge Consumption(H -1) Consumption(H -2) ... Consumption(H -n) Hour of the day(0-1) Hour of the day(0-2) ... Hour in the day(23-24) Day of the week(Sunday) Day of the week(Monday) ... Day of the week(Saturday) Season of the year(Summer) ... Season of the year(Winter)

21 H 0/1 0/1 0/1 0/1 0/1 0/1 0/1 0/1 0/1 0/1 0/1 0/1 0/1 0/1 0/1

Table 3.1: Overall oracle output

have implemented only one of the possible structuring which is done by taking all of the knowledges proposed by the oracle as a single unit. There are other possible structuring which will be looked in the future.

3.5

Predictor

This module consists of the classifiers commonly used in Machine Learning such as the Neural networks. The classifier gets its input from the data selector. It first validates the knowledges proposed by the oracle and then the prediction for the next 24 hours. The detailed implementation of the classifiers are discussed in 4.3.

3.6

Predictor Controller

This module controls the interaction between the different predictors and also the structure of the data going into the predictor. In our implementation we are yet to use this module and the predictor module only consists of one predictor. The validity of each knowledge proposed for the Oracle is done manually, the automated control is this context in yet to be implemented.

Chapitre 4

Experimental Results 4.1

Introduction

The results are organized as follows – First the results obtained after the addition of expert knowledge is discussed. – The choice of a classifier is discussed in brief and the parameters for the same. – Lastly, we give the overall results and also its comparison with some trivial knowledge driven predictors. The scoring is always in terms of accuracy, where accuracy is the number of correct classification to the total number of cases.

4.2

Oracle knowledge results

In this section we validate each of the proposed knowledge in an incremental manner. The results indicate that all the knowledge proposed by the oracle might not result in increase in performance of the prediction system. So for each appliance in a house we need to select a subset of the knowledge proposed by the oracle. Therefore, only knowledge which increase in prediction performance is selected for a particular appliance. We randomly choose a home and an appliance in that house. All the prediction are done using a Neural Network Predictor whose parameters are discussed in 4.3. It must be mentioned, that the knowledge proposed by the oracle are prioritized on the basis of domain knowledge. It can be seen from table 4.1 to 4.14 that due to our incremental approach the knowledge which appears first has a higher chance of getting selected than the next one.

22

Chapter 4.Experimental Results Knowledge Past consumption + Time of the day + Day of the week + Season in the year + Same hour previous 7 day

23 Neural Network Prediction 82.94 83.45 83.73 84.14 83.50

Selected X X X X

Table 4.1: Oracle Result : House- 900 ; Appliance-Lamp

Table 4.1 is the prediction for lamp in House No : 900. The results indicate for this house the relevant knowledge are past consumption, time of the day, day of the week and season of the year. So these are selected for the overall prediction shown in 4.21.

Knowledge Past consumption + Time of the day + Day of the week + Season in the year + Same hour previous 7 day

Neural Network Prediction 94.87 95.00 95.009 95.101 95.05

Selected X X X X

Table 4.2: Oracle Result : House- 900 ; Appliance- Water Heater

Table 4.2 is the prediction for Water Heater in House No : 900. The results indicate for this house the relevant knowledge are past consumption, time of the day, day of the week and season of the year. The same hour previous 7 days does not increase the prediction chances in this case.



Selected X X X

Table 4.3: Oracle Result : House- 911 ; Appliance- Television

Chapter 4.Experimental Results

24

Table 4.3 is the prediction for Television in House No : 911. The results indicate for this house the relevant knowledge are past consumption, time of the day, day of the week. For this appliance in this house the season in the year is also not influential.



Selected X X

Table 4.4: Oracle Result : House- 911 ; Appliance- Lamp.

Table 4.4 is the prediction for Lamp in House No : 911. The results indicate for this house the relevant knowledge are past consumption and time of the day, other knowledge does not influence the prediction of lamp in this house.



Selected X X

X

Table 4.5: Oracle Result : House- 925 ; Appliance- Lamp

Table 4.5 is the prediction for Lamp in House No : 925. The results indicate for this house the relevant knowledge are past consumption, time of the day and Same hour previous days. So these are selected for the overall prediction shown in 4.21. This is among a few cases where same hour previous 7 days knowledge have acquired significance.

Table 4.6 is the prediction for Oven in House No : 925. The results indicate for this house the relevant knowledge are past consumption, time of the day and season of the



Selected X X X

Table 4.6: Oracle Result : House- 925 ; Appliance- Oven

year.



Selected X X X X


Table 4.7 is the prediction for lamp in House No : 932. The results indicate for this house the relevant knowledge are past consumption, time of the day, day of the week and season of the year. So these are selected for the overall prediction.



Selected X X X

Table 4.8: Oracle Result : House- 932 ; Appliance- Oven

Table 4.8 is the prediction for Oven in House No : 932. The results indicate for this house the relevant knowledge are past consumption, time of the day and day of the week. So these are selected for the overall prediction shown in 4.21.



Selected X

Table 4.9: Oracle Result : House- 939 ; Appliance- Washing Machine

Table 4.9 is the prediction for Washing Machine in House No : 939. The results indicate for this house the relevant knowledge are past consumption only. So these are selected for the overall prediction shown in 4.21.



Selected X X X

Table 4.10: Oracle Result : House- 939 ; Appliance- Washing Machine

Table 4.10 is the prediction for Washing Machine in House No : 939. The results indicate for this house the relevant knowledge are past consumption, time of the day and day of the week. So these are selected for the overall prediction shown in 4.21.



Selected X X

Table 4.11: Oracle Result : House- 951 ; Appliance- Cooker


27

Table 4.11 is the prediction for Cooker in House No : 951. The results indicate for this house the relevant knowledge are past consumption and time of the day. So these are selected for the overall prediction shown in 4.21. Knowledge Past consumption + Time of the day + Day of the week + Season in the year + Same hour previous 7 day


Selected X X

Table 4.12: Oracle Result : House- 983 ; Appliance- Electric Heater

Table 4.12 is the prediction for Electric Heater in House No : 983. The results indicate for this house the relevant knowledge are past consumption, time of the day, day of the week and season of the year. So these are selected for the overall prediction shown in 4.21. Knowledge Past consumption + Time of the day + Day of the week + Season in the year + Same hour previous 7 day


Selected X X X X


Table 4.13 is the prediction for Lamp in House No : 951. The results indicate for this house the relevant knowledge are past consumption, time of the day, day of the week and season of the year. The same hour previous 7 days does not increase the prediction chances in this case.

Table 4.14 is the prediction for Television in House No : 986. The results indicate for this house the relevant knowledge are past consumption, time of the day, day of the week and season of the year. So these are selected for the overall prediction shown in 4.21.


28



Selected X X X X

Table 4.14: Oracle Result : House- 986 ; Appliance- Television

4.3

Choice of classifier, parameters and implementation

In this section we discuss the justification of using a neural network classifier for such an application. Our choice of neural networks are more on the basis of past literature than on the initial results seen in table 4.15, where we have compared different non-linear classifiers. The comparison is based on only the past consumption and then prediction for the next hour. We are providing the comparison with other classifiers such as Support vector machines, Naive Bayes and K-nearest Neighbours. It must be mentioned, that at no point we are disregarding the fact that other classifiers may perform better, these are initial results with suitable parameters. In the future we wish to do an extensive study on the choice of the classifier and the parameters for the same. The results are the accuracy for the next hour. Our choice of parameters are also based on past literature and initial results shown in table 4.16 to table 4.18. This study is for the choice of the parameters of the Neural network. The number of hidden layers are chosen to be one and the number of hidden neurons to be half of the number of input nodes. This choice is to avoid the over fitting or under-fitting of the network. Table 4.16 gives the results of using different numbers of hidden neurons. The choice of training algorithm is also shown in 4.17. The results of the choice of architecture is shown in 4.18. The final choice of all the parameters are shown in table 4.19. Appliance

SVM

900 lamp 932 oven

82.40 84.42

Naive Bayes 37.1 84.25

KNN 79.72 83.51

Neural Network 82.94 84.95

Table 4.15: Classifier Comparison

The problem is implemented in two phases. First, the expert knowledge is generated using a C program which gives a CSV file as output. As mentioned earlier the selection of the relevant knowledge for a predictor is done manually, therefore the data selector is yet to be implemented in an automated manner. We train the classifier with 75 percent


29

Hidden Hurons No of input nodes (No of input nodes)/2 (No of input nodes)/4

accuracy 82.76 82.85 83.13

Table 4.16: Neural Network parameter

Training Algorithm Gradient descent BGFS Conjugate entropy

accuracy 57.20 82.94 83.08

Table 4.17: Training Algorithm

Architecture RBF MLP

accuracy 57.20 83.22

Table 4.18: Neural Network Architecture

of the cases and test it with the 25 percent of the cases. In the second phase we use the classifiers to get the prediction from the CSV. The implementation of the predictor is done using the tool Statistica. The scoring is done in terms of accuracy, where accuracy is the number of correct classification to the total number of classification. Two scoring schemes such as the weighted accuracy and average accuracy are used.

Parameter Sampling Method Train sample size Test sample size Network Type Activation function(hidden unit) Activation function(output unit) No of hidden neurons Error Function Training Algorithm Learning Rate

Selection Random 75 25 MLP Tanh Softmax no of input/2 Cross entropy BGFS 0.1

Table 4.19: Neural Network parameters


4.4

30

Overall Result

After the selection of the data, we predict the following 24 hours at each hour and the results in terms of accuracy are shown in 4.20. Here we score the prediction system by two methods, one is by simple averaging all the accuracy for the 24 hours and the second one is a weighted average. The proposed weighing scheme is expressed by the following equation :

⇒

P 2(24−i) 25∗24

∗ Accuracy[i] for i=0,1,...,23

Results contain both the scoring methods and also the results of some trivial knowledge based predictors. The trivial knowledge based predictors are – The predictor that always predicts the appliance wont start : Never starts. – The predictor that always predicts the appliance will start : Always start. – The predictor that predicts “what happens the previous day at the same hour happens the next day”, i.e 24 hour similarity. – The predictor that predicts “what happens the previous week at the same hour happens the next day”, i.e 168 hour similarity. – The predictor that predicts “what happen a random hour back happens the next hour” , i.e random hour similarity. These estimates are important because they give us an overall idea of the performance of our proposed model.

4.5

Discussion

The results indicates that our model works better than other trivial knowledge based predictors. Previous works on appliance usage prediction from consumption data relied heavily on the assumptions expressed in the trivial knowledge based predictor. The assumption of 24 hour or 168 hour similarity is intuitive but other knowledges also need to be incorporated to make the system dynamic. The incorporation and representation of the expert knowledge helps the system to perform better as seen in the tables 4.20 and 4.21. Though our model performs better but the structuring of the expert knowledge could be improved. We can also see in the section 4.2 that with the addition of knowledge incrementally the performance of the system does not increase sufficiently.


31

Now, from the results we get an overall idea about the predictability of the appliance. Though it must be mentioned here that the high prediction for some appliances at homes are due to the fact that some appliances are ON or OFF at most of the time. Our results from the confusion matrix (not listed) is that we are able to predict both the categories (ON and OFF) for the appliances. Appliances which are started very few times seem to require less knowledge for prediction. Among appliances, from the results it indicates that the lamp requires most the knowledge proposed by the expert among other appliances. The applicability of expert knowledge varies not only from appliance to appliance but also from House to House as the user behavior is different. In the future a better structuring and parameter estimation need to be done for more accurate results. The results in tables 4.20 and 4.21 are for the same appliances, so they are separated into two for better readability. The number before the appliance name is the house number. Appliance

Never Start

Always Start

24 hour similarity

168 hour similarity

900 Lamp 983 Electric Heater 925 Lamp 932 Oven 986 TV 951 Cooker 939 Washing machine

43.73 71.76

56.26 28.23

71.25 94.38

66.99 89.39

Random hour similarity 50.36 92.35

32.02 84.46 72.13 93.94 88.88

67.97 15.53 27.86 6.05 11.11

88.014 80.32 71.36 89.72 86.25

83.35 80.76 69.43 89.68 86.79

80.33 72.59 57.30 88.80 78.45

Table 4.20: Overall Result 1

Appliance

900 Lamp 983 Electric Heater 925 Lamp 932 Oven 986 TV 951 Cooker 939 Washing machine

Neural Networks (Average accuracy) 78.25 96.31

Neural Network (weighted accuracy) 78.85 96.68

90.55 86.00 76.64 94.18 89.55

90.82 86.03 77.18 94.21 89.89

Table 4.21: Overall Result 2

Chapitre 5

Conclusion and Perspective To anticipate the energy needed for a service in a home automation system, the system must take into account the uncertainty which can be provided by the user. In this context, a proper prediction of energy demand in housing sector is very important. This work focuses on the prediction of the appliance usage in housing because it is a very important problem in a home automation system. The objective is to construct a model able to predict the appliance usage in housing which help the system to organize energy production and consumption and to decide which appliance will be used at each hour (energy planing). In this work we tried to predict if a particular appliance will be used at a particular hour looking 24 hours in the future. The proposed model tried to formalize expert knowledge using predicate functions and also find a suitable data structuring for the classifier. The model is validated using an IRISE database which contains the consumption record of 100 houses for a period of 1 Year. Our initial results indicate that the approach is useful in appliance usage prediction and its comparison with other trivial knowledge based predictor validates our approach. This model is applied to a wide range of appliances and houses and the initial results are encouraging.

Though we were able to propose a general model for appliance usage prediction, we are far from reaching the overall goal. In the overall project we wish to take into account the user interaction with the system, their behavior characteristics among others. The prediction for certain houses are more difficult than that for the others. This may be due to a variety of reasons among them could be the number of people living in the house. The information about the inhabitants in the house and there general behavior patterns is still lacking. Also, in some cases our implementation have manual intervention. We need to take into account more expert knowledge and deal with appliances separately in this context, for example concepts such as aggregation of 1 hour intervals to 3 hour 32

Chapter 5. Conclusion and perspectives

33

intervals for hours that are difficult to predict may also be considered.

The prediction system is also limited by the nature of the data which gives the consumption at a particular time. Many learning methods of activity modeling could be applied if we are able to know the user activity information. So, the main assumption of our approach is that appliance usage can be predicted from past history. This assumption has its own limitations. Going in the future our aim is to build a general, fully automated and user interactive prediction system for home automation and simulate how the prediction is actually helping energy management in homes.

Bibliographie [1] Seung Ha, Hong Jung, and Yong Oh. Method to analyze user behavior in home environment. Personal Ubiquitous Comput., 10 :110–121, January 2006. ISSN 16174909. doi : http://dx.doi.org/10.1007/s00779-005-0016-9. URL http://dx.doi. org/10.1007/s00779-005-0016-9. [2] Shadi Abras, Stéphane Ploix, Sylvie Pesty, and Mireille Jacomino. A multi-agent design for a home automation system dedicated to power management. In Christos Boukis, Aristodemos Pnevmatikakis, and Lazaros Polymenakos, editors, Artificial Intelligence and Innovations 2007 : from Theory to Applications, volume 247 of IFIP International Federation for Information Processing, pages 233–241. Springer Boston, 2007. URL http://dx.doi.org/10.1007/978-0-387-74161-1_25. [3] S. Abras, S. Pesty, S. Ploix, and M. Jacomino. An anticipation mechanism for power management in a smart home using multi-agent systems. In Information and Communication Technologies : From Theory to Applications, 2008. ICTTA 2008. 3rd International Conference on, pages 1 –6, april 2008. doi : 10.1109/ICTTA.2008. 4530305. [4] Lamis Hawarah, Stéphane Ploix, and Mireille Jacomino.

User behavior pre-

diction in energy consumption in housing using bayesian networks.

In Leszek

Rutkowski, Rafal Scherer, Ryszard Tadeusiewicz, Lotfi Zadeh, and Jacek Zurada, editors, Artificial Intelligence and Soft Computing, volume 6113 of Lecture Notes in Computer Science, pages 372–379. Springer Berlin / Heidelberg, 2010. URL http://dx.doi.org/10.1007/978-3-642-13208-7_47. [5] J.M. House and T.F. Smith. Optimal control of building and hvac systems. In Proceddings of the American Control Conference, Seattle, Washington, 1995. [6] Guo Zhou and Moncef Krarti. Parametric analysis of active and passive building thermal storage utilization. Journal of Solar Energy Engineering, 127 :37–46, 2005.

34

Bibliography

35

[7] M. Muselli, G. Notton, P. Poggi, and A. Louche. Pv-hybrid power system sizing incorporating battery storage : an analysis via simulation calculations. Renewable Energy, 20 :1–7, 2000. [8] L. Pedersen, J. Stang, and R. Ulseth. Load prediction method for heat and electricity demand in buildings for the purpose of planning for mixed energy distribution systems. Energy and Buildings, 40(7) :1124–1134, 2008. [9] Christian Ghiaus and Ion Hazyuk. Calculation of optimal thermal load of intermittently heated buildings. Energy and Buildings, 42, 2010. [10] Peter Palensky and Ratko Posta. Demand side management in private home using lonworks. In Proceedings of the IEEE International Workshop on Factory Communication Systems, 1997. [11] K.P. Wacks. The impact of home automation on power electronics. In Applied Power Electronics Conference and Exposition, pages 3–9, 1993. [12] K. Stum, R. Mosier, and T. Haasl. Energy management systems. Technical report, Porland Energy Conservation Inc.(PECI), 1997. [13] Gregor P. Henze and Robert H. Dodier.

Adaptive optimal control of a grid-

independent photovoltaic system. Transactions of the ASME, 125 :34–42, 2003. [14] Guillaume Foggia. Pilotage Optimal de Système Multi-sources pour le Bˆ atiment. PhD thesis, Grenoble Institute of Technology (INPG), 2009. [15] Julien Eynard. Gestion optimale de l’énergie dans un procédé multi-source pour le chauffage de bˆ atiments. PhD thesis, l’Université de Perpignan Via Domitia, 2010. [16] Ha Duy Long, S. Ploix, E. Zama¨ı, and M. Jacomino. Control of energy consumption in home automation by ressource constraint scheduling. In The 15th international conference on control systems and computer science, Buccarest, Romania, 2005. [17] Long Duy Ha, S. Ploix, E. Zama¨ı, and M. Jacomino. Tabu search for the optimization of household energy consumption. In The 2006 IEEE International Conference on Information Reuse and Intergration IEEE IRI 2006, Waikoloa Hawaii, USA, September 16-18 2006. [18] Yann Riffoneau. Gestion des flux énergétiques dans un système photovolta¨ıque avec stockage connecté au réseau. PhD thesis, Grenoble Institute of Technology (INPG), 2010. [19] Yoseba Penya. Last generation applied artificial intelligence for energy management in building automation. In The 5th International Conference on Fieldbus Systems and their Applications, Aveiro, Portugal, 2003.

Bibliography

36

[20] R. R. Negenborn. Multi-agent model predictive control with applications to power networks. PhD thesis, Technische Universiteit Delft, Delft, Netherlands, december 2007. [21] L. Duy Ha, S. Ploix, E. Zama¨ı, and M. Jacomino. A home automation system to improve household energy control. In The 12th IFAC Symposium INCOM2006, 2006. [22] Long Duy Ha, S. Ploix, E. Zama¨ı, and M. Jacomino. Realtimes dynamic optimization for demand-side load management. International Journal of Management Science and Engineering Management, 3(4), November 2008. [23] Long Duy Ha, Hoang Le Minh, and S. Ploix. An approach for home load energy management problem in uncertain context. In IEEE International Conference on Industrial Engineering and Engineering Management, Singapor, 8-11 december 2008. [24] Long Duy Ha, Stephane Ploix, F. Wurtz, Pierre Perichon, and J. Merten. Energy management system for a photovoltaic grid-connected building. In 24th EU PVSEC and 4th World Conference on Photovoltaic Energy Conversion, Hamburg, Germany, September, 21-26 2009. [25] Long Duy Ha, S. Ploix, M. Jacomino, and Hoang Le Minh. Energy Management, chapter A mixed integer programming formulation of the home energy management problem. ISBN 978-953-307-065-0. INTECH, 2010. [26] S. Abras, S. Ploix, S. Pesty, and M. Jacomino. Informatics in Control, Automation and Robotics III, volume 15 of Lecture Notes in Electrical Engineering, chapter A multi-agent home automation system for power management. Springer Berlin Heidelberg, 2006. [27] S. Abras, S. Ploix, S. Pesty, and M. Jacomino. A multi-agent design for a home automation system dedicated to power management. In The 4th conference on Artificial Intelligence Innovations & Applications, Athens, Greece, september, 1921 2007. [28] S. Abras, S. Ploix, S. Pesty, and M. Jacomino. Artificial intelligence and innovations : from theory to applications, volume 247, chapter A multi-agent design for a home automation system dedicated to power management, pages 233–242. Springer eBooks collection, isbn 0387741615, 9780387741611, edition, 2007. [29] S. Abras, S. Ploix, S. Pesty, and M. Jacomino. Advantages of mas for the resolution of a power management problem in smart homes. In 8th International Conference

Bibliography

37

on Practical Applications of Agents and Multi-Agent Systems, PAAMS’2010, Salamanca, Spain, 26-28 April 2010. Springer Verlag. [30] S. Abras, S. Ploix, and S. Pesty. Housing, Housing Costs and Mortgages : Trends, Impact and Prediction, chapter Managing Power in a Smart Home Using MultiAgent Systems. Number ISBN 978-1-60741-813-9 in Housing Issues, Laws and Programs. Nova Publishers, 2010. [31] Ahmed Mustafa Elmahaiawy, Newal Elfishawy, and Mohamed Nour El-Dien. Anticipation the consumed electrical power in smart home using evolutionary algorithms. In MCIT 2010 conference, 2010. [32] G. Warkozek, M. Jacomino, S. Ploix, and F. Wurtz. Generic formulation of optimization problems for energy management : solving difficulties, practical and mathematical analysis. In Proceedings of the 8th International Symposium on Electric and Magnetic Fields (EMF2009), Mondovi, Italy, May 26th-29th 2009. [33] S. Ploix, Ha Duy Long, S. Abras, M. Jacomino, D. Wurtz, and S. Bacha. patent us/eu 61/326,542 : Environment and method for managing services in a living place, 2010. [34] Smarthouse/smartgrid project - deliverable 1.2 - technology trends for smarthouse/smartgrid. Technical report, EU/ICT project EU FP7-ICT-2007-2 STREP 224628, 2009. [35] Craig Howard Miller. patent us2006/0276938 : Optimized energy management system, 2006. [36] Montford University. Patent gb2448896a : Energy management system, 2007. [37] Michael Hoeynck and Burton W. Andrews. Patent us2010/0025483 : Sensor-based occupancy and behavior prediction method for intelligently controlling energy concumption within a building, 2010. [38] Stephen S. Intille, Kent Larson, J. S. Beaudin, J. Nawyn, E. Munguia Tapia, and P. Kaushik. A living laboratory for the design and evaluation of ubiquitous computing technologies. In In Extended Abstracts of the 2005 Conference on Human Factors in Computing Systems, pages 1941–1944. ACM Press, 2005. [39] Stephen Intille, Kent Larson, Emmanuel Tapia, Jennifer Beaudin, Pallavi Kaushik, Jason Nawyn, and Randy Rockinson. Using a Live-In Laboratory for Ubiquitous Computing Research. 2006.

Bibliography

38

[40] Stephen S. Intille. The goal : smart people, not smart homes. In In Proceedings of the International Conference on Smart Homes and Health Telematics, IOS. Press, 2006. [41] Stephen S. Intille. Designing a home of the future, 2002. [42] Stephen S. Intille, Emmanuel Munguia Tapia, John Rondoni, Jennifer Beaudin, Chuck Kukla, Sitij Agarwal, Ling Bao, and Kent Larson. Tools for studying behavior and technology in natural settings. In In Proceedings of UBICOMP 2003, pages 157–174. Springer, 2003. [43] Emmanuel Munguia Tapia, Stephen S. Intille, and Kent Larson. Activity recognition in the home using simple and ubiquitous sensors. In In Pervasive, pages 158–175, 2004. [44] Eugene A. Feinberg and Dora Genethliou. Load forecasting. In Joe H. Chow, Felix F. Wu, and James Momoh, editors, Applied Mathematics for Restructured Electric Power Systems, Power Electronics and Power Systems, pages 269–285. Springer US, 2005. ISBN 978-0-387-23471-7. URL http://dx.doi.org/10.1007/ 0-387-23471-3_12. [45] Engle R. F., Mustafa C., and Rice J. Modelling peak electricity demand. Journal of Forecasting,, page 241–251, 1992. doi : 10.1002/for.3980110306. [46] Ku-Long Ho, Yuan-Yih Hsu, Chuan-Fu Chen, Tzong-En Lee, Chih-Chien Liang, Tsau-Shin Lai, and Kung-Keng Chen. Short term load forecasting of taiwan power system using a knowledge-based expert system. Power Systems, IEEE Transactions on, 5(4) :1214 –1221, nov 1990. ISSN 0885-8950. doi : 10.1109/59.99372. [47] H.S. Hippert, C.E. Pedreira, and R.C. Souza. Neural networks for short-term load forecasting : a review and evaluation. Power Systems, IEEE Transactions on, 16 (1) :44 –55, feb 2001. ISSN 0885-8950. doi : 10.1109/59.910780. [48] A.G. Bakirtzis, V. Petridis, S.J. Kiartzis, M.C. Alexiadis, and A.H. Maissis. A neural network short term load forecasting model for the greek power system. Power Systems, IEEE Transactions on, 11(2) :858 – 863, may 1996. ISSN 0885-8950. doi : 10.1109/59.496166. [49] D.C. Park, M.A. El-Sharkawi, II Marks, R.J., L.E. Atlas, and M.J. Damborg. Electric load forecasting using an artificial neural network. Power Systems, IEEE Transactions on, 6(2) :442 –449, may 1991. ISSN 0885-8950. doi : 10.1109/59.76685. [50] A. Khotanzad, R. Afkhami-Rohani, and D. Maratukulam. Annstlf-artificial neural network short-term load forecaster generation three. Power Systems, IEEE Transactions on, 13(4) :1413 –1422, nov 1998. ISSN 0885-8950. doi : 10.1109/59.736285.

Bibliography

39

[51] Bo-Juen Chen, Ming wei Chang, and Chih-Jen Lin. Load forecasting using support vector machines : A study on eunite competition 2001. Technical report, 2001.