SlideShare a Scribd company logo
Computational Intelligence
           and Energy Systems: intelligent solutions for
                                    complex problems

                                                Matteo De Felice
                        Unità Modellistica Energetica Ambientale
                                                 UTMEA - ENEA




Tuesday, May 31, 2011                                              1
Sommario

                        Cos’è la Computational
                        Intelligence (CI)?
                        Quali sono le applicazioni della
                        CI ai sistemi complessi?




Tuesday, May 31, 2011                                      2
CI: paradigmi
                                                      Soft
                                    NN              Computing

                          EC                   FS



                                                     IA?
                               SI        AIS



                        Computational Intelligence
Tuesday, May 31, 2011                                           3
Visione d’insieme
                                                      Temi
                                    NN              principali

                          EC                   FS




                               SI        AIS




Tuesday, May 31, 2011                                            4
CI e letteratura
                                   3
                            x 10
                        5
                                   Evolutionary Computation
                                   Swarm Intelligence
                        4          Artificial Neural Networks

                        3


                        2


                        1


                         0
                        1994        1996    1998    2000    2002       2004   2006   2008   2010
                                                                year




                        Dati dalla Thomson Reuters ISI considerando Computer Science &
                        Technology (Gennaio 2010)

                        Due journals sulla CI nei primi 10 in CS (IF 2009)


Tuesday, May 31, 2011                                                                              5
La diffusione della CI
                        Problemi sempre più complessi
                        Più potenza di calcolo
                        disponibile




Tuesday, May 31, 2011                                   6
ma...
                        Assenza di una teoria
                        consolidata
                        Frammentazione degli algoritmi
                        Approccio poco sistematico e
                        confronti poco “robusti”
          PSO APSO CPSO DPSO EPSO FPSO GPSO HPSO IPSO
          LPSO MPSO NPSO OPSO PPSO QPSO RPSO SPSO TPSO
          UPSO VPSO WPSO GA AGA BGA CGA DGA EGA FGA
          HGA IGA KGA LGA MGA OGA PGA QGA RGA SGA
          VGA ...
Tuesday, May 31, 2011                                    7
Applicazioni Principali
                                            Reti Neurali
                                            & Logica Fuzzy


                        1) Modellazione & Forecasting
                        2) Ottimizzazione

                                            Calcolo
                                            Evolutivo

Tuesday, May 31, 2011                                        8
Quadro generale
                                              Reti neurali evolutive con topologia a rete complessa
                            Evolving predictive neural models for complex processes
                                    Evolving Complex Neural Networks




                   2008
                                    Reti Neurali Evolutive


                                                      Ensemble
                        Artificial Neural Networks and Support Vector
                        Machines ensembling: a comparison




Tuesday, May 31, 2011                                                                                 9
2009
                          Ambient temperature modelling with soft computing techniques




                                      Modellazione
                                temperature con NN


              Combining Back-Propagation and Genetic
              Algorithms to Train Neural Networks for Ambient
              Temperature Modeling in Italy




                        2010
                        Ottimizzazione dello start-up
                           centrale a ciclo combinato
                             Combining Back-Propagation and Genetic
                             Algorithms to Train Neural Networks for Ambient
                             Temperature Modeling in Italy

Tuesday, May 31, 2011                                                                    10
Reti Neurali e Load Forecast




                        2011
                         Short-Term Load Forecasting with Neural
                         Network Ensembles: a Comparative Study




                        Climate Variables in Energy Modeling




Tuesday, May 31, 2011                                              11
Altri Progetti
       Identificazione      Reti Neurali      Algoritmi Evolutivi     Ottimizzazione
      Structural System     Evolutive e         Spazialmente        traiettorie missioni
       in Ing. Sismica    Applicazioni alla       Strutturati         interplanetarie
                              Finanza




Tuesday, May 31, 2011                                                                      12
Ottimizzazione
Tuesday, May 31, 2011                    13
Process Optimization
                            Process
                           Parameters     Process     Environment
                               (X)




                                        Measurement




                        Come migliorare la
                        ‘performance’ di un processo
                        tramite i suoi parametri?

Tuesday, May 31, 2011                                               14
Ottimizzazione tradizionale


                        Metodi Line-search and trust-
                        region (serve l’Hessiano!)
                        Metodi Quasi-newton (Hessiano
                        approssimato)
                        Metodi Derivative-free


Tuesday, May 31, 2011                                   15
...ma il real-world è:

                        1) ‘Rumoroso’

                        2) Dinamico

                        3) Difficile da esaminare



Tuesday, May 31, 2011                               16
Evolutionary Computation (EC)



                        Ottimizzazione Black-box
                        Singolo e Multi-Obiettivo
                        Anche funzioni discontinue e non
                        differenziabili
                        Meta-euristica Population-based


Tuesday, May 31, 2011                                      17
Metaeuristica

                        Ottimizzazione Stocastica
                        Algoritmi usati per trovare
                        soluzioni a problemi “difficili”
                        Esempio: Hill-Climbing, Tabu
                        Search, Simulated-Annealing


Tuesday, May 31, 2011                                      18
Real-World problems




Tuesday, May 31, 2011                         19
Metodi di ottimizzazione
                        DIRECT Algorithm
                             Applications


 Taxonomy of Methods




                               Yves Brise   Lipschitzian Optimization, DIRECT Algorithm, and Applications
Tuesday, May 31, 2011                                                                                       20
Applicazione
                        Ottimizzazione dello startup di una centrale a
                        ciclo combinato (CCPP)

                        Minimizzazione del tempo di avvio, consumi,
                        emissioni e stress termico

                        Massimizzazione della produzione di energia
     M. De Felice, I. Bertini, A. Pannicelli, and S. Pizzuti, "Soft Computing based
     optimisation of combined cycled power plant start-up operation with fitness
     approximation methods," Applied Soft Computing, 2011.


              I. Bertini, M. De Felice, F. Moretti, and S. Pizzuti, "Start-Up Optimisation of a
              Combined Cycle Power Plant with Multiobjective Evolutionary Algorithms," in
              Applications of Evolutionary Computation, 2010, pp. 151-160.




Tuesday, May 31, 2011                                                                             21
Procedura
                    1. Definizione di un indice di
                       performance
                    2. Impostazione simulatore sw
                    3. Algoritmo EC tramite simulatore




Tuesday, May 31, 2011                                    22
Indice Performance
                                                      F1

                                     1



                 Informazioni
                                    0.5
                                     0
                                      0   0.5    1    1.5    2    2.5          3
                                                                           4
                                                                        x 10


                 dagli esperti di
                                                      F2

                                     1
                                    0.5



                 processo            0
                                      0   0.5    1    1.5

                                                      F3
                                                             2    2.5
                                                                           5
                                                                        x 10
                                                                               3



                                     1



                 Knowledge
                                    0.5
                                     0
                                      0          5          10             15
                                                                           9
                                                                        x 10


                 modeling con
                                                      F4

                                     1
                                    0.5



                 funzioni fuzzy
                                     0
                                      0   5     10    15    20    25       30

                                                      F5

                                     1
                                    0.5
                                     0
                                      0   50    100   150   200   250      300




Tuesday, May 31, 2011                                                              23
Singolo-obiettivo

                        Algoritmo Genetico
                        operazione di mutazione
                        Gaussiano
                        Funzione di fitness approssimata
                        per velocizzare l’ottimizzazione
                        (da 2070 a 36 ore/CPU)

Tuesday, May 31, 2011                                      24
Risultati

                        Tempo               Prod.                 Stress
                                 Consumi             Emissioni
                         avvio             Energia               Termico

          Esperti       21070    143557    2.5•109      25         10


              GA        16569    115070 1.86•109       18.8       78.4

           Var.
                        -25%      -16%      -16%      -30%        2%
          Norm.



Tuesday, May 31, 2011                                                      25
Multi-obiettivo
                                                  12.65

                                                   12.6

                                                  12.55
                        Emissions (mg s / N m3)




                                                   12.5

                                                  12.45

                                                   12.4

                                                  12.35

                                                   12.3     Real
                                                            NSGA 2
                                                  12.25     WSGA
                                                            RAND
                                                   12.2
                                                      3.9     4      4.1      4.2         4.3       4.4   4.5      4.6
                                                                           Energy Production (KJ)                  9
                                                                                                                x 10




Tuesday, May 31, 2011                                                                                                    26
Modellazione & Forecasting
Tuesday, May 31, 2011                         27
Modellazione con NNs
                                                ||F (x) − f (x)|| < , ∀x

                                 1.2



                                 0.6
                        Y Axis




                                   0
                                        0   1       2       3            4   5   6   6.5


                                 -0.6



                                 -1.2

                                                                X Axis




                                                        y = sin(x)
                                                            NN(x)


Tuesday, May 31, 2011                                                                      28
Modellazione con NNs
                          Disturbances



             Input u(k)                  Output y(k)
                            System



                            Neural
                            Network




                          Errore (MSE)
                          Metodi empirici per decidere la
                          topologia della rete
Tuesday, May 31, 2011                                       29
Regressione con NN
                        Si una una NN per fare
                        regressione non-lineare




Tuesday, May 31, 2011                             30
Time Series Forecasting

                        Possiamo fare una previsione dei
                        dati futuri usando quelli osservati




                                  Altre informazioni utili
                                  (!)


Tuesday, May 31, 2011                                         31
Approcci per le NN
                                                       y(t+1)
                        Input at      Neural           y(t+2)
                                                         ...    Direct Method
                         time t       Network
                                                       y(t+N)




                           Input at
                                                  output t+1
                            time t      Neural
                                        Network
                           output t
                                                                Iterative Method


                                         delay


Tuesday, May 31, 2011                                                              32
Short-Term Load Forecasting
              60

              40
         kW




              20

              0
               0        200    400   600   800   1000     1200   1400   1600   1800   2000
                                                    hours




                              Dati Orari
                              Obiettivo: predizione del carico
                              fino a 24 ore


Tuesday, May 31, 2011                                                                        33
Modelli Seasonal
                           1



                          0.5



                           0



                          0.5
                             0   10     20    30    40    50




                         Implementazione in R

                        ΦP (B s )φ(B)∇D ∇d xt = α + ΘQ (B s )θ(B)et
                                      s


Tuesday, May 31, 2011                                                 34
Modello NN


                   Campioni
                    passati
                                Rete
                                         Previsione
                               Neurale
                Informazioni
                 aggiuntive




Tuesday, May 31, 2011                                 35
Rete Neurale

                                             Funzioni di attivazione f
                                             differenziabili




                   Pesi w   i
                                Pesi w   o




Tuesday, May 31, 2011                                                    36
Backpropagation

                        [Werbos, 1974]
                        Forward phase: il segnale si
                        propaga “in avanti”
                        Backward phase: si calcola
                        l’errore e lo si propaga
                        “all’indietro”, modificando i pesi

Tuesday, May 31, 2011                                        37
Modello NN
                                 36



                                 30
                                                                                  y(k-1)
                                 25                                                     y(k)
                        Y Axis




                                 20
                                                                                           y(k+1)
                                 15


                                 10
                                      0   2   4   6   8   10    12      14   16    18          20   22   24

                                                               X Axis




                                                                        Come scegliere i
                                                                        lags?
Tuesday, May 31, 2011                                                                                         38
Data Analysis
             1. ACF
                                                                        1



                                                                       0.5




             2. Distribution                                            0



                                                                       0.5
                                                                          0   10             20           30       40         50



             3. Multivariate
                analysis                                                                50
                                                                                        45
                                                                                        40
                                                                                                                                   0.25

                                                                                        35                                         0.2
                         60                                                             30




                                                                                   kW
                                                                                        25                                         0.15
                         50                                                             20
                                                                                                                                   0.1
                                                                                        15
                         40
                                                                                        10                                         0.05
             load (kW)




                                                                                        5
                         30
                                                                                                                                   0
                                                                                         1        5   9      13   17    21   24
                                                                                                           hour
                         20

                         10
                                   y = 0.0013*x2 + 0.26*x + 12
                         0
                          0   20     40               60   80    100
                                          occupancy



Tuesday, May 31, 2011                                                                                                                     39
Domanda...
                        Come ridurre la varianza delle
                        reti neurali?




Tuesday, May 31, 2011                                    40
Ensembling




Tuesday, May 31, 2011                41
Ensembling

                1. Calibrazione del modello usando
                   sottoinsiemi dei dati (Bagging)
                2. Uso dei dati pesato per importanza
                   (Adaboosting)
                3. Interazione e cooperazione tra gli
                   stimatori



Tuesday, May 31, 2011                                   42
Ensembling

                        [Hansen  Salomon, 1990]
                        Majority voting (classificazione)
                        Combinazione lineare
                        (regressione) N
                                       1 
                            F (x, D) =       Fi (x, D)
                                       N i=1


Tuesday, May 31, 2011                                       43
Ensembling
                                     Media




Tuesday, May 31, 2011                        44
Applicazioni
                        STLF dell’edificio ENEA Casaccia
                        (C59)
                        Presentato al IEEE Symposium on
                        CI Applications in Smart Grid
        M. De Felice and X. Yao, Neural Networks Ensembles for Short-Term Load
        Forecasting, in IEEE Symposium Series in Computational Intelligence 2011 (SSCI
        2011), 2011




Tuesday, May 31, 2011                                                                     45
Tecniche

                        Predittore naive:

                        modello SARIMA (Seasonal
                        ARIMA):
                        ΦP (B s )φ(B)∇D ∇d xt = α + ΘQ (B s )θ(B)et
                                      s


                        Reti Neurali (NN)

                        NN Ensembles



Tuesday, May 31, 2011                                                 46
Metodologia
                             40
                                                                          24 hours
                             35


                             30

                        kW
                             25   training part


                             20


                             15


                             10
                              2010   2013     2016   2019   2022   2025   2028   2031   2034    2037   2040   2043   2046   2049   2052   2055   2058

                                                                                        hours




                        Dati misurati da Settembre a
                        Novembre 2009
                        Training (13 settimane) e testing
                        (una settimana divisa in T1 e T2)

Tuesday, May 31, 2011                                                                                                                                   47
Misure d’errore

                        Errore Assoluto (MAE e MSE)
                        Error Percentuale (MAPE)
                        Scaled Error (MASE)




Tuesday, May 31, 2011                                 48
Negative Correlation Learning



                        [Liu  Yao, 1999]
                        Modifica alla funzione di
                        backpropagation
                        Penalty term λ
                                  M
                                  
                           ei =         (Fi (xn ) − yn )2 + λpi
                                  n=1



Tuesday, May 31, 2011                                             49
Regularized NCL

                          [Chen  Yao, 2009]
                          NCL con Regolarizzazione

                 M                       M
               1                     2  1
          ei =       (Fi (xn ) − yn ) −       (Fi (x) − F (xn ))2 +
               N n=1                    N n=1

                             T
                        +αi wi wi


Tuesday, May 31, 2011                                                 50
Errori
                                         MAE            MSE

                                      2.34 (0.79)    10.9 (17.88)
                        NN (Media)
                                      2.49 (1.47)   21.67 (59.29)

                                         1.38           2.95
                        NN Ensemble
                                         1.09           2.4
                                         1.47           3.34
                           RNCL
                                         1.07           2.82

                                         2.11           7.61
                           Naive
                                         2.28           6.4

                                         1.89           5.52
                          SARIMA
                                         1.24           2.17

Tuesday, May 31, 2011                                               51
Dati Aggiuntivi
                        Informazioni aggiuntive:
                        occupanti edificio, ora del
                        giorno, giorno della settimana,
                        giorni lavorativi.
                        NN: input aggiuntivi
                        SARIMA: termine lineare
                        addizionale

Tuesday, May 31, 2011                                     52
Dati Aggiuntivi
                                  4
                                       MLP Ensemble   external data       SARIMA   external data
                                  4
                                       MLP Ensemble                       SARIMA
                                  3
                 Absolute error




                                  3
                absolute error
                 absolute




                                  2
                                  2

                                  1
                                  1


                                  0
                                  0
                                   0
                                   0      20
                                          20     40        60        80     100    120      140
                                                                                            140
                                                          forecast window
                                                          forecast window
                                                         Forecasting window




Tuesday, May 31, 2011                                                                              53
Errori – dati aggiuntivi
                                         MAE            MSE

                                      2.46 (0.83)   12.13 (16.80)
                        NN (Media)
                                      2.34 (1.00)   11.61 (10.61)

                                         1.42           3.30
                        NN Ensemble
                                         0.75           1.27
                                         1.33            2.7
                           RNCL
                                         0.92           1.62

                                         2.11           7.61
                           Naive
                                         2.28           6.4

                                         1.91           5.61
                          SARIMA
                                         1.20           2.07

Tuesday, May 31, 2011                                               54
Errori giornalieri
                                              (d) SARIMA T2                                  (e) M

                        Fig. 6.       Univariate approach: 24-hours ahead forecasting absolute er


                                                                            8
                             140                                                    140
                                                                            7
                             120                                                    120
                                                                            6
                             100                                                    100
                                                                            5
                              80                                                     80
                                                                            4
                              60                                                     60
                                                                            3
                              40                                            2        40

                              20                                            1        20


                                  1       5    9      13     17   21   24              1     5
                                               hour of the day

                                               (a) SARIMA                                        (b

Tuesday, May 31, 2011   Fig. 7.       Absolute errors (in kW) made during testing parts T1 and55T
Errori giornalieri
                                (e) MLP Ensembling T2

hours ahead forecasting absolute errors on both T1 and T2. In light grey the area betw


                  8                                                8
                          140                                            140
                  7                                                7
                          120                                            120
                  6                                                6
                          100                                            100
                  5                                                5
                          80                                              80
                  4                                                4
                          60                                              60
                  3                                                3

                  2       40                                       2      40

                  1       20                                       1      20


 21       24                1   5     9      13     17   21   24            1     5
                                      hour of the day

                                    (b) MLP Ensembling

 Tuesday, May 31, 2011                                                                56
Ensemble: altro esempio

                     100


                        80
                kW




                        60


                        40


                        20

                             60    80   100   120     140     160   180   200   220
                                                    testing hours



Tuesday, May 31, 2011                                                                 57
TO-DO

                        Ensemble: usare tutte le stime
                        per creare una pdf
                        Ibridizzazione con metodi
                        statistici classici: analisi
                        multivariate, modelli stagionali,
                        Holt-Winters


Tuesday, May 31, 2011                                       58
The Big View



                            Forecasting 
                              Modeling




Tuesday, May 31, 2011                       59
Passi principali
        1. Definizione target (short-term,
           medium-term, seasonal)
        2. Raccolta dati e analisi
                                            Statistical Analysys
                                            High-dimensionality
                                            Data Mining

        3. Definizione e comparazione
           tecniche
                              Time Series Methods
                              NNs
                              Hybrid Methods

        4. Valutazione       Cost Analysis
                             Performance Measures


        5. Simulazione
                              Software Simulator
                              Multi-Agent Systems


Tuesday, May 31, 2011                                              60
PPSN 2012



                12th International Conference on “Parallel Problem
                Solving From Nature”, Taormina

                Paper submission: 15 Marzo 2012 (Proceedings
                Springer)

                http://www.dmi.unict.it/ppsn2012/

Tuesday, May 31, 2011                                                61
http://matteodefelice.name/research



Tuesday, May 31, 2011                                         62

More Related Content

PPTX
Deep Learning Training at Intel
PDF
Application of artificial_neural_network
PPT
Neural Networks
PPT
PPT
neural networks
PDF
eeglab_jnm03
PDF
Neural Network
PDF
Neural Network Classification and its Applications in Insurance Industry
Deep Learning Training at Intel
Application of artificial_neural_network
Neural Networks
neural networks
eeglab_jnm03
Neural Network
Neural Network Classification and its Applications in Insurance Industry

What's hot (20)

PDF
Artificial Neural Network and its Applications
PDF
NEURAL NETWORK FOR THE RELIABILITY ANALYSIS OF A SERIES - PARALLEL SYSTEM SUB...
PPT
PPT
NEURAL NETWORKS
PDF
Pres Tesi LM-2016+transcript_eng
PDF
IRJET- The Essentials of Neural Networks and their Applications
PPT
Soft Computing-173101
PPT
Artificial neural networks
PDF
Neural networks
PDF
Toward Tractable AGI: Challenges for System Identification in Neural Circuitry
PDF
Basics of Brain-Computer Interface
PPT
Ann by rutul mehta
PPTX
Using uml to model immune system
PPTX
neural-networks (1)
PDF
X trepan an extended trepan for
PDF
ANN-lecture9
PPTX
Deep Learning in Bio-Medical Imaging
PDF
Nature Inspired Reasoning Applied in Semantic Web
PDF
Application of CI in Motor Modeling
PDF
Hand Gesture Recognition using OpenCV and Python
Artificial Neural Network and its Applications
NEURAL NETWORK FOR THE RELIABILITY ANALYSIS OF A SERIES - PARALLEL SYSTEM SUB...
NEURAL NETWORKS
Pres Tesi LM-2016+transcript_eng
IRJET- The Essentials of Neural Networks and their Applications
Soft Computing-173101
Artificial neural networks
Neural networks
Toward Tractable AGI: Challenges for System Identification in Neural Circuitry
Basics of Brain-Computer Interface
Ann by rutul mehta
Using uml to model immune system
neural-networks (1)
X trepan an extended trepan for
ANN-lecture9
Deep Learning in Bio-Medical Imaging
Nature Inspired Reasoning Applied in Semantic Web
Application of CI in Motor Modeling
Hand Gesture Recognition using OpenCV and Python
Ad

Viewers also liked (7)

KEY
Ph.D. Presentation
PDF
Learning by Redundancy: how climate multi-model ensembles can help to fight t...
PDF
Measuring the benefits of climate forecasts
KEY
Gecco 2011 - Effects of Topology on the diversity of spatially-structured evo...
PPTX
IEEE SSCI 2011 Talk - Neural Networks Ensembles for Short-Term Load Forecasting
KEY
ENEA My Activities
PDF
Application of seasonal climate forecasts for electricity demand forecasting:...
Ph.D. Presentation
Learning by Redundancy: how climate multi-model ensembles can help to fight t...
Measuring the benefits of climate forecasts
Gecco 2011 - Effects of Topology on the diversity of spatially-structured evo...
IEEE SSCI 2011 Talk - Neural Networks Ensembles for Short-Term Load Forecasting
ENEA My Activities
Application of seasonal climate forecasts for electricity demand forecasting:...
Ad

Similar to [Italian] ENEA Seminar - Computational Intelligence and Energy Systems: intelligent solutions for complex problems (20)

PPTX
On the Development of a Brain Simulator
PDF
3rd space architecture learning in 3D
PDF
In sequence Polemical Pertinence via Soft Enumerating Repertoire
PDF
IRJET-In sequence Polemical Pertinence via Soft Enumerating Repertoire
PDF
Artificial Neural Network: A brief study
PPTX
Ann model and its application
PDF
A Model-Driven Approach for Deploying Trading-Based Knowledge Representation ...
PDF
Artificial Neural Networking
PDF
Neural networking this is about neural networks
PDF
Early detection of adult valve disease–mitral
PDF
Early detection of adult valve disease mitral stenosis using the elman artifi...
PDF
Early detection of adult valve disease mitral stenosis
DOCX
What is Neural Network?
PDF
Model of Differential Equation for Genetic Algorithm with Neural Network (GAN...
PDF
Analysis of intelligent system design by neuro adaptive control
PDF
Analysis of intelligent system design by neuro adaptive control no restriction
PDF
Intro to Deep Learning for Medical Image Analysis, with Dan Lee from Dentuit AI
PDF
ML_Unit_VI_DEEP LEARNING_Introduction to ANN.pdf
PDF
On the Development of a Brain Simulator
3rd space architecture learning in 3D
In sequence Polemical Pertinence via Soft Enumerating Repertoire
IRJET-In sequence Polemical Pertinence via Soft Enumerating Repertoire
Artificial Neural Network: A brief study
Ann model and its application
A Model-Driven Approach for Deploying Trading-Based Knowledge Representation ...
Artificial Neural Networking
Neural networking this is about neural networks
Early detection of adult valve disease–mitral
Early detection of adult valve disease mitral stenosis using the elman artifi...
Early detection of adult valve disease mitral stenosis
What is Neural Network?
Model of Differential Equation for Genetic Algorithm with Neural Network (GAN...
Analysis of intelligent system design by neuro adaptive control
Analysis of intelligent system design by neuro adaptive control no restriction
Intro to Deep Learning for Medical Image Analysis, with Dan Lee from Dentuit AI
ML_Unit_VI_DEEP LEARNING_Introduction to ANN.pdf

More from matteodefelice (6)

PPTX
Supporting the Energy Union with data & knowledge
PPTX
My presentation at ICEM 2017: From data mining to information extraction: usi...
PDF
Learning from data: data mining approaches for Energy & Weather/Climate appli...
PPT
Anes tdec2012
PPTX
Soft Computing - Lezione di Controlli Automatici
PPTX
EVEN Presentation - April 2009
Supporting the Energy Union with data & knowledge
My presentation at ICEM 2017: From data mining to information extraction: usi...
Learning from data: data mining approaches for Energy & Weather/Climate appli...
Anes tdec2012
Soft Computing - Lezione di Controlli Automatici
EVEN Presentation - April 2009

Recently uploaded (20)

PDF
Agricultural_Statistics_at_a_Glance_2022_0.pdf
PPTX
Digital-Transformation-Roadmap-for-Companies.pptx
PDF
Build a system with the filesystem maintained by OSTree @ COSCUP 2025
PDF
7 ChatGPT Prompts to Help You Define Your Ideal Customer Profile.pdf
PDF
Per capita expenditure prediction using model stacking based on satellite ima...
PDF
Mobile App Security Testing_ A Comprehensive Guide.pdf
DOCX
The AUB Centre for AI in Media Proposal.docx
PDF
Machine learning based COVID-19 study performance prediction
PDF
cuic standard and advanced reporting.pdf
PPTX
KOM of Painting work and Equipment Insulation REV00 update 25-dec.pptx
PDF
Profit Center Accounting in SAP S/4HANA, S4F28 Col11
PDF
Architecting across the Boundaries of two Complex Domains - Healthcare & Tech...
PDF
Peak of Data & AI Encore- AI for Metadata and Smarter Workflows
PDF
Dropbox Q2 2025 Financial Results & Investor Presentation
PDF
Review of recent advances in non-invasive hemoglobin estimation
PDF
Network Security Unit 5.pdf for BCA BBA.
PDF
Diabetes mellitus diagnosis method based random forest with bat algorithm
PDF
Reach Out and Touch Someone: Haptics and Empathic Computing
PDF
Electronic commerce courselecture one. Pdf
PPTX
Programs and apps: productivity, graphics, security and other tools
Agricultural_Statistics_at_a_Glance_2022_0.pdf
Digital-Transformation-Roadmap-for-Companies.pptx
Build a system with the filesystem maintained by OSTree @ COSCUP 2025
7 ChatGPT Prompts to Help You Define Your Ideal Customer Profile.pdf
Per capita expenditure prediction using model stacking based on satellite ima...
Mobile App Security Testing_ A Comprehensive Guide.pdf
The AUB Centre for AI in Media Proposal.docx
Machine learning based COVID-19 study performance prediction
cuic standard and advanced reporting.pdf
KOM of Painting work and Equipment Insulation REV00 update 25-dec.pptx
Profit Center Accounting in SAP S/4HANA, S4F28 Col11
Architecting across the Boundaries of two Complex Domains - Healthcare & Tech...
Peak of Data & AI Encore- AI for Metadata and Smarter Workflows
Dropbox Q2 2025 Financial Results & Investor Presentation
Review of recent advances in non-invasive hemoglobin estimation
Network Security Unit 5.pdf for BCA BBA.
Diabetes mellitus diagnosis method based random forest with bat algorithm
Reach Out and Touch Someone: Haptics and Empathic Computing
Electronic commerce courselecture one. Pdf
Programs and apps: productivity, graphics, security and other tools

[Italian] ENEA Seminar - Computational Intelligence and Energy Systems: intelligent solutions for complex problems

  • 1. Computational Intelligence and Energy Systems: intelligent solutions for complex problems Matteo De Felice Unità Modellistica Energetica Ambientale UTMEA - ENEA Tuesday, May 31, 2011 1
  • 2. Sommario Cos’è la Computational Intelligence (CI)? Quali sono le applicazioni della CI ai sistemi complessi? Tuesday, May 31, 2011 2
  • 3. CI: paradigmi Soft NN Computing EC FS IA? SI AIS Computational Intelligence Tuesday, May 31, 2011 3
  • 4. Visione d’insieme Temi NN principali EC FS SI AIS Tuesday, May 31, 2011 4
  • 5. CI e letteratura 3 x 10 5 Evolutionary Computation Swarm Intelligence 4 Artificial Neural Networks 3 2 1 0 1994 1996 1998 2000 2002 2004 2006 2008 2010 year Dati dalla Thomson Reuters ISI considerando Computer Science & Technology (Gennaio 2010) Due journals sulla CI nei primi 10 in CS (IF 2009) Tuesday, May 31, 2011 5
  • 6. La diffusione della CI Problemi sempre più complessi Più potenza di calcolo disponibile Tuesday, May 31, 2011 6
  • 7. ma... Assenza di una teoria consolidata Frammentazione degli algoritmi Approccio poco sistematico e confronti poco “robusti” PSO APSO CPSO DPSO EPSO FPSO GPSO HPSO IPSO LPSO MPSO NPSO OPSO PPSO QPSO RPSO SPSO TPSO UPSO VPSO WPSO GA AGA BGA CGA DGA EGA FGA HGA IGA KGA LGA MGA OGA PGA QGA RGA SGA VGA ... Tuesday, May 31, 2011 7
  • 8. Applicazioni Principali Reti Neurali & Logica Fuzzy 1) Modellazione & Forecasting 2) Ottimizzazione Calcolo Evolutivo Tuesday, May 31, 2011 8
  • 9. Quadro generale Reti neurali evolutive con topologia a rete complessa Evolving predictive neural models for complex processes Evolving Complex Neural Networks 2008 Reti Neurali Evolutive Ensemble Artificial Neural Networks and Support Vector Machines ensembling: a comparison Tuesday, May 31, 2011 9
  • 10. 2009 Ambient temperature modelling with soft computing techniques Modellazione temperature con NN Combining Back-Propagation and Genetic Algorithms to Train Neural Networks for Ambient Temperature Modeling in Italy 2010 Ottimizzazione dello start-up centrale a ciclo combinato Combining Back-Propagation and Genetic Algorithms to Train Neural Networks for Ambient Temperature Modeling in Italy Tuesday, May 31, 2011 10
  • 11. Reti Neurali e Load Forecast 2011 Short-Term Load Forecasting with Neural Network Ensembles: a Comparative Study Climate Variables in Energy Modeling Tuesday, May 31, 2011 11
  • 12. Altri Progetti Identificazione Reti Neurali Algoritmi Evolutivi Ottimizzazione Structural System Evolutive e Spazialmente traiettorie missioni in Ing. Sismica Applicazioni alla Strutturati interplanetarie Finanza Tuesday, May 31, 2011 12
  • 14. Process Optimization Process Parameters Process Environment (X) Measurement Come migliorare la ‘performance’ di un processo tramite i suoi parametri? Tuesday, May 31, 2011 14
  • 15. Ottimizzazione tradizionale Metodi Line-search and trust- region (serve l’Hessiano!) Metodi Quasi-newton (Hessiano approssimato) Metodi Derivative-free Tuesday, May 31, 2011 15
  • 16. ...ma il real-world è: 1) ‘Rumoroso’ 2) Dinamico 3) Difficile da esaminare Tuesday, May 31, 2011 16
  • 17. Evolutionary Computation (EC) Ottimizzazione Black-box Singolo e Multi-Obiettivo Anche funzioni discontinue e non differenziabili Meta-euristica Population-based Tuesday, May 31, 2011 17
  • 18. Metaeuristica Ottimizzazione Stocastica Algoritmi usati per trovare soluzioni a problemi “difficili” Esempio: Hill-Climbing, Tabu Search, Simulated-Annealing Tuesday, May 31, 2011 18
  • 20. Metodi di ottimizzazione DIRECT Algorithm Applications Taxonomy of Methods Yves Brise Lipschitzian Optimization, DIRECT Algorithm, and Applications Tuesday, May 31, 2011 20
  • 21. Applicazione Ottimizzazione dello startup di una centrale a ciclo combinato (CCPP) Minimizzazione del tempo di avvio, consumi, emissioni e stress termico Massimizzazione della produzione di energia M. De Felice, I. Bertini, A. Pannicelli, and S. Pizzuti, "Soft Computing based optimisation of combined cycled power plant start-up operation with fitness approximation methods," Applied Soft Computing, 2011. I. Bertini, M. De Felice, F. Moretti, and S. Pizzuti, "Start-Up Optimisation of a Combined Cycle Power Plant with Multiobjective Evolutionary Algorithms," in Applications of Evolutionary Computation, 2010, pp. 151-160. Tuesday, May 31, 2011 21
  • 22. Procedura 1. Definizione di un indice di performance 2. Impostazione simulatore sw 3. Algoritmo EC tramite simulatore Tuesday, May 31, 2011 22
  • 23. Indice Performance F1 1 Informazioni 0.5 0 0 0.5 1 1.5 2 2.5 3 4 x 10 dagli esperti di F2 1 0.5 processo 0 0 0.5 1 1.5 F3 2 2.5 5 x 10 3 1 Knowledge 0.5 0 0 5 10 15 9 x 10 modeling con F4 1 0.5 funzioni fuzzy 0 0 5 10 15 20 25 30 F5 1 0.5 0 0 50 100 150 200 250 300 Tuesday, May 31, 2011 23
  • 24. Singolo-obiettivo Algoritmo Genetico operazione di mutazione Gaussiano Funzione di fitness approssimata per velocizzare l’ottimizzazione (da 2070 a 36 ore/CPU) Tuesday, May 31, 2011 24
  • 25. Risultati Tempo Prod. Stress Consumi Emissioni avvio Energia Termico Esperti 21070 143557 2.5•109 25 10 GA 16569 115070 1.86•109 18.8 78.4 Var. -25% -16% -16% -30% 2% Norm. Tuesday, May 31, 2011 25
  • 26. Multi-obiettivo 12.65 12.6 12.55 Emissions (mg s / N m3) 12.5 12.45 12.4 12.35 12.3 Real NSGA 2 12.25 WSGA RAND 12.2 3.9 4 4.1 4.2 4.3 4.4 4.5 4.6 Energy Production (KJ) 9 x 10 Tuesday, May 31, 2011 26
  • 28. Modellazione con NNs ||F (x) − f (x)|| < , ∀x 1.2 0.6 Y Axis 0 0 1 2 3 4 5 6 6.5 -0.6 -1.2 X Axis y = sin(x) NN(x) Tuesday, May 31, 2011 28
  • 29. Modellazione con NNs Disturbances Input u(k) Output y(k) System Neural Network Errore (MSE) Metodi empirici per decidere la topologia della rete Tuesday, May 31, 2011 29
  • 30. Regressione con NN Si una una NN per fare regressione non-lineare Tuesday, May 31, 2011 30
  • 31. Time Series Forecasting Possiamo fare una previsione dei dati futuri usando quelli osservati Altre informazioni utili (!) Tuesday, May 31, 2011 31
  • 32. Approcci per le NN y(t+1) Input at Neural y(t+2) ... Direct Method time t Network y(t+N) Input at output t+1 time t Neural Network output t Iterative Method delay Tuesday, May 31, 2011 32
  • 33. Short-Term Load Forecasting 60 40 kW 20 0 0 200 400 600 800 1000 1200 1400 1600 1800 2000 hours Dati Orari Obiettivo: predizione del carico fino a 24 ore Tuesday, May 31, 2011 33
  • 34. Modelli Seasonal 1 0.5 0 0.5 0 10 20 30 40 50 Implementazione in R ΦP (B s )φ(B)∇D ∇d xt = α + ΘQ (B s )θ(B)et s Tuesday, May 31, 2011 34
  • 35. Modello NN Campioni passati Rete Previsione Neurale Informazioni aggiuntive Tuesday, May 31, 2011 35
  • 36. Rete Neurale Funzioni di attivazione f differenziabili Pesi w i Pesi w o Tuesday, May 31, 2011 36
  • 37. Backpropagation [Werbos, 1974] Forward phase: il segnale si propaga “in avanti” Backward phase: si calcola l’errore e lo si propaga “all’indietro”, modificando i pesi Tuesday, May 31, 2011 37
  • 38. Modello NN 36 30 y(k-1) 25 y(k) Y Axis 20 y(k+1) 15 10 0 2 4 6 8 10 12 14 16 18 20 22 24 X Axis Come scegliere i lags? Tuesday, May 31, 2011 38
  • 39. Data Analysis 1. ACF 1 0.5 2. Distribution 0 0.5 0 10 20 30 40 50 3. Multivariate analysis 50 45 40 0.25 35 0.2 60 30 kW 25 0.15 50 20 0.1 15 40 10 0.05 load (kW) 5 30 0 1 5 9 13 17 21 24 hour 20 10 y = 0.0013*x2 + 0.26*x + 12 0 0 20 40 60 80 100 occupancy Tuesday, May 31, 2011 39
  • 40. Domanda... Come ridurre la varianza delle reti neurali? Tuesday, May 31, 2011 40
  • 42. Ensembling 1. Calibrazione del modello usando sottoinsiemi dei dati (Bagging) 2. Uso dei dati pesato per importanza (Adaboosting) 3. Interazione e cooperazione tra gli stimatori Tuesday, May 31, 2011 42
  • 43. Ensembling [Hansen Salomon, 1990] Majority voting (classificazione) Combinazione lineare (regressione) N 1 F (x, D) = Fi (x, D) N i=1 Tuesday, May 31, 2011 43
  • 44. Ensembling Media Tuesday, May 31, 2011 44
  • 45. Applicazioni STLF dell’edificio ENEA Casaccia (C59) Presentato al IEEE Symposium on CI Applications in Smart Grid M. De Felice and X. Yao, Neural Networks Ensembles for Short-Term Load Forecasting, in IEEE Symposium Series in Computational Intelligence 2011 (SSCI 2011), 2011 Tuesday, May 31, 2011 45
  • 46. Tecniche Predittore naive: modello SARIMA (Seasonal ARIMA): ΦP (B s )φ(B)∇D ∇d xt = α + ΘQ (B s )θ(B)et s Reti Neurali (NN) NN Ensembles Tuesday, May 31, 2011 46
  • 47. Metodologia 40 24 hours 35 30 kW 25 training part 20 15 10 2010 2013 2016 2019 2022 2025 2028 2031 2034 2037 2040 2043 2046 2049 2052 2055 2058 hours Dati misurati da Settembre a Novembre 2009 Training (13 settimane) e testing (una settimana divisa in T1 e T2) Tuesday, May 31, 2011 47
  • 48. Misure d’errore Errore Assoluto (MAE e MSE) Error Percentuale (MAPE) Scaled Error (MASE) Tuesday, May 31, 2011 48
  • 49. Negative Correlation Learning [Liu Yao, 1999] Modifica alla funzione di backpropagation Penalty term λ M ei = (Fi (xn ) − yn )2 + λpi n=1 Tuesday, May 31, 2011 49
  • 50. Regularized NCL [Chen Yao, 2009] NCL con Regolarizzazione M M 1 2 1 ei = (Fi (xn ) − yn ) − (Fi (x) − F (xn ))2 + N n=1 N n=1 T +αi wi wi Tuesday, May 31, 2011 50
  • 51. Errori MAE MSE 2.34 (0.79) 10.9 (17.88) NN (Media) 2.49 (1.47) 21.67 (59.29) 1.38 2.95 NN Ensemble 1.09 2.4 1.47 3.34 RNCL 1.07 2.82 2.11 7.61 Naive 2.28 6.4 1.89 5.52 SARIMA 1.24 2.17 Tuesday, May 31, 2011 51
  • 52. Dati Aggiuntivi Informazioni aggiuntive: occupanti edificio, ora del giorno, giorno della settimana, giorni lavorativi. NN: input aggiuntivi SARIMA: termine lineare addizionale Tuesday, May 31, 2011 52
  • 53. Dati Aggiuntivi 4 MLP Ensemble external data SARIMA external data 4 MLP Ensemble SARIMA 3 Absolute error 3 absolute error absolute 2 2 1 1 0 0 0 0 20 20 40 60 80 100 120 140 140 forecast window forecast window Forecasting window Tuesday, May 31, 2011 53
  • 54. Errori – dati aggiuntivi MAE MSE 2.46 (0.83) 12.13 (16.80) NN (Media) 2.34 (1.00) 11.61 (10.61) 1.42 3.30 NN Ensemble 0.75 1.27 1.33 2.7 RNCL 0.92 1.62 2.11 7.61 Naive 2.28 6.4 1.91 5.61 SARIMA 1.20 2.07 Tuesday, May 31, 2011 54
  • 55. Errori giornalieri (d) SARIMA T2 (e) M Fig. 6. Univariate approach: 24-hours ahead forecasting absolute er 8 140 140 7 120 120 6 100 100 5 80 80 4 60 60 3 40 2 40 20 1 20 1 5 9 13 17 21 24 1 5 hour of the day (a) SARIMA (b Tuesday, May 31, 2011 Fig. 7. Absolute errors (in kW) made during testing parts T1 and55T
  • 56. Errori giornalieri (e) MLP Ensembling T2 hours ahead forecasting absolute errors on both T1 and T2. In light grey the area betw 8 8 140 140 7 7 120 120 6 6 100 100 5 5 80 80 4 4 60 60 3 3 2 40 2 40 1 20 1 20 21 24 1 5 9 13 17 21 24 1 5 hour of the day (b) MLP Ensembling Tuesday, May 31, 2011 56
  • 57. Ensemble: altro esempio 100 80 kW 60 40 20 60 80 100 120 140 160 180 200 220 testing hours Tuesday, May 31, 2011 57
  • 58. TO-DO Ensemble: usare tutte le stime per creare una pdf Ibridizzazione con metodi statistici classici: analisi multivariate, modelli stagionali, Holt-Winters Tuesday, May 31, 2011 58
  • 59. The Big View Forecasting Modeling Tuesday, May 31, 2011 59
  • 60. Passi principali 1. Definizione target (short-term, medium-term, seasonal) 2. Raccolta dati e analisi Statistical Analysys High-dimensionality Data Mining 3. Definizione e comparazione tecniche Time Series Methods NNs Hybrid Methods 4. Valutazione Cost Analysis Performance Measures 5. Simulazione Software Simulator Multi-Agent Systems Tuesday, May 31, 2011 60
  • 61. PPSN 2012 12th International Conference on “Parallel Problem Solving From Nature”, Taormina Paper submission: 15 Marzo 2012 (Proceedings Springer) http://www.dmi.unict.it/ppsn2012/ Tuesday, May 31, 2011 61