15‐06‐2013

SMARTER MODELS
FOR
SMART SYSTEMS
Vladimiro Miranda
IEEE Fellow
Director INESC TEC
President INESC P&D Brasil

Pioneer Portugal
In 2013 about 25.000 MW of wind generation capacity in Iberia
(peak of 65.000 MW)
• About 5000 MW are in Portugal – today’s peak consumption is
around 9500 MW

Portugal: no.2 in the world in 2012 – 20% of electricity from wind –
more capacity installed than Denmark (nº 1)

Only half of the wind
power is monitored at
transmission level

10 million people
5 million wind kW
2

1

15‐06‐2013

Wind generation, load and challenges (case of Portugal)
2.March.2010

3.March.2010

Load

Wind Power greater than load!!!
Wind

wind farms monitored at
transmission level

• Wind power integration in large scale requires reinforced
interconnections
3

WHAT IS THE SMART GRID?
The smart grid is not just more of the same: it requires a new layer
LAYER

CHARACTERISTICS

NATURE

High level
intelligence

Local agents, parallel processing, distributed
computing

Software

Central control

EMS, DMS

Software

Information

Data bases

Software

Data transmission

Protocols

Software

Communication

Dedicated, optical fiber, gsm

Physical

Local control

Sensors, relays

Physical

Physical grid

Generators, lines, transformers

Physical

4

2

 necessary and economically justified Do you have electrical vehicles? Intelligent Grid Do you have explicit reliability costs? Grid islanding Do you have a medium/high degree of distributed generation? Do you have flexile tariff systems or market based tariff system? Microgrids Smart metering Do you have large costs from meter reading with human agents? Telemetering 5 The micro‐grids FROM THE CLEVER INTEGRATION TO THE SMART  SYSTEM 6 3 .15‐06‐2013 Escalating the Concept Ladder The Smart Grid should be built step by step when conditions make it feasible.

Diesel HV Network Capacitor  Bank VSI DFIM Sheddable  Loads MicroGri d Local communication infrastructures. 7 EU countries MoreMICROGRIDS (2006‐2008) 22 partners. Operation Modes: • Interconnected Mode • Emergency Mode MicroGri d MicroGri d CHP Hydro 8 4 . distributed) Standardization of Technical and Commercial Protocols and Hardware Evaluation of the system performance on power system operation  Impact on the Development of Electricity Infrastructures 7 MULTI‐MICROGRIDS  Microgrids Low Voltage distribution  systems with small  modular generation units  providing power and heat  to local loads. A hierarchical  management and control  system. 11 EU countries Work Packages From: Steady State and Dynamic Simulation Tools  Development of Local Micro Source Controllers and Central Controller Development of Emergency Functions Telecommunication Infrastructures and Communication Protocols To: Alternative Control Strategies (hierarchical vs.15‐06‐2013 10 YEARS OF EXPERIENCE IN THE EU MICROGRIDS (2003‐2006) 14 partners.

 may be made bottom‐up: much faster than present day top‐ down practice Major disasters (or terrorist attacks) will no longer cause global electric power supply disruption The dynamic definition of operating conditions and control strategies for  Grids. has been the object of years of research in Europe. EnergyCon 2012 ‐ Florence 9 Bringing reality to the concept ÉVORA. in case of major  blackout.15‐06‐2013 WHAT DO WE GET WITH MULTI‐GRIDS? Added value in reliability: a fault‐tolerant system 1. PORTUGAL:  THE SMART CITY 10 5 . Islanding: the survival of network islands relying on local generation reduces load curtailment 2. Black start: the recomposition of the system. as well as new equipment specification.

000 consumers equiped with the Energy Box Supply points for electric vehicles A communication network in parallel to the power network Distribution controllers concentrate messages LC Electric  appliances MC 11 The PROSUMER: producer and consumer The Energy Box controls load and micro‐generation The Energy Box is the basic element of the InovCity initiative in  Portugal Photo‐voltaic Energy Box Controling loads FROM THE SMART GRID TO THE SMART CITY 12 6 .15‐06‐2013 THE SMART CITY IN PORTUGAL The city of Évora connected as a Smart City • • • • 33.

A team work with highly professional Industrial  partners AND… 4.15‐06‐2013 INTEGRATION WITH SMART METERING 13 InovCity and InovGrid: the formula for an actual success The success of the InovGrid project has its roots in: 1. A robust consortium led by a visionary Power Company 2. An alliance with the excellence of a Research  Institute 3. The consistent support from the political power 14 7 .

59 pr. 146 pr. 189 bay state members National 93% SG projects: national budgets (industry + pub. 2500 M€ for smart‐meter roll‐out) EU SG project budget (industry + EU  financ.): 184 M€ to the JRC (within FP6 and FP7) • Cluster 1 Smart customers:  • Cluster 2 Smart metering:  • Cluster 3 DER integration:  • Cluster 4 Smart Distribution:  92 pr. 113 pr.15‐06‐2013 Smart Grid projects in Europe Statistics 2012 EC (FP6+7) 7% 204 European projects 111 projects with significative demo. Sailing ahead INESC TEC LAB ON SMART GRIDS AND  ELECTRIC VEHICLES 16 8 .): 2500 M€ through the JRC (excl. financ.   15 projects financed by the EC.

15‐06‐2013 INESC TEC Lab on Smart Grids and Electric Vehicles Electric vehicles and battery charging Research in the EU on driving habits UTOPIA Dumb charg./Double tariff The business model will have a decisive impact on the power system. by conditioning the citizens behavior 18 9 .

 conditioned demand and EV batteries. the   integration of EVs will have a negative impact on reliability h/ano ‐/ano LOLE LOLF 19 Charging EVs and network stress No EV Double tariff – 52% EV Dumb charging– 52% EV Smart charging – 52% EV 20 10 . Simulation in 6 European countries: → unless a smart business model is put in place.15‐06‐2013 Impact on reliability: LOLE and LOLF Chronologic Monte Carlo including conventional generation  and wind. storage.

15‐06‐2013 Intelligent V2G INESC TEC Smart charger Bi‐directional communications with the DTC or energy aggregator 11 .

15‐06‐2013 INESC TEC Lab on Smart Grids and Electric Vehicles The lab architecture is based on a conceptual model for actual and future  networks Control and communication requirements • Scenario validation through simulation Multi‐technology Allows combining emulation and physical implementation Integration of distinct entities and information flows – technical and market Integration of a diversity of equipament e and their APIs WISE SISE technology Wind Integration in Smart Environments Solar Integration in Smart Environments W I S E S I S E For weak connection points: • includes an inverter allowing the connection of small wind  generators up to 3 kW to Low Voltage networks • avoids an excessive raise in voltage at the connection point • can track frequency changes and adjust power output • droop control principle 24 12 .

15‐06‐2013 To know better UNVEILING HIDDEN INFORMATION 25 DATA: THE CHALLENGE Added technologies: Distributed generation Distributed vehicle connection Distributed control  Massive data flow Added uncertainty elements:  Renewables Distributed decision factors (V2G. tariffs/prices…)  Deterministic models replaced by probabilistic EnergyCon 2012 ‐ Florence 26 13 .

 when  no breaker status signal arrives at the  Control Centre? Breakers with status  unknown 28 14 .15‐06‐2013 NETWORK OBSERVABILITY State Estimation in the Smart Grid context  Change of paradigm • Decentralized • Clusterized • Based on probabilistic models REPLACEMENT OF THE CONVENTIONAL CONCEPT OF  TOPOLOGICAL OBSERVABILITY BY STATISTICAL OBSERVABILITY 27 UNVEILING THE HIDDEN INFORMATION HYPOTHESIS: the network topology  information is embedded in the  manifold supporting the electrical  data. 18 ~ ~ 22 ~ 23 ~ 16 19 20 ~ 15 24 ~ 13 14 11 12 230 kV 3 9 10 8 138 kV 7 4 5 2 6 ~ 1 ~ ~ We need Information Theory to unveil  the meaning of this “background  micro‐wave radiation” 21 17 ~ How to estimate the current system  (or multi‐microgrid) topology.

  1 N Renyi’s Quadratic Entropy H R 2   log  p 2k k 1  Extension to continuous pdf H R 2   log  fY 2 ( z)dz  30 15 .15‐06‐2013 INFORMATION THEORY NEEDED By simulation: a data set of {electric values M.T) – how much information about T is contained in M? Entropy of the joint pdf I(M.T) = H(M) + H(T) – H(M. topology T} MUTUAL INFORMATION I(M.T) Entropy of the topology pdf Entropy of the electrical data  pdf 29 ENTROPY ENTROPY is a measure of information content N 1 Shannon's Entropy HS   p k log p k k 1 N Renyi's Entropy  1 H R  log p k 1  k 1   0.

 measured by mutual information between each line power flows and breaker status High MI Low MI The flow in some branches is much more informative than in other branches Br.15‐06‐2013 MUTUAL INFORMATION MAP Impact of breaker 6‐10 in system power flow. 6‐10 31 Information serving diagnoses IDENTIFICATION IN TOPOLOGY CELLS 32 16 .

A Manifold for  Topology A B  << A A New  point Manifold for  Topology B B New  point B TOPOLOGY IN BLACK BOXES 16 23 ~ breaker6 Topology cell: a breaker (or switch) with status unknown External electric values known used measurements measurements not available 13 breaker4 breaker5 11 12 9 7 4 breaker3 1 5 2 ~ From power flow simulation… 2 patterns for electrical data: ‐> breaker open ‐> breaker closed 10 8 ~ 3 3 breaker1 ~ Electric pattern B 2 autoencoders to diagnose the inner state of the topology cell ~ 23 14 6 No  information! A min  17 . one of the autoencoders will be “in tune” while  the other will display a large error. corresponding to 2 distinct topologies.  For a new input vector.15‐06‐2013 CLUSTER DISTINCTION WITH AUTOENCODERS ‐ CAM Admit 2 autoencoders trained for 2 distinct data clusters A and  B.

90% 4 18 3982 99.97% 55.33% 99.49% 39 961 96.81% 61.50% 2 10 1990 99.78% 1 2 3 4 5 6 7 8 9 10 10 8 7 3 4 3 breaker3 1 13 19 21 19 13 15 11 11 13 11 Total Hidden  layer Corr.30% 7 33 6967 99. breaker status may be recovered: but… less statistical observability  less precision LOCAL INFORMATION 23 ~ breaker6 10 15 17 15 10 12 9 9 10 9 Wrong  Corr.37% 87.69% No  information! 36 18 .15‐06‐2013 A COMPETITIVE APPROACH: CAM Input to autoencoders selected with Mutual  Information! TOPOLOGY ESTIMATOR  a MOSAIC of local autoencoders New point B A min  results in 10.52% 256 9744 97.44% 35 TOPOLOGY IN BLACK BOXES 16 used measurements measurements not available Efficien.06% 74. Wrong  Corr.69% 100. Efficien.20% Total 266 54734 99.53% 32 968 96.45% 99.00% 99.96% 99. Efficien.000 scenarios including noise Simult.10% 9 47 8953 99. 210 1278 0 194 2555 19 3863 1203 4461 1523 15306 9790 8722 10000 9806 7445 9981 6137 8797 5539 8477 84694 97.80% 8 41 7959 99.77% 84.00% 3 11 2989 99.  Signal reconstructions Topology reconstruction miss.55% 27 973 97. Efficien.00% 100. Output Hidden layer 10 14 12 12 12 10 12 8 8 12 breaker4 breaker5 9 ~ 15 19 17 17 17 15 17 13 13 17 Total 13 11 ~ 1 2 3 4 5 6 7 8 9 10 ~ 23 14 ~ BREAKER Even with remote information only.00% 98.22% 100. 0 131 0 0 0 0 11 67 3 4 216 BREAKER Wrong  12 5 2 breaker1 6 REMOTE INFORMATION Input.39% 84. 1 5 995 99.63% 11 989 98.55% 17 983 98.  signals Wrong  Corr.60% 10 51 9949 99.54% 23 977 97.49% 48 952 95.50% 10 990 99.97% 99.90% 87.00% 100.  Output Input.30% 5 23 4977 99.00% 100.48% 44 956 95.70% 6 27 5973 99.89% 99.50% 5 995 99. 10000 9869 10000 10000 10000 10000 9989 9933 9997 9996 99784 100.00% 98.

28% 198 96. 0 100.88% 279 94.00% 29 99. bus C1 Split bus C2 C3 Total: split bus Global results Subsystem  topology  reconstruction Wrong Effic.00% 0 100.42% 0 100.92% How to improve autoencoder accuracy TRAINING WITH INFORMATION THEORETIC CRITERIA 38 19 .: no Sp.04% 6 99.00% 0 100.42% 86 98.00% 29 99.42% 308 96.15‐06‐2013 TOPOLOGY IN BLACK BOXES External signals may also unveil the hidden topology in  substations… or multi‐microgrids… Split bus y/n L1 No  L2 split bus L3 L4 Tot.

N Renyi’s Entropy  1 log p k 1  k 1   0. • Initializing weigths with values derived from PCA is also a fine strategy [as if the neural network had only linear  activation functions] 20 . trained in unsupervised mode.15‐06‐2013 Maximizing information flow A perfect autoencoder must preserve all information that crosses through ‐ from input to output. should yield at the output the same amount of information as presented at the input. Backpropagation under MSE does not guarantee that. Possible criteria: • Max Mutual Information between input and middle layer • Min Mutual Information among middle layer nodes • Max Entropy at the middle layer supervised unsupervised Maximizing information flow INESC TEC experience: maximizing Entropy at the middle layer is the best criterion overall.   1 H R   • For continuous variables… H R 2   log  fY 2 (z)dz  Entropy is a measure of information quantity. The first half.

15‐06‐2013 Learning macro features Experiments with deep sparse networks have shown that macro features could be learned from data From 10 million image frames collected randomly from youtube videos. a half‐autoencoder trained in  unsupervised mode under an ICA  criterion learned to identify faces and cats! [Quoc et al. 2012] Macro features exist! Faces… bodies… cats… Is there a neuron that captures the essential concept? YES! cat body 21 .

 closed. B.   prediction: open TP FP FN TN Single step 4928 104 5 4963 2‐step MQMI 4915 0 14 5067 Remarkable!    14 in 10. closed. open B. no voltage information) Autoencoder Breaker 2 open Breaker 2 closed Model Single step 2‐step ME 2‐step MQMI Single step 2‐step ME 2‐step MQMI MSE 0. open Active B.0053 0. open.   prediction: closed False Negative: br.000! 43 EXAMPLE There is a reason why 14 answers were wrong: confusion! The active and reactive power flows are close to zero in all 14  FN cases (prediction open when breaker closed). prediction: open True Negative: br.0046 0.0019 0.0015 0.0014 0. closed Reactive 44 22 . closed B.0007 True Positive:   br.15‐06‐2013 EXAMPLE (Only power. open. prediction: closed False Positive:  br.

15‐06‐2013 EXAMPLE – HEURISTIC RULE (HR) What if a heuristic rule is applied.010928 . A very impressive result. for breaker 2 in a CAM‐Blind mode: CAM‐B TP 4613 FP 285 FN 320 TN 4782 Now. neither on the flows through the breaker nor on its status. the generation of False  Positives. i. for breaker 2: HR TP 5048 FP 19 FN 6 TN 4927 Now. 605 wrong predictions in 10. is known? ? A new set of competitive autoencoders was trained without such inputs  (active and reactive power flows through the breaker). in the absence of any information! 46 23 .011940}    Q: {‐0. is in this case a less efficient diagnosing process 45 EXAMPLE: CAM BLIND What if no information. 0.e. often adopted in State Estimation.Q|< Limit  breaker is open”? Picking up narrow intervals for P and Q (from a subset of data) P: {‐0.009872} In this example. 0. declaring as closed 19 cases where the breaker was open! The heuristic rule.000 – approximately 6%. defining that “if |P.010446 .000 – plus. 25 wrong predictions in 10. In this example.

15‐06‐2013 CONCLUSIONS No Smart Grid without smart control No Smart Grid without smart distributed intelligence No Smart Grid without smart information theory No Smart Grid without smart business models No Smart Grid without smart people… 47 VISIT Porto: the Formosa city! 48 24 .

15‐06‐2013 VISIT Porto: the Formosa city! 49 VISIT Porto: the Formosa city! 50 25 .

15‐06‐2013 VISIT Porto: the Formosa city! 51 26 .