Ce diaporama a bien été signalé.
Nous utilisons votre profil LinkedIn et vos données d’activité pour vous proposer des publicités personnalisées et pertinentes. Vous pouvez changer vos préférences de publicités à tout moment.

Intelligence Artificielle et performances énergétiques - 27 sept. LLN

45 vues

Publié le

Les capteurs, contrôleurs, et outils d'analyse combinés à l'Intelligence Artificielle permettront de manager et automatiser les flux d'énergie et deviendront la colonne vertébrale de la gestion intelligente de l'énergie, des bâtiments, smart cities et smart grids.

Publié dans : Technologie
  • Soyez le premier à commenter

  • Soyez le premier à aimer ceci

Intelligence Artificielle et performances énergétiques - 27 sept. LLN

  1. 1. Cluster Technology of Wallonia Energy, Environment and sustainable Development 1 Axis Parc – 27/09/2018Workshop #3 : Intelligence Ar5ficielle
  2. 2. ICT & Energy ? 3 • Fortes synergies : o Concepts : #smart - grids / cities / building / mobility… o Technologies sous-jacentes : (C)EMS, smart metering, … • Collaboration INFOPOLE Cluster TIC / TWEED : o Where ICT meets Energy (2015) o Cartographie TWEED – smartgrids (2016-2017) o Digital Energy Business Club (2018)
  3. 3. Digital Energy Business Club 4 • Objectifs : 1. Développement commercial des entreprises membres des Clusters ainsi que de certains acteurs des secteurs TIC et énergie. 2. Création de synergies, d’innovations et de partenariats ; entre entreprises privées, mais également entre entreprises et investisseurs ou acteurs de R&D. 3. Promotion et développement sectoriel du secteur TIC-énergie. • Outils : o Portail ReWallonia.be : cartographies smartgrids / stockage o Réseaux des Clusters : membres & partenaires o Projets divers : MeryGrid, E-CLOUD, Interests, …
  4. 4. DEBC - Planning 5 • 16/11/2017 : Blockchains • 21/02/2018 : Internet of Things • 07-11/05/2018 : US Matchmaking Mission & Smart Cities Expo NY • 27/09/2018 : Machine Learning - AI • 16-18/10/2018 : IoT, Blockchain & AI Solutions World Congress | Barcelone
  5. 5. Programme d’aujourd’hui 6 • N–Side : IA et optimisation des performances énergétiques, tendances et perspectives • DC Brain : Use case "Réseaux" - Optimisation des flux par l'IA (gaz, électricité et Chaleur) • Yazzoom : Use case "O&M" – l'IA pour booster la maintenance prédictive dans l'industrie • Optimal Computing : Use case "Efficacité énergétique" – Prédiction des besoins énergétiques d'une maison & Optimisation de pompes via IA • Ingestic : Use case "Efficacité énergétique" - Analyse de la consommation énergétique des écoles catholiques francophones • Energis : Use Case "Energis.Cloud", L'outil qui facilite l'accès à l'IA aux professionnels de l'énergie • Pitches : Opinum & Thelis Corporate
  6. 6. 7 Remerciements
  7. 7. Besoin d’aide ? Contactez-nous ! 8
  8. 8. Cluster Technology of Wallonia Energy, Environment and sustainable Development TWEED Asbl Rue Natalis 2 – 4020 Liège – Belgium Bricout Paul Project engineer pbricout@clustertweed.be Olivier Ulrici Project engineer oulrici@clustertweed.be Cédric Brüll Director cbrull@clustertweed.be 9
  9. 9. INFOPOLE Cluster TIC asbl Rue Godefroid, 5-7| B-5000 Namur - Belgium Tél. 32(0)81 72 51 41 | Fax 32(0)81 72 51 43 infopole@infopole.be www.infopole.be A la recherche d’un partenaire TIC ? Arnaud Ligot Président Sandrine Quoibion Directrice Charlie Feron CommunicaRon & Project Manager
  10. 10. IA et optimisation des performances énergétiques: tendances et perspectives Cluster Tweed Mont-Saint-Guibert, September 27th, 2018 1 Olivier Devolder Head of Energy and Industry
  11. 11. How Big Data has delivered for FedEx for 25 years 2 Artificial Intelligence: plenty of success stories among different industries … How BMW uses Artificial Intelligence and Big Data to design and build cars of tomorrow The amazing ways Google uses deep learning AI Big pharma turns to AI to speed drug discovery, GSK signs deal Comment le Deep Learning fait décoller la reconnaissance d’images
  12. 12. 3 Asset Replacement and Maintenance Optimization with Machine Learning Electricity balancing with Artificial Intelligence Dynamic Dimensioning of Reserve with Machine Learning Optimized Flexibility Valorization based on Deep Learning Price Forecasts Wind farms design and operations with digital twins and optimization Situational awareness with geospatial data analytics … with strong potential in the energy sector
  13. 13. 4
  14. 14. 5 Artificial intelligence is a broader concept than machine learning, which addresses the use of computers to mimic the cognitive functions of humans. When machines carry out tasks based on algorithms in an “intelligent” manner, that is AI.”
  15. 15. 7 Machine Learning is a current application of AI based around the idea that we should really just be able to give machines access to data and let them learn for themselves, without being explicitly programmed
  16. 16. 9 Artificial Neural network is one group of algorithms used for machine learning that models the data using graphs of Artificial Neurons, those neurons are a mathematical model that “mimics approximately how a neuron in the brain works”. Neural Network
  17. 17. 10 Deep Learning Deep learning allows computational models that are composed of multiple neural network layers to learn representations of data with multiple levels of abstraction
  18. 18. DESCRIPTIVE ANALYTICS PREDICTIVE ANALYTICS PRESCRIPTIVE ANALYTICS Make Big Data accessible and manageable Make predictions supported by Top Artificial Intelligence/Machine Learning Techniques Take Optimal data-driven decisions supported by cutting-edge optimization Algo. DATA ECOSYSTEM Improve Data Understanding with advanced dashboards, KPIs and Analysis Using Artificial Intelligence for different objectives 11
  19. 19. AI for a successful Energy Transition 12
  20. 20. Flexible Electricity Consumptions 13
  21. 21. 14 Electricity Production in Germany March 12-March 18 2018 Electricity sector is facing a revolution where flexibility becomes a key asset… • Almost all German Demand has to be covered by Conventional Power Plant è Higher Electricity Price • German demand is more than covered by Renewable è Period with negative electricity price and strong export
  22. 22. Flexible Production Processes Models Flexible CHP Models Flexible Auxiliary Processes Models Storage Model 15 Combining different layers of Advanced Analytics to enable flexibility in industrial sites DA Price Forecast Imbalance Price Forecast Energy Production Forecasts DA/Imbalance Spread Forecast Accurate forecasts + Investment Optimization Planning Optimization Nomination Optimization Real-Time Optimization Efficient mathematical modelling Advanced optimization algorithms =+ Customized Energy Flexibility Optimization Platform
  23. 23. Combining different layers of Advanced Analytics to enable flexibility in industrial sites DA Price Forecast Imbalance Price Forecast Energy Production Forecasts Flexible Production Processes Models Flexible CHP Models Flexible Auxiliary Processes Models Storage Model Investment Optimization Planning Optimization Nomination Optimization Real-Time Optimization DA/Imbalance Spread Forecast Accurate forecasts Efficient mathematical modelling Advanced optimization algorithms =+ + ENERTOP: Customized Energy Flexibility Optimization Platform
  24. 24. 17 Belgian Spot Electricity Price Current Week
  25. 25. 3 – PREDICTING Ø Value Ø Probability Distribution 1 - TRAINING Ø Training the model based on hundreds of features 2 - VALIDATION Ø Validation of the model on historical data Present and Past Information • Imbalance prices • MIP,MDP • Reserve volumes and prices • … Real-Time Information • SI, NRV • Regulation volumes • Strategic reserves • … Forecasts • Onshore and offshore wind • Solar per region Least Squares Time series Nearest- Neighbors Neural Networks Model Forecast Forecast Spot Electricity Prices with Regression Algorithms 18
  26. 26. 19 Belgian Imbalance Price
  27. 27. 3 – PREDICTING Ø Class Ø Probability Distribution 1 - TRAINING Ø Training the model based on hundreds of features 2 - VALIDATION Ø Validation of the model on historical data Anticipate high imbalance exposure with Classification Algorithms Present and Past Information • Imbalance prices • MIP,MDP • Reserve volumes and prices • … Real-Time Information • SI, NRV • Regulation volumes • Strategic reserves • … Forecasts • Onshore and offshore wind • Solar per region Decision trees Support Vector Machines Nearest- Neighbors Neural Networks Model Output class IMB >> 0 Class: +1 IMB << 0 Class : -1 IM B ≈ DA Class : 0 20
  28. 28. 21 Scheduler that run on a server Clean Data Forecast DataModel Data Handler Predicting I N T E R F A C E Storage REST API Feature selection Learning Django PostgreSQL x 500 inputs Using Machine Learning is not only about algorithms Training - Validating FORECAST
  29. 29. Combining different layers of Advanced Analytics to enable flexibility in industrial sites DA Price Forecast Imbalance Price Forecast Energy Production Forecasts Flexible Production Processes Models Flexible CHP Models Flexible Auxiliary Processes Models Storage Model Investment Optimization Planning Optimization Nomination Optimization Real-Time Optimization DA/Imbalance Spread Forecast Accurate forecasts Efficient mathematical modelling Advanced optimization algorithms =+ + ENERTOP: Customized Energy Flexibility Optimization Platform
  30. 30. Goals ENERTOP Consume Produce Model & Forecast Modeling of the industrial process and energy markets. Integration of forecasts for electricity price, demand, etc. Optimize Integrated optimisation of the consumption and production Load Shifting Load Scheduling Load Shedding Fuel Switching CHP Modulation By-product Optimization Using Constraint Programming Generate optimal flexibility decisions : Load scheduling with constraint programming 23
  31. 31. 1. Model Generate optimal flexibility decisions : Load scheduling with constraint programming 24 Accurate results Fast running Robust Solution Intuitive planning Optimized planningConstraint store Domain store Variables + Domains Constraint Constraint Constraint Constraint 2. Search Constraint propagation Backtracking tree search
  32. 32. Integrated EU Day-ahead Electricity Markets 25
  33. 33. 26 Electricity Production in Germany March 12-March 18 2018 Intermittent Renewable Electricity Production leads to a growing importance of import/export at EU level Low Renewable Production in Germany but High demand è Germany is importing High Renewable Production in Germany but Low demand è Germany is exporting
  34. 34. 27
  35. 35. Goals Model & Optimize Couple national markets, maximizes the welfare and optimizes the network utilization, while respecting complex constraints High SLA EUPHEMIA Used each day to compute the electricity price of the 23 European countries participating to PCR Math Optimisation 2500 days of European Market Coupling with Euphemia algorithm Market Cutting-edge Optimization algorithms to solve • Large-scale (multi-countries) • Non-linear (complex network representation, complex market rules) • Non-Convex (complex market products) …problems in limited amount of time All over Europe to solve UE market coupling problem average daily value of matched trades Of successful coupling for DA markets 23 countries 10 min 200 M€ 2500 days 28
  36. 36. Smart bidding on electricity markets with reinforcement learning • Model free • Learning optimal sequential behavior/control from interacting with the environment 29
  37. 37. Smart Local Energy Systems Dynamic Sizing of Balancing Reserves 30
  38. 38. 31 Electricity Production in Germany March 12-March 18 2018 The risks in the systems become more variables and needs to be covered by a « dynamic insurance » Risk of Wind Forecast Error: risk of having more/less electricity than expected (Symmetric ?) Risk of PV Forecast Error: risk of having more/less electricity than expected (Symmetric ?) Risk of power plant outage: risk of having less elec. than expected Risk of HVDC cable outage: risk of having more elec. than expected (if exporting)
  39. 39. • Uncertainty in PV and Wind Production • Uncertainty in Load 32 Forecast Uncertainty Failure and Outage • Forced outage of power plants • Failures in the grid (e.g. storms) What size of Reserve is required to cover the risk ? Why Reserve Sizing ?
  40. 40. 33 • Uncertainty in Renewable Production itself depends on D-1 Forecast level • Uncertainty in Load can depend on D-1 expected system state Forecast Uncertainty Failure and Outage • Risk of forced outage of power plants depends on DAM dispatch • Risk of HVDC failures depends on DAM dispatch Incentive to size in a dynamic way instead of on a yearly basis Why Dynamic Reserve Sizing ?
  41. 41. 34 How to map the system conditions to system imbalance? Feature 1 e.g. WIND FORECAST Step 1: sort the imbalance measures depending on features (e.g. PV & WIND) Historical measures of imbalance e.g. SI = -25MW e.g. SI = 122MW e.g. SI = 172MW Feature2e.g.PVFORECAST
  42. 42. 35 How to map the system conditions to system imbalance? Feature 1 e.g. WIND FORECAST Feature2e.g.PVFORECAST STATIC sizing would just use all the data without considering the features and derive an average sizing need
  43. 43. 36 How to map the system conditions to system imbalance? Feature 1 e.g. WIND FORECAST Feature2e.g.PVFORECAST SCENARIO 2 high wind – high PV SCENARIO 1 low wind – high PV SCENARIO 3 low wind – low PV SCENARIO 4 high wind – low PV To make it DYNAMIC, you need to create “several scenarios” • They represent several situations that could occur • To each scenario are associated past imbalance data
  44. 44. How to map the system conditions to system imbalance? Feature 1 e.g. WIND FORECAST Feature2e.g.PVFORECAST This approach seems to work qualitatively… The problem is how to quantitatively compute these scenarios? • Each features is split in 2? 3? 4?... • How to compute each interval? How to design a methodology to compute these parameters smartly? Even if an agreement is found on these parameters, how to automatically update them with new data (lot of data)? More features are needed to properly make the mapping à with this basic method, the number of scenarios will grow exponentially, e.g. if each features is cut in 2: • 2 features à 4 scenarios • 10 features (which is reasonable) à 1024 scenarios These are key concerns that can addressed with MACHINE LEARNING What are the issues & open questions with such approach?
  45. 45. 38 99.9% reliability COMPLEXITY “DISCRETE” MAPPING “Qualitative clustering” “KMEANS ” “KNN” Deep Learning Feature 1 Feature2 Feature 1 Feature2 Feature 1 Feature2 Automatic and smart “clustering” (i.e. scenarios) “CONTINUOUS” MAPPING Local grouping (no predefined scenarios) Machine learning offers powerful tools to smartly map the system conditions to imbalance
  46. 46. 39 Machine learning tool from conception to production - 1 – Algorithm DESIGN OUTPUT theoretical ML model: • # clusters? • Features? • Parameters of the algo… Algo. 1 Algo. 2 Algo. 3 3 Algo. N … KPI 1 KPI k…v Define a list of possible models v TRAIN & VALIDATE on historical data (get the most out of the data) The models are assessed and compared towards several KPIs Algo. x
  47. 47. 40 Feature 1 Feature2 Machine learning tool from conception to production - 1 – Algo. DESIGN OUTPUT theoretical ML model: • # clusters? • Features? • Parameters of the algo… - 2 - TRAINING OUTPUT The practical trained ML model DATA INPUTS Historical data (>1 year) of SI and system features (DA forecast of wind, PV…) … … … Cluster 2 Cluster 1 Cluster 3 Cluster 4 Cluster 5 … Up reserve = xxx Down reserve = xxx
  48. 48. 41 Machine learning tool from conception to production - 1 – Algo DESIGN OUTPUT theoretical ML model: • # clusters? • Features? • Parameters of the algo… - 2 - TRAINING OUTPUT practical trained ML model - 3 - PREDICT ION The reserve sizing for the next day DATA INPUTS Historical data (>1 year) of SI and system features (DA forecast of wind, PV…) DATA INPUTS Day-ahead system conditions for the next day Feature 1 Feature2 For each hour of tomorrow, check in which cluster we are. E.g. Tomorrow at 10am is cluster 3! Cluster 1 Cluster 2 Cluster 4 Cluster 5 Cluster 3 Up reserve = xxx Down reserve = xxx Orange point = features prediction for next day
  49. 49. 42 Gains in reliability, volumes and robustness thanks to AI-based Dynamic Sizing Gain in RELIABILITY Savings of VOLUMES Gain in ROBUSTNESS Robust methodology which remains beneficial & feasible towards the middle and long term system conditions: • Toward 2020 • As well as towards 2027 Positive business case: • Volume reduction more 85%/time • Financial gains expected of more 2M€/y (outweighing the implementation costs) A better reliability management Higher FRR during higher risk periods: proper reliability secured more constantly along the year 1100 1200 1300 1400 1500 1600 1700 static 0,0 50,0 100,0 UPWARD DOWNWARD High BM scenario 2020 Reference Case 2020 LowBM scenario 2020 Post-Nuclear 2027 Volumesavings Study conducted for Belgium by ELIA with N- SIDE support for Machine Learning aspects
  50. 50. From Proof of Concept to Industrialized Tool 43 POTENTIAL ASSESSMENT & PROOF OF CONCEPT INDUSTRIALIZATION CONTINUOUS IMPROVEMENT • Identify the potential • Quantify this potential • Implement & test several algorithms with high level differences • Assess feasibility of the approaches Effort in MACHINE LEARNING • NOT only consists of “translating” the PoC! • Implies fine-tuning, add extra layers of intelligence, hybridation… to extract the full potential Effort TO BUILD A PLATFORM • Automatize data ecosystem management • Robustness (tool needs to run every day!): fall-back solution... • Build an intuitive interface • … • Adapt and improve the tool as the market changes continuously • Improve the method according to the technological improvements 1 2 3
  51. 51. Thank you ! Olivier, Devolder Head of Energy & Industry Tel: +32 472 46 83 44 Email: ode@n-side.com N-SIDE Avenue Baudouin 1er, N°25 B- 1348 Louvain-la-Neuve 44
  52. 52. Sponsored by : Used by : DCbrain makes sense of millions of measures brought to us by sensors distributed in fluid networks such as electricity, gas, or air conditioning. Our software turns data flows into a real time model of physical networks, thanks to Big Data and Artificial Intelligence … DCbrain, Smarter Grids
  53. 53. Secure the network Optimize your operating budget Decrease Consumptions We believe that Artificial Intelligence is the key to answer these stakes Balancing flows into networks = solving the 3 major issues of complex networks managers ü Classical IT tools can not answer perfectly these stakes û Very complex engineering tool, not suitable for distributed networks. û Network management tool concentrated on alerting, not on predicting & optimizing flows and networks û Low ergonomy /flexibility of basic tools : CMS, GIS, SCADA, BIM… û Finally, a still important use of Excel, creating many risks (tracking, maintenance, reliability…) and an important work load.
  54. 54. On top of this technological core, we propose 2 complementary user interfaces Digital Network View: èan exhaustive view on your entire data èA projection of any unit or quantity present in the data-set (easy to custom interface) è Spot incidents in real time èVisualize scenarios Dashboard view: èAutomatize your reporting tasks èCustomize your dashboards to your needs èVisualize information in an optimum way
  55. 55. Artificial Intelligence and flow networks? Our data-driven approach is Unique Sub1A Sub1B Sub2A Customer Customer Source Sub2B Sub1A Sub1B Sub2A Customer Customer Source Sub2B t t t t t t ON/OFF t Learn the map Learn the edge’s behavior Learn the Nods behavior Analyse / Alert Decide t t t Anomalie Detection. Prediction. Classification. Visualization à Target and anticipate actions. à Reduce costs à Scoring for decision making (Risk, costs …) Flow propagation. Non-lineary transfer F°. local/global optimum search à Test hypothesis: Data-driven What if scénario à Chose the best scenario: Optimum Search à Plan your network for efficiency & cost-cutting
  56. 56. Active nodes Deep Neural network (non linear regression) Gaz 1 Steam Water Gaz 2 non linear transfert fonction Our Graph-Based Approach
  57. 57. Benchmark the optimized model of DCbrain with reality on every steam unit in real time.
  58. 58. MiniMix (Minimun Energy Mix) or how do we optimize many related neural networks Learn the objective functions Learn the constraint functions Minimize it ! Key Success factors : • Automatized data cleaning pipeline • Time series modeling Return on investment: AREVA • -9% on energy spent for producing steam : 1M €/year
  59. 59. Take a step back, inject the local optimized models into the graph and do optimum search at the general level
  60. 60. Deep Flow Engine Full Stack Approach : Open source bricks linked together by DCbrain‘s proprietary code & algorithms § The Deep Flow Engine is an expert of flows : we replicate the specificities of topologies in our technocological core (using in particular « graphs of flows ») § We have integrated into this graphe layer analytics functionalities, with in house algorithms
  61. 61. DCbrain, an easy to use tool, dedicated to operational teams • Integration of all flows & Data in real time (integration of assets, maintenance, flows, HR datasets...) in a single data repository • Intuitive and ergonomic visualization interfaces : a BI interface and a “Digital Model” interface Modelization of network evolution and automatic impact analysis o Short term: evolution of charge in the network, planification of maintenance operations, … o Medium term: integration of new assets, “crisis” scenarios,… Visualize Analyse Decide • Use of our proprietary “flow anomaly detection” algorithm to identify non-optimal resource allocation o Non optimized network structure and tuning o Anomalies : leaks, sensors failure,... Facilitated and comprehensive data mining An easy to use simulation tool An optimized network structure and output
  62. 62. Networks simulationReal time optimization 2 uses cases of DCbrain : Real time optimization and simulation Deep Flow Engine • Demand forecast • Diagnosis of the networks : identification of flow propagation anomalies • Optimum finding for production and distribution • Predictive impact analysis • Network evolution modelling (new production points/new clients integration…) • Automatic scenarios generation (re-routing, flow propagation computation) • Optimum identification Our references : Impact on exploitation costs and consumptions Optimization of engineering processes
  63. 63. Context : GRDF regional engineering offices have a legal obligation to study the integration of any Bio-gas projects in less than 2 weeks They do not have the appropriate modelling tool, capable of rapidly modelling a possible integration, estimate resilience and costs. Action: • Models the gas volume propagation and validate its capacity to be consumed by the existing network and clients • Model the existing network and automatically calculate reframing costs, depending on Bio- gas projects localization Modelling of gas network extension program
  64. 64. Demand forecast and distribution management Context: • This Network is composed of 10 ancient networks, meshed together 5 years ago, with one huge incineration plant • The network is evolving at a fast growing rate • The management team of the network is having issues regarding the balancing of the network and the optimization of the global output Action: • Adaptation of our demand forecast model (using external weather sources) • Creation of a propagation model, using 2 years of data • Implementation of the tool
  65. 65. An intelligent software layer to manage VINCI Energies Micro-grid Concessions Learning automatically any mirco-grid topology Maximizing Photo-voltaïque self consumption through meaningful data-analysis and prediction algorithms
  66. 66. Client case : Optimization of network output (steam network) Context: • The “La Hague” Site of Areva uses large quantity of steam to operate its industrial processes (treatment of nuclear Waste) • The production is done through 4 old steam production units • Data was incomplete Return on investment: • -9% on energy spent for producing steam è 1M€/year saved Action: • Analysis of data and recreation of key indicator(eg global network output) • Analysis of each production unit output • Creation of a model correlating demand and output • Use of a model to optimize production according to demand
  67. 67. Mob : +33 7 68276672 Mail : thomas.bibette@dcbrain.com Address : DC Brain, 23 avenue d’Italie, 75013 Paris Thomas Bibette Export Manager Thank you !
  68. 68. creating value from data Operations & Maintenance l'IA pour booster la maintenance prédictive dans l'industrie Alexis Piron – 27/09/2018
  69. 69. vous aide a creer de la valeur a partir de vos donnees 7Annees d’Expertise 11Personnes Science des Donnees I.A – Machine Learning Prescrip<ve Analy<cs Detec<on d’Anomalies Data et Process Mining Ingenierie Mechatronique Physique Vision et Capteurs Régulation Industrielle et contrôle Solutions Consultance et Services Projets Complets Systemes d’aide a la decision Logiciels sur mesure
  70. 70. Traditional monitoring of Industrial assets 2. Descrip7ve sta7s7cs 3. Dedicated sensors 1. Hand-wri>en rules 4. KPI’s and dashboards
  71. 71. Some problems… • You can’t write rules for every issue or sensor: • there’s no 5me • you don’t know what could go wrong • Problems happen in the so=ware execu5on / in the communica5on with someone else’s equipment. • Not every issue is a vibra5on. • Sta5s5cal analysis doesn’t always work (highly dynamical or complex signals)
  72. 72. What changed? Machines are getting more complex More data is available - Sensors, software logs, contextual data, lab measurements… - Easier to collect (IoT etc) - No storage / speed limits Machine Learning and AI are mature techs
  73. 73. brings AI based anomaly detection to any data monitoring platform Yanomaly contains mul.ple anomaly detec.on algorithms • Some are implementa.on of well known techniques, others are proprietary • They are either univariate or mul&variate (or process mining-based)
  74. 74. Compared to tradi,onal monitoring Detect unknow problems, abnormal trends and pa8erns, external sources Detect problems that are not vibrations Scale easily to more data, new sensors, changes, variable components or subsystems… Detect issues in software execution in advanced machinery Handle complex signals with better detection of issues and less false alarms Take context into account (machine state, temperature, raw material proper,es, recipe)
  75. 75. Service O&M • Look at anomaly scores just prior to start of issue • More efficient investigation and troubleshooting • Reduce the mean time to repair of your service team • Real-time detection of anomalies • Incident prevention: avoid down-time • Speed up root diagnostics for faster recovery Two main use cases
  76. 76. Univariate anomaly detection algorithm example • Compute features describing the signal’s “shape” • Compare current feature values with previous one à compute anomaly score Time between pulses longer than usual Abnormal signal amplitude
  77. 77. Mul$variate anomaly detec$on algorithm example • Predict the signal of a sensor based on the signals of mul4ple other sensors • Compare the predic4on with the actual value à compute anomaly score
  78. 78. Unité de Cogénération
  79. 79. Anomaly Detec-on Solu-on
  80. 80. Pourquoi pour • Modèle OEM de licence • U3lisable par des non-data scien3sts • Performance de détec3on vs fausses alarmes
  81. 81. creating value from data Alexis.Piron@yazzoom.com 0486/80.10.43 www.yanomaly.be
  82. 82. Copyright © 2007-2018 Optimal Computing S.P.R.L. All Right Reserved OPTIMAL COMPUTING www.optimalcomputing.be stephane.pierret@optimalcomputing.be Use cases “Efficacité énergétique” – Prédiction des besoins énergétique d’une maison & Optimisation de pompes via IA Digital Energy Business & Technology Club
  83. 83. Copyright © 2007-2018 Optimal Computing S.P.R.L. All Right Reserved Agenda Use case 1 : Prédiction des besoins énergétique d’une maison passive via IA Use case 2 : Optimisation de pompes via IA 2
  84. 84. Copyright © 2007-2018 Optimal Computing S.P.R.L. All Right Reserved Prédiction de la consommation en chauffage d’une maison passive 3 Maison passive équipée de panneaux solaires intégrés au bâtiment Chauffage appoint: Pompe à chaleur air/air dans le living Chauffage et climatisation + Post chauffe électrique sur ventilation 40 mesures principales (+50) Températures intérieures (9) Consommation d’énergie (9) Qualité de l’air intérieur (3), Production d’énergie (4 + 30), Conditions météo (8 + 10) Températures panneaux (10) Supported by the European Commission’s Seventh Framework Programme
  85. 85. Copyright © 2007-2018 Optimal Computing S.P.R.L. All Right Reserved Equipement de mesure http://www.optimalcomputing.be
  86. 86. Copyright © 2007-2018 Optimal Computing S.P.R.L. All Right Reserved Pourquoi prédire la consommation ? 5 1. Quand stocker ou injecter sur le réseau ou consommer du réseau? 2. Détecter des problèmes dans la régulation ou l’utilisation 1. Pour augmenter l’auto consommation 2. Pour diminuer la consommation et donc Diminuer la consommation annuelle Summer WinterWinter SummerSummer
  87. 87. Copyright © 2007-2018 Optimal Computing S.P.R.L. All Right Reserved 6 Feed Forward Artificial Neural Network ∑ A x1 x2 x3 w1 w2 w3 Summation Activation Function Synaptic weights May have many layers Different types of Layers Different types of activation functions Learn the weight using back-propagation Ce que l’on veut prédire? Consommation de la pompe à chaleur J+1 En utilisant quelles données? Température de consigne du Living J+1 Température du Living J Température extérieure J+1 Radiation solaire J+1 Entrainement du réseau de neurones Base de données sur 700 jours Utilisation du réseau pour la prédiction Température de consigne du Living J+1 Température du Living J Température extérieure J+1 (prévision météo) Radiation solaire J+1 (prévision météo)
  88. 88. Copyright © 2007-2018 Optimal Computing S.P.R.L. All Right Reserved Prédiction du réseau de neurones 7 Summer WinterWinter SummerSummer
  89. 89. Copyright © 2007-2018 Optimal Computing S.P.R.L. All Right Reserved Use case 2 : Optimisation de turbomachines via IA 8
  90. 90. Copyright © 2007-2018 Optimal Computing S.P.R.L. All Right Reserved Possible Optimization Methodologies 9 Simulation Design Variables Responses Time consuming, difficult to deal with lots of design variables and responses, …A Make use of Gradient Requires simulation code modification, local method, noise !, Constraints !, multiple objectives !, … Simulation Design Variables Responses B Simulation Design Variables Responses Black box, global optimization, multiple objectives, uncomputable, learningC
  91. 91. Copyright © 2007-2018 Optimal Computing S.P.R.L. All Right Reserved Optimization Algorithm 10 Design of Experiment Simulation DB Neural Network Training Genetic Algorithm Neural Network Inner Optimization Outer Optimization Simulation This is a learning process with 3 key elements DB Neural Network Genetic Algorithm
  92. 92. Copyright © 2007-2018 Optimal Computing S.P.R.L. All Right Reserved CFD Based Shape Optimization 11 Axial Fan External Diameter Fluid Compressibility RPM Flow Rate Flow Rate Range Peak Efficiency Total Pressure Power Rotor blades Stator blades 310 mm Air Incompressible 4500 2,7 m3/s (at peak efficiency) [1,25 m3/s; 5,0 m3/s] 87,26% 4350 Pa (at peak efficiency) 13,4 kW (at peak efficiency) 20 15 CFD Mesh type Rotor Mesh Stator Mesh Flow type Unstructured mesh 55 734 cells 51 344 cells Steady Flow Turbulent
  93. 93. Copyright © 2007-2018 Optimal Computing S.P.R.L. All Right Reserved Workflow Outline 12 Geometry / CAD Free Form Deformation Simulation Connectors Python scripts Optimization Algorithms Neural Network, Genetic Algorithms Original Geometry STL files Pre-processor CFD Post-processor
  94. 94. Copyright © 2007-2018 Optimal Computing S.P.R.L. All Right Reserved Optimization Results 13 Design variables 12 Parameters Free Form Deformation Maximize efficiency at 3 operating points Objectives Key Data # CPU cores Hardware CFD Calculation Time # CFD simulation Optimization time Performance increase 8 Single Desktop PC i7 10 minutes 112 19 h +0,8 % on averaged +2% at high volume flow
  95. 95. Copyright © 2007-2018 Optimal Computing S.P.R.L. All Right Reserved Axial Fan Performance Results 14 +2% -4,5%
  96. 96. Copyright © 2007-2018 Optimal Computing S.P.R.L. All Right Reserved House Multi-Objective Optimization 15 Goal : Optimization the construction cost versus the energy consumption Tools : PHPP software and project PHPP enhanced by a construction cost calculator Xtreme Multi Objective Optimization
  97. 97. Copyright © 2007-2018 Optimal Computing S.P.R.L. All Right Reserved QUESTIONS ? 16 OPTIMAL COMPUTING www.optimalcomputing.be stephane.pierret@optimalcomputing.be
  98. 98. DATA & IT SERVICES TO MAKE YOUR BUSINESS PRODUCTIVEDATA & IT SERVICES TO MAKE YOUR BUSINESS MORE PRODUCTIVE
  99. 99. INGESTIC SPRL Founded on: 05/2011 Size: 30 FTE Offices: Rue de Rodeuhaie 1, 1348 Louvain-La-Neuve services to improve your information systems and make your data productive
  100. 100. COMUNERIS SMART DATA SMART DATA LEGAL ADVISORY Enterprise architecture definition Creation of the future application – together Selection of the software/hardware suppliers – independently Change management Training Functional coordination Technical functional analysis Applicative architecture Help To Define Help To Build Help To Use
  101. 101. COMUNERIS SMART DATA SMART DATA LEGAL ADVISORY Data science Data mining Data quality Machine learning, Artificial Intelligence
  102. 102. OUR EXPERIENCE Utilities Industries 4.0 Industry Big Data Data Science Data Analysis Smart Data Smart Cities Energy Functional Coordination Business Analysis Enterprise Architecture Data Quality Functional Analysis Testing Smart Meter Smart Grid Mobility IA Machine Learning Smart Cities Energy- Efficiency Projet pilote en cours IIoT Smart Data
  103. 103. Data Science : What can we do ?
  104. 104. Case : Calcul du potentiel d’économie d’énergie des écoles
  105. 105. Quel est le potentiel d’économie d’énergie d’une école ? Consommation d’électricité par élève Consommation de gaz par élève Données récoltées pour • 275 000 élèves • Environ 1000 établissements • 49 GWh d’électricité / an • 444 GWh de gaz / an • ~ 100 établissements avec compteur d’électricité quart- horaire
  106. 106. Les différences entre écoles viennent de gestions différentes, mais aussi d’activités différentes Consommation de nuit élevée Consomma0on de nuit basse Réduction le mercredi après-midi Consommation très basse pendant les congés Consommation légèrement réduite pendant les congés
  107. 107. Consommation tard en soirée Consomma.on tôt le ma.n Mercredi bas, aussi en matinée Jours de semaine semblables
  108. 108. Le modèle d’économie d’énergie identifie les meilleures performances énergétiques, en tenant compte des différences d’activités Les quelques données disponibles sur les établissements sont peu informatives Les differences entre niveaux, options, lieux, tailles, sont faibles au regard de la variabilité des données Comme indicateurs des différentes activités, le modèle utilise la consommation de moments spécifiques Features : consommation de jour, de nuit, avant ou après les cours, le week-end, etc. Comme mesure de gestion performante, le modèle utilise certains niveaux de consommation des écoles performantes En combinant ces informations, le modèle produit un profil de consommation idéal par école
  109. 109. Exemple d’école - 1
  110. 110. Exemple d’école - 2
  111. 111. Exemple d’école - 3
  112. 112. Exemple d’école - 4
  113. 113. CONTACT 0476/81.21.81 paul.courtoy@ingestic.be Paul COURTOY Business Developer Thank you
  114. 114. Tweed 27 / 09 / 2018 Energis.Cloud makes AI easy for Energy professionals ENERGIS by Romain Hollanders
  115. 115. Energis.Cloud makes AI easy for Energy professionals Enable rational use of energy © 2017 Energis sa/nv | www.energis.cloud Pag. 2 Scholars Practitioners Develop super powerful but complex algorithms At ease with statistical and technical tools Algorithms often designed for theoretical data sets Have easy access to actual data sets Would greatly benefit from AI tools to analyse their data But will not use it unless it is super easy to use and interpret
  116. 116. What is Energis.Cloud ? Enable rational use of energy © 2017 Energis sa/nv | www.energis.cloud Pag. 3 END-USERS
  117. 117. At Energis, we specialize in ICT solutions. Our aim is to optimise the energy performance of buildings by empowering Energy Experts with the most innovative technologies. 2015 – 2016 Energis 3.0 +Forecast & ICP 2013 – 2014 Energis 2.0 +M&V 2010 – 2012 Energis 1.0 Monitoring 2017 Energis 4.0 Smart Energy Platform Pag. 4 Angelo Santoro CEO and Founder Frederic Wauters Co-founder and Product Manager Lisiane Goffaux Co-founder and CEO of Freemind Mario Rubino Co-founder and Country Manager Italy Enable rational use of energy © 2017 Energis sa/nv | www.energis.cloud A bit of history
  118. 118. Our mission is to bring value in the energy efficiency market by working together with organizations such as: Utilities, ESCOs, engineering companies, energy managers, facility managers, etc. We do NOT sell directly to end users but we position ourselves as your technological partner for you to create a highly competitive energy management offer. Pag. 5Enable rational use of energy © 2017 Energis sa/nv | www.energis.cloud We are your technological partner
  119. 119. 16/10/2018 (c) 2018 Energis SA | www.energis.cloud | EBM3 22032018 6 Who are we ? ENTER ENERGIS.CLOUD
  120. 120. Energis.Cloud Advanced Analysis Optimisation & Control Bi-directional Data Communication Standard Analysis Pag. 8Enable rational use of energy © 2018 Energis sa/nv | www.energis.cloud Energis Overview Invoice Analysis Invoices/ Contracts DSO MMR AMR Data Loggers BMS/ SCADA IoT Raspicy Weather Services File Upload
  121. 121. Advanced Analysis
  122. 122. What is a model in Energis.Cloud ? Enable rational use of energy Pag. 10 Output data (typically consumption data) Input data (temperature, occupancy, ...) !(#) %& # , %( # , … *! # = ,(%& # , %( # , …) ≈ !(#)
  123. 123. What is the purpose of a model ? Enable rational use of energy © 2017 Energis sa/nv | www.energis.cloud Pag. 11 Quantify and verify energy savings (Measurement & Verification) Track performance of site or equipment Make forecasts and budgets Interrogate to answer questions (past, present, future) Alert about missing/outlier data and correct the data Optimisation & control of systems ?
  124. 124. Types of models White-Box Model (physics-based) Grey-Box Model (semi physics-based) Black-Box Model (machine learning) Enable rational use of energy © 2017 Energis sa/nv | www.energis.cloud Pag. 12 !" # = %('( # , '* # , …) !" # = %('( # , '* # , …)!" # = %('( # , '* # , …) general structure is known parameters are unknown fully known unknown
  125. 125. IPMVP ID Card Enable rational use of energy © 2017 Energis sa/nv | www.energis.cloud Pag. 13
  126. 126. Enable rational use of energy © 2017 Energis sa/nv | www.energis.cloud Pag. 14 Model Identification within a few clicks 1 Select the Y data 2 Select the modelling period 3 Select the granularity 4 Select the X data 5 Identify
  127. 127. ? Enable rational use of energy © 2017 Energis sa/nv | www.energis.cloud Pag. 15 What Energis.Cloud does next HDD Humidity Irradiation Occupancy … ? Which input variables ? ? Which functions ? ? Which regimes ? Many possibilities Machine Learning magics f HDD, Humidity, Irradiation, Occupancy, ...
  128. 128. Enable rational use of energy © 2017 Energis sa/nv | www.energis.cloud Pag. 16 We go much further than linear regression...
  129. 129. Enable rational use of energy © 2017 Energis sa/nv | www.energis.cloud Pag. 17 Many different patterns can be recognized
  130. 130. Enable rational use of energy © 2017 Energis sa/nv | www.energis.cloud Pag. 18 We also recognize many recurrent patterns Seasonal effect Weekly effects And more...
  131. 131. Enable rational use of energy © 2017 Energis sa/nv | www.energis.cloud Pag. 19 Result is evaluated using the IPMVP criteriaResult is fast and precise
  132. 132. Advanced Analysis Optimisation & Control Bi-directional Data Communication Standard Analysis Invoice Analysis It is ready to use
  133. 133. Customer cases
  134. 134. Energy Performance Contract Spotlight: AZ Sint Lucas Hospital Floor surface: 40.000 m² Number of beds: 412 Electricity: Gas: Cost: 7 GWh / yr 4 GWh / yr 750 k€ / yr Goal: • -15% gas • -10% electricity Contract: 5 years (2017 – 2021) Energy efficiency project with low investment “No cure no pay”
  135. 135. Follow-up of photovoltaic installations 80+ monitored sites 30+ identified models 300+ production alerts per year 60% of which have led to an intervention A few key values
  136. 136. Energis.Cloud makes AI easy for Energy professionals
  137. 137. CONTACT US Thank you ! VISIT www.energis.cloud CONTACT ME Romain Hollanders romain.hollanders@energis.cloud or: info@energis.cloud WHERE TO FIND US Brussels (BE) Ottignies-Louvain-la-Neuve – HQ (BE) Caserta (IT)
  138. 138. Enable rational use of energy © 2017 Energis sa/nv | www.energis.cloud Pag. 26 Energis.Cloud decomposes the data into 3 sets Training set Validation set Testing set Outliers Training set used to evaluate individual models Validation set used to select the best model Testing set used to evaluate the IPMVP criteria
  139. 139. Data Platform for Energy & Environmental Actors IA et optimisation des performances énergétiques Alexis ISAAC
  140. 140. Opinum contributes to a better environment and to the energy transition with a digital platform letting energy and environmental actors leverage the untapped potential of data
  141. 141. aaScalable aaFlexible aaSecure Backbone of your digital transformation
  142. 142. Eoly Eoly is the power utility company of Colruyt. The company is in charge of the group zero carbon emission target as well as selling energy their produce through renewable power. As the Energy Management Platform (EMP) of Eoly, Opisense : • Improve the efficiency of the operation team by better predict issues on renewable assets • Drastically extend data analytics capabilities of the energy trading team • Foster innovation in terms of energy management of their assets (building, EV, local renewable power) • Accelerate the development of new digital services for their client Eoly is an happy customer of Opisense since June 2017 More than ~6000 data sources are connected to Opisense combining data from Wind turbine, invoices, sub metering, solar, biomass, open data… Creation of ~200.000.000 data points each month 553 stores 687.000 m² 30.000 employees 581 Structures affiliated (BE & FR)
  143. 143. üR coding environment üPreview result directly in the platform üConnect to your coding tool üDeploy your own processing environment Your own algorithms Calculated Variables Datasources </code R>
  144. 144. Total is transforming its business to become a global energy company. In this context, they massively invest in new business line such as renewable and power & gas supply through their division GRP (Gas Renewable & Power). Total decided to opt for Opisense to leverage the untapped potential of data produced by either their own assets or clients. Call EMP (Energy Management Platform) by Total, the platform handle metering and invoice data from SAP, Salesforces, sub-metering system, smart meters, solar, … The solution is use by Total and by subsidiaries like Lampiris, Greenflex and BHC. Total is an happy customer of Opisense since June 2016 Hundreds of thousands of data sources are connected to Opisense combining data from smart meters, invoices, sub metering, solar, open data…
  145. 145. üRestful API üPowerful authorization scheme handle by the platform üOpen source application to kick start your portal https://github.com/opinum Build on top of Opisense Build on API A P I Sitecare…
  146. 146. MERCI ! Contact : Alexis ISAAC Head of Business Development Rue Emile Francqui, 6 – 1435 Mont-Saint Guibert Phone +32 (0)2 340 19 23 Mobile +32 (0)486 69 69 89 Mail ais@opinum.com Web www.opinum.com
  147. 147. RESEAU IA Le Collectif pour la Wallonie • Accélérer le développement économique de la Wallonie • Réponses simples et pragma7ques • Pour les acteurs privés et publics
  148. 148. RESEAU IA Le Collectif pour la Wallonie • Rassembler l'exper-se • Rendre visible (vitrine wallonne de l’IA) • Rendre lisible (l'écosystème) • Renforcer la chaine de valeurs • Experts de spécialités complémentaires • Lien entre ini-a-ves et organismes existants • Simplifier l'accès à l'informa-on
  149. 149. RESEAU IA Le Collectif pour la Wallonie • Définir les axes importants • Forma2ons • Priorités régionales • Economiques et Éthique • Présenter l'offre de forma2on • Ini2ale et Con2nuée • Technique et Business • Partage d’expériences
  150. 150. RESEAU IA Le Collectif pour la Wallonie • Présence digital • Label « Reseau IA » sur Digital Wallonia • Site internet www.reseauia.be • Groupe LinkedIn • Conférences • IniBée par le groupe (14 Novembre à Liège) • Relais vers l’agenda de l’IA • Proposer orateurs (UCM en janvier) • Ateliers • Entre experts IA (partage d’expériences) • TransformaBon business
  151. 151. RESEAU IA Le Collectif pour la Wallonie • Membres= experts de l’IA • Technique • Business • Légal • Financement • Fiscalité Adhésion par la signature de la charte et NDA • Les clients de l’IA • Entreprises • Communes • Politiques Soumettre sa question, son projet (NDA possible)
  152. 152. RESEAU IA Le Collectif pour la Wallonie Site internet: www.reseauia.be Mail: contact@reseauia.be Tél: 081/402891 Localisa>on:

×