[
  {
    "chunk_id": "db43f402-dfb9-43b1-ae56-be1ea40d0b6b",
    "text": "The Concept of the Deep Learning-Based System\nArtificial Dispatcher to Power System Control and\nDispatch Nikita Tomin Michael Negnevitsky\nand Victor Kurbatsky School of Engineering and ICT\nElectric Power System Department University of Tasmania\nMelentiev Energy Systems Institute Hobart, Australia\nIrkutsk, Russia 664033 Email: Michael.Negnevitsky@utas.edu.au\nEmail: tomin@isem.irk.ru",
    "paper_id": "1805.05408",
    "title": "The Concept of the Deep Learning-Based System \"Artificial Dispatcher\" to Power System Control and Dispatch",
    "authors": [
      "Nikita Tomin",
      "Victor Kurbatsky",
      "Michael Negnevitsky"
    ],
    "published_date": "2018-05-07",
    "primary_category": "cs.CY",
    "arxiv_url": "http://arxiv.org/abs/1805.05408v1",
    "chunk_index": 0,
    "total_chunks": 10,
    "char_count": 384,
    "word_count": 47,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "bdb3cc9e-c930-401a-9bb1-777359c6086d",
    "text": "Abstract—Year by year control of normal and emergency complicated by the low time reserve, the dispatcher makes\nconditions of up-to-date power systems becomes an increasingly a large body of errors. According to some estimations, upMay\ncomplicated problem. With the increasing complexity the existing to 50% of the dispatchers actions, first of all in the stress\n7 control system of power system conditions which includes opersituations, prove to be either insufficiently effective in contents ative actions of the dispatcher and work of special automatic\ndevices proves to be insufficiently effective more and more or timeless.\nfrequently, which raises risks of dangerous and emergency Thus, in the context of such multi-dimensionality and\nconditions in power systems. The paper is aimed at compensating uncertainty the problem of searching and choosing the refor the shortcomings of man (a cognitive barrier, exposure to quired solutions becomes in some cases unsolvable without\nstresses and so on) and automatic devices by combining their\nthe corresponding computer maintenance of the dispatchers strong points, i.e. the dispatchers intelligence and the speed of[cs.CY]\nautomatic devices by virtue of development of the intelligent intellectual activities. Hence, the highly topical problem of\nsystem Artificial dispatcher on the basis of deep machine learning optimal coordination of the capabilities of man and automation\ntechnology. For realization of the system Artificial dispatcher in arises to implement an effective control of the modern power\naddition to deep learning it is planned to attract the game theory system conditions. At the time being, the effective solution to\napproaches to formalize work of the up-to-date power system as\nthis problem seems possible by creating artificial intelligence a game problem. The gain for Artificial dispatcher will consist in\nbringing in a power system in the normal steady-state or post- systems on the basis of the latest practical achievements in\nemergency conditions by means of the required control actions. this sphere, primarily on the basis of the models of deep\nmachine learning. Such intelligent systems allow the effective\nsimulation of the mans actions when solving complicated\nI. INTRODUCTION system problems, in particular, of control of bulk technical\nDevelopment of modern power systems sharply compli- systems, which is confirmed by specific introductions of smart\ncates problems of their control and survivability support. The programs to many important branches: medicine, aviation,\nfunctions of power system condition control are performed electric power industry, etc.\nprimarily by automatic load frequency and generation control The paper is aimed at compensating for the shortcomings of\nmulti-year operating practice as well as a set of studies reveal automatic devices by combining their strong points, i.e. the\nlimited capabilities and disadvantages of the existing automa- dispatchers intelligence and the speed of automatic devices\ntion systems which are stipulated by their low intelligence by virtue of development of the intelligent system Artificial\nlevel, absence of coordination, low fault tolerance. This makes dispatcher on the basis of deep machine learning technology.\nit necessary to improve the principles of automatic control The recent practical developments vividly demonstrate that the\nof normal and emergency conditions in power systems. At smart models first of all on the basis of deep machine learning\nthe same time the modern power systems cannot operate reli- can both successfully solve highly complicated system probably without an active participation of dispatchers in solving lems and rank over the man in some capabilities. This fact is\nthe problems of their condition control. Specific features of clearly illustrated by the recent victories of the smart programs\nmodern power system operation are predetermined to a great in complicated game problems.\nextent by availability of the continuous flow of disturbances, Development of new forms of learning (e.g. on the basis\nwhich lead to ever present transient processes.",
    "paper_id": "1805.05408",
    "title": "The Concept of the Deep Learning-Based System \"Artificial Dispatcher\" to Power System Control and Dispatch",
    "authors": [
      "Nikita Tomin",
      "Victor Kurbatsky",
      "Michael Negnevitsky"
    ],
    "published_date": "2018-05-07",
    "primary_category": "cs.CY",
    "arxiv_url": "http://arxiv.org/abs/1805.05408v1",
    "chunk_index": 1,
    "total_chunks": 10,
    "char_count": 4129,
    "word_count": 615,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "3e95c429-9f9b-49cc-aa23-5d1bfbf4ebaa",
    "text": "Uncertainty of the Boltzmann machine, elastic weight consolidation) and\nand severity of emergency consequences create high risks new architectures of models for machine learning (deep neuin decision making. In such conditions which as a rule are ral networks, convoluted neural networks, etc.) gave rise to intuitive thinking and memory effect of such systems. This microgrids management [11], enhancement of relay protection\nmade it possible to learn smart programs for solving several [12] and some others. Meanwhile, it seems promising to\nproblems (multi-problem learning) simultaneously, acquiring develop machine intelligence systems on the basis of deep\nthe capability of intuitive understanding of the winning situa- learning in a complex of power system security control probtion, as called by the experts. Presently all these factors allow lems (decision making in emergency situations, restoration\nthe current algorithms of machine learning to solve extremely after emergency of complex power systems, on-line security\ncomplicated problems, such as control of a bulk technical assessment, etc.) which can hardly be solved by the existing\nsystem, better than by man. methods, including those intelligent ones, because it is difficult\nto formalize the process and its high rate. STATE OF THE ART\nIII.",
    "paper_id": "1805.05408",
    "title": "The Concept of the Deep Learning-Based System \"Artificial Dispatcher\" to Power System Control and Dispatch",
    "authors": [
      "Nikita Tomin",
      "Victor Kurbatsky",
      "Michael Negnevitsky"
    ],
    "published_date": "2018-05-07",
    "primary_category": "cs.CY",
    "arxiv_url": "http://arxiv.org/abs/1805.05408v1",
    "chunk_index": 2,
    "total_chunks": 10,
    "char_count": 1308,
    "word_count": 193,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "c98ec873-a029-4067-b732-e494adbac714",
    "text": "THE ARTIFICIAL DISPATCHER CONCEPT The main directions in the enhancement of dispatching and\nemergency control have become automation of dispatching Most of the software systems and systems used to autocontrol and coordination of automatic actions on the basis of mate the management of power systems (products SK-2007,\ncontemporary mathematical methods and up-to-date control PowerOn Fusion DMS, Network Manager SCADA / DMS,\nmeans, computation equipment and automation [1]. An indi- etc.) are designed either as algorithmically rigid programs or\nvidual direction in the enhancement of on-line power system using relatively simple production rules based on a specially\ncontrol has become the development of the so called dispatcher created knowledge base. This does not allow us to consider\nadvisers (DA) or decision making systems that represent a them fully expert, since intellectual processes in them are not\nman-machine system which enables power system dispatcher separated into a separate category and methods for their impleto apply the data, knowledge, and on-line models to analyze mentation have not been proposed. As a result, such systems\nand solve control problems in the power systems under the are not always able to adapt effectively to the variety of types\nconditions of uncertainty and small time reserve. of power systems and the conditions of their operation. The\nThe evolution of artificial intelligence methods made it need for training and adaptation to specific conditions often\npossible to considerably accelerate and automate the process of forces developers to make changes or completely, to process\nsolving the problem of power system security analysis [2]. For algorithms at the programming level.\nthe dispatcher adviser systems this allowed the implementation The paper proposes the concept of a new intelligent tool of\nof an on-line analysis of emergency consequences when the a new type - the \"Artificial Dispatcher\" system, which will\nmathematical model of power system is based on the results combine the functions of operational dispatch and automatic\nof measurements that come in a real-time mode. The earliest control of the power system based on deep learning techintelligent technology for dispatcher adviser represented expert nology. The key aspect of this technology is the simulation\nsystems that were certain continuation and development of a of human decision making with significantly higher speed\nsituational modeling methodology [3,4]. The development of and accuracy, which is achieved by the original principles\ndispatcher adviser intelligent systems involved the approaches of learning the models of deep neural networks and the\nbased on the machine learning algorithms [1,5] as well as a ability to operate with a big data that is inaccessible to a\nmethodology of distributed artificial intelligence, first of all, human due to limitations of his memory. Moreover, unlike the\nmulti-agent systems [6]. algorithms of conventional predictive models, which are laid in\nHowever, despite all the indicated advantages of the intelli- the traditional dispatcher advisors, deep neural networks do not\ngent methods when applied to solve the problem of power simply take into account the factors indicated by programmers,\nsystem operation control, many smart models have some but reveal these factors themselves.\ncertain downsides (retraining, curse of dimensionality , one- For realization of the system Artificial dispatcher in addition\ntask learning, etc.), which hinders their full adoption in current to deep learning it is planned to attract the game theory\noperation of operating and automated control [2]. In fact this approaches to formalize work of the up-to-date power system\nmeans that such technologies as multi-agent systems, expert as a game problem with complete and/or incomplete informasystems, conventional machine learning do not allow us to tion [13]. In this case the gain for Artificial dispatcher will\nobtain a fully-functional machine intelligence, which though consist in bringing in a power system in the normal steadywith a certain error still imitates the intelligence of man state or post-emergency conditions by means of the required\n(in our case power system dispatcher).",
    "paper_id": "1805.05408",
    "title": "The Concept of the Deep Learning-Based System \"Artificial Dispatcher\" to Power System Control and Dispatch",
    "authors": [
      "Nikita Tomin",
      "Victor Kurbatsky",
      "Michael Negnevitsky"
    ],
    "published_date": "2018-05-07",
    "primary_category": "cs.CY",
    "arxiv_url": "http://arxiv.org/abs/1805.05408v1",
    "chunk_index": 3,
    "total_chunks": 10,
    "char_count": 4244,
    "word_count": 640,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "7e82f4c2-4292-4cb8-b4a5-5f2bab17b930",
    "text": "However, the recent control actions. Thereby, the imaginary enemies of Artificial\ndevelopments in the field of deep machine learning which dispatcher will be both arising unpredictable situations in\nhave already found wide application in some sectors (aviation, power systems (equipment overload, shutdowns, short circuits\nmedical industry, security, etc.) allow the establishment of and the others), and purposeful malicious actions (terrorism,\nmachine intelligence systems that in some cases are capable hacker attacks).\nto solve complex system problems even better than man [7]. Structure\nIn the past years the deep learning models have started to\nfind their use in the problems of power industry, including the The intelligent system Artificial dispatcher should include:\nproblems of emergency control [8], energy equipment diag- • module for collection, processing, formalizing events\nnostic [9], system transient stability assessment [10], energy from SCADA/EMS/DMS systems; 2) Intelligent machine support for the dispatcher actions Power Grid\nSCADA/EMS (\"open-loop control mode\" or \"dispatcher advisor\"\nmode), in which control actions are generated, which\nSystem Dispatcher can then be implemented by the dispatcher.\n3) Automatic intelligent control of the regime of realArtificial Dispatcher System time (\"'closed-loop control mode\"'), in which optimum\nSystem interface preventive and / or emergency control actions are auModule for collection, The making decisions module tomatically implemented without operator verification.\nprocessing, formalizing\nIn closed-loop mode the confirmation of the commands events Search for solutions/actions\n(PolicyDNN/CNN) by the user is not needed. As soon as the results are The system state analysis\nmodule available the commands are automatically build and\nEstimate of the value of the\nSystem state evaluation current state issued without any user interaction.\n(ValueDNN/CNN) 4) Combined control (\"'open-closed loop control mode\"'), Pattern recognization\n(Onlinedecisiontrees)\nGenerating decision in which part of the actions can be performed by the\nGoal formation (MonteCarloTreeSearch) dispatcher, while others determine the machine intelligence. The big database of The big database The big database dispatcher s experience B.",
    "paper_id": "1805.05408",
    "title": "The Concept of the Deep Learning-Based System \"Artificial Dispatcher\" to Power System Control and Dispatch",
    "authors": [
      "Nikita Tomin",
      "Victor Kurbatsky",
      "Michael Negnevitsky"
    ],
    "published_date": "2018-05-07",
    "primary_category": "cs.CY",
    "arxiv_url": "http://arxiv.org/abs/1805.05408v1",
    "chunk_index": 4,
    "total_chunks": 10,
    "char_count": 2275,
    "word_count": 314,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "c111f216-7fc4-42ba-840a-9a449fd13e02",
    "text": "Realization of power grid of power grid and actions of automation real behavior simulations devices The advantages of deep learning technology are achievable\nwith the use of specialized hardware training based on the\nsystem of multi-core graphic processors (MGP), which allow\nFig. 1. A general structure of the system \"Artificial dispatcher\".\nboth high-performance computing and high speed processing\nof big data at an adequate time. Therefore, the MGP-based\nsystem will represent the hardware part of the \"Artificial\n• a system state analysis module designed to identify a\nDispatcher\", on the basis of which an original set of specialclass of power system states, evaluate the current system\nized programs will be implemented, including models of deep\nstate patterns and form a goal. To solve these problems,\nneural networks trained to manage electrical networks.\nwe use the offline big data and/ or online data of\nThe \"Artificial Dispatcher\" system will be designed in\nSCADA/EMS/DMS systems. This module will use online\nsuch a way as to be able to integrate into existing\ndecision trees algorithms to monitor and evaluate power\nSCADA/EMS/DMS software platforms. Some of this platgrid states in real time.\nforms is open for the integration of applications from differ-\n• a decision making module based on the\nent developers based on international protocols. As a result,\ndeep/convolutional neural networks (DNN/CNN), which\nthe software-hardware system \"Artificial Dispatcher\" will be\nallows to provide the analysis of control capabilities\nimplemented as an additional basic module (application) and\nof the power grid (security constraints, availability\nuse the functional of the operated SCADA/EMS/DMS (calcuof control devices and remote control, sensitivity\nlation, analysis, data collection / transmission, remove control,\nsettings of relay protection, etc.), search and issuance\netc.) for the generation and realization of control actions.\nof preventive and / or corrective control actions based\nAt the same time, the machine intelligence of the system\non the global descriptors obtained. For this module the\n\"Artificial dispatcher\" will make it possible in many cases\npower system control and dispatch will be formalized\nto replace the real dispatcher on the principle of \"autopilot\"\nas the multi-objective gaming problem with potential\nin order to improve the efficiency of power grid management.\nvirtual players (uncertainties of renewable power output,\nThe positive effects of the implementation (integration)\noutages, hacker attacks etc.) as the main opponents of\nof the system \"Artificial Dispatcher\" as an additional basic\nthe \"Artificial Operator\".\nmodule (application) into the existing SCADA/EMS/DMS\n• three big databases obtained on current/retrospective\nsystem are:\nmeters, simulations, real dispatcher's experience (dis-\n• neutralizing a negative aspect of human factor in the patch instructions, rules for preventing and elimination\npower system control and dispatch; alarm/dangerous states, etc.)and actions of automation\n• increasing the power system management efficiency (with devices\nvoltage regulation, optimization of power flow, security • interface \"Artificial dispatcher\", containing tools for gencontrol, etc.), defined as the original properties of the erating and managing a dialogue with the user (disDNN/CNN models for quickly finding optimal solutions, patcher)\nand the ability to operate with bigdata, they are inaccessiThe main functions performed by the \"Artificial dispatcher\" ble to the dispatcher due to the physiological limitations\nwill be: of his memory;\n1) Online system state monitoring with predicting possible • increase in the speed of the power system management\nalarm states associated with the original properties of the DNN/CNN 0.4\nActual L−Index\nTraditional Approach 0.3\nMachine Learning Approach\n0.2 Fig. 3. The curves Lsum before and after corrective control actions.\n0 100 200 300 400 500 600 700 800 900 1000\nLoad Factor of IEEE 118 and the speed of automatic devices within the unified softwareFig. 2. Comparison results of testing different approaches to IEEE 118 test hardware system to power system management, solving thus\npower system using corrupted\" test set.\nthe above topical problem of the man-automatic device integration. Such a system will make it possible to coordinate\nmodels to quickly find optimal solutions in the automatic actions of the security and emergency control systems on the\ncontrol range (fractions of seconds); basis of solutions obtained by the machine intelligence. As\n• reduction of financial costs due to optimal power system a result the risk of erroneous and timeless control actions\nmanagement (energy loss saving, emergency reduction, will decrease essentially and the reliability and survivability\netc.) of modern power systems will improve substantially. The scientific novelty of the work consists in development\nC.",
    "paper_id": "1805.05408",
    "title": "The Concept of the Deep Learning-Based System \"Artificial Dispatcher\" to Power System Control and Dispatch",
    "authors": [
      "Nikita Tomin",
      "Victor Kurbatsky",
      "Michael Negnevitsky"
    ],
    "published_date": "2018-05-07",
    "primary_category": "cs.CY",
    "arxiv_url": "http://arxiv.org/abs/1805.05408v1",
    "chunk_index": 5,
    "total_chunks": 10,
    "char_count": 4914,
    "word_count": 727,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "b4970421-ebb4-4eba-a381-c84fd6594346",
    "text": "Case Study\nof an innovative system for the up-to-date power system\nWe have devised an innovative on-line machine learning control which is a set of programs of the artificial intelligence\nmethod for voltage stability control of power system, using the having no analogs in the world electric power industry until\ntechnology of online decision trees [1, 14]. This models were now. The system Artificial dispatcher is supposed to be able to\ntrained to determine in real time the voltage stability indicator completely or partially substitute a real dispatcher simulating\n(L-index) for the security assessment of an entire system, and his actions at a high speed and efficiency. This will become\nthe required reactive power injections, when determining the possible owing to the original properties of the advanced algoplace and magnitude of corrective actions.",
    "paper_id": "1805.05408",
    "title": "The Concept of the Deep Learning-Based System \"Artificial Dispatcher\" to Power System Control and Dispatch",
    "authors": [
      "Nikita Tomin",
      "Victor Kurbatsky",
      "Michael Negnevitsky"
    ],
    "published_date": "2018-05-07",
    "primary_category": "cs.CY",
    "arxiv_url": "http://arxiv.org/abs/1805.05408v1",
    "chunk_index": 6,
    "total_chunks": 10,
    "char_count": 858,
    "word_count": 133,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "663ac5dc-cada-4988-8df9-dc7b9fc2441f",
    "text": "We will use this rithms of machine learning, first of all deep neural networks\nmethod to develop the system \"Artificial Dispatcher\". which allow the optimal solution to be found on the basis of\nThe experiments have showed the efficiency of proposed ap- the experience in the analysis of enormous number of schemes\nproach. First of all, \"'bad data\"' IEEE 118 simulating showed and conditions, which is physically impossible for the human\nthat the machine learning algorithms are robust to various brain.\nfailures and gaps in the initial data (Fig. 2). At the same time,\nthe traditional algorithmic calculation of this indicator for \"bad REFERENCES\ndata\"' produced significant errors for the security assessment. Moreover, the calculations show that the machine-learning [1] Voropai N.I. et al., A set of intelligent tools for prevention of large-scale\napproach provides lower errors (rootmeansquare error of order emergencies in electric power systems, Novosibirsk: Publishing House\nNauka, 2015 - 332 p.\n13% for IEEE 118) and high speed of solving process (about [2] M. Negnevitsky, Artificial intelligence: a guide to intelligent systems, 3rd\ncentiseconds for each steady state of IEEE 118 compared to edn, Addison Wesley, Harlow, England, 2011. 504 p.\n30-40 minutes in the traditional approach) when determining [3] N.A. New information\ntechnologies in the problem of operational control of electric power\nthe additional reactive power injections. The obtained values systems - Yekaterinburg: UB RAS, 2002. Pp. 43-59\nof additional injections from machine learning model were [4] M. Negnevitsky, An Expert System Application for Clearing Overloads,\nused for reactive power compensation by using reactive power Int.",
    "paper_id": "1805.05408",
    "title": "The Concept of the Deep Learning-Based System \"Artificial Dispatcher\" to Power System Control and Dispatch",
    "authors": [
      "Nikita Tomin",
      "Victor Kurbatsky",
      "Michael Negnevitsky"
    ],
    "published_date": "2018-05-07",
    "primary_category": "cs.CY",
    "arxiv_url": "http://arxiv.org/abs/1805.05408v1",
    "chunk_index": 7,
    "total_chunks": 10,
    "char_count": 1714,
    "word_count": 260,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "5b6f1690-1066-4495-986b-13e2aa03bb25",
    "text": "J. of Power and Energy Systems, Vol.15, No.1, pp. 9-13, 1995.\nsources, which decreased the sun of local indices Lsum, whose [5] Diao R. at [al]. Decision tree-based online voltage security assessment\nusing PMU measurements, IEEE Trans. on Power Systems Vol. 24, no.\nincrease is indicative of even greater proximity of voltage 2, May 2009. Pp. 832-839.\ncollapse (Fig. 3) [6] R. Zarrabian, Experimental Implementation\nof Multi-Agent System Algorithm to Prevent Cascading Failure after NIV. CONCLUSION 1-1 Contingency in Smart Grid Systems, IEEE Power & Energy Society\nGeneral Meeting. – 2015\nThis project is devoted to development of the system [7] Li Deng, Dong Yu, et al. Deep learning: methods and applications. Artificial dispatcher which will combine the mans intelligence Foundations and Trends in Signal Processing, 7(34):197387, 2014. Wang et al, A multilevel deep learning method for big data analysis\nand emergency management of power system,2016 IEEE International\nConference on Big Data Analysis (ICBDA), 2016\n[9] Yixing Wang, Meiqin Liu, Zhejing Bao, Deep learning neural network for\npower system fault diagnosis, 35th Chinese Control Conference (CCC),\n[10] K.Jothinathan , S. Ganapathy, Transient security assessment in power\nsystems using deep neural network, International Journal of Applied\nEngin. Research, Vol. 10, No. 15, 2012. - Pp 787-790\n[11] V.",
    "paper_id": "1805.05408",
    "title": "The Concept of the Deep Learning-Based System \"Artificial Dispatcher\" to Power System Control and Dispatch",
    "authors": [
      "Nikita Tomin",
      "Victor Kurbatsky",
      "Michael Negnevitsky"
    ],
    "published_date": "2018-05-07",
    "primary_category": "cs.CY",
    "arxiv_url": "http://arxiv.org/abs/1805.05408v1",
    "chunk_index": 8,
    "total_chunks": 10,
    "char_count": 1366,
    "word_count": 207,
    "chunking_strategy": "semantic"
  },
  {
    "chunk_id": "77604ca7-5559-48c1-b19e-5b9375c33d7e",
    "text": "Fonteneau, Deep Reinforcement Learning Solutions for Energy Microgrids Management, European\n[12] Igor Kogan, K. Herrmann, Deployment of advanced machine learning models to protection relays, Proc. of the Relay Protection\nand Automatics, Sankt-Petersburg, Russia, 2017\n[13] Shengwei Mei Wei et al. On engineering game theory with its application\nin power systems, Control Theory and Technology, February 2017, Vol.\n15, Issue 1, pp 112\n[14] N. Negnevitsky,\nDevelopment of automatic intelligent system for on-line voltage security\ncontrol of power systems, IEEE PES PowerTech 2017, Manchester, UK.",
    "paper_id": "1805.05408",
    "title": "The Concept of the Deep Learning-Based System \"Artificial Dispatcher\" to Power System Control and Dispatch",
    "authors": [
      "Nikita Tomin",
      "Victor Kurbatsky",
      "Michael Negnevitsky"
    ],
    "published_date": "2018-05-07",
    "primary_category": "cs.CY",
    "arxiv_url": "http://arxiv.org/abs/1805.05408v1",
    "chunk_index": 9,
    "total_chunks": 10,
    "char_count": 594,
    "word_count": 84,
    "chunking_strategy": "semantic"
  }
]