Energy Speaker Series - Module 3: Energy Storage and Power System Control with AI - MATLAB
Video Player is loading.
Current Time 0:00
Duration 1:11:47
Loaded: 0.23%
Stream Type LIVE
Remaining Time 1:11:47
 
1x
  • descriptions off, selected
  • en (Main), selected
    Video length is 1:11:47

    Energy Speaker Series - Module 3: Energy Storage and Power System Control with AI

    Dr. Pietro Raboni and Guido Recalcati, NHOA
    Prof. Francisco Gonzalez-Longatt, UniversitetetiSørøst-Norge
    Dr. Francisco Sánchez, Loughborough University
    Matteo Galgani and Marco Magrini, ENEL

    Overview

    Session 3.2: Development of State-of-the-art Power Plant Controllers for Energy Storage Applications,
    Dr. Pietro Raboni and Guido Recalcati, NHOA

    NHOA, formerly Electro Power Systems - Engie EPS, is one of the top global players in energy storage and e-mobility with the aim of enabling the paradigm shift in the global energy system towards clean energy and sustainable mobility.

    This presentation aims to showcase how MATLAB®, Simulink® and Simulink PLC Coder™ sped up the development of the EEPS’s proprietary Power Plant Controller (PPC). The speech introduces the PPC as the key brick for any utility scale renewable, storage and Vehicle-to-Grid plant. Indeed, these applications have in common the need of coordinating fast acting power electronics converters alongside with TSO requests and the longer-term decisions of an Energy Management System or a Market Aggregator. Moreover, the PPC plays a vital role for large scale plants in terms of grid code compliance, as well as for preserving and optimally managing the batteries in BESS and V2G applications. 

    In the last 3 years EEPS upgraded its PPC development approach embracing Model Based Design and using Simulink. This eased the design of more complex control structures, their testing and integration with the rest of our library. Moreover, the integration of Git with Simulink makes easier tracking the versioning and code development in a growing company. The Simulink projects are then automatically converted to different industrial hardware control platforms on a project base, leveraging on Simulink PLC Coder™ and a set of in-house scripts. This revolution accelerated the whole PPC design and turned out a key-advantage during the COVID-19 months for the remote commissioning of the plants. The presentation is based on simulation and field recordings, collected during design, commissioning and operation of our BESS and V2G projects.

    Session 3.3: Deep Reinforcement Learning-Based Controller for SOC Management of Multi-Electrical Energy Storage System,
    Prof. Francisco Gonzalez-Longatt - UniversitetetiSørøst-Norge and Dr. Francisco Sanchez, Researcher - Loughborough University

    Electrical energy storage systems (EESSs) have become increasingly attractive to provide fast frequency response services due to their response times. However, proper management of their finite energy reserves is required to ensure timely and secure operation. This presentation shows the implementation of a deep reinforcement learning (DRL) based controller to manage the state of charge (SOC) of a Multi-EESS (M-EESS) and provide frequency response services to the power grid.

    The DRL based controller decides when to charge or discharge the M-EESS to provide the frequency service while simultaneously controlling the SOC of the M-EESS to reach the desired level. The DRL agent is trained using an actor-critic method called Deep Deterministic Policy Gradients (DDPG) that allows for continuous action and smoother SOC control of the M-EESS. MATLAB and Simulink are used as the modelling and simulation framework for the controller. Battery energy storage, flywheel and ultra-capacitor energy storage models have been implemented using Simulink together with the environment used to define observation and actions; also, the agent has been developed in Simulink taking advantage of the Reinforcement Learning Toolbox (RLT). The training process was implemented using MATLAB live script, making it easy to understand and use the RLT. The proposed controller is compared to benchmark DRL methods and other control techniques, i.e., Fuzzy Logic and a traditional PID control. Simulation results show the effectiveness of the proposed approach.

    Session 3.4: Diagnosis and control of geothermal plants with MATLAB
    Matteo Galgani and Marco Magrini, ENEL Italy

    Enel Green Power GEO deals with the production of electricity from geothermal sources. The fleet currently has 35 plants for a total of 750MW of installed power. This presentation focuses on the calculation architecture adopted by Enel Green Power for the execution of algorithms for both diagnostic and control purposes.

    The calculation system is composed of two parts: the first part is the execution of monitoring and diagnostic algorithms, residing in the central Mainframe; the second part for the execution of control algorithms, residing in the PLC of the production.

    The heart of the system is MATLAB, which is used for the development, testing and execution of the developed algorithms. The data is acquired from the production plants with an average sampling time of 5 seconds and stored in the central Mainframe. This data stream is accessible by MATLAB for the development of algorithms based on historical data. Through the use of the MATLAB Production Server we are also able to execute the algorithms developed in real time directly on the data acquired by the plants.

    In this presentation, we will show plant signature algorithms we used for monitoring & diagnostics and control algorithms we developed and implemented on a PLC to allow a direct and much faster action directly on a centrifugal compressor.

    About the Presenters

    Session 3.2

    Dr. Pietro Raboni received his B.S. and M.S. degrees in Electrical Engineering from Università degli Studi di Pavia, Italy, in 2008 and 2011, and the Ph.D. in Energy Technology from Aalborg University, Denmark, in 2016. He is currently Head of System R&D at ENGIE-Eps, Milan, Italy. From 2014 to 2017 he was inverter and plant modelling specialist with ABB, either for PV or BESS products. He is currently member of CEI CT-316 and industrial lecturer at Politecnico di Milano, Italy. He is an expert in modelling and control of large-scale PV and BESS plants, as well as microgrids. His research interests span from inverter to renewable power plant controllers and include EMS for microgrids and V2G applications.

    Guido Recalcati received the M.S degree in Electrical Engineering from Politecnico di Milano in 2018. Since then, he has been working with Engie EPS as part of the R&D team. Recently he has joined Free2Move eSolutions, a JV between Stellantis and ENGIE Eps. His main activities include simulation and development of control systems for energy storage and V2G plants.

    Session 3.3

    Prof. Francisco Gonzalez-Longatt is a full professor in electrical power engineering at Instituttfor elektro, IT ogkybernetikk, UniversitetetiSørøst-Norge, Norway. His research interest includes innovative (operation/control) schemes to optimize the performance of future energy systems.

    Dr. Francisco Sánchez received his BS degree in Electrical Engineering from Simon Bolivar University, Caracas, Venezuela, in 2011 and master’s degree in Renewable Energy Technologies from Polytechnic University in Madrid, Spain, in 2013. He recently completed his PhD. degree in Electrical Engineering at Loughborough University in the UK. His research focuses on the development of artificial intelligence techniques for power system analysis and energy management applications.

    Session 3.4

    Matteo Galgani received his BS degree in computer engineering and specialized in robotics and artificial intelligence.  Mr. Galgani has been with ENEL for six year, where he focuses on technology innovations and data analysis.

    Marco Magrini received his BS degree in Information Technology from University of Pisa and he also holds a MBA from the Politecnico di Milano. Mr. Magrini has been with ENEL since 1991, where he focuses on Green Power such as the operation and maintenance of geothermal power plants. In the past, Mr. Magrini worked on physical models of the geothermal reservoir, implementation of relational databases, SAP customizations, remote control and monitoring of power plants, development of computer networks and advanced data analysis related projects.

    Recorded: 1 Dec 2021

    Hi, ladies and gentlemen. This is Pietro Raboni, head of the system R&D at NHOA.

    And this is Guido Recalcati, system R&D engineer at Free2Move eSolutions.

    Thanks, Guido. And thank you, MathWorks for arranging this event. Today, we will present to you our work related to power and plant controller for energy storage applications. We will start with a brief introduction about NHOA. Then we will have a look at energy management system and our development of power plant control. Then we will jump to show you some field results collected from our better energy storage and vehicle-to-grid applications.

    Let's start from NHOA. In 2005, we were just two spin-off coming from Politecnico di Milano and Politecnico di Turino are among the best Universities of Southern Europe. Then in 2015, we were listed on the Paris Stock Exchange market, while in 2018 ENGIE acquired the majority of the stakes. The same, in the last month, we are taken over by TCC, a Taiwanese industrial player. And it is worth to mention also that in the same month, we signed also the joint venture with Stellantis under the name of Free2Move eSolutions.

    We are certainly an energy storage and eMobility player. We count nowadays around 200 employees. And we are all based in Italy but with offices also in the US and in Australia. With reference to our iconic projects, we have to mention our utility-scale battery energy storages, most of all in the US, but also our off-grid microgrids in Africa and the eMobility products that we deliver in the last year.

    With respect to our offer, let's say that it relies on Hyess, our vertically-integrated technology platform. It spans from the power conversion system, so let's say the inverter, that we design and also assembled internally to energy management system, as well as to our digital platform for eMobility. With this set of products, we are capable to satisfy our business lines.

    In particular with respect to the stationary storage, we used to split into two product lines-- giga storage that are basically utility-scale standalone battery energy storage, and solar-plus-storage, and the industrial solution, and under this set, full off-grid microgrids, but also on-grid communities. We have also storages at the service of large industries as well as the activities played in the context of an hydrogen integrator.

    In this slide, we pasted a snapshot of our most important plants, starting from the first microgrids in Africa, passing to the first pilots of grid-connected battery energy storages at the service of interconnected power systems, so for providing ancillary services. Then there are also application with hydrogen, application in remote islands, as well as the latest project in the US and related to eMobility. Now I pass the floor to Guido that will tell you more.

    Thank you, Pietro. As Pietro mentioned earlier, Free2Mode eSolutions is the joint venture between NHOA and Stellantis that was born earlier this year. Our aim is to bring the experience of NHOA in storage system and power electronics together with the industrial history of Stellantis to build innovative eMobility products and services. Our main business lines are charging solutions for business, residential, and public customers; Charging-as-a-Service, providing customers an easy and flexible charging plan; and advanced energy services like second-life battery usage and vehicle-to-grid, which I will tell you more about later.

    Now I give the floor again to Pietro to tell you more about the EMS.

    Thanks, Guido. And now let's have a look at the PROPHET EMS. But first of all, let's clarify what do we mean for an energy management system? Because if you Google it, you will find thousands of papers related to EMS, as well as hundreds of products.

    Nevertheless, according to IEC, we can refer to an EMS as a system operating and controlling energy resources and loads. Basically, if you go on the market, we find EMS that deal with the management of large interconnected grids, as in the case of TSO and large utilities. We may find EMS that, through the operation of microarrays as well as virtual power plants, and also EMS that optimize the generation assets. Evidently, in this case there are EMS a little bit out of IEC definition.

    But if you look at the EMS from NHOA perspective, we have to speak about our PROPHET EMS, that is our proprietary network of smart control cabinets, that orchestrates multiple assets as well as the balance of plant. And it is represented by a modular hardware and software architecture. Nowadays, PROPHET EMS allows operating utility-scale battery energy storage, PV-plus-storage, and vehicle-to-grid applications.

    But what has been the evolution of our EMS? Let's say that in our early days, our EMS was represented by a single layer of PLC that were carrying out either real-time operation and fuel minimization in the context of off-grid microgrids. But at the end of 2017, we recognized that this was an outdated approach, and we needed a new optimization layer.

    To do so, we started a four-year project with Politecnico di Milano just under the name of PROPHET Project for the design of such optimization layer in order to deal with the off-grid microgrids, but also grid-connected battery energy storages providing ancillary services, as well as in the last year, we worked about the design of vehicle-to-grid stations. From an industrial perspective, in parallel NHOA had been working to the technological transfer and industrial developments.

    This means-- this basically means identify a new EMS structure where the optimization layer proposed by the Politecnico di Milano could be mounted, as well as we try to use the EMS for building purposes, so as a kind of a planning tool. We added the new components as K-Wize that is our remote monitoring. We carried out our during the loop test with the former colleagues of ENGIE research.

    We also changed completely the hardware platform. As well as more recently, we passed from a system approach to a product approach to our EMS and also to a product targeting vehicle-to-grid.

    Nowadays, PROPHET EMS is represented by this architecture. We have five main components. The real EMS that is the optimization layer that is fed by a forecasting module, it can be forecaster of load but also PV production, both, let's say, a day-ahead forecaster but also a shorter-term forecaster. Then we have the so-called power plant controller. And two further layers represented by SAMS and local monitoring system.

    This basically deals with the safety and the data collection in order to pass the data to K-Wize, our post-processing platform. Focusing on the power plant control, [AUDIO OUT] we should understand what a power plant controller is. In particular, why is it so important?

    Well the plant controller represents the intermediate chain between a DNS that operates over minutes time scale with forecasting on off days-- it may be also longer times-- and also with the PCS, so the inverter, that operates with the switching frequency of kilohertz. So you can figure out that an intermediate layer is useful.

    This intermediate layer deploys, let's say, the automation function, as well as it allows to perform a closed loop control at the point of common capping of the plant, as well as it plays a kind of interface with respect to the DSO, and TSO, and utilities. Moreover, it's in charge sometimes for the grid code compliance, because the inverter alone may not be enough.

    Side activities with respect to these are the one related to the batteries. Indeed, we should take into account that in a battery energy storage plant, it's necessary to speak with the so-called battery management system, so let's say the management system provided by the battery supplier. But as well, the batteries-- the battery bank must be equalized or must be optimized. This kind of functionality are carried out by the power plant controller.

    Last but not least important is the fact that in any battery energy storage and vehicle-to-grid plant, there is a plant controller, while the EMS, so let's say the optimization layer, may be missing. This is normally due to the fact that some customers demanded to use their own platform, for example their own aggregator. On our end, as NHOA we can provide the full package or just a part of it.

    But how can we define a roadmap in order to achieve our power plant controller product? It's fair to say that power plant controller is something well known among renewables integrators. Nevertheless, the introduction of battery energy storage in the plants, so their hybridization, made them much more complex. Moreover, some TSOs started to demand the validation of power system model against field data. And if this validation fails, plants are not allowed to connect, and thus the whole economics of the plants are jeopardized.

    For this reason, it's important to have a real product approach. Nevertheless, as you likely know, the battery energy storage and vehicle-to-grid markets are booming. And thus, it is difficult to define real specs that last over time. At NHOA, we conceived a hybrid approach in order to create a product. In particular, we try to sketch a minimum valuable product, and then we add the features on top of that, taking into account the needs of the execution as well as the R&D product roadmap.

    It's also worth to mention that sometimes, we have to kill our darlings. This means basically that we can discontinue the maintenance of outdated features if it is not cost effective.

    Well, but how MathWorks came into the game? Well, on our hand, we have been using MathWorks toolboxes since 2017. But at the time, we were just trying to write a piece of code and test it in Simulink environment. But then the same was written again from scratch, let's say in the PLC coding environment.

    Only in 2018 we started to carry out closed loop testing, while in 2019 we made the huge step forward because we automatically compiled the Simulink Power Plant controller model into the target code of our control platform. More recently, in 2020, we integrated the Simulink project with grid, as well as in the last month we have streamlined our offering, reducing it to just to standalone battery energy storage, PV-plus-storage, and vehicle-to-grid, as well as we repeated, let's say, the compiling approach but with respect to the most common power system tools.

    In this framework, we now explain how we use MathWorks products for the design of our power plant controllers. Evidently for confidentiality reasons, we cannot disclose the underlying scheme of our power plant controller. Nevertheless, it may be simplified to the one that we see at the top of this slide, where a set of blocks that are in charge for defining the power set points, then a set of regulators possess them. And at the end, a final set of blocks that are in charge for dispatch in the set points among the underlying inverters.

    All these blocks, functional blocks, are contained in our own library, Simulink library, the one that you see at the center of this slide. We have two further repositories also. One is used for containing the final solution, so let's say for keeping track of the final project that will be deployed, as well as the scripts used for the automatic testing. And the third folder contains the script for automatic compiling the Simulink blocks into the code that runs behind our control hardware platform.

    At NHOA, we have been developing our power plant controllers considering the so-called V model, a methodology very common nowadays among engineers that deploy model-based design. Indeed, we start from this packet in order to shape the control scheme, the functionalities of the power plant controller. Then we carry out unit testing. We passed them to system testing in the simulation. Then we pass to system testing without running the loop. And then we go on field. In this way, we reduce the likelihood of bugs in our code.

    But as I mentioned a few slides ago, at NHOA in the last month we have been trying to replicate the model obtained in this kind of type of the model that you see in this slide that allows to obtain the so-called model-to-x. Indeed, starting from the same Simulink project that we used for obtaining the final power plant controller, we are capable to obtain the same code in the most common power system tools as PSCAD and exceeding power factory that are normally used for carrying out grid integration studies.

    Now let's have a look to the fields results collected from our battery energy storage platform. I'm going to start from Lifou, New Caledonia, 5-MVA, 5 megawatt-hour battery energy storage at the service of the local off-grid microgrid. In this case, as I mentioned a few slides ago, for the first time we deployed the automatic conversion of the Simulink code to the control hardware platform thanks to the PLC toolbox.

    Moreover, in 2019, for the first time a GPS deployed its own preforming inverters, in particular all the time operating in this way. This is extremely relevant, because it means that either we are connected to a synchronous generator or we are the grid forming unit in the system, we continue to operate the inverter in the same way. This allows to perform a seamless transition, as we see in the results of this slide.

    On the left in particular, you can see that we can achieve 100% renewable penetration around midday. In particular, we are capable to switch off the genset, while in the results on the right, you see that we obtain small voltage and frequency perturbation whenever we switch off and switch on the genset, taking into account that this is a kind of pioneer plant, because in 2019 there were only a couple of companies that were proposing grid-forming inverters in parallel for the operation of microgrid.

    Still related to these functionalities we see in this slide that the frequency profile improves significantly in presence of the battery energy storage, while without it we see the characteristic oscillation on/off of these generators around a nominal frequency, while on the right we see the positive effects of a battery energy storage whenever there is a load change. In particular in the top plot, we see that the frequency profile, the frequency recovers slowly and with some oscillation with just the genset, while whenever we have also battery energy storage in parallel, basically there are no differences across the load change.

    The second plant I'm going to tell you about is Leini in Italy, a 7-MVA 5 megawatt-hour BESS combined with a pre-existing combined-cycle power plant. In this case, the goal of the battery energy storage is optimizing the efficiency of the gas turbine. But on the other hand, we have to integrate, interface our power plant control with the grid system, the DCS, that was already owning the operation of the power plant.

    The interface process has been extremely successful. And this is confirmed by the oscilloscope snapshot that we see in this slide that shows, basically, the step response from a signal sent out by the DCS respect to the response of the inverter. The dynamic is shorter than 100 milliseconds, confirming that NHOA battery energy storage are extremely fast. And in particular, they will be capable to participate to the Fast Reserve market launched by the Italian transmission system operator, Terna, for next year.

    The third planet I show you the results is Sol de Insurgentes, a 27-megawatt solar plant combined with a 5-MVA almost 3 megawatt-hour battery energy storage with the same point of common capping. This plant is extremely relevant, because it is connected to the Baja California power system that is a kind of large off-grid microgrid, because even if it is managed by the Mexican TSO, CENACE, it is not interconnected with a mainland grid nor interconnected with the US-- America and California, let's say.

    And for this reason, it makes it extremely sensitive to load and generation perturbation. For these reasons, CENACE set very strict requirements in terms of ancillary services. As we can see here, for example, they demanded the ramp rate limitation of the PV power plant in line with the prerequisites of the Puerto Rico Grid Code. They demand primary frequency regulation with the kind of dynamic depend. They demand the plant to obey to external power limitation, so let's say to set a kind of ceiling defined by TSO, as well as the plant may provide automatic generation controls, so let's say secondary frequency regulation.

    Now I'm going to comment these services, looking at this plot. In particular on the left, we see that varying with step changes of the frequency, it's possible to have step changes of the output power. This is a test carried out-- we're going in site acceptance test witnessed by the TSO.

    We have also the results of AGC. With this reference. It's important to point that you must satisfy AGC. Either you are in excess of PV production or in deficit of PV production. So let's say that the battery energy storage should complement the production of the PV in order to satisfy the request of the TSO.

    And then the last slide that I'm going to show you today, we have the results of PV smoothing. On the left, we see a very significant clouding event. In particular, we had a fast, thick cloud passing above the PV plant. This determined a huge drop of the PV production, but at least we had a battery energy storage that tried to compensate the deficit within its power limitation. In this case, the BESS was downrated with respect to the PV.

    On the right, we see, let's say, a traditional operation, where PV production is affected by scattered clouds passing above the panels, but the output power of the plant at the point of interconnection represented by the red line is smooth. And in particular, it complies with this new rate demanded by the TSO. Now I give the floor again to Guido.

    Thank you, Pietro. It is now time to talk about V2G. When we first approached V2G, we had the idea to replicate the structure of battery energy storage systems in which we are experts. This way, we planned a bidirectional charging infrastructure with a centralized AC-to-DC conversion system that uses cars as stationary batteries. In some projects, we also included an actual storage system to improve flexibility.

    Our goal with this project is to reduce total cost of ownership and extract new value out of the vehicle without impacting its life and on vehicle availability for the final owner of the car, all this while offering a large variety of grid services. This application is particularly interesting in case of what we call fleets. And the idea is to interact directly with a car maker or with companies that make use of a large number of EVs, for example logistic companies, to exploit the period in which the EVs are parked without being used.

    Our first two plants are located in the parking lot of FCA in Turin. The first plant, which we can see in the picture, is a 2-megawatt pilot project with 64 EVSCs for a total of 2 megawatts, 2.6 megawatt-hours. It is tailored for FIAT 500e, and it is equipped with a single-stage AC-to-DC conversion system formed by four of our 500 kilowatt-ampere inverters. Each inverter DC bus is connected to 16 vehicles, which batteries are connected in parallel.

    The second plant will also be located in Turin, but it will be much larger. It will include up to 560 EVSCs and then integrated stationary battery for a total of 28 megawatts, 28 megawatt-hours. It is equipped with a double-stage power conversion system formed by 14 of our 2 megawatt-amperes inverters and the DC/DCs of the EVSCs. These DC/DCs will allow us to operate with different kind of vehicles from small city cars, like the 500, to larger utility trucks, like the Fiat Ducato, which has a higher battery voltage.

    Let's talk about the first phase, the pilot project. As we mentioned earlier, the pilot project's peculiar architecture implies that the car batteries are connected in direct parallel with one another. This means that when a new car arrives in the plant, its SOC has to be brought to the same level of their sister cars connected to the same cluster to the same DC bus. Moreover, as per request by FCA supply chain, when a car is disconnected to be shipped to the final customer, it has to be brought to 35% SOC, which is the level that minimizes aging.

    How to do so? How do we do so? We make use of a fifth PCS, 70 kilowatt-amperes in power, to charge or discharge each car to the desired SOC. Each car is connected to either the main PCS or to the smaller auxiliary PCS with a double bus part and conductor systems, which you can see in the picture.

    Given its power, the auxiliary PCS is able to charge or discharge maximum two cars at a time. So we had to define a queuing method to ensure we meet the disconnection and connection request by FCA.

    How does it work? Car request times are collected from FCA with a web service. Given current SOC and target SOC, processing duration is computed for each car and updated every 100 milliseconds. We represent it in a sort of Gantt chart, which you can see in the picture.

    If two or more processing periods are overlapping, a queue must be form processing one car at a time. This is what we call Method X1. Intervals can only be moved to an earlier time to meet the request. With this method, we built a first plant, which is the one you can see in the picture. If, like in this case, start time of the first period is negative, this means that we are too late, that we should have started processing cars in the past.

    So another method must be used to solve the problem. This method is Method X2. With Method X2, we make a new queue with two processes at a time, completely overlapping the intervals that are closest to one another. Processing two cars at a time is slightly slower than processing just one, due to PCS power limitation. The equivalency rate is around 0.8.

    If start time of the first period is negative, as for X1, Method X3 must be used. Method X3 implies that all cars are prepared at the same time, making use of the main PCS to charge or discharge them. At the end of the charge or discharge process, only cars that were actually requested by FCA are disconnected. The other remain connected to the system, and they are available for grid services.

    In the second phase of the project, the 28-megawatt project, we remove many of the main constraints of the problem. In fact, having the DC/DC converter inside EVSCs, it is not necessary to connect batteries in parallel anymore. And so the SOC must not be leveled among the different cars. This also means that cars can enter or leave the system independently from one another.

    The remaining constraints are lighter. We want to balance SOC between cars in the long run to optimize plant capability. And we still have to bring a car back to 35% when it is required by FCA. We plan on doing so, keeping at a minimum the exchange of energy between cars to minimize losses. This means the power set point of the cars must have the same sign. We are developing a sorting algorithm to reach these two goals.

    In the picture, you can see a simulation for four clusters. Each line is the SOC of a car. And at the bottom, you can see a typical profile for grid services. You can observe that in the end, it is possible to align SOC of a car in a cluster. And a similar logic would be used for car delivery.

    Last but not least, these two plants will provide the latest frequency regulation service defined by the Italian TSO Terna, the so-called Fast Reserve. We developed a logic block and an energy model which, according to the regulation defined by Terna, provides a power set point for the plant reacting to grid frequency variation and applies an SOC management system to ensure the availability of the plant.

    In the lower part of the picture, you can see a test frequency profile we gave as input to the model with the two main thresholds defined by Terna. In the upper part, we reported the resulting power profile divided into contributions for reacting to frequency variation, SOC management, and deramp, as well as plant SOC trajectory.

    All of our developments make large use of MATLAB and Simulink tools, as you could see. We thank you for your attention. And feel free to contact us for any questions you might have.

    Hello, everyone. Welcome to this MathWorks Energy Speaker Series. And I have a video and a presentation. And the title is "Deep Reinforcement Learning-Based Controller for a State-of-Charge Management of Multi Energy Storage System." And there are two presenters here, Francisco Sanchez from Loughborough University and myself, Francisco Gonzalez, from University of South Eastern. Norway.

    Francisco Sanchez received his bachelor's degree in electrical engineering from University Simon Bolivar in Caracas, Venezuela in 2011, master's degree in renewable energy technology from Polytechnic University in Madrid, Spain in 2013. And he recently completed a PhD degree in electrical engineering at Loughborough University in the United Kingdom. His research interest is mainly artificial intelligence techniques for power system analysis and energy management application.

    And myself, I am Francisco Gonzalez-Longatt. I am currently a full professor in electrical power engineering at the University of South Eastern Norway. Well, the agenda is very, very short. We will start with a short background about the frequency control in energy systems. Then we will discuss some basics about reinforced learning to go into the deep reinforcement learning-based state-of-charge controller that we present here, and then a couple of study cases, and finally a summary.

    As you must be aware, several economies around the world, they are in a very fast transition to reducing the CO2 emissions. And the GB system is one of them. However, one of the challenges when we are in the transition to reducing the CO2 emission is that the integration of the renewable energies, they cause a reduction in the physical rotational inertia.

    And as a consequence, the frequency control inside the power system start to be challenged. The frequency is the main indicator that define the balance between generation and demand. And when the rotational inertia is reduced, then the frequency starts to change faster and deeper.

    One of the possible solutions to cope with the reduced inertia is using what we call battery energy storage, or even combining several energy storage technologies to cope with the variability and the imbalance between generation and demand. However, one of the main challenges related with having energy storage inside the power system is to control the state of charge of those energy storage devices.

    And in the GB case, National Grid ESO has created several services in order to keep the balance between the generation and demand. One of them is called enhanced frequency response. And the enhanced frequency response is dedicated to provide or absorb active power from the power system there is a power imbalance. In this case, we are dealing here with several energy storage technologies, in fact ultra capacitor, flywheel, and battery energy storage, in order to compensate the frequency fluctuation by injecting or absorbing active power.

    However, the main task here is that we want to develop a controller that deliver these enhanced frequency response but at the same time keep a control, keep a track about the state of charge, in order to make the asset available during the long term. Well, now I will proceed to allow Francisco Sanchez to continue with the presentation.

    Thank you, professor, for this nice introduction. In this section, I will review some basic concepts related to the growing field of reinforcement learning. These concepts are necessary to fully understand the proposed state-of-charge controller, which will be examined in later sections. After that, I will discuss the most important aspects of the various case studies, analyze, and finally provide some concluding remarks.

    Reinforcement learning is a category of machine learning that is concerned with finding the best sequence of actions to generate the optimal outcome. In the reinforcement learning concept, a computerized agent learns to take actions at discrete time steps to maximize a numerical reward from the environment. If an action selected by the agent in a given state gives a low reward, then in a subsequent instance of that state the agent may choose an action that yields a higher reward.

    In this framework, as shown in the figure, there is an entity that learns and takes decisions, i.e. the agent, and an environment which comprises everything outside the agent. Generally, the actions performed not only influence the immediate reward signal but also impact all subsequent rewards. The process gets underway when the agent receives an observation containing the state of the environment, and based on that, it performs an action which prompts the state of the environment to change.

    Besides the state, the environment also provides the agent with feedback associated with the action made in the form of a reward signal. The set of rules that the agent follows in mapping the states to a probability distribution over the actions is termed the policy. The goal of reinforcement learning is to provide the agent with a policy that maximizes its total future rewards.

    At this point, it is important to stress the distinction between the reward given from the environment as a consequence of an action and the value of a given state, which refers to the total expected reward that the agent can get from that state. The value function includes the effects of all subsequent rewards that the agent may accrue starting from any given state. The goal of the agent is to maximize its value, which is shown on the equation below.

    Here, Th is the time horizon, and gamma is the discount factor for future awards which is bounded between 0 and 1. The discount factor defines the degree to which the reward is delayed in time. That is, discount factors closer to 0 give more weight to short-term rewards, whereas values closer to one place more value on long-term rewards.

    Depending on the size of the observation and action spaces, the value function can be modeled with different techniques. Systems with action spaces that are discrete and with a few observations can store the value functions in a tabular representation. However, for larger systems, particularly those with continuous action and observation spaces, such as the multi-electrical energy storage system considered in this talk, it becomes impractical to represent the value functions using a tabular scheme. Deep neural networks, typically used as universal function approximators, are ideally suited to estimate the value function, forming what is known as Deep Reinforcement Learning, or DRL.

    When using Deep Neural Networks, or DNNs, to represent the value function, there are three main approaches to arrive at an optimal policy. These DRL training structures are known as policy-based methods, value-based methods, and actor-critic methods. In policy-based methods, a DNN is set up called the actor network that maps the current state of the environment to the best action that the agent can take to maximize the total expected reward. These methods arrive directly at an optimized policy by differentiating the expectation of the value function with respect to the neural network parameters, and changing them in such a way that this quantity is increased. Optimization techniques are used to find the set of actor network parameters that maximizes the expected reward.

    Since policy-based methods use the gradient of the expected reward function to find an optimal policy, they might converge to a local instead of a global maximum. And unlike value-based methods, they can only learn from their own experience.

    Value-based methods, on the other hand, do not deal directly with the gradient of the expected reward. Instead, they use another network termed the critic network to approximate the Q-function and infer an optimal policy from the state-action pair. Once the Q-function is accurately estimated, an optimal policy would consist of performing the action which maximizes the Q-function at each state.

    Because they use another DNN to arrive at its policy, these methods are also known as off-policy. They are powerful, because they can learn by looking at others' actions while trying to achieve any goal. However, they are constrained to discrete action spaces, as the agent must explore each action to find the one with the highest value, which is computationally impractical in most cases of interest.

    A blend of both policy- and value-based methods is known as the actor-critic approach. It uses one DNN, the actor, to predict the best action given the current state, and another DNN, the critic, to estimate the Q-function for the state-action pair taken by the actor. First, the actor outputs an action based on the state as in policy-based methods. Then the critic outputs an estimate Q-value for that state-action pair and uses the reward signal from the environment to adjust the accuracy of the estimate by changing its set of parameters. Finally, the actor network updates its set of parameters based on the critic's response to improve the actions taken on each state.

    Since the critic only needs to output an estimate Q-value for the action taken by the actor, this method can work on continuous action spaces. However, one downside of actor-critic methods is that the Q-function is overestimated, which may lead to slow convergence.

    This section presents the structure of the proposed DRL-based state-of-charge controller for the multi-electrical energy storage system providing EFR to the power system. The environment is represented by a Simulink model with various subsystems, including the EFR controller of each energy storage asset, which sets the power reference to inject or to absorb it each time step, and the assets dynamic model. Based on these signals, the environment outputs an observations vector and the scale of reward.

    In the following sections, each of the components of the environment is explained in detail. The frequency controller implemented for this talk is based on the EFR service proposed by the ESO GB. EFR makes up a dynamic service in which the active power varies in proportion to changes in the system frequency. This service was explicitly designed to be provided by energy storage assets, as it includes a zone solely for state-of-charge management purposes.

    Furthermore, its small dead band allows it to improve the frequency regulation capacities or prefault state of a system. As shown in the figure, the EFR service comprises three distinct regions. For frequencies at or below the under frequency threshold, the asset should follow a unique power reference proportional to the frequency and cannot manage its state of charge. Conversely, for frequencies at or above the over-frequency threshold, the asset should also follow a unique power reference in proportion to the frequency.

    For all other frequency values, the asset can manage its state of charge by changing the control reference parameter rho. The output power reference of the asset in terms of rho can be written as shown below. When rho equals 1, the asset operates in its upper envelope, and thus it is primed to inject more power into the grid, which reduces its state of charge. On the other hand, when rho equals minus 1, the asset operates in its lower envelope. Therefore, it is primed to absorb more power from the grid, which leads to an increase of its energy level.

    One of the metrics defined by the ESO to assess the quality in the provision of EFR and incentivize good performance is termed the Service Performance Measure, or SPM. This metric assumes that a perfect delivery of the service is unrealistic. Therefore, if the asset operates within the specified envelopes, an SPM of 1 is granted. Otherwise, it is calculated as per the expression shown below, where P is the actual normalized response from the storage system. When the asset operates outside of the specified envelope, it is penalized with a lower SPM, which results in a proportional reduction in payment.

    A mix of three electrical energy storage systems of different technologies has been considered for this talk. That is, a battery, an ultracapacitor, and a flywheel energy storage system. The dynamic models used to represent each energy storage asset take as an input the power reference set by the EFR controller and outputs the actual power response from the asset and its state of charge. A lithium-ion generic model with similar charging-discharging dynamics has been used to represent the battery storage system.

    The open circuit voltage is modeled as a function of the battery current in its state of charge. In the flywheel energy storage system dynamic model, the state of charge is proportional to the square of the rotating speed and the inertia moment, while torque and rotational speed limits are considered by including a saturation block. The ultracapacitor model used resembles the battery storage system. However, the charge-discharge dynamic block is replaced by the Stern equation to represent the fast charge-discharge capabilities achieved with this technology.

    The components of the observation vector of an energy storage asset at time step t is shown in the figure. The observations fed to the DRL agent at time step t only include the actual normalized response from the electrical energy storage system and its state of charge at that time step, but also comprise the charging reference parameter rho sampled at the previous time step. Besides the current value of the state-of-charge mismatch, the observations also include its previous case samples in the accumulated value up to time step d, estimated numerically using the forward Euler method.

    The reward structure proposed in this talk develops a DRL agent that minimizes the state-of-charge error and control effort while keeping the SPM as high as possible. By penalizing large values of rho, the agent is deterred from behaving as a discrete on/off switch. At the same time, the SPM element ensures that the asset manager is not penalized economically by the ESO from over- or under-producing.

    Specifically, the instantaneous reward scheme proposed contains a continuous and a discrete component. Both components are combined at each time step to grant the agent the mixed reward signals shown in the slide. The discrete component varies discontinuous with changes in the observations from the environment. And it is used to steer the agent away from undesirable states, such as states with the state of charge below the minimum or above the maximum values recommended by the asset manufacturer.

    The term C delta SOC is the component of the discrete reward because of the state-of-charge error, and CB is a penalty awarded when the state of charge of any storage asset reaches its minimum or maximum allowable values. When the state of charge of any asset is within a predetermined tolerance band, epsilon, a positive reward equal to K delta SOC, is given to the agent. Conversely, an asset with a state-of-charge mismatch beyond the epsilon band yields no reward for the agent.

    The final element of the discrete reward corresponds to a penalty for operation outside its allowable state of charge range. When the state of charge of the asset is outside the safe operation range, a large penalty, KB, is applied to the agent. The continuous component, on the other hand, improves convergence during training as the agent receives constant feedback on its actions, which translates into simpler network structures to represent the value and action-value functions.

    This component is based on the quadratic regulator function to drive the state-of-charge mismatch to 0 with minimal control effort. The first term penalizes the state of charge of each asset, whereas the remaining two terms penalize, respectively, the control action and SPM of the storage assets. The term k delta SOC is a weight term for state-of-charge mismatch. k SPM is a weight term for the deviation between the target SPM and its actual value. And k rho is a weight term for the previous control action.

    In this formulation, all the weight terms are positive. To further visualize the reward structure, it is useful to combine the continuous and discrete reward components related to the state of charge into an aggregate state of charge reward function, r SOC, shown on the right-hand side of the screen. It highlights that a penalty proportional to the square of the state-of-charge mismatch steers the agent towards the asset's reference state of charge, while rewarding actions leading to values of state-of-charge mismatch smaller than a predefined threshold.

    The DRL algorithm used to train the agent corresponds to the actor-critic category. As discussed previously, in this kind of learning framework two independent neural networks are required, that is, one to represent the actor that chooses the action based on the state, and another to represent the critic that estimates the expected long-term reward from that state-action pair. The critic network comprises two paths, that is, the state and action paths, and is depicted on the right-hand side of the slide.

    The dimensions of the input layer for the state and action paths correspond to the dimensions of the state and action vectors. The state path of the critic network comprises two fully connected layers, critics state FC1 and FC2, whereas the action path comprises a single fully connected layer, critic action FC1. An addition layer sums the outputs from the two paths element-wise, forming a common path that contains a fully connected layer, critic hidden output and the final output layer, which yields the expected Q-value.

    The actor network, shown on the left-hand side of the slide, has a more straightforward structure, since it is only concerned with the action to take given the observations. It comprises three fully connected layers, actor FC1, FC2, and FC3, and an output layer with a hyperbolic tangent activation function to bound the actions. The hidden layers in both networks were activated with rectified linear units ReLu.

    The agents were trained using the MATLAB implementation of the DDPG algorithm. Each training episode simulated the operation of the multi-electrical energy storage system during two hours, long enough to observe the effects of the agent's actions while short enough to allow sufficient training episodes in a reasonable time given the available hardware. During training, the initial and target state of charge of each asset was randomized to ensure that on each episode, the agent was exposed to a different set of initial conditions.

    Since large grids typically show a symmetric frequency characteristic around the nominal value, the target state of charge for each asset was drawn from a uniform distribution between 40% and 60% and kept constant during each training episode. Initially, the agent was trained with the grid frequency fixed at 50 hertz to arrive at a sufficiently optimal policy. The only variation in the environment between separate training episodes arose from the different sets of initial and target state-of-charge values.

    The agent was trained for 5,000 episodes. And the performance curve is shown in the figure, alongside the normalized reward from agents trained with other popular learning algorithms, namely Q-learning and DDQN. The DDPG agent improves its policy continuously during the first 1,000 episodes, and then it stabilizes around a null reward. In contrast, the DDQN achieves a consistently lower average reward, and the Q-learning algorithm is unable to learn a sufficiently optimal policy.

    In this section, the proposed DRL-based state-of-charge controller is compared with two conventional control techniques, namely PID and fuzzy logic. Moreover, the two case studies devised to test the performance of the DRL agent are explained. The first case study corresponds to the multi-electrical energy storage system under normal operating conditions, that is, slight frequency deviations around a nominal value.

    The signal used to test the proposed controller corresponds to the frequency of Great Britain during the first seven days of June 2019 with 1-second resolution and unseen during the training phase. The maximum and minimum values of the testing sample are 50.25 hertz and 49.75 hertz respectively, ensuring that the multi-electrical energy storage system is inside the state-of-charge management region of the EFR service.

    The figures on the left show the evolution of the state of charge during the test period and under three different controllers. Though the controllers can manage the energy level inside the permissible range, the proposed DRL controller follows more closely the target state of charge of the assets. Hence, it leaves some capacity available to operate it securely in case of a sudden period of low or high frequency.

    Such a period of high frequency occurred during the early hours of June 5, 2019. During this period, the EFR controller signaled the energy storage assets to absorb power and bring down the frequency, thus raising their energy levels. The state of charge of the assets nearly reached 80% when managed by the PID in the FLC, whereas it capped at 56% with the proposed DRL controller.

    The second case study investigates the performance of the DRL agent after a significant system frequency disturbance, and thus with the multi-electrical energy storage system operating in the post-fault low-frequency region of the EFR service. In this case, the input frequency signal is set to that recorded in the GB power system during the under-frequency event of August 9, 2019. Outside the frequency management region, there is a single reference for the storage asset to provide support to the grid. Therefore, post-fault behavior is similar among the different controllers, and only the DRL agents response is shown.

    The input frequency signal is presented above, whereas the response of the multi-electrical energy storage system is illustrated below. While the frequency remains inside the state-of-charge management region, the output of the energy storage assets is regulated by the DRL controller to keep the state of charge within its predefined range and depends on the initial state of charge of the assets. However, when the frequency drops below the under-frequency threshold, in this case 49.75 hertz, the multi-electrical energy storage system is directed to inject power to the grid at its nominal rate. After the frequency is restored to within the state-of-charge management region, the DRL agent adjusts the output of the multi-electrical energy storage system according to the control policy developed.

    The control action from the DRL agent quickly reduces the state-of-charge error from its initial value and then keeps it tightly within a narrow band related to the reward tolerance band parameter. Both the PID and FLC attain a broader control of the state-of-charge mismatch, within plus or minus 0.3 PU, whereas the proposed DRL controller achieves a narrower control range, around plus or minus 0.05 PU, with over 95% of the time above 0.02 PU as depicted in the above histograms.

    The box plots presented below present a comparison between the different state-of-charge control strategies among the multi-electrical energy storage system. It is evidenced that the proposed DRL strategy keeps the state of charge closer to its target value than the other controllers. Both the PID and FLC are centered around zero state-of-charge mismatch for each asset. However, they have broader distributions when compared to the DRL controller.

    The next group of slides present the sensitivity of the reward function to changes in the environment in agent settings during training. A proper selection of the learning algorithm's hyperparameters, though a time-consuming activity, is crucial for the learning process and the agent's performance. The first environment parameter considered is the input frequency signal to the EFR controller.

    When trained with a fixed input frequency signal, as described in the preceding slides, the normalized average reward after 2,000 training episodes is slightly lower than 1. In contrast, when a more realistic frequency signal is used, similar to the frequency that the controller would experience in actual operation, the reward drops down almost to minus 3. In this case, the DRL agent is exposed to a more adverse yet more realistic training environment which impacts its learning development. The choice of training signal would depend on several factors, for instance, availability of the agent to continue training while deployed in the field.

    The second environment parameter considered is the number of past delta SOC observations, k, also known as the control memory. Low values of k yield acceptable results and require less training episodes to reach learning stability. In contrast, when the agent receives signals from several time steps in the past, it's not able to learn a suitable control policy. In this case, the complexity of the actor and critic networks may be increased by, for instance, adding more layers or more neurons per layer so that the agent can achieve learning stability.

    This figure shows the normalized agent reward for three different discount factors. Here, a large discount factor of gamma equals 0.99 achieved superior learning performance. A slightly lower discount factor of gamma equals 0.95 achieved similar performance during the first 3,000 episodes. However, it cannot stabilize as the previous case. An even lower discount factor of gamma equals 0.7 makes the agent excessively focused on short-term rewards, and thus not able to reach a stable control policy.

    This figure highlights the impact of the batch size on the agent's reward. Though the values assessed yields similar average rewards, the scenario with a batch size of 512 can learn with a reduced number of episodes, around 300. In contrast, smaller batch sizes increase the number of episodes required to reach an acceptable policy. Larger batch sizes require longer training episodes, as the agent updates its policy using more samples. However, the time required to achieve a specified performance threshold is reduced.

    The last hyperparameter explored in this talk is the agent's learning rate. This critical parameter regulates the magnitude of the deep neural network changes during training. For DDPG, in its original paper its developers recommend that the critic's learning rate should be larger than the actor's learning rate. This figure shows the performance of the agent for different combinations of actor and critic learning rates.

    It is evidenced that large values impact the learning performance negatively, whereas small values lead to long training time, that is, more training episodes required to achieve a specific average reward. In this case, actor and critic learning rates of 10 to the minus 3 and 10 to the minus 2 respectively appear to produce a superior training performance.

    This talk presented a controller based on deep reinforcement learning to manage the state of charge of a multi-electrical energy storage system, providing enhanced frequency response to the grid. The controller, designed with MATLAB and Simulink, employs a continuous and model-free DRL algorithm termed deep deterministic policy gradients to manage each assets energy level within a pre-specified tolerance band. It was evidenced that a proper state-of-charge management is essential to keep energy storage systems available when required, thus allowing the asset manager to fulfill its frequency service commitments in a timely and secure manner.

    As the proposed controller bases its actions solely on the measured state of charge sampled at different time steps, it is relatively immune to environment perturbations, such as frequency measurement noise. Artificial intelligence techniques such as reinforcement learning show potential in power system control tasks and are certain to become more prominent in the coming years.

    Thank you all for your attention. For any suggestions or questions, please do not hesitate to reach out. Our contact details can be found on the links shown below.

    Hello, everyone. In this presentation, we will illustrate the Enel Green Power Project for diagnostics and advanced control of geothermal production plants. My name is Marco Magrini. I work at Enel Green Power in the specialist technical support team. And together with Matteo Galgani, we will be the speakers of this presentation.

    Enel Green Power is a utility that produces electricity from renewable sources. We work specifically in operation and maintenance of geothermal Italy. In this presentation, we will focus on geothermal production plants. And now, we speak about the location. You will probably know Tuscany for a lot of things, for example for its historic city such as Florence, for the beautiful landscape, and for sure for the good wine.

    In fact, Tuscany is also famous throughout the world for being the first place where, over 100 years ago, electricity was produced with the use of geothermal energy. The geothermal cycle is very simple. Through drilling, the steam is sought underground. This steam is directed to the production plants through a complex distribution network.

    Once it arrives on the plants, it is used to produce electricity. The steam at this point is transformed into water, which in turn is directed into the injection wells through a complex network of aqueducts. The injection wells inject cold water into the subsoil. And thus, the cycle closes.

    All these various stages of production are monitored and controlled through a digital architecture and a very complex communication network. Currently, the installed power is over 760 megawatts with a fleet of 37 production units. All systems are automated and autonomous. For the monitoring, there is an automation center, 24 hours a day, seven days a week.

    It is anticipated all the production phases are automated and the monitoring takes a complex communication network and control and data collection devices. A long and complex optical fiber network allows the acquisition of over 50,000 sensors installed in the field. This data is then collected in a large central database and used for monitoring and analysis and diagnosis purposes. This project will develop a new advanced diagnosis and data analysis tools.

    Historically, we have always used the data collected by the plants for analysis and monitoring. In recent times, we did event about fast and distributed processing technologies and advanced analysis and diagnostic algorithms. We have found a limit in our traditional platforms. This did not allow us to develop a complex data analysis algorithms. The biggest challenge was to find a flexible, powerful, and easy-to-use product on the market which was also fully integrated into our traditional system and data acquisition and management platforms.

    So the solution was MATLAB. Specifically, the introduction of the MATLAB Production Server and the PLC Coder. These are the tools, combined with the intrinsic power of MATLAB, allowed has a great flexibility in the development of complex algorithm and solution, together with the complete integration with our traditional system.

    The calculation architecture introduced is simple but with some particularities. Traditionally, our systems are monitored by specialized technicians. The introduction of specialized personnel in the creation of complex algorithms was important to incentivize the decision to develop the project. The new architecture introduced two parts-- the central computing server where the diagnostic algorithms are developed and next executed in the devices on the edge, where the control algorithms are released.

    The central computing ecosystem is composed of three parts that communicate-- the server where the data is collected, OSIsoft, machine for developing algorithms, and a machine where the developed algorithms are executed in real time. All this is possible using the MATLAB Production Server. In fact, the developed algorithms are released on the production server, which takes care of executing them in real time by collecting the data and writing the results directly into the central database.

    The system is very flexible and scalable. Currently developed algorithms are listed below-- monitoring of the steam transport network developed through the use of fluid dynamics and optimization libraries; the prediction of signal in the future developed through the use of prediction techniques based on stochastic signal TPL of the economy; the diagnostic of plant components and processes developed with machine learning techniques and identification of dynamic systems; predictive maintenance of auxiliary components with the use of breakdown event identification algorithms. As we can see from the representation, the advantages are many. Avoiding accidental breakages and optimizing production are the main ones.

    The plant control ecosystem is also made up of three parts-- control algorithm development server, the diagnostic PLC where the algorithms are released and execute, and the main plan PLC that is in charge of controlling the plant directly through the actuators. This is possible using the Simulink PLC Coder. The main PLC communicates with the diagnostic PLC by sending the data collected by the plant real time and acquiring the results of the algorithms performed. This system, too, is very flexible and scalable.

    Currently developed algorithms are listed below-- compressor control to avoid pumping. This series of algorithms are developed with adaptive control techniques and recognition of events typical of pumping conditions. Digital twin and physical modeling of the compressor using the same Simscape tool for physical modeling of systems. As we can see from the presentation, the advantages are many, avoid system blockages due to the pumping condition of compressor the main one.

    Here, the list of project key results specific to geothermal technology-- flexible steam control and monitoring in networks; increased ancillary monitoring; new diagnostic and control algorithms. Here, the lists of the horizontal key results of the project that can also be used in other technologies-- MATLAB as an algorithm development ecosystem that can fully integrate into the typical workflow of data scientists and universities; range of innovative algorithms developed; PLC diagnostic as an add-on to the main control system for the execution of complex algorithms.

    In this slide, we show how the algorithm development processes will be managed in the future, divided between the horizontal and vertical ones. The horizontal algorithms are present in the literature. The vertical ones will be developed by our technicians and the engineers. The data scientists will deal with translating this knowledge into programs that can be run on MATLAB. The advantages will be greater digitalization of systems, an increase in efficiency and optimization of the geothermal resources.

    The project lasted about two years in its development and implementation phase. Now other algorithms are being developed. However, not everything was linear in the development phase.

    The lessons learned during the development of the project. Having a lot of historicized data is of great value and enables advanced analysis for process optimization. Structuring data to make it usable by algorithms, however, requires time and resources. The development of the algorithms must be done in cooperation between data scientists and process technicians. Since the geothermal market is very niche, the literature is sparse of developed techniques. We often have to develop them from zero.

    Using complex algorithms that have dozens of inputs requires a robust and error-proof data acquisition and validity chain. Otherwise, you risk having more disadvantages than advantages. Thanks, everyone, for your attention.

    View more related videos