Machine Learning for Solar Energy Forecasting || Mr. Eroshenko Stanislav Andreevich || ITC 2020

the acronym reductant [Music] [Music] you bingo all I am a man occur from team i triple e BB MSP and I welcome you back to the function of the International tech Conclave ITC before going ahead participants kindly keep a note certificates will be provided only those who attend all the sessions of fight is I request to like the videos and also subscribe to the channel the same time if you have any queries or the speaker you can use the chat sections and our moderators will forward it to the speaker today for the upcoming session we have with us mr. Stanzler Electrico arashan cosa would be talking about machine learning for solar energy forecasting now I would like to call dr. McKay shrimp itself associate dean in hiring D at EVM Engineering College take the proceedings further good afternoon doctor Tempe currently choristers at villa whisper karma Mahakali welcome all the participants to the international specialist Conclave 0 to Z now I welcome our speakers we serve any club hero sinker of our today's session on machine learning for solar energy forecasting he's a senior lecturers leading and inner in Europe our engineering Institute visual Federal University it specialization reason power engineering is having 10 years experience in teaching is having scientific community membership in Russia and involved in many R&D projects he has published more than 25 papers is repeated high index journal like web of science purpose and it's just he has published four books and participate in more than 30 international conferences he is a leader and developer of the modules power and electrical engineering within the framework of federal university educational program we welcome you sir now I thought it was special to imagine I request that annulus please take over the further sessions I requested to unmute is my thank you sir yeah yeah thank you very much so I hope it's okay with video and audio connection it's my great pleasure to introduce myself and the university that I represent I am from Ural Federal University and my today's topic would be devoted to machine learning in solar energy forecasting well actually it's a project of my life one of the projects of my life and we have a very powerful research team and our University and actually there are lots of collaborations in terms of this project with other universities from all over the world but the main ones are from India from China and from France well let me firstly speak very briefly about the concept of today's energy sector well according to the let's call its public events like secret session which was held in 2018 the power industry and the energy sector today is becoming more oriented towards digital technologies and towards the final customer yes and the final customer the load becomes to play a more and more increasing role in the life of for the energy sector and today's energy sector is represented by such let's call it novel phenomena like overall distributed generation integration development of renewable energy sources or cybersecurity issues and big data applications because there are lots of monitoring and control systems nowadays that are producing huge amounts of data which can be used in a proper practical way and produce very good results the Internet of Things are the Internet of energy it's actually one of the greatest means oriented toward the final customer and there are lots of noble services which appear in the energy markets and give the customers more opportunities in terms of the load regulation and their profits and benefits of course a very big share of publications and a very big share of topics in the research community are devoted to smart grids it's a highly integrated grids with information technologists which provide flexibility reliability and security of the final energy supply well my today's topic is not so devoted to renewable energy sources development and well maybe a couple of years ago maybe four or five years ago it was not so urgent information because we did not have too much renewable energy because of the climate you know the climate is quite severe and according to the statistics of the recent years so you can see it on the slide starting from 2016 the installed capacity of photovoltaic power plants is growing rapidly and actually there are some strategic development plans and documents which provide us with the strategic goals of the share of renewable energy sources to be provided in the final balance of the fields for the country and in terms of renewable energy sources let's call it we some hydropower a small hydropower the total share should be for approximately five percent by 2024 yes and now we have the total amount of 1439 megawatts of photovoltaic power plants which are operated and integrated in the United power system of Russia well actually if taking the big the total installed capacity of the generation in our country the total chair is not so impressive but anyway taking into account the climatic and weather specifics of the country all of these generation capacities are concentrated in southern regions of our country and some of the regions are let's call it overpopulated with PV and for this reason this produces some power system operation problems so I will talk briefly about some power system impacts which are introduced by photovoltaic power plants so this figure presents a share of the United power system which is supplied by 220 kilovolt power transmission lines or transmission lines and actually there is a very big share of renewable energy sources allocated throughout the territory and the total solar solar capacity is about 3,300 megawatts moreover the total dweet installed capacity is about 90 megawatts so we have about 400 megawatts of renewable energy operated for this share of the power system actually if speaking about the reliability and let's call it assurance of the current carrying capacity of the power transmission lines it was always thought that we are to ensure reliable iteration of the power system in the conditions of thermal power plant generation outage and for this year of the power system the typical outage of thermal power plant generation is one generation unit having the capacity of 250 megawatts and now have the scenario that the total installed capacity of Sun and total installed capacity of winged together they are higher than the United capacity of thermal power generation plant unit and for this reason there are actually some new power operation conditions which should be taken into account and there are some new regulations and rules should be provided to ensure the maximum possible power flow for this power transmission lines because if we have unpleasant weather producing lots of clouds maybe rain we have a group reduction of the power produced by photovoltaic power plants and this can exceed the maximum calculate generation outage for thermal power plant thus it can provide the overloading of the transmission lines which connect this share of the power system with the big power system of the country and actually on this slide there are some diagrams providing the measurements of possible power production reduction of the photovoltaic power plants and here we can see that these measurements are provided for these 220 powerful smishing lines and we see the 100 megawatts reduction of the power produced by photovoltaic power plants for a 10 minutes period actually photovoltaic power plants are one of the generation sources which have very low inertia yes and the speed of power production decrease can be very high and it could be hundreds of megawatts per second maybe per minute depending on the climatic conditions which a cure for this reason the big share of photovoltaic power plants allocated throughout the territory of southern regions impose a very serious problem connected with the power system operation module liability insurance and for this reason some let's call it serious actions should be taken to ensure reliable power system operation and reliable supply the final customers if this is the another region of our country having the photovoltaic power plant and the thermal power plant installed in the worst loss center and we see here that the transmission for transmission lines and the power flow for these transmission lines depends greatly on the transmission and so it depends greatly on the generation of the power at solar power plant and it can be seen from these figures because when we have day time and the power plant photovoltaic power plant produces a certain amount of energy so we can see the fluctuations for the power transmission lines and we see that the power flow for these transmission lines depends greatly on the amount of energy produced by renewable energy source well actually this mainly conditions the urgency of the problem of photovoltaic energy forecasting well when dealing with this problem we find out that it's not an easy thing to do because the power produced by photovoltaic power plants is highly stochastic and it depends on a very big amount of external factors mainly meteorological and weather factors which cannot be easily predicted highly accurately to produce the reliable forecasts of the energy for some period ahead and if we speak about the physics of solar irradiation distribution in the atmosphere of the earth so we should take into account lots of factors there is there's own lawyer while there are some water molecules in the air and there are some clouds which consume the energy of solar radiation in the atmosphere of the earth moreover if we have the angle so if we do not have the direct angle of falling grease so we'll have some energy losses to some radiation would be reflected some radiation would be diffused and all these factors are to be taken into account because there is a very small share of useful energy which can be extracted from the total energy of solar irides it's only the one side of the problem because the second side of the problem deals with the photovoltaic plant itself so we have different types of PV panels they can be polycrystalline mono crystalline or amorphous so we have a certain circuit connection that the solar power plant so we have some converters inverters and they highly nonlinear and they produce fluctuating efficiency so we have some alternating current side of the photovoltaic power plants so we have losses the power transformers and power transmission lines we have some restrictions associated with external power system and all of these criteria had to be taken into account when producing a forecast of the PV power plant production moreover so if we take Walton period characteristics of solar power plant will see that the amount of the power produced by the PV panel depends greatly on the level of solar radiation so and moreover it depends greatly on the outside temperature and it's surprising but from the point of view of the ambient temperature the best place to install solar power panels is the northern pole but they are we have very short someday for this reason it's in the majority of cases it's not too much efficient but from the point of embed temperature we actually provide some coolant bling for the semi conductors to make them work more efficient moreover so if we consider consider the geometry of the Sun movement in the sky and we have a very complicated sinusoidal model giving us a basic knowledge about the amount of solar irradiation which can be usefully extracted for certain periods of time throughout the day well if we take photovoltaic power plants forecasting problem well it is not firstly solved all over the world of course and there are lots of research projects devoted to this problem and in the majority of cases the models the mathematical models can be subdivided into adaptive machine during paste some used analytical models and some use mixed models because they produce let's call it the golden needle between the adaptive metals and statistical approaches but the general research community provides us with the following data and which depends greatly on the perspective of forecasting because if we consider different problems like very short-term forecasting is like intraday and intra our forecasting if you consider short-term or middle term forecasting these problems are very different and different initial data is required to succeed and to get the accurate result of photovoltaic plant energy production and so in my presentation we mostly deal with short-term energy forecasting in very short terms because when we speak about middle term and long term forecasting we mostly deal with the global Meteorological models which produce us with the sort of evening assessments of the energy which can be produced by photovoltaic and this information can be used for strategic planning and the goal of our work was to find the solution the reliable solution to provide our system a curator power system operator with the tool to produce the forecasts of power system operation mode for tomorrow and for intraday planning and moreover so we produce some recommendations in terms of power system reserve capacities because when we have a big share of renewable energy sources there is a very big need of extra reserve capacity thermal plant reserve capacities which covered the fluctuations of solar power plants output so we've developed the methodology which was built on risk approach which gives us the year of the risk of the balance deviation in the power system and provides us with certain measures to be taken to eliminate these disbalance and ensure reliable operation of the power system well if coming back to business in terms of the short-term scenario of photovoltaic plants forecasting so this diagram gives a very generality of course we have some input data which is somehow protests and it's being prone to filtration but we have some adjacent power system parameters it's like power transmission line power transformers so we have to ensure that our photovoltaic power plants is able to produce all of the power which can be extracted from solar irradiation we have some weather data there is an external data source and this data is extracted from with the providers which produce us with the forecasts for the head and our head perspective and they give us the values of clouds the temperature the pressure humidity and other factories some of the way the providers producing we see variables like cloudy partly cloudy or maybe for B and so we have geographical data it's very important because the geographical data actually gives us the pilot point which can be used to to calculate the basic basic possible solar radiation for this geographical region we have avoid the monitoring system data this system is installed at the power plant facility and it gives us again the ambient temperature pressure humidity and solar radiation because we have installed parameter and this data is very useful for us to produce the intraday forecasting and into our forecasting it makes it very accurate so and of course we have power equipment best for data in terms of inverters and rotors PV panels sometimes PV panels are quite different because there can be different panels installed at one power plant facility so in this should be taking into account us then we produce the forecast of solar irradiation and based on the forecasts of solar irradiation who produces forecasts of solar power plant output and this makes us a two-stage procedure and these two stage procedure gives us some extra flexibility because the second model which calculates solar power plant output is mainly electrical and in electrical model we can take into account the state of the panels the state of the inverters because some of the equipment can be put out of decoration for repair and maintenance for some reason and we may have some let's call it constraints from our system operator and all of these features have to be taken into account in the second part of the system but from my point of view personally the most interesting is the first one because it provides us with the opportunity to produce a forecast of solar irradiation based on the weather data which is extracted from external weather providers yep so this slide gives us a very general idea of the of the mathematical approach which was used to produce solar radiation forecast and so we used the so called transparency index the atmosphere which gives us the relation between the solar irradiation which comes to the atmosphere and which is measured at the horizontal surface at the earth yes and the transparency index of the atmosphere is always between zero and one we are one who is the Yogi's but we can never reach this because we always have some clouds because all these caps are water molecules which extract some energy from solar energy from the irradiation race but in the majority of cases the weather is sunny and we can provide a very reliable forecast so of course one of the pilot points of the algorithm is the calculation of the irradiance at the border of the atmosphere because it is the basic value which is subsequently multiplied by the transparency index so it is the way we can get the value of soar or the aliens at the surface of the earth we use multi regression analysis and I'll tell you a little bit about the history of this mathematical model because this model was provided for the industrial company to provide forecasting of power system operation mode and we had some several iterations and by improving the model with the initial data sets the model itself and so and so forth but actually it was the first version the model looked like and it was implemented in the industrial software so you can see here the screenshot of this software and I will give you a brief description of the application of this software like first results of the application of this software in real operation conditions actually this soft was tested on a real power plant which was installed in southern region of our country and well we had actually two-year initial data set and the data had one hour resolution sometimes we get half hour resolution and not work but here we see the result of the sunny day forecast actually the sanity forecasting is quite reliable for all the possible versions of the model because it is mostly based on the value of the solar irradiance at the border of the atmosphere so we get some mean transparency index introduced the forecast for a sunny day quite reliably and here we see the results for 20th of September of 2017 and the total output forecasts of energy was produced to be 86 thousands of kilowatt hours and the actual value was 80 thousand of kilowatt hours so the second case study is the partly cloudy day well this condition is quite more complicated because we have some characteristics of clouds for these day P and actually if we look at the screen shot a do-over part forgive me it's in Russian but we see the graph of the clouds yes so in the orange one is the actual value of the clouds and the yellow one it before cast its body of the clouds because yeah we know we have the data of actual clouds which is provided by the weather provider as it is so if this hour passes so we can use this data and say it's actual data but we have some forecasts and the forecasts are updated each hour it means that if we produce a day hit forecast for example after 24 hours we will have 24 forecasts of clouds for each specific hour and actually we calculate the mean value of the clouds yes and make a sort of the forecast for our system - well quite from this figure it can be seen that forecast is still reliable and we can use it for power system operation the next case study is cloudy conditions so again we see the cloudy graph at the bottom side of the strain shot the orange one if the actual value and the yellow one is a forecast as well you so we see that situation is a little bit more tricky from the point of view of the accuracy of forecasting but from the point of view of the energy market again let's state in terms of electrical energy production so actually it almost matches and the actual forecast sorry the actual value of the energy production was about 3 30 30 thousands of kilowatt hours and we forget these 30 thousands of kilowatt hours well but there are some days we'll call it extremely noisy days which produced us we a very unreliable forecast and so there were some first taken to cope with this problem because so firstly let's speak about the possible reasons of this problem occurrence because there are some cloudy and rainy days which cannot be forecast if a person from one point of view it's not too much important because if we have a rainy or maybe a snowy day it means for us that maximum possible power production of photovoltaic power plant is very low so for example if we see look at this figure for example yeah so this power plant has 15 megawatts of installed capacity so theoretically it can produce 15 megawatts but for this day yes so it produces only one megawatt it's a very small share and it can be negligible was in some cases from the point of view of power system operation mode planning but if you produce the forecast of 3 megawatts yes so the error is significant but the total share for the power system is not to much significant anyway we have to cope with this problem and firstly we try to identify the reasons so what the reason of course in the majority of cases the reason is the noisy data we used different forecasts from different weather providers and you know different providers first of all they use different prediction models of of the weather I mean numerical weather forecast models its first point the second point different models have different resolutions and in the majority of cases the models which produce free forecasts for the internet they have resolution I mean the cell yes of prediction is 100 kilometers versus 100 kilometers so it's a very big point and it means for us that for itself having white and heights of 100 kilometers will have a similar weather forecast but this territory is quite big the area is big and the weather can differ greatly and if we see the conditions the geographical position of our photovoltaic power plant on the map we'll see here that actually the closest the nearest metrological station it is a located in the airport is 45 kilometres from the photovoltaic power plant so we have some sort of source of error in our weatherdata.forecasts it's first point the second point actually it's different where the providers for these different data in terms of the type because let's say Yandex weather could use cloudiness forecast in percentage when whatever the ground was a provider produces cloudiness forecasts with the figure and with linguistic description like foggy rainy partly cloudy and some other characteristics are above 20 possible characteristics which can be used but it's quite complicated to protest this characteristics because we need to make a sort of the scale of factors and taking to account distinguish the variable Street yeah on this light we see the bottlenecks but mainly it's not the algorithm bottlenecks but the initial data bottle next when we compare let's call it different cases with the different noise scenarios for example the case one which is showed on the left side of the slide is characterized by the following the first the left one and our so it's given for 95 percentage of clouds but about 500 watts of the installation but if we consider the right a part of the figure so we see that clouds of 50 percent have the irradiation of one hundred and seventy but from logical point of huge should be vice versa yeah and but for this scenario the great cloudiness correspond to the greater radiation so and this with a very noisy case of PV power plant forecast at the right side of the slide we see the second case it's when irradiation of 500 watts per square meter is observed when the clouds are 80% but irradiation of 80 watts per square meter is observed from the cloud for 95% the difference is very small in terms of the cloud and its coverage but the amount of irradiation which was measured by the thermometer difference dramatically it's five times different and this scenario actually means that there is a certain lack of features in the initial data which can give us some extra useful information to characterize particular weather conditions and the first thing we made to eliminate this bad days forecasts is the customization of initial data so on this slide we see the characteristics of the initial data step of the two-year initial data set produced for our photovoltaic power plant and we see different times of the day it's like morning evening before and afternoon the noon time and on the vertical axis we see the transparency index and horizontal axis corresponds to the cloudiness it can be easily seen that different transparency is or a different cloudy conditions characterized by very big very wide range of transparency index for example for morning evening the cloudiness of 50 percent response to the range of transparency index is starting from 0 to 0.9 actually this is a very wide range which means that we request an extra feature or DT customization to provide more reliable worthless so what have we done we have used Kamen's data set authorization to produce some useful clusters characterizing one or another climatic and weather conditions so on this slide provides us with the step by step procedure of producing the clusters of initial dataset and we had some initial clusters that we had iteration 1 2 3 4 and finally we'll come up to the Foley in clusters of the initial data so we have the following the following features used in our algorithm its solar altitude angle actually it can be said that it the height of the Sun the horizontal plane then we have cloudiness in percentage and then we have some declination angle it's seasonal characteristics of the Sun and this produced us with the following clusters for solar altitude angle its morning evening value it's close to morning evening value some transient values then close to noon and noon then for cloudiness we have almost sunny low clouds middle and high clouds and for seasons we have winter and summer but we do not have spring or autumn because in our case it's mainly offseason period because the weather is highly unstable and this time mainly starting from February till May in spring is characterized by extreme least elastic weather and in cotton it's typically starts from September to November when the member when the rider fluctuates and we are really challenged to produce a reliable 40 so and actually the clusters give us the possibility to provide some filtration because we've calculated some dispersion and standard deviations for each of the clusters and we have illuminated some run-outs which were initially given in the initial data set well and after that it's provided us much more pleasant picture because we had the determination index of 0.88 when in the previous case yes so we had 0.65 which gives us the opportunity to produce much more reliable photovoltaic power plant forecast yeah and actually this is the example of the PV power plant production forecast in orange columns we see the actual data that we see the predictive data using multiple regression and they predicted data with the dotted line with multiple regression improved by pipelining so and the final result the mean result of the forecasts was about 90 percent and actually it suits our system operator in terms of power system operation planning because when the power system operation planning is carried out we typically take 20% reserve for uneven fluctuations of electrical load and fluctuations of renewable energy production yeah and as I have already said this methodology mathematical approach was implemented in industrial software process systems and actually we have the error less than 60 percent for the head forecast in approximately 80 percent of cases and a roar not exceeding 13 percent in 92 percentages cases which provides us with quite valuable result well the second problem that we were dealing with is a very short term forecasting because the system operator provides each hour provides the update of the electrical load forecast and in terms of the development of renewable energy sources so if you have to provide the update of renewable energy production forecast and we have taken the problem of very short term forecasting and the main let's call it the core perspective is one hour ahead but you can produce forecasts for six hour ahead with quite attractive accuracy and so I'm sorry for this what we have used in our model so we use the algorithm of extreme extreme gradient boosting trees and we implemented this model in Python but there is some extra data which should be taken into account when compared with the algorithm implemented for day ahead forecasting actually what we use intraday forecasting sorry when we take intraday forecasting so we have some extra data which can be useful for us to adjust the short-term forecast and to make it more reliable one of these types of the data is the measured transparency index because we have the horizontal in structural ammeter on our photovoltaic power plants and it provides us with online measurements of the solar irradiation so we can use this reliable data to adjust our forecast for one overhead perspective more or do you make our forecast more reliable so we introduce another feature into very short term forecasting model so we introduced the short term Ferguson so if we have a short term forecast value and the measured transparency index the algorithm tries to find the equilibrium between the produced short-term forecasts and the actual irradiation to minimize the total error of one hour ahead for Justin and we had three different input datasets the first one we call it learning without history actually in the from point of view of taking into account the weather conditions it's absolutely meaningless but we have faced some extreme situations when the internet connection failed and the weather forecasting data cannot be extracted from the database of the weather provider and we do not have with the forecasts of the clouds of the temperature of the humidity so what shall we do now and this reason learning without history model do exists because it produces us with the forecast of the sunny day so it doesn't take into account the history it doesn't take into account the meteorological data but it produces us the mostly physically viable solution when we do not share the required data in terms of the weather and online monitoring systems the second case study is learning with history when we have measured transparency index and for transfer guests and the final case is learning with history and actual meteorological data which is extracted from the database of the weather provider yes as I have already said we use gradient boosting trees gradient boosting trees provides us a very powerful tool because it belongs to the class of the ensamble algorithms and the trees in the humble abode for the best solution and actually it's each subsequent tree for the brilliant boosting trees algorithms provide us with the solution minimizing the error of the previously previously obtained decision tree and we have tested this problem and get very pleasant results because now we have four one hour ahead perspective we have the accuracy of 97% and in 92% of cases the award does not exceed 7 percent of the installed capacity of the solar power plant and if we give a little bit more attention to this diagram we can see some very interesting problems dealing with the initial data because the red line gives out the idea of the data which characterizes the cloudiness it's the cloud coverage which is expressed in percentage they exist it's the right one and we see that the short term forecasts which was produced for the day head perspective it is the orange columns actually they strictly follow the value of the clouds yes and when the clouds cloudiness groups the forecast of the produced energy decreases but we see that the actual data gives us what another result because we do not have a severe reduction of the energy production when we have the cloudy conditions but this drawback introduced by short-term forecasting algorithm can be easily caught by a very short term forecasting well introducing current measurement yes and one of the important things that I have to say about the application sphere of course well some of the some of my colleagues from my University say that I'm rather material meteorologist yes rather than the electrical engineer but but I am certainly an electrical engineer and I D of the problem of PV power plant forecasting in application to power industry and one of the goals of the research work was to produce the model which could provide us with a certain knowledge of the energy produced by renewable energy sources to optimize the spinning reserves of the power system because nowadays actually without the instruments of renewable energy sources forecasting the installed capacity of solar and wind power plants they are completely taking into account the resource capacity and in case of the United power system of Russia so if we have one gigawatt of installed capacity of renewable energy sources we have one gigawatt of extra reserve capacities but it's not effective from the point of view of the energy market because it is let's call it highly efficient thermal power plants or hydro power plants ensuring very fast regulation but in the majority of cases the reserves in the amount of the installed capacity of renewable energy are not required and we introduce a methodology which gives us very idea of the risk of balance deviation in the power system and actually it gives us the result of measures what should be taken to eliminate this imbalance in the power system and coming back to the initial State and this slide provides quite a theoretical example but it can be compared with the big industrial region of our power system and so let's assume we have 1000 watt megawatts of combined heat and power production and 50 megawatts of photovoltaic energy are introduced into the power system so and we finally produce the risk distribution curves and when the photovoltaic power plant is introduced the risks of imbalance is increased by 5% yes and let's fix that photovoltaic installed capacity is 15 megawatts but in order to eliminate the risk of the imbalance of the power system we are to introduce only 7 megawatts of extra reserve capacities lost 50 megawatts only 7 to put the risks to the initial value when we had the shear of the power system without photovoltaic generation and of course it is it can only be applied if the certain forecasting procedure is introduced because we made a very big amount forecasts and finally we've got the statistical distribution of the forecasting in work or they had an intraday perspective and having the statistical characteristics of the error distribution we are able to find the value of possible risk and the value of the active power which is needed to be introduced to eliminate mysteries well actually in terms of the meaning part is my presentation that's all but I'd like to introduce the project team of course firstly it's me but last but not the least person in the project is my colleague she's Alexander Hislop she's the show's associate professor of our University and she works with me in the sphere of renewable energy sources forecasting and well it's like an acknowledgement page for me and her role in the project is very very significant for research work to some information should be given about the project support actually because this project takes its roots in the international brand which we are still participating in terms of the European Union we have a consortium of ten universities and there are universities from Russia the universities from Vietnam and three universities from Europe we are providing producing the educational program of smart energy assistance in Russian and Japanese universities which will be implemented I hope in 2021 in all of the universities of the consortium and one of the interesting parts of the project is the course the discipline artificial intelligence and machine learning in smart energy systems and the problem of renewable energy sources integration in terms of the application of machine learning and artificial intelligence is a very important share of this course and discipline and for this reason it relates to the project and is developed within the framework of the project tube for this topic I mean the topic of machine learning the power system so we have the internship in Ural Federal University so we have the associated the corresponding associated partners of the internship these are the main the leading industrial partners and companies of our university in the sphere of power and energy and some famous companies like Gazprom and Russian reads for our city and as well it is the producer of power transformers and the soft systems corporation it's the corporation which produced industrial automation and it was the first company really interested in the development of the program software and tools to produce renewable energy source forecast so we'd like to thank our University for the opportunity to share our knowledge with our partners and our friends from other international universities and it's my great fortitude to our international education problem programs department because it's there recently that we produce the course of machine learning and artificial intelligence in summer school University for the students coming from Indiana University central French universities and I'd like to share some very good experience of collaboration with the parallel University of India because the students coming from that University from the summer school and the academia from that University after the summer school we've made a sort of the collaborative project and now we are producing scientific projects and improve the results which were obtained during the project of renewable energy sources forecasting and now we are have submitted and I Triple E transactions paper and I think it's a very important scientific results for us as a professor staff and for the students to well that's all for me it's was a great pleasure to introduce myself and the topic of my research thank you for the floor and if there are any questions I would love to answer them so thank you very much for your attention thank you so now moving you in the chat section thanks to operation yes well taking something yeah so I have one remarked that countries with multiple season need accurate forecasts for solarization it's it's true actually because we have some solar power plants allocated in different regions of the country and even we have a solar power plant allocated in Yakutsk and it's a far northern part of our country and it could be up to let's call it minus sixty degrees in the winter and actually it is not efficient I think it's not efficient in winter because the there are some polar nights and during the polar nights the solar power energy is not produced but the summer time for the northern regions of our country are typically very hot and they have very long days and for this reason there is a very big shear or diffused solar radiation and they produce a huge amount of electrical energy and it could be really it should be very carefully addressed for different regions which it's true yes please give your opinion on hybrid wind battery systems actually we had some experience of hybrid battery systems and the rest and some colleagues from other Russian universities working with us on that matter and we've calculated hybrid battery systems for different regions too and I think it's effective but it depends it depends greatly upon the conditions because we have to address very carefully the energetic potential of wind energy or maybe solar energy to use it with the battery and it's the first point and the second point we have to very accurately address the sphere of application of our system so what for are we going to use the wind in the bedroom so if we are going to use it for send alone operation just to cover the energy demand of the separate power system maybe small power system and we region yes so it's a one case but when we consider for example the reserve capacities it could be used at the backup source of power supply it's another case and for example in our case we have calculated the case for the far north region yeah when there is a network connection of the load but the network connection sometimes well not sometimes but often fails due to the distance because it's was about sixty kilometres of 10 kilovolts power transmission lines and there are forests swamps and very hard to access territories and the period of service and maintenance is very big too so and once it for the new year a failure and it was 11 days of recovery of the power supply and we suggested the application of hybrid wind battery systems to provide the energy backup for the period of the energy shortage and I think it depends on the conditions and first thing we are to think about is the sphere of application and the energetic potential of the energy source that we are going to use what kind of machine learning algorithms have you applied in conditioner monitoring of transformer well yeah I know the answer actually it's quite out of the scope of my today's presentation but anyway in terms of the condition monitoring of the power transformer we used several approaches of machine learning firstly we made it using let's call it by the numeral interference and now we are using extreme gradient boosting trees to you and actually now we found out that the gradient boosting approach outperformed the fuzzy one because you know the fuzzy logic approaches they require a lot of manual labor yeah so we have to provide some membership functions some rules we have to do it expertly and to provide all of these conditions for our system and firstly it's very time consuming for the operator of the software tool well secondly it's at least at initial stages it gives us a very significant airport in power cook state estimation so brilliant boosting trees in this case outperform so we use so now we are working on other transformer monitoring approaches and other equipment approach is to actually because if we do not consider a power transformer or maybe we consider a circuit breaker it's another case and it's another model because it's a very important power equipment you need to and technical state estimation of a circuit breaker is carried out according to the national legal documentation but there are no tools I mean software tools that produce state estimation of circuit breakers and there are a very interesting feature that can be taken into account in circuit breakers for example like short circuit currents the calculated values of short circuits can be used to improve the assessment of circuit breaker technical state actually if the topic mainly of Aleksandra if you are interested so you can email me and I'll give you further contacts and she will share her knowledge and achievements in the sphere Oh will future big data database supported by artificial intelligence be used in precise forecasting in order to verify petrological projections well it's a very interesting question actually um in terms of precise forecasting to verify future projections hmm let's call it it if we if dealing with the project of renewable energy source for testing at the moment we do not have a database which can be called big data because we have some historical databases of a couple of photovoltaic power plants allocated to different regions of our country operated for five years and well let's say five years and one our resolution of data it gives us about fifty thousands of data samples but it's not too much for protesting but anyway we're moving forward to Big Data database is supported by artificial intelligence I think some other approaches can be used to provide the problem solution like deep learning approaches which are well required to be further investigated on this case in theory means servants in Greece due to the solar rooftop system in generation of power plant well actually yes but it depends again so when we consider the global process of installing the rooftop of photovoltaic power plants there are some issues which should be taken into account very carefully the first issue is degreed one because initially power grid is designed to provide sufficient power flow for the final customer so let's assume you have the peak energy energy load of let's call it five kilowatts for example and you have the corresponding power network facilities which gives you the opportunity to consume these five kilowatts from the external power Network but when you install your own generation source for example the photovoltaic one so you [Music] produce they unload yes of the external Network and it's good from one point of view because well there are no severe and overloading conditions of the external Network but from the other point of view the economical point of view the grid company suffers from these measures because they lose the useful output of their network but they still have to provide maintenance and service for the power transmission lines which were initially designed for the full load it's a little bit the problem of the economic sphere because from my point of view it's a little unfair because the power network is designed to cope with the peak loads but the peak loads are not required and the grid company has to provide service and maintenance of the grid facilities in the first point and the second point deals with the security because when we have some disconnections in the external network we have to ensure the secure operation of the engineering staff of the power Network company which provides maintenance and services about mission lines but if the Greek company doesn't know that you have your own generation source that produces a certain amount of voltage yes so it produces a threat for the operation personal working on the power transmission line in terms of recovery and maintenance and anyway these protests can be regulated somehow I think but it is the problem of power system operation personal security too and the third point actually is the balance point I mean the balance of consumption and production of electrical energy it there are some countries for example like our where the roof top power generation are not so popular and it does not produce any threat for power system operation but if we take a certain share of the photovoltaic power plants covering the load let's call it 20% maybe 30% of load so it can produce some problems in terms of the operation and control which can be coped with by using adaptive methods information telecommunication systems and the corresponding machine learning approaches in terms of the control reliability and security of the power system well if you a way to increase intensity of solar radiation of all on a photovoltaic cell like magnified focusing well I think actually there are such ways but the most efficient way for personal use application let's call it is to keep the surface of the photovoltaic panel clean in the first measure to be taken the second measure to be taken is maybe to find the optimal angle towards the Sun ray so if you do not have a solar tracker which follows the Sun so there is a certain calculation algorithm which gives you the year how to select the optimal angle of installing the solar panel to produce more energy and finally if you are able to provide some cooling environment for a photovoltaic power plant may be some weedy conditions on the rooftop it's quite valuable too because if you provide good cooling conditions for a photovoltaic panel it gives you a very good efficiency characteristics how can we check the feasibility of solar predictor model well actually I did not get the idea of the question because what is the feasibility of solar predictor model and the in terms of the physiology there should be some visibility criteria because when we speak for example about the electrical load forecasting the feasibility criteria the forecasting error again but the let's call it top value of this Aurora is very low we nowadays Russian power system produces the electrical load forecasting with the Aurora not exceeding 1/2 of 1% yes but when he produced the forecast of renewable energy sources so we cannot ensure the forecasting accuracy of 1 percent or maybe 1/2 percent and for this reason actually we are to check ourselves with the feasibility criteria so what are the facility criteria and in our case with both criteria swarthy let's call it elimination of 20% aurorus of solar power plant forecasting and it was reached for this reason we say that the solar predictor model is feasible and maybe for some other power systems there are some other constraints or limitations and they should be treated differently and maybe for this reason we should introduce some improvements or are the mathematical methodologies to make this model feasible and one of the ways which should be taken into account actually so when we speak about solar power plants we cannot calculate the error the forecasting error related to the installed capacity of the solar power plant because the installed capacity is constant and it is not changing throughout the time but if we consider the maximum theoretical impossible amount of electrical energy produced by a solar power plant for each hour it differs and there are well very few hours which can provide our solar power plant with the opportunity to produce the installed capacity and for example if we consider the morning or evening conditions we can produce only let's say 10 maybe 20% of the installed capacity but it is theoretical in maximum value and we cannot produce more in any conditions and it doesn't mean it does mean that we are to check the prediction Aurora what was the installed capacity value but the value of the maximum theoretical possible power produced by a solar power plant for a certain hour it's a sort of the trick but it's very logical because it produces us with the physics of the protis of electrical energy production from photovoltaic power plants yeah so can you please guide us all synchro phasers actually I think that this question likes life's little beach outside the scope of the presentation because it's a very urgent actually but a problem dealing with the power system operation mode calculations so then I pass to the second question the equipment is to be used in ideal temperatures but in countries like India and the ones near equator have high temperatures and there ways to still get optimum output yes actually if we take into account international standards regarding the approaches of design of solar power plants there are some recommendations of which deal with the power plan design approach so firstly we are to select the site of solar power plant location and when we select the site the area so there are very big amount of criteria which would be taken into account first of all it's the price of the earth yeah let's say it's very typical it's very tight secondly so if the potential of solar energy well in Russia it's quite urgent question because there are not too much territory with the high potential of solar energy maybe for India there is a little bit another situation but in this case you have to consider the areas with the highest winds so and if you have high wind conditions it provides natural cooling of photovoltaic panels and semiconductor devices making them work more efficient and there are some constructions I mean the structures the steel structures used for photovoltaic panel mounting they have some sort of modifications we ensure initial circulation of air and thus providing us with more pleasant cooling conditions so did the case actually the wind is also have to be taken into account when designing solar power plants as well what are the parameters used to forecast about solar energy well in our model firstly we had lots of parameters which can be used to calculate these parameters can be totally can be classified to separate separate classes yeah one class of the parameters is the parameter that can be calculated so we just use the laws of geometry yes so how the earth moves around the Sun and how the Sun moves in the sky so we use complicated a sensorial model to get the maximum theoretical possible value of the photovoltaic energy sorry both of the solar irradiation which goes into the atmosphere it's one share of initial data which is the second share of initial data is meteorological data which is used to take into account current weather conditions so in our case we use for solar radiation forecasting we use only clouds but when we come to the step number two of the algorithm if you remember is the electrical part of the algorithm we calculate the power output of each solar panel and each solar panel output is completed taking into account the humidity the wind speed in the ambient temperature so and practically we take into account all the parameters which are provided by weather company's event and the third share of data class of data is online data which is extracted from the metering systems installed installed to the power plant itself firstly it's online whether data like online irradiation online temperature line wind speed and secondly its online measurements of the electrical meters of the power produced or by the solar power plant what is exactly circle of transparency index well the transparency index used in the algorithm actually it describes the relation of the of the solar irradiance at the border of the atmosphere yes before passing through the atmosphere with the value of the solar irradiation measured by the parameter installed on the horizontal surface of the earth thus it's takes into account the loss of energy of solar radiation in the atmosphere lower so it never gives out the United I'll you because even for a very sunny day we have some dust some worried molecules some invisible clouds yes and but it can be around zero point nine ninety five for a sunny day how load balances in wind energy grid or how load balanced in wind energy grid and solar energy well if we speak about balancing of electrical loads in the renewable energy sources based power networks we have to ensure some some energy storage facilities firstly it's very important so it could be some hydro accumulating plants or maybe it could be some electrical storage it's first point and the second point to my opinion there should be a certain share of thermal generation which is controllable and provides us a certain value of frequency a certain value of power output to balance the deviations of renewable energy sources at unload or of course some electrical energy storage how high the variation in the outputs of solar panel plants can be reduced in order to synchronize it with the power grid actually there are no problems with synchronizations of solar power plant with the power grid because you know that solar panel produces the power at direct current and then we use some converters and inverters producing a certain value of the power output at a given frequency yes and for this reason I don't think that the variation of the output of solar power plants somehow influences the synchronization with the power Network but in terms of the balance and power system activation mode it plays a very important role because there are some mmm that's called practical examples yeah when we had some energy districts isolated from the power system after the failure and the emergency and these regions were supplied only by photovoltaic power plants actually I haven't seen the measurements of frequency in this power system but I know that solar power plants produced energy at a very low frequency anybody produced somehow but anyway there is no problem of synchronization of power Network sorry of the power plant with the power Network well if there are no questions let me thank you again for the question session and for the opportunity to speak and thank you very much it was my greatest pleasure okay thank you so much sir hello everyone I partial trader I I Triple E medium store entrance first of all I would like to thank mr. Stanislav and spa sir for giving things valuable time to deliver such a great and informative session on machine learning for solar energy for fasting and taking us through the topic from a basic to an intermediate level I'm sure it will has many Oh fun curious mind prison here so we definitely look forward to have you with us in the near future thank you for your sharing thank you for sharing your knowledge thank you very much anxious now I would like to call dr. Kashi capital man associate Dean's international collaboration program to say a few words hello parth am i audible am i audible to you yes well and visible - yeah okay fine thank you good evening to one and all present over here I'm dr. Kashi capital associate dean international collaboration for bbm our students do come to Ural Farrell for summer schools and their experiences at the University is excellent over there you would like to have more research collaboration for faculties and students as well we surely are looking for forward for that with that first of all I would like to thank mr. stanislav yoshi n ku4 the informative session on machine learning for solar energy forecasting in the era of the renewable energy it will surely helpful to everyone to learn something new actually I need to stay as a special thanks to professor Yana director international relation for academic collaboration and speakers provision I would like to thank our principal dr. energy potential for supporting us in all the sense I also thank dr. - in the lower e-grants councilor I Triple E bbm and our host humans should Tucker along with his moderator team Ibrahim Pasha for handling the every session during the ITC 2020 I would like to thank the team I Tripoli bbm for organizing this event and all of the dear participants without whom we would not be able to spread the knowledge I hope everyone enjoyed this session and we'll meet soon again in our next session in the next session tomorrow stay safe and stay home thank you thank you man and before ending this session I would like to remind everyone about our upcoming flagship event bloggers on a national level blog writing competition where the topic is not limited to only technical blog but you can also submit blog return on the non technical aspects of the word and it yes why not you can also write a blog about what you have learned from today's session and build your ideas on it so friendly boomers and registers as soon as possible for the results are given in the chat box I once again and thank you [Music] you

Channel: IEEE BVM SB

Views: 3,852

Rating: 4.9795918 out of 5

Keywords: ITC, 2020, International, Tech, Conclave, IEEE, BVM, SB, Student, Branch, Quarantine, Education, Webinar, Session, Technology, Science, Space, NASA, CERN, Learning, Lectures, Video, Conference, Effect

Id: 5B067bBWcSk

Length: 90min 16sec (5416 seconds)

Published: Tue Apr 28 2020