the ministry of emergency situations is mastering neural networks, russian rescuers are now monitoring natural fires using high technology. the intelligent system was created at skultekh and is already being used in practice. deputy chairman of the russian government dmitry chernyshenko spoke about this at the st. petersburg international economic forum. the neuron automatically predicts the situation 5 days in advance. russia with an accuracy of more than 84%. this solution is already being used by the ministry of emergency situations in the irkutsk region and krasnoyarsk territory, helping to save forests, money and lives. natural fires are an eternal problem summer season in russia. every year, the fire element costs the country's budget billions of rubles. people are dying, houses are burning down, and air quality is deteriorating. it is not surprising that the most advanced technology available has been involved in predicting and preventing fires. artificial intelligence, but

not only fire threatens russians, catastrophic downpours, squalls, heat waves, cold snaps, all these are the realities of the new climate era, but is it possible to predict these phenomena using neural networks, how does the model work? predicting wildfires, and whether it can warn about other natural disasters with the same accuracy, take into account the goal in climate change calculations, and whether the neronci will ever replace real living meteorologists, this is a question of science, i ... associate professor , faculty of geography, moscow state university, and today our guest is professor, head of the skaltech center for artificial intelligence, leading researcher at the institute evgeniy burnaev, evgeniy vladimirovich, hello, hello, about identification forest fires, well, here’s a short and accessible way of how the algorithm works? it works as follows: firstly, in this task, as usual, in many artificial intelligence and machine learning tasks , the actual data needs to be collected and prepared. what data is this? firstly, these are data

from thermal points, that is, where the temperature was exceeded, in what area of ​​space, at what point in time, these are satellite data, secondly, these are data that are also obtained from satellites about the underlying surface , various vegetation indices, of course, this data on the relief, of course, is data on the distance to populated areas, because in general, in predicting fires , you need to understand that the human factor unfortunately sometimes happens and... does occur, and finally, these are various data describing the climate, meteorology, that is, this is a weather forecast, the weather that was known at previous points in time, some data on soil moisture if it is not available, or if somewhere, for example, this soil was specially moistened, such procedures also exist, so this data collected, this is like input data that certain factors, certain combinations of which can... characterize that

the probability of a fire occurring at a certain point in time, a day, two, three days ahead, in a certain area of ​​space, is potentially high , yeah, this is what, so to speak, is fed to this model as an input, that array of data that is analyzed at each specific moment in time to make a forecast, but the model needs to be trained, these knobs need to be adjusted, and for this there is, so to speak , there must be markings, that is , historical information about those fires where they actually occurred has already been collected. this array of data in our case is 10 years old, we have collected such an array of data from all regions of russia, there are some thousands of fires, well, there are about 10,000 fires of different situations, we have collected it over the historical period, and nerosit actually learns to understand, what combinations of input factors that she can identify from this data set are most likely to correspond to a fire. m, well, i see, yes, then

there is a person who is not able to sort through all this multitude of options, well, of course, there are all sorts of models out there, such as the nesteru index, the soil moisture index, which naturally in... i know these problems very well, well, we all know very well the problems with generally quite poor amount of data, and i’m not talking about weather stations at all, especially in siberia, there are not enough of them even now, despite all the development, but even if we take satellite data, and some parameters, well, first of all

, the average, upper atmosphere is good, here the importance of the soil, because it is still reconstructed from satellite data with large errors, but if you aggregate it into your model, it’s not... well, i understand that it is one of the main parameters from it, yes, probably for you a very important parameter, but if it is bad according to satellite data initially, how much does it harm you, or do the other parameters outweigh it? it turns out quite well? well, fortunately, we have access to the model, we add data that represents, firstly, a characteristic weather here now, historical data about the weather there at previous points in time, because this model is fed with data for... the last conditionally 7 days, yes, because we need to track the dynamics, and secondly, that is, this is data about weather, secondly, this is data on the weather forecast, which is considered in a special way as a calculation model, and you somehow calculate the weather forecast in your own way or take the results, we use the modeling results of the russian hydrometer in

this case, but unfortunately, there is no resolution there very good, well these are those the data that is available inside the circuit, because we did this work for the ministry of emergency situations, is the data that is inside the circuit in order to carry out some of our daily tasks, if, of course, we had data on the weather and weather forecasts with greater detail, then we would even achieve much higher accuracy, well, yes, there, if i understand correctly, there are generally scales, you probably need spatial resolution,

literally kilometers are needed for this to be good, because there is a step global model grid 25 km, probably nothing why for such a delicate task as fires, the question is, you need to understand that... we are not talking now about predicting what will happen in a specific, i don’t know, area of ​​the forest at the edge of the forest or some specific building there , the hut will burn down, after all , we are now talking about such a global ranking of the forces of means on the scale of russia, and this is a slightly different story, in this sense, there are several tens, a couple of tens of kilometers - this, as they say, is the right scale for this task, yeah , yeah, what i'm saying is that we could make the accuracy significantly increase, well, it’s not 80%, but...

some you can brag about? no, well , practically the result is completely concrete, we built this type of model, one that predicts a fire 5 days in advance, in detail, that is, every day in the corresponding grid cells that cover the entire territory of russia, we predict the corresponding probability, this is possible, this model has been introduced into the internal counter of the ministry of emergency situations, passed these tests there, now they use it in everyday work, according to information from the ministry of emergency situations, the model is very useful for them, they use it just to rank the strength of funds, and then we are interested in,

firstly, increasing the accuracy, there are different ways to do this, and this can be discussed separately, the second is based on this model, we can continue to solve some subsequent problems. problems, well, for example, if we roughly understand how the risk is distributed over the territory, we can actually solve the optimization problem, not just roughly estimate with experts where, that these very forces of means can be mixed, and to do this more accurately, well... mathematically, do you already feel feedback from forest fire services, from rescuers, that is , things have already started, whether such a pilot is beneficial so far, i can say everything, well, this is not a pilot state, we have already passed the pilot stage, the question is that now this is in the work of a specific department, and from them we received positive communication, and we will continue this interaction, because first we are interested in and make other models according to other types. for example, floods to implement them, on the other hand, of course, you are absolutely right,

indeed, a lot still depends on how this forecast is used, they, of course, can optimize their own activities there and confirm that for it was useful for them, but then there are some other departments, including some files, raives, some other government agencies, so how will this affect their work?

are not adapted to the specific data that observed here, now on a specific river, to a lesser extent, at least,

adapted, but they internally include, as it were, physical equations, well , how our world as a whole is structured, due to this combination of such models and models that trained at this university will allow you to obtain much higher forecast accuracy, that is, here you will need some additional work with algorithms, a combination of models. based on physics based on non-networks, we have such technologies right in the center we are developing artificial intelligence, but that’s great, this is just such a smart hybrid, yes a smart hybrid, the second comment here is, the next step that could be taken here, in fact, there is a very large field for fundamental research, which can then develop into something important specifically for the industry, namely the construction of what is now called a fundamental model of artificial intelligence, well, everyone has seen these cats that... draws uh models like kandinsky and others, yes, yes, that’s it

it’s a little bit too much to say that it’s a bomb, but, but, but, it looks cool, but so, these, this type of technology, they can and should be adapted for the data that we are now discussing, and this can be very big benefits, moreover, abroad, there in the usa, in china, this work is actively being carried out, we need to start it here too, well, yes, of course, in russia there is a real boom in erositey. right now, other models are being developed to estimate the likelihood of natural disasters. another one, by the way, also from otskoltech, is engaged in long-term flood risk assessments. more details about it in our review. we load climate model calculations, weather station readings, reanalysis data into the machine, add information about the relief and voila. artificial intelligence provides an answer to the question of where, with what probability , rivers may overflow their banks in the foreseeable future. this cannot be fully called a forecast; it is a fact of flooding in a specific area. the negyroset will not be able to predict, but the creators of the system, skoltech specialists, say that even in its current form the results of the work

artificial intelligence will be useful for financial companies and large businesses. the model will be more accurate, its metrics will be higher if we model, for example, a year in advance, by month. this format seems optimal from the point of view of model quality. the first story is who this could be for, where it could be used, this is definitely the insurance industry. colleagues can use. or for the purpose of clarifying the risk of natural disasters and, accordingly, this will be a useful tool for them. second the industry for which the use of our models is relevant is banking. that is. we can reduce investment risks, including credit risks. the third industry for which our models may be relevant is business as such, that is, for example, for industries such as agriculture and mining industries. the technology is applicable to other emergency situations, in addition to

floods, artificial intelligence can calculate droughts, heat waves, increased winds, convective processes, that is, powerful precipitation, as well as... degradation of the long-term world zloty, for which a separate model has been developed, however, no matter what natural disasters are assessed by the neural network, it is always necessary to make adjustments for climate changes, due to which the data entered into the system must be constantly adjusted, one of levels of input data into our model are climate scenarios, which means they are formed on the basis of greenhouse gas emissions and are developed, accordingly , climatologists in our case always take an ensemble

toring to form accurate forecasts does not have enough data; sometimes, in order for the machine to calculate at least something, you have to resort to tricks, even modeling virtual gauging stations on rivers, if there are not enough real ones. the wider the network of meteorological and hydrological observations, the sooner it will be possible to integrate artificial intelligence into forecasting specific, rather than hypothetical, emergencies . well, let's go back to the studio, here is evgeny vladimirovich, well, my favorite topic is about relevance, about the fact that, as it were, systems natural. does not stand still, our climate is changing quite rapidly, we will not, of course, be avarmists, we will not say that this is a disaster, but nevertheless, that’s how much, we just talked about all kinds of twist handles in models, to what extent in fact, this problem is acute, when you generally, in principle, we assess any risks, no matter fire or flood, it is important to emphasize that

the forecast that we make is not what is here...

we are all we do it, the more we do it later potential reliability of such a system. well , now i just want to talk about what just caused an information boom among meteorologists and climatologists, you of course heard this at the end of autumn, so to speak, advanced british minds, but all this was done under the company google, they created a neural network that began to make better weather forecasts than the hydrodynamic models of the european center for medium-range forecasts, and of course. with much less computing resources, it's all cheaper, it's all better, and of course, immediately, as happens in such cases, a thought immediately arose, some meteorologists are so decadent that that’s it, meteorologists are no longer needed, hydrodynamic models are no longer needed, that’s it, artificial intelligence, well, i ’m much more careful about this,

what about you? i also treat such alarmist statements very carefully: a person is not going anywhere in the near future, especially a person with good expert knowledge, simply a person - armed with such models based on artificial intelligence, he can solve his problems there faster, more accurately and more. number of tasks in the same time, if we talk specifically about models from google, or, for example, there is the huawei company, which , based on its capabilities and developments, also made such a panga climate model, which is in some sense a similar type of fundamental, so what are we talking about, first, again, data is collected, and this is data as numerical modeling using these complex models, which take a long time to read, of the earth's system, which. on the basis of the equations take a long time to calculate, they are launched many times, for different initial game conditions large data arrays are generated, how different climatic meteorological variables

depend on each other, and this is the first part of the data array, the second naturally adds an array of real measurements of climate data from weather stations there and - in general, everything is everything that is available and then there is such a fundamental model - this is a very large non-network, which of course is computationally quite difficult to train. or is it much faster than if you again calculate this big climate model for the whole earth, which you do year by year, of course, one without the other still won’t go anywhere in the sense that

the story is absolutely correct when you somehow then you calculate a rough estimate, a rough forecast of such a computationally heavy model, and then refine it using an artificial intelligence model, which is additionally adapted to a specific region, the important thing here is that it’s still an expert... people are fundamental scientists, but then it’s possible not just to generate cats, which we all see on the internet now, there are very beautiful pictures, but

still this is not the same, it makes our life better, without cats, life is bad, but in fact, i also want some industrial applications, and this is where it should go next, even such a reasonable hybrid of artificial intelligence with hydrodynamic models, a huge number. ensemble calculations, everything, everything, all the data that exists, but it still doesn’t save from the vorenets butterfly, yes, if we talk about long-term weather forecast, that is , the problem of determinism, yes, the weather forecast, they will not solve for sure, that is, the most minimal error in the initial data, the response can be anything, in general, yes , this is probably a question, a debatable question, which means the point is that in the case of non-rosity, for example, it is still possible to make an assessment of the uncertainty of the forecasts. see in which situations it is very large, where it is smaller, this is the first comment, the second comment, of course, if we assume that somewhere locally, for example, there is some

kind of factory there, something happened somewhere, something polluting began to be released into the atmosphere, this greatly influenced some local microclimate , it is clear that this cannot be predicted, there is no climate model, that is, this is some kind of external intervention, well, it is clear that nothing can be done here, but if we are talking about the fact that there is simply the situation that exists now, it is just like- then continues...

answered today by evgeniy burnaev, professor, head of the skoltech artificial intelligence center. evgeniy vladimirovich, thank you very much for the interesting conversation. well, perhaps, see you again, thank you, see you again, thank you. here, if it’s a sport, then with records, if it’s a holiday, then it’s a national one, we love traditions, honor our history, value family and... strong

relationships, we admire how the country is blossoming, there are still more achievements ahead of us, fall in love with the country stronger at the russia exhibition.

used to watching videos online, stopped working, install, open, watch, russian channels, all series, movies and cartoons, educational programs and documentaries... movies, watch, watch in the app or on the website.

in russia, june 22 is a day of memory and sorrow, the anniversary of the beginning. soviet union. a fierce bloody war began, which ended with the complete defeat of nazi germany and the countries that supported it. the soviet union made the greatest contribution to the victory and suffered the greatest losses. more than 27 million soviet citizens died on the battlefields in captivity and occupation.