 Okay, our next presenter today is Isabela Suza Sierra from University of Texas at El Paso. The title of her presentation, this is a mouthful, you're working me, integrating hydrologic and machine learning models to predict lake water temperature profiles and release temperatures throughout the Red River Basin USA. I'm Isabela Suza Sierra and I'm graduating of the University of Texas at El Paso and today I'm going to present some results with my advisor, who is here, thank you so much for your support about how to integrate hydrological models and machine learning techniques to understand wire systems, basically. So I'm going to start with a motivation, why we want to make hydrological models and we want to make hydrological in Calibria hydrological models for the Red River Basin. There is a really crucial hydrological system in the owner of the United States and they are essential for making informed decisions regarding with different key factors like reservoir fissioners and stream flows and societal wired needs. Also machine learning techniques could be applied to enhance hydrological model performance that I'm going to explain, how can we relate both of them, machine learning and hydrology. Here, accurate wire temperatures predictions lead to more effective wire management tasks. Like for example reservoir fissioning tasks and optimized size like reservoirs inflows and also ensure like wire supply for different systems. So it's like a really big thing that can make like a tool, right? So lastly, the successful integration of machine learning into hydrological models has potential to revolutionize like wire sources planning and how we are making decisions not just like in the Red River Basin that is my study area, also we can apply it to other basins by different approaches, right? So what are the objectives of this research? First is to develop and calibrate a hydrological daily model in a software called WIP that is Wired Evaluation and Planning System for the Red River Basin. Second, we want to explore how machine learning can help us to predict temperatures in lakes and reservoirs for find the temperature releases and vertical profiles. And third, we want to explore how to incorporate the machine learning techniques into a new variable of a developed hydrological model WIP to enhance its capabilities for better wire resources management and planning. So after like this introduction of what we want to do this, I want to present the study area. So in the pictures, you can see the Red River Basin. The Red River Basin we see in the map to the right, there is a really big area between Texas and Oklahoma and it's started in New Mexico and it's finished in the Mississippi River and it is a major tributary of the last one. So we can see also the dots that are like the main reservoirs. They have different purposes starting like for wire supply, for diversions and for fishes. And also it's the far longest river in the United States and it has like a really different diverse climate conditions with a really marked gradient to the west that is dry, really dry to the east that is wet. So in the pictures, we can see the Red River Basin in Texas, then in Oklahoma and then close to Louisiana. So we see like the ecosystems are really different and the amount of water that is in the river too. OK, so now I'm going to talk a little bit about how we achieve it, what we wanted to achieve. So I'm going to start with WIP, a model of the Red River Basin. So WIP is a sophisticated user and friendly to that is have been usually wired on the network to define water manage parameters. Right. So it is a little like creating scenarios. So here I have my WIP model. I know it is like a busy map. It has a lot of components, but it reflects how complex is the system in the Red River Basin. We have 90 catchments. Catchment means like subbasin. So we have like the big one that is the Red River Basin, but we have like small washes into the systems that we have 90 catchments. We have also feed in one rivers. We have some groundwater systems connected by the blue lines. And we have also some United U.S. Observe gauge to calibrate the stream flows. So like this is how it looks, like how we build the model, right? The conceptual part. So we have the hydrology, right? We have a water sheet when then we want to what to join the water hydrology, the water sheet hydrology with machine learning. Right. So we are going to join these two to find what the release temperatures for the green dots that I showed you before that are like the reservoirs. So, yeah. So main goal hydrology, machine learning, combine a new variable for in the WIP system is what we want to do. OK, so a key advantage of this methodology is like we can incorporate climate production, production, sorry. So you're asking like why we want to do this because we can incorporate like new climate scenarios and see how the water releases are going to be in the future in the difference and we can try the flow right from the West to the East. OK, so now talking a lot. I'm going to talk a little bit about the how we made the hydrology. So in WIP, there is like a method called soil moisture that it is like a one dimensional two compartments are bucket soil moisture that account empirical equations to find the surface rain. So we have different inflows that is going to be the precipitation. We have also irrigation and we have also like ET. We have some equations for the base floor and interflow. So in the parameters that we see here, like, for example, the roots on conductivity, the conductivity are parameters that we use it for calibrate or model, because we, as I told you, we have like 52 observed stream flow gauges. We were able to calibrate 25 and we are still something that I want to talk later, but we're still working on that, right? So with these soil parameters, we made model calibration to find the best fit for the model versus the observed. So in the next image to the right, we have the conceptual model that someone made for the river basin. So we can see like the interaction between groundwater and the base floor of the stream that is the main one, the river basin. So it's like a general concept, but I'm going to explain later some funds. We caught up on it, but yes. So then I'm going to I'm going to explain a little bit about random forest, that is the technique that we're using. So we have six different training features for our model, right? So we have air temperature, we have solar radiation, effective precipitation for ET, like volume, wiring, inflow and measure depth. So what is, what is random forest is a machinery model that allows us to make decisions based on regression trees. So it makes like, it makes like kind of questions and it goes by paths. So at the end, so like, for example, when you put a new training data set or you put a new data set to find the prediction, it's going to ask to each node that you see the dots. So it's going to add, oh, is this condition met or is that idea? So, oh, this is condition is met. So I'm going this path, if not, I'm going to the other path. And then I will have like an average of predictions. The average of prediction is good because it reduce the overfeeding. So the model is not like kind of replicated and it's learning. So it's like a really good method to and have been used for a lot of tasks in water management. So, yeah, so, and, you know, all final goal is get what? The water temperature by depth, yep. So now I'm going to talk about the results and discussions. So I'm going to start with analysis about the Renault coefficient. But what is Renault coefficient? The Renault coefficient is the relationship of the amount of water that is from precipitation that become like Renault, basically. So when we have like Renault coefficient close to zero, it means that there is no surface right now. There is no water right in the stream. But if we have like higher Renault coefficients or water is becoming like Renault in the surface, right? So one interesting thing is that we found out that it's starting to the west to the east, there is like a train when there were no coefficient one almost zero until it was almost zero point five. So we can see in the precipitation climatology map over the river basin that we have at the west, we have really dry conditions compared to the to the east part of the basin. That is wet between Louisiana and Arkansas. So what we want to say with is that calibrating this hydrological well was really challenging, like, because it's not just like one hydrological system with these conditions, it's just like you have a lot of like soup conditions into a really big basin. So, so, yes, so this is yeah. And so some I'm going to present like some results about the calibration of the hydrological model. So I want to focus on two. One of the west part of the basin, you can see in the map, the the white dot. It represents a dry point. So we can see that the metrics, the national coefficient and the key year are positive. They are no more than zero point five. That is the body that we want to get to say, oh, this is doing good. But they are positive and the bias is low. So it's what we want to get. And also we can see that the runoff coefficient is almost zero. And in the in the graph, we can see the observer versus the model stream flows. Yes. So now I want to show you the difference between wet point, that is the Cado River. We can see that the performance here is higher. Like we have a national coefficient in KGA, almost 0.5. That is a good value. And the bias is less than 5 percent. It means that this system, that this kind of wet system can be very model in terms of our method so far. But we still need to continue working on these. Yes. So real quick about the machinery performance. So remember, we had like air temperature, solar radiation, precipitation, boiling in flow and depth as a training or as an input. To get the temperature released. So in this plot, this is carpo. We can see the performance for the random forest model. We see like good statistics. We see that all the values almost fell down between the 10 percent error bounds and also it's good to know that in the next graph, we have the receivables, the error. So a good, a good goal when you're modeling water temperatures like half errors between minus five and five. So this is the grade here. There are between minus three and four, which means that the model is having good performance. So after all this, our key point is this, like how we can integrate hydrology plus machine learning, right? So you're looking in the image, like the web user interface, right? So we spend a lot of time trying to figure it out how we can make a script because web is just you can just like write scripts in Visual Basin and Java. So it has like the Python theme, but it required work, right? So here it's just like I show you the code and how we in red, we can see how we call the function that we created. And we can see in the next image, like how the model is going to simulate the water temperature by changing the climate projections. So it means what the model is making water predictions in temperature. We need to improve it a lot, but it is working right now, right? So we saw that now the system has the ability to track like hydrology reservoirs, water temperatures. And one key point thing here is like this is a Windows based model, right? But these type of models are the tools that people are on the work use for water planning, because these tools are open, are like user friendly. You can understand easy. So it's how water management is been doing along the way. Yes, so I think I finished it. I want to make like a summary about the research. So first we develop it in Calvary, a daily hydrological model web for the reservoir basin. And we also explore random forests to integrate machine learning into the web ecosystem, and we so simple integrate this model into the hydrology, and it can be done for water punishment. Planning, sorry. Future work, a lot of it, validation and stream floods. We need to validate because we have the calibration, right? For some gauges, but we need to validate them. The validation period has to be one that we didn't use, so 2010 to 2020, and also the water temperatures observations. We need to validate the water temperatures. The machine learning model, we need to validate it. Calibrate the reservoirs. We have 38 reservoirs. We need to calibrate all of them. And also the additional stream flow that we didn't calibrate because of the amount of time. And for the machine learning model, we need to improve it a lot. So we need to add some variables like wind, man depth, and also the climate variables needs to be averaged by seven days time step. And yes, that's all. Thank you so much for listening me. Thank you so much to Jerry for giving me the opportunity to be here. It was amazing, my amazing mentor. Thank you so much, all my friends. It was a really nice experience. Thank you. Good luck with questions. I think she answered them all along the way. Good job. Do we have any questions from the audience? Yeah. Oh, you're fine. Thanks for that presentation. That's really great. Can you explain to me as someone who doesn't know a lot about water management at all, the release temperature? I understand that that's the temperature profile of the water that gets through your river. How how is that used in water management? Like how what do we use that for? I can imagine that people would want to know like a volume of flow or availability of water. Like what exactly is the temperature used for? Do you know? Yeah. Like, for example, we the main goal is like the vertical profile, right? So with the vertical profile, we can make like planning for fishing, right? Like, oh, how these species are going to survive to this temperature and ecosystem, things like that. So we we call it is like water release because I don't know. I need to like to I don't know how to. I need to probably draw, but it's like the profile. I don't know how to do this. So the profile, right? So we want to see like the model. We want to create the variable that makes like some physical. So it's going to be at that point, how much water that I'm releasing from the reservoir is going to be the temperature body, the vertical temperature. Oh, this step, right? Is the release that body is going to be the vertical of all the reservoir. I don't know if that makes sense. Yeah, yeah, that makes sense to me. I guess I I didn't expect the water temperature in a river to be that variable that it would. Oh, it's going to be for like the water that from the reservoir, the temperature I'm giving to the river. OK, OK, I get that. Thanks. I'll say they strategically release water from a reservoir to meet a downstream temperature target. Oh, see, there you go. Yeah, I did not understand that. Thank you so much. And what should I repeat? We they make a release. They release of a specific temperature to meet a downstream temperature target. They make a release temperature from a reservoir to certain elevation in the reservoir or a controlled release to meet a downstream temperature target for fisheries. Yeah, you know, cold water fisheries. Yes, if I may ask another question, you talked about the runoff coefficient and you said it was the ratio between precipitation and no, so maybe can you repeat that? They they a stream flow versus the precipitation. OK, perfect. Thank you. So why did you decide to use random forest instead of another method for machine learning? OK, so we we explore it like before this I explored some methods for wire temperature and it was the best performance between support vector regression, LSTM and a gradient boosting bot. The LSTM there is like a neural network method was we wanted to use it because it was it has like a close performance to the random forest, but we were we were enabled into the whip environment. So we're using random forest because got the best performance, basically, of the other models. Yeah, any other question? Great presentation. My question is so going back to your results, a lot of your data is from the late 1990s to about 2010. And I was just wondering if like you're trying to find have you considered finding a way to verify those results with observed data that may have been collected instead of predicted? And if so, like, I mean, it may be not even be accessible like for those years. So I'm just curious if you've thought about that. Oh, yeah, we will do a validation between 2010 and 2020. So this is like the collaboration data, the collaboration times and the validation is going to we already have it. So we need to do the validation. Yeah. OK. Yeah, very interesting work, Isabella. Another question. OK, great. Thank you so much.