Urban streams in rapidly growing cities are increasingly susceptible to sudden rises in water levels during intense precipitation events due to the lack of public investment or extreme natural events. This work investigates water level peaks and significant flooding events in the Mburicaó stream in Asunción, Paraguay. We develop two predictive models using ten-minute interval data from three weather stations: a support vector machine classifier (SVM) to detect threshold-exceeding events and a multilinear regression model to estimate peak water levels. We employ mutual information analysis combined with cross-correlation between stream level and lagged precipitation to inform the selection of predictive input features. Our results indicate that with rainfall data from 40 minutes before the peak of a flood event, the regression model obtains a coefficient of determination (R2) of 0.8209, an RMSE of 0.3509 meters, and a MAPE of 31.27%. While the classification model, with a 50-minute prediction horizon, targeting occurrences with peaks larger than 1.0 meters, achieves an F1-score of 0.66 with 85% recall and 54% precision. The classification performance still shows potential for improvement, especially in lowering false positives by having longer time series of data or considering hydrological models. These models could be the basis of early warning systems in data-scarce urban environments, offering a practical tool for risk mitigation in Paraguay.

Dengue fever remains a persistent and growing public health concern in many tropical and subtropical regions, where warm climates and urbanization create ideal conditions for mosquito proliferation, such as in Asunción, Paraguay, our focal city. Epidemic control strategies targeting the control of mosquito populations, particularly those of the Aedes aegypti (AE), are essential for mitigating outbreaks and protecting vulnerable communities. This work aims to tackle this problem by introducing a computer vision system that utilizes the You Only Look Once (YOLO) architecture for counting mosquitoes and classifying their sex in real-time, while also addressing issues related to out-of-focus and incomplete specimens. The distilled YOLOv8-nano model reached an mean Average Precision (mAP)50of 88.7% for female mosquitoes and 94.1% for male mosquitoes, with inference times around 4.3 milliseconds on Graphics Processing Unit (GPU) and 1 second on a Raspberry Pi 5. The results indicate that the model can be integrated into embedded systems for automated vector surveillance.

Accurate prediction of river levels is essential to anticipate extreme events such as floods and low flows. This paper presents a short-term forecasting model of the Paraguay River level for the port of Asunción, based on Gated Recurrent Unit (GRU) networks reinforced with hydro-meteorological covariates and an adaptive retraining scheme. To achieve the above, this study incorporates cumulative flows and rainfall from multiple stations, and a Bayesian search was applied to select the optimal combinations of variables and adjust the model's hyperparameter.The approach was evaluated on daily data between 1995 and 2022, using temporal cross-validation with a sliding window. Compared to a baseline model already superior to traditional methods, the proposed model achieved an Nash-Sutcliffe model Efficiency coefficient (NSE) of 0.9455 and reduced the Error Porcentual Absoluto Medio (MAPE) by half. In addition, it maintained higher predictive stability over a 28-day horizon in complex hydrological periods. These results demonstrate the potential of GRU models enriched with multi-source information as practical tools for water management in vulnerable watersheds.

In the CONNIE experiment, 14 charge-coupled device (CCD) sensors record particle interactions near a nuclear reactor in Brazil. Muons, originating from cosmic rays, create background noise that can hinder the detection of neutrino interactions. This study proposes an intelligent system for re-constructing muon traces within the sensor data. To achieve this, a Convolutional Neural Network (CNN) model based on YOLOv8 was developed. Two datasets containing real experimental images were prepared: one for training the model to identify and classify muonic events, and another for calibrating an algorithm to predict muon trajectories based on the impact characteristics in the detectors. The system achieved a detection success rate exceeding 83 % for single muons in real data, demonstrating its potential for mitigating noise and enhancing particle tracking in physics experiments.

This study aimed to develop accurate short-term air quality forecasting models for Asunción, Paraguay using machine learning algorithms. The dataset, which consisted of PM concentrations, air quality indices, meteorological variables, and traffic information, spanned a duration of 18 months. Two machine learning algorithms, Support Vector Regression (SVR) and XGBoost, implemented for window-based regression, were optimized using a grid search and Bayesian optimization process. The optimization explored the impact of the look-back window$(w)$, sampling frequency$(f)$, and prediction horizon$(h)$on model performance. The models were evaluated through a hold-out scheme and trained and evaluated for all stations in the network. Results showed that the XGBoost model outperformed the SVR model, with a mean absolute error of 2.774, 4.643, and 7.819 for AQI 2.5 predictions using 6, 12, and 24-hour horizons, respectively. The optimization process significantly improved the performance of both models compared to using default settings. The study's findings have implications for public health decisions and could be used to issue alerts or advisories when air pollution levels are expected to be high.

Short-term electricity demand forecasting is a fundamental part of the decision-making process of entities involved in electricity consumption management, since it allows the development of strategies to deal with variations in electricity demand in short periods of time. Developing a highly-accurate predictive model is necessary to understand and reflect the consumption behavior, as well as to adjust the generation program essentially to meet the demand at each moment. Therefore, a performance comparison has been made between statistical, machine learning and deep learning models for short-term forecasting. The deep learning models are based on recurrent neural networks, incorporating attention mechanisms in some of them. Hyperparameter tuning was also applied using a Bayesian optimization algorithm. A dataset was also developed including historical electricity demand and external factors such as weather and calendar variables recorded in Paraguay from 2009 to 2022. Models were evaluated from a set of numerical experiments using classical error metrics such as: mean squared error, root mean square error, mean absolute error and mean percentage absolute error. In addition, new special measures were introduced to analyze the error in this type of applications: the percentage error at peak hour and the maximum percentage error of the day to analyze the error in certain events.

CONNIE experiment installed 12 charge- coupled devices (CCDs) sensors near the Angra II nuclear reactor in Angra dos Reis (Brazil) aiming to detect low energy antineutrinos produced in the core of the reactor. For two years, these sensors recorded particle images catching mainly muons and other particles such as electrons and alphas that will be considered as external radioactive background that must be removed. The images were created from the data taken every 3 hours, generating in this way a vast catalog of detected events on each CCD. In this work we propose an instance segmentation and a classification model in order to study the variation of the muon rate produced by those particles. For this purpose we developed two models: a Convolutional Neural Network (CNN) for event classification, and an instance segmentation model for identifying overlapped events. The classification model demonstrated exceptional efficacy, achieving an impressive accuracy of 0.8. Furthermore, precision and recall values of 0.85 and 0.92, respectively, were achieved for the particles of interest. Within the domain of bounding box detection, our model exhibited remarkable recall and precision rates of 71 % and 69 %, respectively, further underlining its adeptness in accurately localizing objects. Our results not only illuminate the potential of these models but also contribute to a deeper understanding of muon rate variations within the experimental context of the CONNIE setup. Index Terms- muon, deeplarning, yolo V8, yolo, CONNIE, neutrine.

Emerging infectious diseases such as COVID-19, caused by the SARS-CoV-2 virus, require systematic strategies to assist in the discovery of effective treatments. Drug repositioning, the process of finding new therapeutic indications for commercialized drugs, is a promising alternative to the development of new drugs, with lower costs and shorter development times. In this paper, we propose a recommendation system called geometric confidence non-negative matrix factorization (GcNMF) to assist in the repositioning of 126 broad spectrum antiviral drugs for 80 viruses, including SARS-CoV-2. GcNMF models the non-Euclidean structure of the space using graphs, and produces a ranked list of drugs for each virus. Our experiments reveal that GcNMF significanlty outperforms other matrix decomposition methods at predicting missing drug-virus associations. Our analysis suggests that GcNMF could assist pharmacological experts in the search for effective drugs against viral diseases.

Dengue fever represents a public health problem and accurate forecasts can help governments take the best preventive actions. As the volume of data provided continuously increases, machine learning and deep learning (DL) models have become an attractive approach. However, it is difficult to perform accurate predictions in areas with fewer cases. In this work, we compare traditional approaches such as LASSO Regression (LR), Random Forest (RF), Support Vector Regression (SVR) vs DL models based on long short-term memory (LSTM), considering weekly dengue incidence and climate, in 217 cities in Paraguay. Several city models may present heterogeneous behaviors and poor accuracy. To mitigate this problem, a clustering analysis between time series is performed based on silhouette scores and measuring how well an observation is clustered. Our results indicate the hierarchical clustering combined with Spearman correlation is the most appropriate approach. Then several LSTM models are compared on subgroups of similar time series. The root mean squared error (RMSE) confirms that the LSTM clustered models improve the accuracy by 31.6% approximately. The main contribution of this work is that LSTM clustered models can perform predictions in cities with low incidence by combining information from similar time-series and weather data.