Improving Recommendation Techniques by Deep Learning and Large Scale Graph Partitioning

BackgroundMore than 79.2 million confirmed COVID-19 cases and 1.7 million deaths were caused by SARS-CoV-2; the disease was named COVID-19 by the World Health Organization. Control of the COVID-19 epidemic has become a crucial issue around the globe, but there are limited studies that investigate the global trend of the COVID-19 pandemic together with each country’s policy measures.ObjectiveWe aimed to develop an online artificial intelligence (AI) system to analyze the dynamic trend of the COVID-19 pandemic, facilitate forecasting and predictive modeling, and produce a heat map visualization of policy measures in 171 countries.MethodsThe COVID-19 Pandemic AI System (CPAIS) integrated two data sets: the data set from the Oxford COVID-19 Government Response Tracker from the Blavatnik School of Government, which is maintained by the University of Oxford, and the data set from the COVID-19 Data Repository, which was established by the Johns Hopkins University Center for Systems Science and Engineering. This study utilized four statistical and deep learning techniques for forecasting: autoregressive integrated moving average (ARIMA), feedforward neural network (FNN), multilayer perceptron (MLP) neural network, and long short-term memory (LSTM). With regard to 1-year records (ie, whole time series data), records from the last 14 days served as the validation set to evaluate the performance of the forecast, whereas earlier records served as the training set.ResultsA total of 171 countries that featured in both databases were included in the online system. The CPAIS was developed to explore variations, trends, and forecasts related to the COVID-19 pandemic across several counties. For instance, the number of confirmed monthly cases in the United States reached a local peak in July 2020 and another peak of 6,368,591 in December 2020. A dynamic heat map with policy measures depicts changes in COVID-19 measures for each country. A total of 19 measures were embedded within the three sections presented on the website, and only 4 of the 19 measures were continuous measures related to financial support or investment. Deep learning models were used to enable COVID-19 forecasting; the performances of ARIMA, FNN, and the MLP neural network were not stable because their forecast accuracy was only better than LSTM for a few countries. LSTM demonstrated the best forecast accuracy for Canada, as the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) were 2272.551, 1501.248, and 0.2723075, respectively. ARIMA (RMSE=317.53169; MAPE=0.4641688) and FNN (RMSE=181.29894; MAPE=0.2708482) demonstrated better performance for South Korea.ConclusionsThe CPAIS collects and summarizes information about the COVID-19 pandemic and offers data visualization and deep learning–based prediction. It might be a useful reference for predicting a serious outbreak or epidemic. Moreover, the system undergoes daily updates and includes the latest information on vaccination, which may change the dynamics of the pandemic.

Accurate forecasting of diabetes burden is essential for healthcare planning, resource allocation, and policy-making. While deep learning models have demonstrated superior predictive capabilities, their real-world applicability is constrained by computational complexity and data quality challenges. This study evaluates the trade-offs between predictive accuracy, robustness, and computational efficiency in diabetes forecasting. Four forecasting models were selected based on their ability to capture temporal dependencies and handle missing healthcare data: Transformer with Variational Autoencoder (VAE), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and AutoRegressive Integrated Moving Average (ARIMA). Annual data on Disability-Adjusted Life Years (DALYs), Deaths, and Prevalence from 1990 to 2021 were used to train (1990–2014) and evaluate (2015–2021) the models. Performance was measured using Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). Robustness tests introduced noise and missing data, while computational efficiency was assessed in terms of training time, inference speed, and memory usage. Statistical significance was analyzed using ANOVA and Tukey’s post-hoc tests. The Transformer-VAE model achieved the highest predictive accuracy (MAE: 0.425, RMSE: 0.501) and demonstrated superior resilience to noisy and incomplete data (p < 0.01). LSTM effectively captured short-term patterns but struggled with long-term dependencies, while GRU, though computationally efficient, exhibited higher error rates. ARIMA, despite being resource-efficient, showed limited capability in modeling long-term trends, indicating potential benefits in hybrid approaches. While Transformer-VAE provides the most accurate diabetes burden forecasting, its high computational cost and interpretability challenges limit its scalability in resource-constrained settings. These findings highlight the potential of deep learning models for healthcare forecasting, while underscoring the need for further validation before integration into real-world public health decision-making.

Fractional-order models (FOMs) have been recognized as superior tools for capturing the complex electrochemical dynamics of lithium-ion batteries, outperforming integer-order models in accuracy, robustness, and adaptability. Parameter identification (PI) is essential for FOMs, as its accuracy directly affects the model’s ability to predict battery behavior and estimate critical states such as state of charge (SOC) and state of health (SOH). In this study, a hybrid deep learning approach has been introduced for FOM PI, representing the first application of deep learning in this domain. A simulation platform was developed to generate datasets using Sobol and Monte Carlo sampling methods. Five deep learning models were constructed: long short-term memory (LSTM), gated recurrent unit (GRU), one-dimensional convolutional neural network (1DCNN), and hybrid models combining 1DCNN with LSTM and GRU. Hyperparameters were optimized using Optuna, and enhancements such as Huber loss for robustness to outliers, stochastic weight averaging, and exponential moving average for training stability were incorporated. The primary contribution lies in the hybrid architectures, which integrate convolutional feature extraction with recurrent temporal modeling, outperforming standalone models. On a test set of 1000 samples, the improved 1DCNN + GRU model achieved an overall root mean square error (RMSE) of 0.2223 and a mean absolute percentage error (MAPE) of 0.27%, representing nearly a 50% reduction in RMSE compared to its baseline. This performance surpasses that of the improved LSTM model, which yielded a MAPE of 9.50%, as evidenced by tighter scatter plot alignments along the diagonal and reduced error dispersion in box plots. Terminal voltage prediction was validated with an average RMSE of 0.002059 and mean absolute error (MAE) of 0.001387, demonstrating high-fidelity dynamic reconstruction. By advancing data-driven PI, this framework is well-positioned to enable real-time integration into battery management systems.

The precise forecasting of groundwater levels significantly influences plant growth and the sustainable management of ecosystems. Nonetheless, the non-stationary characteristics of groundwater level data often hinder the current deep learning algorithms from precisely capturing variations in groundwater levels. We used Variational Mode Decomposition (VMD) and an enhanced Transformer model to address this issue. Our objective was to develop a deep learning model called VMD-iTransformer, which aims to forecast variations in the groundwater level. This research used nine groundwater level monitoring stations located in Hangjinqi Ecological Reserve in Kubuqi Desert, China, as case studies to forecast the groundwater level over four months. To enhance the predictive performance of VMD-iTransformer, we introduced a novel approach to model the fluctuations in groundwater levels in the Kubuqi Desert region. This technique aims to achieve precise predictions of the non-stationary groundwater level conditions. Compared with the classic Transformer model, our deep learning model more effectively captured the non-stationarity of groundwater level variations and enhanced the prediction accuracy by 70% in the test set. The novelty of this deep learning model lies in its initial decomposition of multimodal signals using an adaptive approach, followed by the reconfiguration of the conventional Transformer model’s structure (via self-attention and inversion of a feed-forward neural network (FNN)) to effectively address the challenge of multivariate time prediction. Through the evaluation of the prediction results, we determined that the method had a mean absolute error (MAE) of 0.0251, a root mean square error (RMSE) of 0.0262, a mean absolute percentage error (MAPE) of 1.2811%, and a coefficient of determination (R2) of 0.9287. This study validated VMD and the iTransformer deep learning model, offering a novel modeling approach for precisely predicting fluctuations in groundwater levels in a non-stationary context, thereby aiding sustainable water resource management in ecological reserves. The VMD-iTransformer model enhances projections of the water level, facilitating the reasonable distribution of water resources and the long-term preservation of ecosystems, providing technical assistance for ecosystems’ vitality and sustainable regional development.

This study evaluates the performance of traditional statistical methods, specifically the Seasonal Autoregressive Integrated Moving Average (SARIMA), in comparison to advanced deep learning models—Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and Convolutional Neural Network (CNN)—for wind speed forecasting across various regions in Saudi Arabia: Al-Jouf, Abha, Al-Ahsa, and Al-Dawadami. The historical wind speed dataset covering the year 2018 was utilized, with data preprocessing conducted using the Exploratory Data Analysis (EDA) method to ensure consistency and quality. The models were assessed based on three primary error metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Square Error (RMSE). Results show that CNN consistently outperforms SARIMA and the other deep learning models, particularly when forecasting stable wind speed patterns. LSTM demonstrates an ability to handle fluctuating wind speeds effectively, while BiLSTM offers advantages in capturing complex bidirectional temporal dependencies. On the other hand, SARIMA generally exhibits lower performance compared to deep learning approaches. The superior performance of CNN is likely attributed to its strength in local feature extraction and handling spatial patterns effectively, which are beneficial for short-term forecasting. These findings provide valuable insights into model selection for wind energy forecasting and can contribute to optimizing renewable energy integration and planning strategies. Future work may explore hybrid model approaches to further enhance forecasting accuracy.

This work evaluates machine learning (ML) and deep learning (DL) models for Photovoltaic (PV) power output prediction based on two real-world datasets. Eight models – Support Vector Regression (SVR), K-Nearest Neighbors (KNN), Linear Regression (LR), Random Forest Regression (RFR), Gradient Boosting Regression (GBR), Decision Tree (DT), one Dimensional Convolutional Neural Network (1DCNN) and Artificial Neural Network (ANN) – were evaluated using Root Mean Squared Error (RMSE), R-squared (R²), Mean Absolute Error (MAE), and Mean Squared Error (MSE). For Dataset1 (region of Cacak) 1DCNN achieved the optimal performance, with the reduced errors and the largest R² values at three spatial locations, and it was closely followed by RFR and GBR. Likewise, in Dataset2 (region of Kraljevo), 1DCNN, ANN, and RFR give the best results. Conventional models, such as LR and DT, performed poorly in both data tribes. The results highlight the ability of state-of-the-art ensemble and DL models in learning nonlinear patterns in solar energy data and thus, the importance of choosing the right prediction tools to provide accurate and reliable PV power forecasts. &nbsp;

Deep learning (DL) models are increasingly used to forecast water quality variables for use in decision making. Ingesting recent observations of the forecasted variable has been shown to greatly increase model performance at monitored locations; however, observations are not collected at all locations, and methods are not yet well developed for DL models for optimally ingesting recent observations from other sites to inform focal sites. In this paper, we evaluate two different DL model structures, a long short-term memory neural network (LSTM) and a recurrent graph convolutional neural network (RGCN), both with and without data assimilation for forecasting daily maximum stream temperature 7 days into the future at monitored and unmonitored locations in a 70-segment stream network. All our DL models performed well when forecasting stream temperature as the root mean squared error (RMSE) across all models ranged from 2.03 to 2.11°C for 1-day lead times in the validation period, with substantially better performance at gaged locations (RMSE = 1.45–1.52°C) compared to ungaged locations (RMSE = 3.18–3.27°C). Forecast uncertainty characterization was near-perfect for gaged locations but all DL models were overconfident (i.e., uncertainty bounds too narrow) for ungaged locations. Our results show that the RGCN with data assimilation performed best for ungaged locations and especially at higher temperatures (&amp;gt;18°C) which is important for management decisions in our study location. This indicates that the networked model structure and data assimilation techniques may help borrow information from nearby monitored sites to improve forecasts at unmonitored locations. Results from this study can help guide DL modeling decisions when forecasting other important environmental variables.

Air Quality Index prediction using an effective hybrid deep learning model

&amp;lt;p&amp;gt;Accurate water level (WL) forecasting is important for water resources management and planning purposes in the Great Lakes. The objectives of this research are two-fold.&amp;amp;#160; The first objective is to apply machine learning (ML) (i.e., random forest (RF) and support vector regression (SVR)) and hybrid convolutional neural network(CNN)-long-short term memory (LSTM) deep learning (DL) models for multi-step (i.e., one-, two- and three-monthly step ahead) WL forecasting in the Great Lakes (Michigan and Ontario). The second objective is to integrate the boundary corrected (BC) maximal overlap discrete wavelet transform (MODWT) with SVR, RF, and CNN-LSTM models to improve the performance of the individual models. By employing a BC-wavelet decomposition method, the &amp;amp;#8216;future data&amp;amp;#8217; issue (i.e., data from the future that is not available), often overlooked in the literature and a major barrier to achieving realistic forecasting performance is overcome.&amp;amp;#160;&amp;lt;/p&amp;gt;&amp;lt;p&amp;gt;For Lakes Michigan and Ontario, 1212 monthly WL (m) records (spanning Jan 1918&amp;amp;#8211;Dec 2018) were used to develop the models. For the non-wavelet-based models (SVR, RF, and CNN-LSTM), candidate model inputs included the WL recorded over the previous 12 months.&amp;amp;#160; For the BC-MODWT-based models (BC-MODWT-SVR, BC-MODWT-RF, and BC-MODWT-CNN-LSTM), the lagged input time series were decomposed into BC-wavelet and scaling coefficients by using different mother wavelets (Haar, Daubechies, Symlets, Fejer-Korovkin and Coiflets), filter lengths (from two up to 12) and decomposition levels (from one up to seven).&amp;amp;#160; For each method (SVR, RF, and CNN-LSTM), mother wavelet, and decomposition level a model was generated.&amp;amp;#160; For both wavelet- and non-wavelet-based models, the particle swarm optimization (PSO) method was used to select the most appropriate inputs to include in the proposed multi-step WL forecasting models.&amp;lt;/p&amp;gt;&amp;lt;p&amp;gt;The datasets were partitioned into calibration and validation subsets. After calibrating the models, various performance evaluation metrics, e.g., coefficient of determination (R&amp;lt;sup&amp;gt;2&amp;lt;/sup&amp;gt;), root mean square error (RMSE), mean absolute error (MAE), root mean square percentage error (RMSPE), mean absolute percentage error (MAPE) and the Nash-Sutcliffe efficiency coefficient (NSC) were used to assess model accuracy.&amp;lt;/p&amp;gt;&amp;lt;p&amp;gt;Of the ML models, the SVR outperformed RF while the DL models outperformed the ML models for each forecast lead time (one-, two-, and three-step(s) ahead). Results from this case study indicate that not all wavelet families and decomposition levels perform equally and in some cases, the wavelet-based models do not improve performance over the non-wavelet-based models. However, the BC-MODWT-CNN-LSTM using suitable mother wavelets (e.g., Haar) outperforms the individual ML and BC-MODWT-ML-based models. More accurate forecasts were obtained for Lake Michigan although the performance in both Great Lakes was accurate. The outcomes of this research indicate that the BC-MODWT-CNN-LSTM model is a promising tool for generating accurate WL forecasts.&amp;lt;/p&amp;gt;

This study investigates the use of deep learning algorithms to predict the discharge coefficient (Cd) of contaminated multi-hole orifice flow meters with circular opening. Datasets (MHO1 and MHO2) were obtained from computational fluid dynamic simulations for two circular multi-hole orifice flow meters of different geometries. To evaluate the performance and generalization capabilities of different models, three distinct scenarios, each involving different dataset configurations and normalization techniques were designed. For each scenario, three deep learning models (feedforward neural networks, convolutional neural network, and recurrent neural network) were implemented and evaluated based on their performance metrics, including mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), and the coefficient of determination (R2). For all three scenarios eight models for each neural network model were developed (FFNN – four models, CNN – two models, RNN – two models). The same structure of models was used across all scenarios to ensure consistency in the evaluation process. Key input parameters include geometrical and flow variables such as β – parameter, contamination thickness, radial distance, Reynolds number, and orifice diameters. Results demonstrate the effectiveness of deep learning in accurately predicting discharge coefficient for different contamination conditions and different geometries. This study showed that deep learning models can be used for prediction of discharge coefficients for multi-hole orifice flow meters of similar geometry, based on data obtained from one orifice flow meter for different contamination parameters.

Tropical cyclones (TCs) pose signif­icant risks, particularly in coastal regions, making accurate prediction of their track and intensity is crucial for effective disaster preparedness and response. Traditional numerical models struggle with balancing accuracy and computational efficiency, although TC track prediction has achieved substantial progress, challenges remain in forecasting TC intensity, especially rapid intensification (RI). This study aims to (1) develop a Transformer-based deep learning (DL) model to predict TC track, intensity, and 24-hour future intensity change simultaneously, and (2) investigate the relative importance of input variables to the contribution of improving model forecast ability. Based on 2001–2021 best track data and ERA5 reanalysis data over the western North Pacific (WNP), we develop an optimal model called OWZP-Transformer, which leverages the multi-head self-attention mechanism and incorporates the 13 input parameters categorized into four factors: basic, environmental, gradient, and structural. Specifically, our study incorporates structural parameters, which is represented by Okubo-Weiss-Zeta Parameter (OWZP). Our OWZP-Transformer model achieves competitive results for track prediction and shows excellent performance in intensity forecasting for the next 6 hour, with an overall root mean square error (RMSE) of 0.91 m/s. This result represents an improvement of 61.9% to 68.4% compared to existing DL models, which generally have RMSE values above 2 m/s. In addition, our model demonstrates superior performance in predicting 24-hour future intensity change, achieving an overall lower mean absolute error (MAE) of 1.57 m/s, which is 63.5% to 74.6% lower than existing DL models. Furthermore, our model successfully identifies all 11 RI events out of 30 TCs samples from WNP test dataset during 2020-2021. We further evaluate the contributions of each parameter for the first time using two explainable feature importance methods: DeepLIFT and DeepLiftShap. The results indicate that self-contributions and the basic factors play a dominant role in short-term forecasts, while the OWZ parameter plays a significant role following them. This study is the first attempt to comprehensively predict a broad range of TC forecasting tasks using a single DL model, highlighting the potential the OWZP-Transformer model as a reliable tool for enhancing both the accuracy and efficiency of TC predictions.

Introduction: The intermittency of solar energy necessitates accurate short-term forecasting for effective energy grid management. Pyranometer Irradiance Recorded Point (PIRP) values — representing instantaneous solar irradiance measurements (in W/m²) — can be predicted using deep learning models. This study introduces a hybrid Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) architecture to forecast PIRP values from sequential sky images, historical PIRP data, and cyclical time-based features, providing a scalable and low-cost alternative to conventional sensor-dependent methods. Objectives: The objectives of this research are to develop a hybrid CNN-LSTM model that leverages spatial features from sky images and temporal dependencies from historical PIRP values, to enhance forecasting accuracy by integrating cyclical time features, and to demonstrate a scalable and cost-effective forecasting approach for smart grid operations, solar farms, and urban energy planning. Methods: The model utilizes a TimeDistributed CNN to extract spatial features from sequences of sky images and LSTM layers to capture temporal dependencies. Timestamp features are encoded using sine and cosine transformations. An attention mechanism is integrated to focus on the most informative parts of the sequence. The curated dataset, comprising 13,818 PIRP records and over 39,000 images, was preprocessed through timestamp alignment, filtering, and feature scaling. Training employed Adam optimization and Huber loss, with extensive evaluation through MAE, RMSE, Filtered MAPE, and R² Score metrics. Results: The proposed model achieved a Mean Absolute Error (MAE) of 14.46, a Root Mean Squared Error (RMSE) of 30.03, a Filtered Mean Absolute Percentage Error (MAPE) of 9.85%, and an R² Score of 0.9737. It significantly outperformed baseline models such as CNN-only, LSTM-only, and Linear Regression, validating its effectiveness in short-term PIRP forecasting. Comparative evaluations confirmed the superiority of the hybrid CNN-LSTM approach with cyclical time feature integration and attention mechanism. Conclusions: The hybrid CNN-LSTM model demonstrates robust performance by accurately capturing spatio-temporal dependencies inherent in PIRP values. Its ability to predict 15-minute-ahead irradiance values with high accuracy highlights its potential for real-world deployment in renewable energy forecasting. Future work will focus on expanding datasets across diverse geographical regions, integrating additional meteorological parameters, and exploring transformer-based architectures for further enhancements.

Globally, wind energy if properly harnessed, could serve as a source of energy generation in Africa. This study compared the performance of two Machine Learning (ML) algorithms (Linear regression and Random Forest) in predicting wind speed in five major cities in Africa (Yaoundé, Pretoria, Nairobi, Cairo and Abuja). Wind data were collected between January 1, 2000, and December 31, 2022, using the Solar Radiation Data Archive. The data preprocessing was carried out with 80% of the data used for training and 20% for validation. The performance of these ML algorithms was evaluated using Mean Square Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE) and coefficient of determination (R2). The result shows that Nairobi (3.814795 m/s) closely followed by Cairo (3.606453 m/s) has the highest mean wind speed while Yaoundé (1.090512 m/s) has the lowest. Based on the performance metrics used, the two Machine Learning algorithms were competitive. Still, the Linear Regression (LR) algorithm outperformed the Random Forest Algorithm in predicting wind speed in all the selected major African cities. In Yaoundé (RMSE = 0.3892, MAE= 0.3001, MAPE =0.5030), Pretoria (RMSE=1.2339, MAE=0.9480, MAPE=0.7450) Nairobi (RMSE= 0.4223, MAE =0.6499, MAPE =0.1872), Nairobi (RMSE=0.6499, MAE=0.5171, MAPE =0.1872), Cairo (RMSE =1.0909, MAE =0.8544, MAPE =0.3541) and Abuja (RMSE = 0.70245, MAE =0.5441, MAPE= 0.4515) the Linear regression algorithms was found to outperformed Random Forest Regression. Therefore, the Linear regression algorithm is more reliable in predicting wind speed compared with the Random Forest regression.

Improved prediction of hourly PM2.5 concentrations with a long short-term memory and spatio-temporal causal convolutional network deep learning model

Monte Carlo (MC)-based estimation of dose-averaged linear energy transfer (LETd) in proton therapy is accurate but computationally intensive. Although deep learning (DL) models offer efficient alternatives, few have been optimized for multi-source inputs or validated on anatomically complex tumors, such as nasopharyngeal carcinoma (NPC). This study developed and evaluated a multi-source DL framework for voxel-level LETd prediction in proton therapy for NPC, aiming to identify optimal input configurations and assess both prediction accuracy and clinical reliability. This study analyzed 100 patients with NPC, yielding 500 proton therapy fields. Ground-truth LETd maps were generated using GPU-based Monte Carlo simulations. A deep learning model was trained with different combinations of four input features: dose distribution, CT image, beam mask, and radiological depth, resulting in eight configurations (Mdose, Mdose+ct, Mdose+beam, Mdose+depth, Mdose+beam+depth, Mdose+ct+depth, Mdose+ct+beam, and Mdose+ct+beam+depth). The dataset was randomly divided into training (300 fields, 60 patients), validation (100 fields, 20 patients), and testing (100 fields, 20 patients) sets. Model performance was evaluated using mean absolute error (MAE), peak signal-to-noise ratio (PSNR), normalized cross-correlation (NCC), structural similarity index measure (SSIM), and 3D Gamma analysis. A 3D U-Net baseline was retrained under identical conditions for comparison. Model interpretability was examined with Gradient-weighted Class Activation Mapping (Grad-CAM), and prediction confidence was assessed using uncertainty maps. The Mdose+ct model achieved the best performance (MAE 0.21±0.03keV/µm, PSNR 28.99±1.33, NCC 0.98±0.01, SSIM 0.971±0.007, Gamma 3%/3mm: 94.69±3.69%), outperforming the 3D U-Net baseline (MAE 0.38±0.52keV/µm). Grad-CAM visualizations confirmed that the model focused on physically relevant regions, while uncertainty maps highlighted high-variability zones at LETd transitions. The model generated voxel-level LETd predictions in approximately 7 s per case, which was substantially faster than GPU-based MC simulations (30 s per case). The proposed multi-source deep learning framework enables rapid and accurate voxel-level LETd prediction for NPC. Combining dose with CT yielded the most reliable input configuration, and incorporating interpretability and uncertainty analyses enhances confidence in its suitability for biologically guided proton therapy planning.