An interpretable DeePMD-kit performance model for emerging supercomputers

Accurate prediction of solar irradiation, particularly Diffuse Horizontal Irradiance (DHI), is essential for optimizing solar energy systems. This paper presents a hybrid framework that integrates satellite imagery and satellite-derived data to improve the precision of DHI forecasts. Two key datasets were employed: ground-based Solar Irradiation Measurement (SRM) and satellite-based Solar Irradiation (SSR). The proposed methodology utilizes machine learning models along two paths. The first path processes satellite imagery using advanced techniques, including pixel imputation with a modified Random Forest (RF) and Generative Adversarial Network (GAN), noisy region detection with Self-Organizing Maps (SOM), and noise removal through a Latent Diffusion Model (LDM). The second path handles tabular data through a diffusion probabilistic model designed for missing data imputation. These features are then merged and fed into a Long Short-Term Memory (LSTM) network, enhanced to capture seasonality and trends for accurate DHI predictions. The paper makes significant contributions by introducing a robust hybrid framework that leverages both satellite imagery and tabular data, incorporating novel preprocessing methods such as GAN-based pixel imputation and diffusion-based noise removal. The LSTM model is further adapted to handle seasonal and trend components, resulting in enhanced DHI forecasting accuracy. The proposed model demonstrates superior performance, achieving a Mean Squared Error (MSE) of 8.170 W/m2, Root Mean Squared Error (RMSE) of 1.749 W/m2, Mean Absolute Error (MAE) of 0.693 W/m2, and an R-squared value of 2.574, significantly outperforming conventional methods such as Artificial Neural Networks (ANN), Support Vector Regression (SVR), and traditional LSTM models. Additionally, the model records a low Mean Absolute Percentage Error (MAPE) of 0.8183% and a computation time of 35.4 s, highlighting its efficiency. These results contribute to more accurate and reliable solar energy resource management by offering enhanced DHI forecasts across various spatial and temporal scales.

Load forecasting is a key element in the performance of an electrical power system because of the associated economic constraints. The amount of load consumed depends on a number of factors including the environment, time of the day and day of week etc. Such complex features make the load forecasting task a tricky one, warranting efficient techniques for the problem. In this paper, four machine learning models named: K-Nearest Neighbors (KNN) based regression, Deep Neural Network (DNN) based regression, Random Forest (RF) based regression and Decision Tree (DT) based regression have been employed for the load forecasting purpose on an hourly historical load consumption dataset of a household in Islamabad, Pakistan, aggregated over a period of four years. The results show that the DT model gives the best forecasting results as compared to the RF, DNN and KNN algorithms, yielding a very low Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE).

Rainfall is the height of rainwater that falls on a flat area, assuming it doesn't evaporate, doesn't seep, and doesn't flow. Rain levels are measured in mm (millimeters). The target of the research being conducted is in Badung Regency, Bali because Bali is a tourist area that is often visited by tourists and from Indonesian itself, so predictions of meteorology, such as rainfall will greatly impact tourism. In this test, predictions use the Long Short Term Memory (LSTM) method, using daily weather data from the BMKG from 2010 to 2020 as training data and daily weather data for 2021 as prediction data. Based on the test results above, the results show that the two LSTM tests with LSTM Model 128.64 and LSTM Model 64.32 have low MAE and MAPE error values. From First Scenario, the Mean Absolute Error (MAE) value is 8.97246598930908 and Mean Absolute Percentage Error (MAPE) value is 1.7657206683278308%. From Second Scenario, the Mean Absolute Error is 9.706669940783014 and Mean Absolute Percentage Error is 1.9028466692362323%. From the MAE and MAPE values obtained in these two scenarios, it can be proven that from the evaluation results of Rainfall predictions in Badung Regency, Bali, the predictions can be said to be very accurate because they have an error value of less than 10.

This study showcased the Markov switching autoregressive model with time-varying parameters (MSAR-TVP) for modeling nonlinear time series with structural changes. This model enhances the MSAR framework by allowing dynamic parameter adjustments over time. Parameter estimation uses maximum likelihood estimation (MLE) enhanced by the Kim filter, which integrates the Kalman filter, Hamilton filter, and Kim collapsing, further refined by the Nelder–Mead optimization technique. The model was evaluated using U.S. real gross national product (GNP) data in both in-sample and out-of-sample contexts, as well as an extended dataset to demonstrate its forecasting effectiveness. The results show that the MSAR-TVP model improves forecasting accuracy, outperforming the traditional MSAR model for real GNP. It consistently excels in forecasting error metrics, achieving lower mean absolute percentage error (MAPE) and mean absolute error (MAE) values, indicating superior predictive precision. The model demonstrated robustness and accuracy in predicting future economic trends, confirming its utility in various forecasting applications. These findings have significant implications for sustainable economic growth, highlighting the importance of advanced forecasting models for informed economic policy and strategic planning.

Power forecasting models offer valuable insights on the electricity consumption patterns of clients, enabling the development of advanced strategies and applications aimed at energy saving, increased energy efficiency, and smart energy pricing. The data collection process for client consumption models is not always ideal and the resulting datasets often lead to compromises in the implementation of forecasting models, as well as suboptimal performance, due to several challenges. Therefore, combinations of elements that highlight relationships between clients need to be investigated in order to achieve more accurate consumption predictions. In this study, we exploited the combined effects of client similarity and causality, and developed a power consumption forecasting model that utilizes ensembles of long short-term memory (LSTM) networks. Our novel approach enables the derivation of different representations of the predicted consumption based on feature sets influenced by similarity and causality metrics. The resulting representations were used to train a meta-model, based on a multi-layer perceptron (MLP), in order to combine the results of the LSTM ensembles optimally. This combinatorial approach achieved better overall performance and yielded lower mean absolute percentage error when compared to the standalone LSTM ensembles that do not include similarity and causality. Additional experiments indicated that the combination of similarity and causality resulted in more performant models when compared to implementations utilizing only one element on the same model structure.

The objective of this research is to develop a solution for predicting BRI stock prices using Long Short-Term Memory (LSTM) models. The LSTM model was selected for its capacity to process extensive time series data and discern latent temporal patterns. In this study, a BRI stock dataset obtained from Yahoo Finance is employed for the training and testing of an LSTM model. The evaluation results demonstrate that the LSTM model exhibits excellent predictive performance, with a mean absolute percentage error (MAPE) of 1.58768% and a root mean square error (RMSE) of 81.88216%. The Google test results demonstrate a low mean absolute percentage error (MAPE) of 1.5%, indicating a strong correlation between the predicted and true values. In other words, the RMSE values indicate the absolute error level in predictions, indicating the extent to which the model performs well when predicting a value that takes into account the context of the data. In conclusion, the proposed LSTM model shows promise for use in stock price prediction applications. The precision of these models can be tested by using them to make predictions, which would validate the decision-making supported by data. This research suggests that there is room for improvement of these models using techniques such as hyperparameter optimization or ensemble methods (bagging with other weak learners, etc.) to improve their accuracy.

The rapid development of technology has led to an ever-increasing demand for electrical energy. In the context of Timor-Leste, which still relies on fossil energy sources with high operational costs and significant environmental impacts, electricity load forecasting is a strategic measure to support the energy transition towards the Net Zero Emission (NZE) target by 2050. This study aims to utilize historical electricity load data for the period 2013–2024, as well as data on external factors affecting electricity consumption, to forecast electricity load in Timor-Leste in the next 10 years (2025–2035). The forecasting results are expected to support efforts in energy distribution efficiency, reduce operational costs, and inform decisions related to the sustainable energy transition. The method used in this study consists of two main approaches: the causality method, represented by the econometric Principal Component Analysis (PCA) model, which involves external factors in the data processing process, and the time series method, utilizing the LSTM, XGBoost, and hybrid (LSTM+XGBoost) models. In the time series method, data processing is combined with two approaches: the sliding window and the rolling recursive forecast. The performance of each model is evaluated using the Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). The model with the lowest MAPE (

Sales forecasting is essential for businesses to predict future demand and inform strategic and operational planning, especially in the building materials retail industry. Accurate sales prediction supports inventory management, cost control, and supply chain efficiency. This study compares the performance of 3 deep learning models, Convolutional Neural Network (CNN), Multilayer Perceptron (MLP), and Long Short-Term Memory (LSTM), in forecasting daily iron sales at PT Surya Aneka Bangunan from 2016 to 2020. The models were trained on 80% of the historical data and tested on 20%. Model performance was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Coefficient of Determination R². The results show that the CNN model achieved the best performance with an MAE of 0.293, RMSE of 0.357, MAPE of 0.081, and R² of 0.9989, indicating high accuracy and stability. The MLP model produced higher errors, while the LSTM model had the lowest MAPE but greater error variability. These findings suggest that the CNN model is the most reliable for capturing temporal patterns in iron sales data. The study contributes to the development of adaptive sales forecasting systems and opens opportunities for applying similar methods in other retail sectors to support data driven decision making.

This paper focuses on the Least Square (LS) regression using the mean and Quantile (M) regression analysis using median which is based on “well-Known” parametric estimation methodologies. Data from Oregon and California highway bridges were used for the comparison of the two methods. Relationships were developed to predict the unit cost of FRP repair work and FRP cost was found to have a high degree of correlation with FRP area for both Oregon and California. It was observed that the Cost Estimating Relationships (CERs) obtained by Quantile (M) regression method had the smaller Mean Absolute Deviation (MAD) values and lower Mean Absolute Percentage Error (MAPE) values than Least Square (LS) regression. The stuudy showed that Quantile Regression is much less sensitive to outliers than Least Squares Regression.

Geographic location and neighbourhood attributes are major contributors to residential property values. Automated valuation models (AVM) often use hedonic pricing with location and neighbourhood attributes in the form of numeric and categorical variables. This paper proposed a novel approach to automated property valuation using a machine learning model with a convolutional neural network (CNN), fully connected neural network layers with numeric and categorical variables. In this study we compare the results of a fused model, which treat geographical data as an input with the performance of the baseline neural network model with only numerically or categorically represented data. Furthermore, the residential valuation by the proposed fused model was tested with actual sold price data in Greater Sydney, Australia. The study found that the fused model produced valuations with a significantly lower mean absolute percentage error (MAPE) (8.71%) than the MAPE of the baseline model (11.59%). The results show that the fused model with CNN significantly improves the accuracy for residential valuation, reducing spatial information loss by data manipulation and distance calibration.

Accurate estimation of water acidity is essential for characterizing acid mine drainage (AMD) and designing effective remediation strategies. However, conventional approaches, including titration and empirical estimation methods based on iron speciation, often fail to account for site-specific geochemical complexity. This study introduces a high-accuracy, site-specific empirical model for predicting acidity in AMD-impacted waters, developed from field data collected at the Trimpancho mining complex in the Iberian Pyrite Belt (Spain). Using multiple linear regression (MLR), a robust predictive relationship was established based on Cu, Al, Mn, Zn, and pH, achieving a coefficient of determination (R²) of 99.2%. The model significantly outperforms the standard Hedin method, with a lower mean absolute percentage error (13% vs. 29%). Results also reveal strong spatial and seasonal hydrochemical variability, underscoring the limitations of generalized acidity models in such environments. This work demonstrates the applicability of site-calibrated multivariate models as practical tools for enhancing acidity prediction in complex AMD systems.

Globally, buildings are responsible for an estimated 40% of energy consumption and 33% of CO2 emissions. In a bid to reduce CO2 emissions and hence, global warming, it has become necessary to ensure the energy efficient construction and operation of buildings. Understanding how a building utilises energy is a critical step to increase its efficiency. In this study, we leverage on an open-source data obtained from UCI data repository. Exploratory data analysis and feature engineering were used to eliminate non-contributing features while identifying key attributes of the data for model training. Linear Regression (LR) and Support Vector Regression (SVR) were employed as the machine learning techniques for the study. The models were trained using a repeated cross-validation technique. The models’ performance was evaluated on an independent data set segregated for testing. The LR model was trained with nine out of thirty-three features, while the Support Vector Regression (SVR) model used twenty-eight features for its training. The SVR model had a higher variance (0.48), accuracy (92.41%), and lower Mean Absolute Percentage Error (MAPE) of 7.59% compared to the LR model's variance of 0.26, accuracy of 91.87%, and MAPE of 8.13%. The SVR model was more accurate in predicting energy consumption, as it showed better accuracy on the test set with lower MAPE and higher R-squared value. Both models outperformed a relatively complex and computationally expensive model in a previous study. It also identified areas with high energy consumption which could be used to inform the building's energy management strategy.

Background and ObjectiveBuilding on our previous success in artificial intelligence–driven food classification, this study aims to enhance portion size estimation methods, with a particular focus on rice and congee, using milk as a low-glycemic index reference in selected Taiwanese hospitals.MethodsWe used a RealSense D435i depth camera to capture both 2-dimensional (2D) and 3-dimensional (3D) images of rice, congee, and milk served in standard hospital bowls. The images were labeled and classified by a convolutional neural network, Xception, whereas food weight and volume were estimated using 25 depth points with multiple perceptron (MLP) models. MLP models were trained on 80% of the data set and tested on the remaining 20%, with predictions refined through a weighted ensemble of selected MLP models. We validated the MLP model performance through 3 comparisons: 1) MLP with traditional statistical methods (linear regression), 2) MLP with clinical dietitians, and 3) MLP performance across 3 different hospital settings.ResultsThe Xception model accurately identified foods with high glycemic index and recognized milk as a low-glycemic reference, achieving a training accuracy of 1.0 with the loss approaching zero by epoch 22. Test accuracy stabilized early with consistently low loss. MLP models effectively predicted food weights, with models E, F, and G yielding the lowest mean squared errors for rice, congee, and milk, respectively. Across all food types, MLP models consistently outperformed traditional linear regression, showing lower mean absolute percentage error (MAPE: 0.3%–3.3%) than linear regression (MAPE: 4%–9.2%). When compared with estimates from 6 experienced dietitians (MAPE: 39%–41%), the MLP models also demonstrated markedly higher accuracy. Performance remained stable across 3 hospital sites, indicating strong generalizability.ConclusionsThe Xception and MLP models effectively quantify foods with high glycemic index and identify milk as a low-glycemic reference, demonstrating strong predictive performance and generalizability. Further studies are warranted to extend this approach to other food types.

With an estimated offshore potential wind energy capacity of 575 gigawatts, the Great Lakes region is a promising area for future wind energy development. Electric utilities that use wind energy rely on accurate day-ahead wind energy forecasts, mainly informed by predicted wind speed, to account for the variability of wind energy production. Since wind turbines have not yet been developed in the Great Lakes, observational wind data is collected at very few sites. Hence, utilities may seek to use synthetic wind data to pre-train accurate day-ahead wind speed prediction models. Yet, prior studies have only utilized synthetic data in predicting sub-hourly winds speeds. We hypothesized that training long short-term memory neural networks to predict day-ahead wind speeds on synthetic wind data instead of observational wind data would increase accuracy since synthetic wind data is available over longer time spans than observational wind data. We used observational data from the Lake Michigan Wind Assessment and synthetic data from the National Offshore Wind Great Lakes dataset and found that networks trained on synthetic data had a lower mean absolute percentage error score in day-ahead wind speed prediction than networks trained on observational data. We also optimized additional parameters of the networks for both synthetic and observational network types, further improving accuracy. The availability of a wind speed prediction model trained on synthetic data will reduce reliance on historical observational data at future sites of wind energy infrastructure, allowing utilities to swiftly adapt accurate prediction methods to new sites.

Network Attention of tourist attractions serves as a crucial basis for tourism analysis and development. Accurate prediction of Network Attention can provide data support for enhancing the management level of tourist attractions. This paper proposes a Prophet-LSTM hybrid model for forecasting the network attention of tourist attractions. Using network attention data from five attractions in Nanjing, Jiangsu Province, spanning 2017 to 2023, we first analyze seasonal fluctuations and the impact characteristics of special events (e.g., the pandemic) to help the subsequent model construction. Based on this, the Prophet model extracts the linear components of the network attention data series, while the Long Short-Term Memory (LSTM) network models the nonlinear components. Finally, grid search is employed to determine the hybrid weights. The study finds that the Prophet-LSTM hybrid model achieves significantly lower Mean Absolute Percentage Error and Root Mean Square Error compared to ARMA, standalone Prophet, and standalone LSTM models in data fitting, alongside a higher Coefficient of Determination; In predicting future network attention, the model’s predictions exhibit lower MAPE and RMSE than individual Prophet and LSTM models, with a relative prediction error ranging from 1.0044% to 30.0157%. Thus, this model enables more accurate forecasts for the network attention of tourist attraction, providing robust support for enhancing attraction management and optimizing resource allocation.