Mean Average Percentage Error Research Articles

The Prediction of Abrasion Resistance of Mortars Modified with Granite Powder and Fly Ash Using Artificial Neural Networks

This paper predicts the abrasion resistance of a cementitious composite containing granite powder and fly ash replacing up to 30% of the cement weight. For this purpose, intelligent artificial neural network (ANN) models were used and compared. A database was built based on mix composition, curing time, and curing method. The model developed to predict the abrasion resistance of the cementitious composites containing granite powder and fly ash was shown to be accurate. It was proved by the very high values of the accuracy parameters that were above 0.93 in the case of the coefficient of the determination R2 and very low values of the errors, which were about 10% in the case of mean average percentage error. This method can be used especially for designing cement mortars with granite powder and fly ash additives replacing cement in a range from 0 to 30% of its weight. These mortars can be used for floors in industrial buildings.

Bitcoin Forecasting Performance Measurement: A Comparative Study of Econometric, Machine Learning and Artificial Intelligence-Based Models

Bitcoin is a type of Cryptocurrency that relies on Blockchain technology and its growing popularity is leading to its acceptance as an alternative investment. However, the future value of Bitcoin is difficult to predict due to its significant volatility and speculative behavior. Considering this, the key objective of this research is to assess Bitcoins’ explosive behavior during 2013–2022 including the most volatile COVID-19 pandemic and Russia–Ukraine war period and to forecast its price by comparing the predictive abilities offive different econometric, machine learning and artificial Intelligence methods namely, ARIMA, Decision Tree, Random Forest, SVM, and Artificial Intelligence Long Short-Term Memory Network (AI-LSTM). The precision of such methodologies has been assessed using root mean square error (RMSE) and mean average per cent error (MAPE) values. The findings confirmed that the AI-LSTM model performs better than other forecast models in predicting Bitcoins’ opening price on the following working day. Therefore, Bitcoin traders, policymakers, and financial institutions can use the model effectively to better forecast the next day’s opening price.

Automatic Prediction of Paediatric Cardiac Output From Echocardiograms Using Deep Learning Models

<h2>Abstract</h2><h3>Background</h3> Cardiac output (CO) perturbations are common and cause significant morbidity and mortality. Accurate CO assessment is crucial for guiding treatment in anaesthesia and critical care, but measurement is difficult, even for experts. Artificial intelligence methods show promise as alternatives for accurate, rapid CO assessment. <h3>Methods</h3> We reviewed paediatric echocardiograms with normal CO and a dilated cardiomyopathy patient group with reduced CO. Experts measured the left ventricular outflow tract diameter, velocity time integral, CO, and cardiac index (CI). EchoNet-Dynamic is a deep learning model for estimation of ejection fraction in adults. We modified this model to predict the left ventricular outflow tract diameter and retrained it on paediatric data. We developed a novel deep learning approach for velocity time integral estimation. The combined models enable automatic prediction of CO. We evaluated the models against expert measurements. Primary outcomes were root-mean-squared error, mean absolute error, mean average percentage error, and coefficient of determination (<i>R</i>²). <h3>Results</h3> In a test set unused during training, CI was estimated with the root-mean-squared error of 0.389 L/min/m², mean absolute error of 0.321 L/min/m², mean average percentage error of 10.8%, and <i>R</i>² of 0.755. The Bland-Altman analysis showed that the models estimated CI with a bias of +0.14 L/min/m² and 95% limits of agreement -0.58 to 0.86 L/min/m². <h3>Conclusions</h3> Our model estimated CO with strong correlation to ground truth and a bias of 0.17 L/min, better than many CO measurements in paediatrics. Model pretraining enabled accurate estimation despite a small dataset. Potential uses include supporting clinicians in real-time bedside calculation of CO, identification of low-CO states, and treatment responses.

Enhanced LiteHRNet based sheep weight estimation using RGB-D images

Sheep farming is a strategic sector of livestock husbandry, and its production has large market demand in many countries. The live weight of sheep provides important information about the health state and the time point for marketing. Manual weighing sheep is time-consuming for farmers even with the help of a ground scale. With the development of Artificial Intelligence (AI) and smart sensors, non-contact sheep weighing methods have gradually been used to estimate weight. However, the performance of prior studies tends to degenerate with varying postures and light conditions in practical natural environments. In this study, we propose a sheep live weight estimation approach based on LiteHRNet (a Lightweight High-Resolution Network) using RGB-D images. Class Activation Mapping (CAM) guided the design of efficient network heads embracing visual explanation and applicability in practical natural environments. Experiments are conducted on our challenging dataset (of 726 sheep RGB-D images, weight range between 19.5 to 94 kg). Comparative experiment results reveal that the lightweight Convolutional Neural Network (CNN) model trained on RGB-D images can reach an acceptable weight estimation result, Mean Average Percentage Error (MAPE) is 14.605% (95% confidence interval: [13.821%, 15.390%], t test) with only 1.06M parameters. Our works can be viewed as preliminary work that confirms the ability to use lightweight CNNs for sheep weight estimation on RGB-D data. The results of this study are potential to develop an embedded device to automatically estimate sheep live weight and would contribute to the development of precision livestock farming.

Prediction of the daily nutrient requirements of gestating sows based on sensor data and machine-learning algorithms.

Precision feeding is a strategy for supplying an amount and composition of feed as close that are as possible to each animal's nutrient requirements, with the aim of reducing feed costs and environmental losses. Usually, the nutrient requirements of gestating sows are provided by a nutrition model that requires input data such as sow and herd characteristics, but also an estimation of future farrowing performances. New sensors and automatons, such as automatic feeders and drinkers, have been developed on pig farms over the last decade, and have produced large amounts of data. This study evaluated machine-learning methods for predicting the daily nutrient requirements of gestating sows, based only on sensor data, according to various configurations of digital farms. The data of 73 gestating sows was recorded using sensors such as electronic feeders and drinker stations, connected weight scales, accelerometers, and cameras. Nine machine-learning algorithms were trained on various dataset scenarios according to different digital farm configurations (one or two sensors), in order to predict the daily metabolizable energy and standardized ileal digestible lysine requirements for each sow. The prediction results were compared to those predicted by the InraPorc model, a mechanistic model for the precision feeding of gestating sows. The scenario predictions were also evaluated with or without the housing conditions and sow characteristics at artificial insemination usually integrated into the InraPorc model. Adding housing and sow characteristics to sensor data improved the mean average percentage error by 5.58% for lysine and by 2.22% for energy. The higher correlation coefficient values for lysine (0.99) and for energy (0.95) were obtained for scenarios involving an automatic feeder system (daily duration and number of visits with or without consumption) only. The scenarios including an automatic feeder combined with another sensor gave good performance results. For the scenarios using sow and housing characteristics and automatic feeder only, the root mean square error was lower with Gradient Tree Boosting (0.91 MJ/d for energy and 0.08g/d for lysine) compared with those obtained using linear regression (2.75 MJ/d and 1.07g/d). The results of this study show that the daily nutrient requirements of gestating sows can be predicted accurately using data provided by sensors and machine-learning methods. It paves the way to simpler solutions for precision feeding.

Incorporating stochastic volatility and long memory into geometric Brownian motion model to forecast performance of Standard and Poor's 500 index

&lt;abstract&gt; &lt;p&gt;It is known in the financial world that the index price reveals the performance of economic progress and financial stability. Therefore, the future direction of index prices is a priority of investors. This empirical study investigated the effect of incorporating memory and stochastic volatility into geometric Brownian motion (GBM) by forecasting the future index price of S&amp;amp;P 500. To conduct this investigation, a comparison study was implemented between twelve models; six models without memory (GBM) and six models with memory (GFBM) under two different assumptions of volatility; constant, which were computed by three methods, and stochastic volatility, obeying three deterministic functions. The results showed that the best performance model was for GFBM under a stochastic volatility assumption using the identity deterministic function $ \sigma \left({Y}_{t}\right) = {Y}_{t} $, according to the smallest values of mean square error (MSE) and mean average percentage error (MAPE). This revealed the direct positive effect of incorporating memory and stochastic volatility into GBM to forecast index prices, and thus can be applied in a real financial environment. Furthermore, the findings showed invalidity of the models with exponential deterministic function $ \sigma \left({Y}_{t}\right) = {e}^{{Y}_{t}} $ in forecasting index prices according to huge values of MAPE and MSE.&lt;/p&gt; &lt;/abstract&gt;

A comparative study on daily evapotranspiration estimation by using various artificial intelligence techniques and traditional regression calculations.

Evapotranspiration is an important parameter to be considered in hydrology. In the design of water structures, accurate estimation of the amount of evapotranspiration allows for safer designs. Thus, maximum efficiency can be obtained from the structure. In order to accurately estimate evapotranspiration, the parameters affecting evapotranspiration should be well known. There are many factors that affect evapotranspiration. Some of these can be listed as temperature, humidity in the atmosphere, wind speed, pressure and water depth. In this study, models were created for the estimation of the daily evapotranspiration amount by using the simple membership functions and fuzzy rules generation technique (fuzzy-SMRGT), multivariate regression (MR), artificial neural networks (ANNs), adaptive neuro-fuzzy inference system (ANFIS) and support vector regression (SMOReg) methods. Model results were compared with each other and traditional regression calculations. The ET amount was calculated empirically using the Penman-Monteith (PM) method which was taken as a reference equation. In the created models, daily air temperature (T), wind speed (WS), solar radiation (SR), relative humidity (H) and evapotranspiration (ET) data were obtained from the station near Lake Lewisville (Texas, USA). The coefficient of determination (R2), root mean square error (RMSE) and average percentage error (APE) were used to compare the model results. According to the performance criteria, the best model was obtained by Q-MR (quadratic-MR), ANFIS and ANN methods. The R2, RMSE, APE values of the best models were 0,991, 0,213, 18,881% for Q-MR; 0,996; 0,103; 4,340% for ANFIS and 0,998; 0,075; 3,361% for ANN, respectively. The Q-MR, ANFIS and ANN models had slightly better performance than the MLR, P-MR and SMOReg models.

A DYNAMIC SIMULATION MODEL FOR OPTIMAL DEEP-LEVEL MINE COOLING MANAGEMENT AND OPERATIONAL DECISION-MAKING FOR ESKOM’S LOAD CURTAILMENT

The South African mining sector depends on intensive energy usage to extract minerals. Most of this energy is used to enhance underground working conditions through cooling. Electricity costs have been on the rise, reducing the profitability of mines. Mines must invest in management systems that optimise energy usage while maintaining underground conditions. Dynamic simulation models are critical for systems analysis, optimal management, and improved decision-making. This paper presents an integrated dynamic simulation model for deep-level mine cooling systems. The model was developed using Process ToolBox simulation software, and a Mean Average Percentage Error of 3.53% was achieved. The impact of load curtailment on underground services was simulated. Eight working areas exceeded the legal limit of 32.5 °C, and the steady state simulation underestimated the impact, as only three working areas were affected. The simulation was sensitive to the ambient conditions; high temperatures were noted in the working areas when the ambient temperatures peaked.

Estimation of Cardiorespiratory fitness using a using a chest mounted accelerometer

Abstract Cardiorespiratory fitness expressed as maximal oxygen consumption (V̇O2 max) is a strong predictor of cardiovascular health. However, quantification of cardiorespiratory fitness by cardiopulmonary exercise (CPX) assessment is complex and costly and therefore not suitable in most clinical settings. In the current study we develop and validate an algorithm for estimation of V̇O2 max using seismocardiography (SCG-V̇O2 max). SCG is measurement of precordial vibrations using an accelerometer. Methods SCG recordings and results from ergometer CPX testing of V̇O2 max in 300 subjects from six clinical studies were combined in a database. 83 subjects underwent repeated measurements sessions across several days. SCG was obtained using a sensitive accelerometer, located on the lower sternum at the Xiphoid protrusion and subjects were placed in a supine position. 5 subjects were excluded due to cardiovascular disease, or missing data. A machine learning algorithm was devolved for estimation of V̇O2 max in a training set including 221 subjects. The remaining 74 subjects were included in a test set for validation. Correlation and accuracy between SCG-V̇O2 max and ergometer CPX V̇O2 max assessed and day to day variation was assessed in subjects who underwent multiple ergometer CPX sessions. Results In 144 recordings from 74 test set subjects SCG-V̇O2 max was 45.0±9.3 ml/min/kg which was comparable to the ergometer CPX V̇O2 max at 44.3±10.1 ml/min/kg (p=0.09) and correlation between ergometer CPX V̇O2 max and SCG-V̇O2 max was r=0.875. Mean average percentage error was 8.7%. Day to day variation measured as the within-subjects-standard-deviation was 1.2 ml/min/kg for SCG-V̇O2 max and 1.7 ml/min/kg for ergometer CPX. Conclusions SCG based estimation of VV̇O2 max is a novel, low cost, easy to use and accurate estimation method of CRF with a high level of reproducibility. SCG-V̇O2 max can potentially facilitate that cardiorespiratory fitness becomes an integrated part of modern health assessment. Funding Acknowledgement Type of funding sources: Private company. Main funding source(s): VentriJect A/S DK

An intelligent big data analytics method for two-dimensional non-residential building energy forecasting

The wide use of sensor and information technologies in buildings resulted in the massive generation of data related to its operation. Thus, there is a need for identifying patterns in data that they may use for the optimal operation of the buildings. The case of non-residential building characteristically provides a big volume of data whose analysis requires computational efficient methods. In this paper, we introduce a new big data analytic method that is applicable to forecasting energy demand in non-residential buildings. The goal is to make energy forecasts in a two-dimensional (2D) space defined by i) the electricity load and ii) gas demand. The proposed method combines the matrix profile (MP) method with a Long-Short Term Memory (LSTM) neural network. The combination of the above tools provides an efficient method in 2D hour ahead forecasting with the big data environment of smart buildings. In specific, with respect to mean average percentage error (MAPE) the combined MP-LSTM method provides a concurrent forecast of electricity and gas around 3% and 4%, respectively.

Forecasting HFMD Cases Using Weather Variables and Google Search Queries in Sabah, Malaysia

HFMD is a viral-mediated infectious illness of increasing public health importance. This study aimed to develop a forecasting tool utilizing climatic predictors and internet search queries for informing preventive strategies in Sabah, Malaysia. HFMD case data from the Sabah State Health Department, climatic predictors from the Malaysia Meteorological Department, and Google search trends from the Google trends platform between the years 2010-2018 were utilized. Cross-correlations were estimated in building a seasonal auto-regressive moving average (SARIMA) model with external regressors, directed by measuring the model fit. The selected variables were then validated using test data utilizing validation metrics such as the mean average percentage error (MAPE). Google search trends evinced moderate positive correlations to the HFMD cases (r0-6weeks: 0.47-0.56), with temperature revealing weaker positive correlations (r0-3weeks: 0.17-0.22), with the association being most intense at 0-1 weeks. The SARIMA model, with regressors of mean temperature at lag 0 and Google search trends at lag 1, was the best-performing model. It provided the most stable predictions across the four-week period and produced the most accurate predictions two weeks in advance (RMSE = 18.77, MAPE = 0.242). Trajectorial forecasting oscillations of the model are stable up to four weeks in advance, with accuracy being the highest two weeks prior, suggesting its possible usefulness in outbreak preparedness.

A study on airflows induced by vehicle movement in road tunnels by the analysis of bulk data from tunnel sensors

A novel analytical formula is introduced for calculation the average longitudinal airflow velocity in tunnels as a result of the piston effect. Existing correlations were analyzed and reworked to better represent the piston effect of vehicles travelling in bulk, which allows relating the airflow velocity to not only average vehicle velocity, but also the traffic intensity and the traffic structure. The approach was validated using the recorded bulk data for an urban road tunnel located in Gdańsk (Poland). Air velocity, weather conditions, wind parameters at tunnel portals, traffic intensity and structure as well as average traffic speed by lane were recorded at 1 s intervals, and analyzed in form of 10 s and 5 min averages. Since ambient conditions may affect the flows inside the tunnel, their impact on flow measurements was studied with a numerical model of the tunnel portals. Despite its simplicity, the proposed model accurately predicted the air velocity due to vehicle movement. The significance of this approach is related to the ease of achieving the initial data, which allows for easy implementation in tunnel ventilation design, risk assessment methods, determination of initial conditions for fire related CFD analyses, continuous tunnel management or estimation of the potential of energy recuperation from airflow within the tunnel. The mean absolute error (MAE) of calculated air velocity to measurements was 0.28–0.38 m/s, with mean average percentage error (MAPE) of 7.8–10.9 %. The determination coefficient for the relation with the measured data was significantly over 0.9.

Dynamic Time Warping Techniques for Time Series Clustering of Covid-19 Cases in DKI Jakarta

The number of positive cases of Covid-19 in DKI Jakarta has contributed to the national issues, reaching 25% of the total cases in Indonesia. The research examined and modeled the distribution pattern of Covid-19 positive cases in DKI Jakarta based on 44 districts spreading over six administrative areas. The data were regarding positive Covid-19 cases in DKI Jakarta for the past year, from April 2020 to April 2021. The research related to the pattern of positive Covid-19 distribution in 44 districts was carried out by time series clustering through Dynamic Time Warping (DTW) distances and agglomerative hierarchical methods. Then, the effectiveness of the clustering process is evaluated by comparing the predicted value of Covid-19 cases between clustering and non-clustering forecast results at the city level for the next 14 days through the Autoregressive Integrated Moving Average (ARIMA) model. The results group 44 districts into 6 optimal clusters based on the pattern of positive cases of Covid-19 in each district. The highest distribution rate is in cluster A, and the lowest is in cluster F. Geographical characteristics are also indicated by clusters A, B, E, and F. Then, the results show that the Mean Average Percentage Error (MAPE) value of the clustering model ranges from 16% to 20%. The difference between MAPE values to the non-clustering model implies that the forecasting accuracy is not far apart, which is in the round of 5%−6%.

Performance evaluation of deep learning and boosted trees for cryptocurrency closing price prediction

The emergence of cryptocurrencies has drawn significant investment capital in recent years with an exponential increase in market capitalization and trade volume. However, the cryptocurrency market is highly volatile and burdened with substantial heterogeneous datasets characterized by complex interactions between predictors, which may be difficult for conventional techniques to achieve optimal results. In addition, volatility significantly impacts investment decisions; thus, investors are confronted with how to determine the price and assess their financial investment risks reasonably. This study investigates the performance evaluation of a genetic algorithm tuned Deep Learning (DL) and boosted tree-based techniques to predict several cryptocurrencies' closing prices. The DL models include Convolutional Neural Networks (CNN), Deep Forward Neural Networks, and Gated Recurrent Units. The study assesses the performance of the DL models with boosted tree-based models on six cryptocurrency datasets from multiple data sources using relevant performance metrics. The results reveal that the CNN model has the least mean average percentage error of 0.08 and produces a consistent and highest explained variance score of 0.96 (on average) compared to other models. Hence, CNN is more reliable with limited training data and easily generalizable for predicting several cryptocurrencies' daily closing prices. Also, the results will help practitioners obtain a better understanding of crypto market challenges and offer practical strategies to lower risks.

Day-ahead heat load forecasting during the off-season in the district heating system using Generalized Additive model

The paper presents the results of heat demand forecasting during the off-season (summertime and transitional period) in a real district heating system. The hourly heat demand in the analyzed system ranges from 16 to 28 MW, depending on the month and hour of the day. Short-term prediction of the heat demand in the day-ahead horizon is needed for effective operation of production units in the Cogeneration Heat and Power Plant (CHP). The method of an hourly heat load predictions using a data-driven model is presented. The algorithm is based on the Generalized Additive Model with hour of the day, day of the week and day of the year as predictors. A way of increasing the accuracy of forecasts using the autoregressive model for residuals of the model was discussed. The influence of the training dataset size on the model error was also examined. The several versions of the model were evaluated by comparing forecasted results with real values during the district heating system's operation from June to the end of August. For the most accurate one, Root Mean Square and Mean Average Percentage Error were 0.84 MW and 3.3 %, respectively.

Comparative Evaluation of Data-Driven Approaches to Develop an Engine Surrogate Model for NOx Engine-Out Emissions under Steady-State and Transient Conditions

In this paper, a methodology based on data-driven models is developed to predict the NOx emissions of an internal combustion engine using, as inputs, a set of ECU channels representing the main engine actuations. Several regressors derived from the machine learning and deep learning algorithms are tested and compared in terms of prediction accuracy and computational efficiency to assess the most suitable for the aim of this work. Six Real Driving Emission (RDE) cycles performed at the roll bench were used for the model training, while another two RDE cycles and a steady-state map of NOx emissions were used to test the model under dynamic and stationary conditions, respectively. The models considered include Polynomial Regressor (PR), Support Vector Regressor (SVR), Random Forest Regressor (RF), Light Gradient Boosting Regressor (LightGBR) and Feed-Forward Neural Network (ANN). Ensemble methods such as Random Forest and LightGBR proved to have similar performances in terms of prediction accuracy, with LightGBR requiring a much lower training time. Afterwards, LightGBR predictions are compared with experimental NOx measurements in steady-state conditions and during two RDE cycles. Coefficient of determination (R2), normalized root mean squared error (nRMSE) and mean average percentage error (MAPE) are the main metrics used. The NOx emissions predicted by the LightGBR show good coherence with the experimental test set, both with the steady-state NOx map (R2 = 0.91 and MAPE = 6.42%) and with the RDE cycles (R2 = 0.95 and nRMSE = 0.04).

Estimation of Cardiorespiratory fitness using a using a chest mounted accelerometer

Abstract Cardiorespiratory fitness expressed as maximal oxygen consumption (V̇O2 max) is a strong predictor of cardiovascular health. However, quantification of cardiorespiratory fitness by cardiopulmonary exercise (CPX) assessment is complex and costly and therefore not suitable in most clinical settings. In the current study we develop and validate an algorithm for estimation of V̇O2 max using seismocardiography (SCG-V̇O2 max). SCG is measurement of precordial vibrations using an accelerometer. Methods SCG recordings and results from ergometer CPX testing of V̇O2 max in 300 subjects from six clinical studies were combined in a database. 83 subjects underwent repeated measurements sessions across several days. SCG was obtained using a sensitive accelerometer, located on the lower sternum at the Xiphoid protrusion and subjects were placed in a supine position. 5 subjects were excluded due to cardiovascular disease, or missing data. A machine learning algorithm was devolved for estimation of V̇O2 max in a training set including 221 subjects. The remaining 74 subjects were included in a test set for validation. Correlation and accuracy between SCG-V̇O2 max and ergometer CPX V̇O2 max assessed and day to day variation was assessed in subjects who underwent multiple ergometer CPX sessions. Results In 144 recordings from 74 test set subjects SCG-V̇O2 max was 45.0±9.3 ml/min/kg which was comparable to the ergometer CPX V̇O2 max at 44.3±10.1 ml/min/kg (p=0.09) and correlation between ergometer CPX V̇O2 max and SCG-V̇O2 max was r=0.875. Mean average percentage error was 8.7%. Day to day variation measured as the within-subjects-standard-deviation was 1.2 ml/min/kg for SCG-V̇O2 max and 1.7 ml/min/kg for ergometer CPX. Conclusions SCG based estimation of VV̇O2 max is a novel, low cost, easy to use and accurate estimation method of CRF with a high level of reproducibility. SCG-V̇O2 max can potentially facilitate that cardiorespiratory fitness becomes an integrated part of modern health assessment. Funding Acknowledgement Type of funding sources: Private company. Main funding source(s): VentriJect A/S DK

Toward High-Performance Map-Recovery of Air Pollution Using Machine Learning

Mobile and pervasive sampling of urban air pollution has been increasingly valued as a sustainable method, in terms of economic and operational factors, for surveying atmospheric environment with high space-time resolution. Specifically, fine-granular air quality (AQ) inference provides fundamental progress toward data-driven urban management, as it estimates grid-level pollutant concentrations constantly using pollutant measurement data collected from fixed and mobile sensors. In this paper, we propose a tree-based multicascade space-time learning model (MCST-Tree) for AQ inference to recover pollution maps by exploiting multisource AQ samples (fixed and mobile) and heterogeneous urban feature sets (land-use, meteorology, population, traffic, etc.). This is implemented and evaluated in a study case of Chengdu (4900 km2, 14 June to 14 July 2018), which achieves map-recovery of PM2.5 distribution based upon the sparse measurements (ca. 16.2% space-time coverage) with high-performance (symmetric mean average percentage error (SMAPE) (%) = 14.13%; R2 = 0.94). Detailed evaluations are presented through the analysis of model performance, space-time coverage of mobile sampling, and AQ inference. We conduct a series of sensitivity analyses of mobile sampling coverage, and the experimental results show that it is a critical issue to enhance the model trust, which contributes to improve the R-square from 0.81 (fixed data + 10% mobile data) to 0.94 (fixed data + 100% mobile data). The results show that the mobile sampling significantly improves the space-time modeling capability, and our proposed model has great potential to achieve map-recovery for air pollution at high spatial-temporal resolution with high performance.

Prediction of Sea Surface Temperature by Combining Interdimensional and Self-Attention with Neural Networks

Sea surface temperature (SST) is one of the most important and widely used physical parameters for oceanography and meteorology. To obtain SST, in addition to direct measurement, remote sensing, and numerical models, a variety of data-driven models have been developed with a wealth of SST data being accumulated. As oceans are comprehensive and complex dynamic systems, the distribution and variation of SST are affected by various factors. To overcome this challenge and improve the prediction accuracy, a multi-variable long short-term memory (LSTM) model is proposed which takes wind speed and air pressure at sea level together with SST as inputs. Furthermore, two attention mechanisms are introduced to optimize the model. An interdimensional attention strategy, which is similar to the positional encoding matrix, is utilized to focus on important historical moments of multi-dimensional input; a self-attention strategy is adopted to smooth the data during the training process. Forty-three-year monthly mean SST and meteorological data from the fifth-generation ECMWF (European Centre for Medium-Range Weather Forecasts) reanalysis (ERA5) are collected to train and test the model for the sea areas around China. The performance of the model is evaluated in terms of different statistical parameters, namely the coefficient of determination, root mean squared error, mean absolute error and mean average percentage error, with a range of 0.9138–0.991, 0.3928–0.8789, 0.3213–0.6803, and 0.1067–0.2336, respectively. The prediction results indicate that it is superior to the LSTM-only model and models taking SST only as input, and confirm that our model is promising for oceanography and meteorology investigation.

Physics-based digital twins for autonomous thermal food processing: Efficient, non-intrusive reduced-order modeling

One possible way of making thermal processing controllable is to gather real-time information on the product's current state. Often, sensory equipment cannot capture all relevant information easily or at all. Digital Twins close this gap with virtual probes in real-time simulations, synchronized with the process. This paper proposes a physics-based, data-driven Digital Twin framework for autonomous food processing. We suggest a lean Digital Twin concept that is executable at the device level, entailing minimal computational load, data storage, and sensor data requirements. This study focuses on a parsimonious experimental design for training non-intrusive reduced-order models (ROMs) of a thermal process. A correlation (R = − 0.76) between a high standard deviation of the surface temperatures in the training data and a low root mean square error in ROM testing enables efficient selection of training data. The mean test root mean square error of the best ROM is less than 1 Kelvin (0.2% mean average percentage error) on representative test sets. Simulation speed-ups of Sp ≈ 1.8 × 104 allow on-device model predictive control. Industrial relevanceThe proposed Digital Twin framework is designed to be applicable within the industry. Typically, non-intrusive reduced-order modeling is required as soon as the modeling of the process is performed in software, where root-level access to the solver is not provided, such as commercial simulation software. The data-driven training of the reduced-order model is achieved with only one data set, as correlations are utilized to predict the training success a priori.