Desenvolvimento de um modelo neural para estimar a massa de sólidos acumulada na filtração de água na irrigação localizada

In this work, the development of an Artificial Neural Network (ANN) based soft estimator is reported for the estimation of static-nonlinearity associated with the transducers. Under the realm of ANN based transducer modeling, only two neural models have been suggested to estimate the static-nonlinearity associated with the transducers with quite successful results. The first existing model is based on the concept of a functional link artificial neural network (FLANN) trained with mu-LMS (Least Mean Squares) learning algorithm. The second one is based on the architecture of a single layer linear ANN trained with alpha-LMS learning algorithm. However, both these models suffer from the problem of slow convergence (learning). In order to circumvent this problem, it is proposed to synthesize the direct model of transducers using the concept of a Polynomial-ANN (polynomial artificial neural network) trained with Levenberg-Marquardt (LM) learning algorithm. The proposed Polynomial-ANN oriented transducer model is implemented based on the topology of a single-layer feed-forward back-propagation-ANN. The proposed neural modeling technique provided an extremely fast convergence speed with increased accuracy for the estimation of transducer static nonlinearity. The results of convergence are very stimulating with the LM learning algorithm.

Frequency-weighted variant of Kalman filter (FWKF), reported previously improves kinematic state estimates by reducing the effect of high frequency noise components. However, this introduces time lag in estimates, which may not be acceptable for some specific application environments. Again, a target tracking algorithm employing an artificial neural network (ANN) in cascade with a standard KF (KF-ANN) has been reported to be promising in improving the quality of estimates without introducing any appreciable lag in the estimates. Further improvement of the KF-ANN estimates has been discussed by employing a synergic approach of FWKF and KF-ANN, where the estimates from FWKF has been post-processed by an appropriately trained ANN. However, the study was carried out for limited number of samples due to inadequacy of the ANN learning algorithm (back propagation) to generalize a scenario with large dynamic range of data. This problem has been alleviated by using Levenberg-Marquardt (LM) learning algorithm in the present case. The current study presents comparative results of KF, FWKF and their ANN-aided variants, viz. KF-ANN and FWKF-ANN. It has been shown that by using LM learning algorithm, improved estimates from FWKF-ANN algorithm has been obtained with reduced high frequency error for larger duration of flight.

Available neural network-based models for predicting the oil flow rate (q<sub>o</sub>) in the Niger Delta are not simplified and are developed from limited data sources. The reproducibility of these models is not feasible as the models’ details are not published. This study developed simplified and reproducible three, five, and six-input variables neural-based models for estimating q<sub>o</sub> using 283 datasets from 21 wells across fields in the Niger Delta. The neural-based models were developed using maximum-minimum (max.-min.) normalized and clip-normalized datasets. The performances and the generalizability of the developed models with published datasets were determined using some statistical indices: coefficient of determination (R<sup>2</sup>), mean square error (MSE), root mean square error (RMSE), average relative error (ARE) and average absolute relative error (AARE). The results indicate that the 3-input-based neural models had overall R<sup>2</sup>, MSE, and RMSE values of 0.9689, 9.6185x10<sup>-4 </sup>and 0.0310, respectively, for the max.-min. normalizing method and R<sup>2</sup> of 0.9663, MSE of 5.7986x10<sup>-3</sup> and RMSE of 0.0762 for the clip scaling approach. The 5-input-based models resulted in R<sup>2</sup> of 0.9865, MSE of 5.7790×10<sup>-4</sup> and RMSE of 0.0240 for the max.-min. scaling method and R<sup>2</sup> of 0.9720, MSE of 3.7243x10<sup>-3</sup> and RMSE of 0.0610 for the clip scaling approach. Also, the 6-input-based models had R<sup>2</sup> of 0.9809, MSE of 8.7520x10<sup>-4</sup> and RMSE of 0.0296 for the max.-min. normalizing approach and R<sup>2</sup> of 0.9791, MSE of 3.8859 x 10<sup>-3</sup> and RMSE of 0.0623 for the clip scaling method. Furthermore, the generality performance of the simplified neural-based models resulted in R<sup>2</sup>, RMSE, ARE, and AAPRE of 0.9644, 205.78, 0.0248, and 0.1275, respectively, for the 3-input-based neural model and R<sup>2</sup> of 0.9264, RMSE of 2089.93, ARE of 0.1656 and AARE of 0.2267 for the 6-input-based neural model. The neural-based models predicted q<sub>o</sub> were more comparable to the test datasets than some existing correlations, as the predicted q<sub>o</sub> result was the lowest error indices. Besides, the overall relative importance of the neural-based models’ input variables on q<sub>o</sub> prediction is S>GLR>P<sub>wh</sub>>T/T<sub>sc</sub>>γ<sub>o</sub>>BS&W>γ<sub>g</sub>. The simplified neural-based models performed better than some empirical correlations from the assessment indicators. Therefore, the models should apply as tools for oil flow rate prediction in the Niger Delta fields, as the necessary details to implement the models are made visible.

In the present work, Artificial Neural Network (ANN) model has been developed to predict the energy and exergy efficiency of a roughened solar air heater (SAH). Total fifty data sets of samples, obtained by conducting experiments on SAHs with three different specification of wire-rib roughness on the absorber plates, have been used in this work. These experimental data and calculated values of thermal efficiency and exergy efficiency have been used to develop an ANN model. Levenberg-Marquardt (LM) and Scaled Conjugate Gradient (SCG) learning algorithm were used to train the proposed ANN model. Six numbers of neurons were found with LM learning algorithm in the hidden layer as the optimal value on the basis of statistical error analysis. In the input layer, the time of experiments, mass flow rate, ambient temperature, mean temperature of air, absorber plate temperature and solar radiation intensity have been taken as input parameters; and energy efficiency and exergy efficiency have been taken as output parameters in the output layer. The 6-6-2 neural model has been obtained as the optimal model for prediction. Performance predictions using ANN were compared with the experimental data and a close agreement was observed. Statistical error analysis was used to evaluate the results. Citation:  Ghritlahre, H. K. (2018). Development of feed-forward back-propagation neural model to predict the energy and exergy analysis of solar air heater. Trends in Renewable Energy, 4, 213-235. DOI: 10.17737/tre.2018.4.2.0078

The objective of this study is to compare the predictive ability of Bayesian regularization with Levenberg–Marquardt Artificial Neural Networks. To examine the best architecture of neural networks, the model was tested with one-, two-, three-, four-, and five-neuron architectures, respectively. MATLAB (2011a) was used for analyzing the Bayesian regularization and Levenberg–Marquardt learning algorithms. It is concluded that the Bayesian regularization training algorithm shows better performance than the Levenberg–Marquardt algorithm. The advantage of a Bayesian regularization artificial neural network is its ability to reveal potentially complex relationships, meaning it can be used in quantitative studies to provide a robust model.

The current study looks at the effectiveness of the removal of nickel (II) from aqueous solution using an adsorption method in a laboratory‐size reactor. An artificial neural network (ANN) and an adaptive neuro‐fuzzy inference system (ANFIS) were used in this study to predict blend hydrogels adsorption potential in the removal of nickel (II) from aqueous solution. Four operational variables, including initial Ni (II) concentration (mg/L), pH, contact duration (min), and adsorbent dose (mg/L), were used as an input with removal percentage (%) as the only output; they were studied to assess their impact on the adsorption study in the ANFIS model. In contrast, 70% of the data was used for training, while 15% of the data was used in testing and validation to build the ANN model. Feedforward propagation with the Levenberg–Marquardt algorithm was employed to train the network. The use of ANN and ANFIS models for experiments was used to optimize, construct, and develop prediction models for Ni (II) adsorption using blend hydrogels. The adsorption isotherm and kinetic models were also used to describe the process. The results show that ANN and ANFIS models are promising prediction approaches that can be applied to successfully predict metal ions adsorption. According to this finding, the root mean square errors (RMSE), absolute average relative errors (AARE), average relative errors (ARE), mean squared deviation (MSE), and R 2 for Ni (II) in the training dataset were 0.061, 0.078, 0.017, 0.019, and 0.986, respectively, for ANN. In the ANFIS model, the RMSE, AARE, ARE, MSE, and R 2 were 0.0129, 0.0119, 0.028, 0.030, and 0.995, respectively. The adsorption process was spontaneous and well explained by the Langmuir model, and chemisorption was the primary control. The morphology, functional groups, thermal characteristics, and crystallinity of blend hydrogels were all assessed.

Prediction of a wellhead separator efficiency and risk assessment in a gas condensate reservoir

The method of Levenberg-Marquardt learning algorithm was investigated for estimating tropospheric ozone concentration. The Levenberg-Marquardt learning algorithm has 12 input neurons (6 pollutants and 6 meteorological variables), 28 neurons in the hidden layer, and 1 output neuron for the Ozone (O3) estimate. The Multilayer Perceptron Model (MLP) performance was found to make good predictions with the mean square error (MSE) less than 1 µg/m3 (0.002 µg/m3). In addition, the correlation coefficient ranges from 0.74 to 0.95 in The Levenberg-Marquardt learning. The Levenberg-Marquardt learning algorithm that a multilayer perception method of Artificial Neural Network (ANN) has performed well and an effective approach for predicting tropospheric ozone. Ozone concentration was influenced predominantly by the nitrogen oxide (NOx, NO2, NO), SO2 and temperature. The model did not predict solar radiation to ozone with sufficient accuracy.

The rheology of drilling fluid is a crucial component during drilling operations to achieve optimum performance and avoid non-productive time resulting from drilling problems. In field operations, measuring the rheological and filtration properties of drilling fluids necessitates a significant amount of preparation time and experimental work. Often, such data analysis is not conducted on a footage basis. However, some surface parameters, such as mud weight using the mud balance, marsh funnel viscosity using the Marsh funnel, and mud flow line temperature, can be measured in real-time. This study introduces physics-based machine learning models for real-time prediction of both rheological and filtration properties. The machine learning algorithms used include support vector machines (SVM), extreme gradient boosting (XGB), random forests (RF), and multilayer perceptron networks (MLP). These models predict the rheological and filtration properties based on collected field experimental data. The model inputs are mud weight, marsh funnel viscosity, and flow line temperature, while the targets are apparent viscosity (AV), yield point (YP), plastic viscosity (PV), shear stress based on Fann readings at 600, 300, 200, 100, 6, and 3 shear rates in terms of revolutions per minute (RPM), and filtration loss volume. The developed machine learning models underwent hyperparameter tuning based on cross-validation to select the optimum model for each algorithm using the coefficient of determination (R2) and absolute average relative error (AAPRE). Furthermore, all machine learning model predictions were tested and validated using various datasets. The SVM models' performance ranged from 0.85 to 0.97 in terms of R2, with most AAPRE values less than 2%, reaching a maximum of 6%. The developed RF models ranged from 0.94 to 0.97 in terms of R^2, with AAPRE values between 2 and 3%. XGB models ranged between 0.95 and 0.98 in terms of R^2, with AAPRE values from 1 to 2%. The MLP models ranged from 0.991 to 0.99 in terms of R^2, with AAPRE values less than 6%. The developed machine learning models exhibited high accuracy in predicting the mud's rheological properties in terms of PV, YP, and n, with AAPRE values less than 1.94%, 1.77%, and 4%, respectively.

In this study, machine learning (ML) models were developed to predict permeability (k), porosity (φ) and water saturation (Sw) using 1241 datasets obtained from well-logs data in the Niger Delta. The datasets were screened to remove incomplete sets and outliers and make them suitable for adequate training using the maximum-minimum normalization approach. Three multiple-input multiple-output (MIMO) machine learning methods, namely artificial neural network (ANN), decision tree (DT) and random forest (RF), were used to train the datasets. Five performance metrics, coefficient of determination (R2), correlation coefficient (R), mean absolute error (MAE), average absolute relative error (AARE), and average relative error (ARE), were used to evaluate the performance of the developed models. The results indicate that the MIMO neural-based model had overall MSE and R values of 1.9801×10-3 and 0.9866, while the DT model had 2.2540×10-3 and 0.98281, and the RF model had 5.1490×10-3 and 0.95989. The ANN model predicted k resulted in R2, R, MAE, ARE, and AARE of 0.95740, 0.97847, 2.0677, -0.0011, and 0.0343, respectively, while the predicted φ had R2 of 0.96336, R of 0.98151, MAE of 0.0055, ARE of -0.0006, and AARE of 0.0185. The predicted Sw had an R2 of 0.98430, R of 0.99212, MAE of 0.0265, ARE of -0.0045, and AARE of 0.0521. Also, the developed DT model predicted k resulted in R2, R, MAE, ARE and AARE of 0.95250, 0.97596, 0.0277, 5.6981 and 0.0382, respectively, while the predicted φ had R2 of 0.9380, R of 0.9685, MAE of 0.0276, ARE of -0.5796 and AARE of 5.8199. The predicted Sw had R2 of 0.99039, R of 0.9518, MAE of 0.0182, ARE of -0.49969 and AARE of 5.0452. Furthermore, the developed RF model predicted k resulted in R2, R, MAE, ARE, and AARE of 0.88438, 0.94041, 0.0552, -6.8754 and 15.8391, respectively, while the predicted φ had R2 of 0.90377, R of 0.95067, MAE of 0.0504, ARE of -5.3429 and AARE of 12.8260. The predicted Sw had R2 of 0.95495, R of 0.97722, MAE of 0.0469, ARE of -25.1422 and AARE of 32.6698. The relative importance of the ML input parameters on the predicted outputs is RES>D>GR>VSh>RHOB>NPHI>CALI. Based on the statistical indicators obtained, the predictions of the developed ML-based models were close to the actual field datasets. Thus, the ML-based models should be used as tools for predicting k, φ and Sw in the Niger Delta.

Optimization of operating conditions for compressor performance by means of neural network inverse

The aim of this study is to demonstrate the comparison of an artificial neural network (ANN) and an adaptive neuro fuzzy inference system (ANFIS) for the prediction of the coefficient of performance (COP) for a water purification process integrated in an absorption heat transformer system with energy recycling. ANN and ANFIS models take into account the input and output temperatures for each one of the four components (absorber, generator, evaporator, and condenser), as well as two presures and LiBr+H2O concentrations. Experimental results are performed to verify the results from the ANN and ANFIS approaches. For the network, a feedforward with one hidden layer, a Levenberg-Marquardt learning algorithm, a hyperbolic tangent sigmoid transfer function and a linear transfer function were used. The best fitting training data set was obtained with three neurons in the hidden layer. On the validaton data set, simulations and experimental data test were in good agreement (R2>0.9980). However, the ANFIS model was developed using the same input variables. The statistical values are given in as tables. However, comparaison between two models shows that ANN provides better results than the ANFIS results. Finally this paper shows the appropriateness of ANN and ANFIS for the quantitative modeling with reasonable accuracy.

Knowledge of surface reflection of an object is essential in many technological fields, including graphics and cultural heritage. Compared to direct multi- or hyper-spectral capturing approaches, commercial RGB cameras allow for a high resolution and fast acquisition, so the idea of mapping this information into a reflectance spectrum (RS) is promising. This study compared two modelling approaches based on a training set of RGB-reflectance pairs, one implementing artificial neural networks (ANN) and the other one using multivariate polynomial approximation (PA). The effect of various parameters was investigated: the ANN learning algorithm—standard backpropagation (BP) or Levenberg-Marquardt (LM), the number of hidden layers (HLs) and neurons, the degree of multivariate polynomials in PA, the number of inputs, and the training set size on both models. In the two-layer ANN with significantly fewer inputs than outputs, a better MSE performance was found where the number of neurons in the first HL was smaller than in the second one. For ANNs with one and two HLs with the same number of neurons in the first layer, the RS reconstruction performance depends on the choice of BP or LM learning algorithm. RS reconstruction methods based on ANN and PA are comparable, but the ANN models’ better fine-tuning capabilities enable, under realistic constraints, finding ANNs that outperform PA models. A profiling approach was proposed to determine the initial number of neurons in HLs—the search centre of ANN models for different training set sizes.

In this study, some physical and mechanical properties of yellow pine wood (Pinus sylvestris), which is used extensively in furniture industry, were tested after heat treatment. The findings obtained were modelled by artificial neural network (ANN) and interval values related to temperature and time variations were tried to be estimated. This study, which makes it easier to reach intermediate values, aims to save the relevant researchers from trial load all of the heating parameters during the furniture design/production stages. In the study scotch pine samples were heat-treated at 150, 160, 170, 180, 190 and 200 °C for 2, 4 and 6 hours, under normal atmosphere conditions. Color changes, weight losses and compression strength parallel to grain values of heat-treated samples were determined. After experimental study, modelling procedure was performed by ANN using two different learning algorithm- Levenberg-Marquardt (LM) and Scaled Conjugate Gradient (SCG) algorithm- 15 different hidden neurons. The best model was obtained from 2-7-6 structure using LM learning algorithm. Mean absolute percentage error (MAPE) of the best model was found below 8.0% for estimated color parameters. The weight loss and compression strength parallel to grain were 5.79% and 1.50%, respectively. It was concluded that ANN can be used successfully to predict all studied parameters of heat-treated wood samples.

A new approach based on artificial neural networks is successfully introduced to determine the characteristic parameters of a coplanar waveguide (CPW) sandwiched between two dielectric substrates. Neural models were trained with eight different learning algorithms to obtain better performance and faster convergence with a simpler structure. The best results were obtained from the models trained with Levenberg-Marquardt and Bayesian regulation learning algorithms. The results obtained from the neural model are in very good agreement with theoretical and experimental results available in the literature. The presented neural model is valid for both conventional and sandwiched CPWs.