No full-text available
To read the full-text of this research,
you can request a copy directly from the author.
Neural Network Based System Identification Toolbox
... The type of ANN considered (Pachepsky et al., 1996;Schaap et al., 1998). Tamari here is called the multilayer perceptron (Nørgaard, 2000). A and Wö sten (1999) gave a review on ANN and their network with an input vector of elements x l (l ϭ 1, . . . ...
... These in-nonlinear regression. The purpose of this paper is to are determined from a set of data through the process of training (Nørgaard, 2000). The training of a network is accomplished using an optimization procedure (such as nonlinear least squares). ...
... 5.3 with freeware toolbox NNSYSID dubious data were discarded, which left us with 862 soil samver. 2.0 (Nørgaard, 2000). This is the neuro-p method, which ples. ...
... In this study, the proposed model is built based on feedforward neural network model with one hidden layer to provide the third-floor behavior, as presented in Figure 3. The parametric identification models are used to detect the behavior of structures [43][44][45]. The autoregressive (AR) and autoregressive moving average (ARMA) are among the common and simplified parametric models [28,43]. ...
... The parametric identification models are used to detect the behavior of structures [43][44][45]. The autoregressive (AR) and autoregressive moving average (ARMA) are among the common and simplified parametric models [28,43]. In this study, the two methods are evaluated with feedforward neural network solution. ...
... In this study, the two methods are evaluated with feedforward neural network solution. The components of the proposed model are described and explained with details in [43]. A multi-input single-output (MISO) three-layer model is designed for building behavior analysis (Figure 3). ...
This study evaluates the performance of passively controlled steel frame building under dynamic loads using time series analysis. A novel application is utilized for the time and frequency domains evaluation to analyze the behavior of controlling systems. In addition, the autoregressive moving average (ARMA) neural network are employed to identify the performance of the controller system. Three passive vibration control devices are utilized in this study, namely: tuned mass damper (TMD), tuned liquid damper (TLD) and tuned liquid column damper (TLCD). The results show that the TMD control system is a more reliable controller than TLD and TLCD systems in terms of vibration mitigation. The probabilistic evaluation and identification model showed that the probability analysis and ARMA neural network model are suitable to evaluate and predict the response of coupled building-controller systems.
... The type of ANN considered (Pachepsky et al., 1996;Schaap et al., 1998). Tamari here is called the multilayer perceptron (Nørgaard, 2000). A and Wö sten (1999) gave a review on ANN and their network with an input vector of elements x l (l ϭ 1, . . . ...
... These in-nonlinear regression. The purpose of this paper is to are determined from a set of data through the process of training (Nørgaard, 2000). The training of a network is accomplished using an optimization procedure (such as nonlinear least squares). ...
... 5.3 with freeware toolbox NNSYSID dubious data were discarded, which left us with 862 soil samver. 2.0 (Nørgaard, 2000). This is the neuro-p method, which ples. ...
... The supervision and control system CITECT and EROS (Eros 1997), were used for data collection for the experiments that were carried out with the aim to identify the process, which enable measuring all the variables of interest for the study and analysis of the post-combustion sub-process. The software MATLAB (MathWorks Inc., MA) and the identification toolbox with neural networks developed by (Nørgaard 2000) were used for the analysis of ANN in this study. ...
... This has been proved through numerous practical applications (Larsen et al. 2014;Li and Chan 2017;Nabavi-Pelesaraei et al. 2016). The class of MLP network considered here has a single hidden layer with a hyperbolic tangent activation function (f) and in the output layer a linear activation function (F) (see equation 3 (Nørgaard 2000)). ...
... The idea is to select the regressors based on inspiration from linear system identification and then determine the best possible network architecture with the given regressors as inputs. The toolbox provides six different model structures; in this case, the non-linear autoregressive model NNARX (equation 8 is the predictor, and equation 9 is the regression vector (Nørgaard 2000)) was selected because the basic rule of thumb is however to use this type model whenever possible: ...
In a nickel-producing multiple hearth furnace, there is a problem associated to the automatic operation of the temperature control loops in two of the hearths, since the same flow of air is split into two branches. A neural model of the post-combustion sub-process is built and served to increase the process efficiency of the industrial furnace. Data was taken for a three-months operating time period to identify the main variables characterizing the process and a model of multilayer perceptron type is built. For the validation of this model, process data from a four-months operating time period in 2018 was used and prediction errors based on a measure of closeness in terms of a mean square error criterion measured through its weights for the temperature of two of the hearths (four and six) versus the air flow to these hearths. Based on a rigorous testing and analysis of the process, the model is capable of predicting the temperature of hearth four and six with errors of 0.6 and 0.3 °C, respectively. In addition, the emissions by high concentration of carbon monoxide in the exhaust gases are reduced, thus contributing to the health of the ecosystem.
... Regarding the selection of the structure of the model, it is not only necessary to select a set of regressors, but also the architecture of the network. The procedure used is described in [10]. The idea is to select the regressors based on the identification of linear systems and then determine the best possible network architecture with the regressors given as input. ...
... The Levenberg-Marquardt was used as a training algorithm. The MATLAB ® program and the identification toolbox with neural networks developed by M Nørgaard [10] were used as a software to carry out the research. ...
In the muti-hearth furnace, there is a problem related to the automatic operation
of the loops of temperature regulation in hearths four and six, since the same flow of
air diverged into two branches. In this work, the authors take advantage of the capacity
of artificial neural networks for the learning of complex relationships, starting from a
set of examples. A neuronal model of the post-combustion sub-process in an Indus-trial
furnace, which will serve to raise an automatic control strategy, is obtained. Experiments
were carried out with binary pseudo-random sequences of modulated amplitude on the
flow of ore, and the openings of the regulating valves of air flow to hearths mentioned
before, to determine their effect on the temperature. The trial and error process enabled
to obtain an artificial neural network of multilayer perceptron type, capable of predicting
the temperature of hearth four with errors less than 0.5%, and 0.9% for the hearth six.
... Regarding the selection of the structure of the model, it is not only necessary to select a set of regressors, but also the architecture of the network. The procedure used is described in [10]. The idea is to select the regressors based on the identification of linear systems and then determine the best possible network architecture with the regressors given as input. ...
... The Levenberg-Marquardt was used as a training algorithm. The MATLAB ® program and the identification toolbox with neural networks developed by M Nørgaard [10] were used as a software to carry out the research. ...
... Para o ajuste dos parâmetros nos modelos neurais caixa preta foi utilizado o algoritmo de Levenberg-Marquadt, implementado no pacote Norgaard Toolbox (Norgaard, 1997). Esse algoritmo ajusta os pesos da rede de maneira a minimizar a norma L 2 do vetor de resíduos, ou seja, o vetor de erros de predição de um passo à frente ao longo dos dados dinâmicos de treinamento Z. ...
... Os procedimentos de treinamento foram os mesmos apresentados na Seção 2. Para a obtenção do modelo caixa preta, foram treinadas 100 redes, utilizando o algoritmo de Levenberg-Marquadt, implementado no pacote Norgaard Toolbox (Norgaard, 1997). ...
This work presents a new approach for gray-box identification using neural network models. Information about the process steady state is used as auxiliary knowledge during the training of the neural models, which is performed in a bi-objective framework. Evolutionary algorithms were implemented to solve the optimization problem. The approach was applied to two experimental processes: a pilot hydraulic pumping system and an industrial gas-lift offshore oil well. The proposed gray-box identification technique was compared to black-box approaches particularly in operating regimes that were not available in the dynamical identification data sets. Results show that the gray-box procedure yields models with better performance than the ones obtained by the black-box approach in at least one of the objectives: static function error or dynamic test data prediction error, where the dynamic data cover a broader operating range. It is shown that the implementation of this gray-box approach is justifiable when the black-box procedure does not achieve a model with good static performance.
... The identification of the ANN models was done based on NNSY-SID toolbox, developed by Prof. Magnus Nørgaard from the Department of Automation of the Technical University of Denmark in 2000. The MATLAB Version 2.0 of this toolbox, released in the year 2000, was used in this work (Nørgaard, 2000). Even though NNSY-SID is 20 years old, it is considered a complete and powerful tool to build classical ANNs models. ...
... This procedure is similar to a surgery where the unnecessary neurons are eliminated and it is called General Pruning. The pruning algorithm used here was the Optimal Brain Surgeon (OBS) available within NNSY-SID, which is currently the main pruning strategy for ANNs (Nørgaard, 2000). After OBS removes each unnecessary weight, the network is retrained, and this goes on until all unimportant weights determined by OBS are eliminated. ...
Syngas is one of the main sources available for the production of pure H2 and synthetic fuels, among others. Pressure swing adsorption (PSA) is considered to be an efficient alternative for pre-treatment of syngas. However, it displays very complex dynamical behaviour. This work proposes the development of different Artificial Intelligence based models for the prediction of the dynamic behaviour of several process output variables. A classical model of ANNs, a machine learning model and a deep learning model was here developed. It was found that Deep Learning networks were the only ones capable of fully representing the dynamic behaviour of the PSA unit, whereas the other models were only partially capable of predicting it. Thus, it is proposed a reliable real-time soft sensor for a PSA unit based on Deep Leaning strategy. This strategy provides bases to overtake several problems associated to this processes control, operation and optimization.
... Fig. 17 shows the obtained static curve (after applying a second order polynomial fit) considering only the relation between downhole pressure (PT1) and the gas-lift flow (FT4). A black-box neural network was trained using only dynamical data with the Norgaard's toolbox (Norgaard, 1997), and gray-box neural networks were obtained using static and dynamical data, by solving the bi-objective problem. The graybox model which presented the lower dynamical error on validation data was selected from the Pareto front. ...
Downhole pressure is an important process variable in the operation of gas-lifted oil wells. The device installed in order to measure this variable is often called a Permanent Downhole Gauge (PDG). Replacing a faulty PDG is often not economically viable and to have an alternative estimate of the downhole pressure is an important goal. Using data from operating PDGs, this paper describes a number of issues dealt with in the development of soft sensors for several deepwater gas-lifted oil wells. Some of the tested models include nonlinear polynomials, neural networks, committee machines, unscented Kalman filters and filter banks. The variety of model classes used in addition to the diversity of oil wells considered brings to light some of the key-problems that have to be faced and reveal the strengths and weaknesses of each alternative solution. A major constraint throughout the
work was the use of historical data, hence no specific tests were performed at any time. The aim of this work is to discuss the procedures, pros and cons of the tested solutions and to point to possible future directions of research
... Where n y is the number of previous output samples used, n u is number of previous control signal samples used and td is the time delay of the system. During the identification and control tasks the NNSYSID [12] and NNCTRL [13] toolboxes for MATLAB were used. Figure 9 shows the training and test signals used for training the models, separated by the hashed vertical line. ...
... It is also uncertain why the most common tool used for training ANN, Matlab, has not had a Levenberg-Marquardt version in its toolbox for several years. The first LM version diffused was in a toolbox freely distributed (Nørgaard, 1996a), which was a result of a PhD thesis (Nørgaard, 1996b) concluded in 1996. Only in the following decade did Matlab propose a LM version in its Neural Network toolbox. ...
Artificial Neural Networks are used in a variety of problems occurring either in
research or in the industry. The first step is to train a network to perform a desired
function, which requires a training algorithm. Levenberg-Marquardt is a second order
algorithm which outperforms Backpropagation and is currently available in most Neural
Network toolboxes. This paper tests two toolboxes, Neural Network Toolbox of MatLab
and Neural Network System Identification Toolbox, in order to demonstrate that the
implementations differ according to the toolbox used and that Matlab obtains better
results in all the datasets used. This paper also explains the differences between the
implementations of each tool and the advantages/disadvantages of each toolbox.
... When this initial learning is completed, the resulting model is pruned. The non robust pruning algorithm used is the OBS algorithm as implemented in [11]. In a first time, only one initial learning is performed with the robust criterion in order to obtain a satisfactory model before pruning (MSE in validation of 0.0054 with robust learning compared to 0.2992 with non robust learning for same initial weights). ...
Two ideas are combined in the parameters estimation and the structure determination of a one hidden layer perceptron for nonlinear model building. The first one is to start from an overparametrized structure and to use a weight elimination method to remove spurious parameters. The second one is to use a learning criterion robust to outliers. The ability of neural networks to model nonlinear relationships presents the risk to incorporate the noise as well as the gross errors in the model. A robust weighted criterion is thus introduced in a second order initial learning procedure and, for structure reduction, in the Optimal Brain Surgeon (OBS) algorithm. A simulation example illustrates the interest of the proposed approach, particularly a very stable residual criterion during pruning and good generalization until final structure selection.
... The training was performed using the neural network toolbox of Matlab (Beale, et al., 2011) and another toolbox that runs with Matlab, NNSYSID (Nørgaard, 2000), to check which solution produced the best results. The information about the data used for training and test is presented in tables 2 and 3. ...
The capacity to predict accurately the energy production in a wind park is
extremely relevant both from an economical point of view and to control the stability of
the electrical grid. Many different models have been used for this purpose, such as
physical, statistical, neuro-fuzzy and artificial neural networks. The data available from
wind parks is usually noisy and has measurements that are unexpected regarding the
available inputs. Dealing with this data and determine system characteristics is not an
easy task but, in this work, Artificial Neural Networks are used to predict power
generation based on in site wind measurements. The results show that Artificial Neural
Networks are a tool that should be considered under these difficult conditions, since they
provide a reasonable precision in the predictions. When compared with the results
presented in the literature the obtained results are in the same order of values, although
the data used in this work seems to have a large amount of outliers.
... The main advantage of MWOBS-parallel algorithms is that it requires fewer evaluations of the H -inverse and when coupled with a parallel approximation of the H -inverse, more speedup is achieved. Furthermore, we compared the performance of the pruning variants implemented in DANNP against the OBS implementation in NNSYSID MATLAB toolbox (Norgaard, Ravn & Poulsen, 2002). The NNSYSID is a toolbox for the identification of nonlinear dynamic systems with ANNs, often used as a benchmark, which implements several algorithms for ANN training and pruning. ...
Background
Artificial neural networks (ANNs) are a robust class of machine learning models and are a frequent choice for solving classification problems. However, determining the structure of the ANNs is not trivial as a large number of weights (connection links) may lead to overfitting the training data. Although several ANN pruning algorithms have been proposed for the simplification of ANNs, these algorithms are not able to efficiently cope with intricate ANN structures required for complex classification problems.
Methods
We developed DANNP, a web-based tool, that implements parallelized versions of several ANN pruning algorithms. The DANNP tool uses a modified version of the Fast Compressed Neural Network software implemented in C++ to considerably enhance the running time of the ANN pruning algorithms we implemented. In addition to the performance evaluation of the pruned ANNs, we systematically compared the set of features that remained in the pruned ANN with those obtained by different state-of-the-art feature selection (FS) methods.
Results
Although the ANN pruning algorithms are not entirely parallelizable, DANNP was able to speed up the ANN pruning up to eight times on a 32-core machine, compared to the serial implementations. To assess the impact of the ANN pruning by DANNP tool, we used 16 datasets from different domains. In eight out of the 16 datasets, DANNP significantly reduced the number of weights by 70%–99%, while maintaining a competitive or better model performance compared to the unpruned ANN. Finally, we used a naïve Bayes classifier derived with the features selected as a byproduct of the ANN pruning and demonstrated that its accuracy is comparable to those obtained by the classifiers trained with the features selected by several state-of-the-art FS methods. The FS ranking methodology proposed in this study allows the users to identify the most discriminant features of the problem at hand. To the best of our knowledge, DANNP (publicly available at www.cbrc.kaust.edu.sa/dannp ) is the only available and on-line accessible tool that provides multiple parallelized ANN pruning options. Datasets and DANNP code can be obtained at www.cbrc.kaust.edu.sa/dannp/data.php and https://doi.org/10.5281/zenodo.1001086 .
... Unfortunately, this is not the case in real applications, where the matrix is generally singular [see also observations following (2.10)]. To overcome such problems, different authors (see [27] for instance) evaluate the effective number of parameters by counting the number of non-null weights and biases. This procedure is definitely not correct and either (3.12) or (3.14) should be used. ...
The Internet of Things paradigm is supporting-and will support-an ever increasing number of services and applications impacting on almost every aspect of our everyday life. The current trend is forecasting IoT to connect tens of billions of objects by 2020 yielding a very high volume of data to be acquired, transmitted, and processed. IoT typically relies on cloud computing to process, analyze, and store the data acquired by IoT entities. Unfortunately, the need to transmit all data from the information producing objects to the cloud for a subsequent processing/analysis phase would require a large bandwidth and increase the latency in the decision making process whenever decisions/reactions must be promptly made by the IoT units. The fog computing paradigm aims to address these problems by extending cloud computing toward the edge of the network. In this direction, this article introduces a novel FC-IoT paradigm designed to move computing, storage, and applications/services close to IoT objects so as to reduce communication bandwidth and energy consumption as well as decision making latency. The proposed IoT-based solution has been designed to have intelligent and autonomous IoT objects that are integrated with an FC and fog networking approach. The distinguishing features of the intelligent FC-IoT platform are low latency, self-adaptation, low energy consumption, and spectrum efciency.
... The project was partially supported by Automatic Control Laboratory of the Swiss Federal Institute of Technology in Zurich. The neural network models have been computed using freeware NNSYSID Toolbox (Nørgaard, 1997), see also http://www.iau.dtu.dk/research/control/nnsysid.htm . Maria Silkey substantially improved the readability of this paper. ...
System identification methods are applied to predict avalanche hazard. A tenyear record of avalanche hazard and snow cover from Davos and Weissfluhjoch in the Swiss Alps is analyzed. Several models are estimated and validated. Dynamic models present a substantial improvement in the quality of hazard predictions compared to predictions derived from a commonly used static nearest neighbor model structure. Further possible modeling improvements are briefly discussed.
... As the measurements have noise the signals were filtered from high frequency noise. During the on-line training the NNSYSID (Nørgaard, 1996b) and NNCTRL (Nørgaard, 1996a) toolboxes for MATLAB were used. ...
The Levenberg-Marquardt algorithm is considered as the most effective one for training Artificial Neural Networks but its computational complexity and the difficulty to compute the trust region have made it very difficult to develop a true iterative version to use in on-line training. The algorithm is frequently used for off-line training in batch versions although some attempts have been made to implement iterative versions. To overcome the difficulties in implementing the iterative version, a batch sliding window with Early Stopping version, which uses a hybrid Direct/Specialized evaluation procedure is proposed and tested with a real system.
... The other three approaches are known as simplified models for building simulations. The linear parametric and neural network models have extensively applied in building simulations [221,[227][228][229][230]. This method has a shortcoming that intensive experimental data should be available in order to identify involved parameters. ...
In tropical climates, conventional air-conditioning systems contribute more than
70% to the electricity demand of buildings. This is mainly due to the high ambient humidity and the applied conventional method of cooling the air to below a necessary dew point in order to reach the desired dehumidification. Handling dehumidification load and sensible cooling load separately is a potential solution to considerably reduce the electricity demand for air-conditioning in the tropics if the dehumidification
process is driven by heat energy (e.g. solar thermal or waste heat) instead of
electricity.
This thesis focuses on theoretical and experimental analyses and optimizations of a
novel solar/waste heat assisted air-conditioning system for applications in tropical
climates. The proposed system comprises a solar/waste heat-driven two-stage
dehumidification system using membrane and desiccant technologies, and an
electricity-efficient unit for sensible cooling.
Detailed mathematical models for a membrane-based moisture and heat exchanger
and an evaporatively cooled desiccant dehumidification system (ECOS system) are
developed. Experiments with such a novel air-conditioning system were carried out at
SERIS, Singapore for comparing experimental observations and model based
simulation results.
Through simulation studies, the performance of individual system components and
a complete two-stage dehumidification and energy exchange system is assessed, and
the most influential design and material parameters of the membrane and the ECOS
units are identified. The membrane and desiccant materials are characterized
experimentally in order to provide materials properties for the simulation of the
systems.
Detailed optimization and sensitivity analyses are applied in order to generate
general information for the design of cost-effective systems with desirable performance for applications in tropical conditions. The best performance of the two-stage
dehumidification and energy recovery system is achieved if, as much as possible,
dehumidification and cooling are done by the membrane unit. On the other hand, the
adsorption based ECOS system is essential in order to reach the required level of
dehumidification, while the evaporative cooling process in the ECOS system has
significant impact on the improvement of the dehumidification performance and the
reduction of the air temperature.
Sensitivity analyses affirm that materials research based potential advances in mass
and heat transfer properties of membrane and adsorbent materials can significantly
further reduce the investment and operating costs of the system. The electricity
demand of the advanced air-conditioning system described in this thesis is reduced
significantly compared to that of a comparable fully electrically powered conventional
system under Singapore climate conditions. Under the assumption that sufficient
cheap heat energy at a temperature level of 70 - 80 °C is available for the regeneration
process, an assessment of the systems indicates that the total lifetime operating cost is
reduced by about 50% compared to that of a conventional air-conditioning system.
... Cloudless conditions were selected from the cloud-screened Level-1.5 dataset of one co-located Cimel sunphotometer belonging to NASAs AERONET (AErosol RObotic NETwork). 3 ANN DESCRIPTION Artificial neural networks are implemented here using a combination of custom-designed MATLAB functions [12] in conjunction with several routines developed elsewhere [13]. A standard multilayer perceptron (MLP) architecture with three fully interconnected layers (input, hidden and output) is employed, as shown in Figure 1. ...
Visibility is traditionally needed for air quality monitoring or air traffic control, and has become a key input to determine the transmission losses of solar radiation propagating between heliostats and the receiver of solar tower power (STP) plants. Recent studies suggest that haze can reduce visibility and increase these losses up to 25% compared to clear conditions. Monitoring visibility would thus needed for proper design and operation of STPs, but this is usually not done at all potential sites. Here, visibility's magnitude and variability are analyzed in terms of more common atmospheric variables: temperature, humidity, pressure, wind speed and precipitable water. To that effect, 1-min observations from a visibility meter located in Huelva (southwestern Spain) are analyzed over a 2-month period. Relative humidity is linearly correlated with visibility and explains 40% of its variability. This correlation is strongest under cloudless and daytime conditions. Using standard statistical techniques, no significant correlation is found between visibility and other atmospheric variables. Artificial neural networks (ANN) are thus investigated here for mapping the complex and non-linear relationships between visibility and multiple atmospheric inputs. This improves results significantly, increasing the explained visibility variability up to 72% and reducing RMSE from 40% to 30%. Moderate improvements in visibility estimation are further obtained when radiometric information (direct and diffuse solar irradiances) are added as ANN predictors. These findings show that visibility can be estimated from local atmospheric and radiometric observations using ANN, despite the complex and non-linear relationships between them.
... ANNs are implemented here using a combination of custom-designed MATLAB functions [15] in conjunction with several routines developed elsewhere [16]. A standard multilayer perceptron (MLP) architecture with three fully interconnected layers (input, hidden and output) is employed, as shown in Fig. 4. The hyperbolic tangent transform is chosen as the nonlinear activation function in the hidden layer, and the identity function is selected as the activation function for the output layer. ...
Visibility has become a key input to determine the transmission losses of solar radiation propagating between heliostats and the receiver of solar tower power (STP) plants. Recent studies suggest that haze can reduce visibility and increase these losses up to 25% compared to clear conditions. Monitoring visibility would thus be needed for proper design and operation of STPs, but this is usually not done at all potential sites. In this work, the dependence of visibility's magnitude on relative humidity (RH) and aerosol optical depth (AOD) at three different wavelengths is analyzed. To that effect, 1-min observations from a visibilimeter located in Huelva (southwestern Spain) are analyzed during the winter season. RH is linearly correlated with visibility and explains 46% of its variability. A complex non-linear relationship between visibility and AOD is observed with also dependence on RH. Artificial neural networks (ANN) are thus investigated here for mapping the complex and non-linear relationships between visibility, RH and AOD at multiple wavelengths. This improves results significantly, increasing the explained visibility variability up to 68% and reducing RMSD from 30% to 22% with almost zero bias. The ANN analysis shows that the visibility-AOD relationship is not sensitive to the specific wavelength at which AOD is measured. These findings show that visibility can be estimated from local observations of RH and AOD at only a single wavelength using ANN.
... Comparing the performance indices for training and test set (recall phase), it has been observed that the network overfits the data. This means that the selected model structure contains too many weights [11]. In order to find the best structure of NARX, the OBS strategy for pruning the neural network model is then used [12]. ...
The nonlinear autoregressive network with exogenous input (NARX) is used to perform hourly solar irradiation and wind speed forecasting, according to a multi-step ahead approach. Temperature has been considered as the exogenous variable. The NARX topology selection is supported by a combined use of two techniques: (1) a genetic algorithm (GA)-based optimization technique and (2) a method that determines the optimal network architecture by pruning (optimal brain surgeon (OBS) strategy). The considered variables are observed at hourly scale in a seven year dataset and the forecasting is done for several time horizons in the range from 8 to 24 h ahead.
... where n u and n y are the input and output lags, which correspond to n b and n a in [14] if the function F in the above equation is linear. Assume that a data set of N input-output pairs exists and the following regressor matrix can be formulated: ...
... Artificial neural networks are implemented here using a combination of custom-designed MATLAB functions (MatLab, 1999) in conjunction with several routines developed by Nørgaard (1997). A standard multilayer perceptron (MLP) architecture with three fully interconnected layers (input, hidden, and output) is employed, as shown in Fig. 2. The hyperbolic tangent transform is chosen as the nonlinear activation function in the hidden layer, and the identity function is selected as the activation function for the output layer. ...
This work analyses the influence of water vapor on the atmospheric transmission loss of solar radiation between heliostats and the receiver of solar power towers. To this purpose, an atmospheric transmission code (MODTRAN) is used to generate values of direct normal irradiance (DNI) reaching the mirror and the receiver under different geometries (including sun position, tower height and mirror-to-receiver slant range) and atmospheric conditions related to water vapor and aerosols. These variables are then used as inputs to an artificial neural network (ANN), which is trained to calculate the corresponding DNI attenuation. Two different aerosol scenarios are simulated: an ideal aerosol-free atmosphere and a widely different one corresponding to hazy conditions. The developed ANN model is then able to provide the DNI attenuation under any value of the input variables considered here with root mean square differences of only 0.8%. The transmission loss due to water vapor is found to be dependent on the sun position. The simplicity and accuracy of the algorithm are its great strengths, allowing its easy inclusion into the actual energy simulation codes currently used for solar tower plant design.
... Each clustering that is formed can be viewed as a class of objects from which rule can be derived. In clustering problems, you want a neural network to group data by similarity [4], [5], [6]. For example: market segmentation done by grouping people according to their buying patterns; data mining can be done by partitioning data into related subsets; or bioinformatics analysis such as grouping genes with related expression patterns. ...
... [9] : ( 1 ) 0 1 1 w w y l k m i k ki i k k b w x ( ضاتغٝ زض وٝ 1 ) k y ٘ ٛ ذطٚجی، ضٖٚ فؼاَ تاتغ ٘ ؾاظی ٛ ضٖٚ، m قثىٝ، ٚضٚزی پاضأتطٞای تؼساز i x ٘كا٘سٞٙسٜ i ٚضٚزی، پاضأتط أیٗ ki w ٘ ٞط ٚظٖ ٛ ٚ ضٖٚ k b ٔی ٚضٚزی پاضأتطٞای تایاؼ فؼاَ تاتغ تاقٙس. ضا ؾاظی ٔی تهٛضت تٛاٖ وطز تؼطیف ظیط ؾیٍٕٛئیس تاتغ : ( 2 ) x e x 1 1 ) ( قثىٝ زض ٟٓٔ ٔؿاِٝ ٘ تؼساز الیٝ چٙس ٔهٙٛػی ػهثی ٞای ٛ الیٝ x Z x f x Y ( ضاتغٝ زض 3 ) f(x) ٚ ضٌطؾیٖٛ تاتغ Z(x) پطزاظـ ٚاضیا٘ؽ ٚ نفط n p n p x b x b x b x b F F y x r M x 1 ( 8 ) F F x r F x r x r x s T T T 1 / 1 1 2 1 1 تاال ضٚاتظ زض x b x b x b M p ,..., , 2 1 زض ٔسَ ٔاتطیؽ ٘ظط، ٔٛضز ٔجَٟٛ ٘مغٝ y F F F T T 1 1 1 ٚ ضٌطؾیٖٛ تاتغ تٝ ٔطتٛط ضطایة تطزاض n x x x x x r , ...
In this paper 3 methods used to modeling and estimate of velocity and strain fields in Iran. Artificial neural network (ANN), kriging interpolation and polynomial fitting have been studied as accurate methods in crustal deformation. GPS measurements of Iranian permanent GPS network (IPGN) and Azerbaijan network (local network) is used to obtain the results. In IPGN, 4 and Azerbaijan network, 2 stations selected for testing the accuracy of methods. Using ANN, average relative error for northern component (VN), eastern component (VE) and height component (VZ) are computed 10.10%, 13% and 15.18%, respectively. For Azerbaijan network VN, VE and VZ are calculated respectively 18.99%, 8.61% and 21.53%. Similarly, for 3-D strain components average relative error in 4 test stations, exx, eyy and ezz obtained 13.29%, 12.56% and 17.97% respectively. In Azerbaijan network for 2 test stations, exx, eyy and ezz computed 10.76%, 13.52% and 16.01%. The results show the capability and efficiency of ANN method in comparison with kriging and polynomial models. As well as, it is found that polynomial and kriging methods required many computational points in adjustment step.
... If a linear model does a decent job, one should not bother wasting time on fancy neural network based model structures. To identify a linear OE model, the System Identification Toolbox from The MathWorks, Inc. is employed [10]. Fig. 7 & 8 show validation of linear model. ...
... Diseño de un controlador neuronal utilizando el modelo inverso. 4 En el diseño la red neuronal computa directamente la variable de control para hacer que la salida actual de la planta, siga el valor deseado (set-point). Teniendo esto en cuenta el objetivo del aprendizaje consiste en modelar la dinámica entre la entrada y la salida de la planta. ...
El artículo aborda el uso de las Redes Neuronales (RN) para el diseño de un controlador. En este caso se controla la velocidad del rotor de un motor de corriente directa (CD) para que siga una trayectoria deseada especialmente cuando las características del motor y la carga no son conocidas. La dinámica desconocida del motor y la carga son “identificadas” por la red neuronal artificial en la fase de entrenamiento y una vez que dicha red está entrenada se evalúa la capacidad del controlador neuronal para seguir una referencia deseada. Estos algoritmos fueron evaluados en simulaciones sobre Matlab [4] y probados en un motor de CD a nivel de laboratorio. Con este trabajo se demuestra la efectividad de un controlador neuronal en sistemas de control donde no se conocen sus parámetros y las características de retardo.
... e recursive algorithm autoregressive-moving average with the exogenous inputs (RARMX) method is a model training based on the Gaussian Newton method [32]. First, the unit sampling interval can be defined as t � 1, 2, . . ...
Data-driven damage identification based on measurements of the structural health monitoring (SHM) system is a hot issue. In this study, based on the intrinsic mode functions (IMFs) decomposed by the empirical mode decomposition (EMD) method and the trend term fitting residual of measured data, a structural damage identification method based on Mahalanobis distance cumulant (MDC) was proposed. The damage feature vector is composed of the squared MDC values and is calculated by the segmentation data set. It makes the changes of monitoring points caused by damage accumulate as “amplification effect,” so as to obtain more damage information. The calculation method of the damage feature vector and the damage identification procedure were given. A mass-spring system with four mass points and four springs was used to simulate the damage cases. The results showed that the damage feature vector MDC can effectively identify the occurrence and location of the damage. The dynamic measurements of a prestress concrete continuous box-girder bridge were used for decomposing into IMFs and the trend term by the EMD method and the recursive algorithm autoregressive-moving average with the exogenous inputs (RARMX) method, which were used for fitting the trend term and to obtain the fitting residual. By using the first n-order IMFs and the fitting residual as the clusters for damage identification, the effectiveness of the method is also shown.
... Many techniques are performed for detection of the motor faults [21][22][23][24][25][26][27][28][29][30][31][32][33][34][35][36][37][38]. The previous procedure are deal with the linear model of the induction motor and deal with the online diagnostics of the motor fault detection, from the previous work we find that many factors are lead to motor faults such that bearing faults induce 40% of the motor faults , 38% are due to the stator winding , 10% are due to rotor faults and 12% other faults . ...
... A NARX model is obtained for the tall structure using Neural Network System Identification Toolbox [23]. The data shown in Fig. 12 is used for this purpose. ...
This article presents the use of neural network predictive controller as a novel technique for vibration control of tall structures employing single degree of freedom active tuned mass damper (ATMD). Additionally, the proposed technique is compared with two modern control techniques: pole-placement controller and adaptive neuro-fuzzy inference controller. A scaled down laboratory model is used to validate the control techniques. A linear and a nonlinear auto-regressive exogenous (ARX) models are identified for the constructed structure. A neural network predictive controller is designed using the nonlinear ARX model. Polynomial and state-space pole-placement controllers are designed using the linear ARX model. A fuzzy logic controller is designed for the structure and trained using adaptive neuro fuzzy inference system (ANFIS). Hardware-in-the-loop implementation of these controllers demonstrates that the neural network predictive controller combines the advantages of both pole-placement and the ANFIS controllers and reduces the settling time of the mass damper six times with same amplitude mitigation.
... Then, the obtained values of neurons, which create an initial response by combining together, enter a process known as back-propagation and begin to modify the primary weights by comparing the error rate and calculating the impact of each entry in all layers. Reaching the least error rate between the output data of the model and the target data entered as input into the model is the stop condition of the process [45]. ...
Irregularities in electron density usually correlate with ionospheric plasma perturbations. These variations causing radio signal fluctuations, in response, generate ionospheric scintillations that frequently occur in low-latitude regions. In this research, the combination of an artificial neural network (ANN) with the genetic algorithm (GA) was implemented to predict ionospheric scintillations. The GA method was considered for obtaining the ANN model's initial weights. This procedure was applied to GNSS observations at GUAM (13.58°E, 144.86°N, 201.922H) station for the daily prediction of ionospheric amplitude scintillations via predicting the signal-to-noise ratio (S4) or via prediction of the rate of TEC index (ROTI). Thirty-day modeling was carried out for three months in January, March, and July, representing different seasons of the winter solstice, equinox, and summer solstice during three different years, 2015, 2017, and 2020, with different solar activities. The models, along with ionospheric physical data, were used for the daily prediction of ionospheric scintillations for the consequent day after the modeling. The prediction results were evaluated using S4 derived from GNSS observations at the GUAM station. The designed model has the ability to predict daily ionospheric scintillations with an accuracy of about 81% for the S4 and about 80% for the ROTI.
UAV control system is a huge and complex system, and to design and test a UAV control system is time-cost and money-cost. This paper considered the simulation of identification of a nonlinear system dynamics using artificial neural networks approach. This experiment develops a neural network model of the plant that we want to control. In the control design stage, experiment uses the neural network plant model to design (or train) the controller. We use Matlab to train the network and simulate the behavior. This chapter provides the mathematical overview of MRC technique and neural network architecture to simulate nonlinear identification of UAV systems. MRC provides a direct and effective method to control a complex system without an equation-driven model. NN approach provides a good framework to implement MEC by identifying complicated models and training a controller for it.
A procedure for the short-term prediction of the thermal energy consumption of an hospital is shown in this paper. At first, linear ARX models are built to get information on the influence of the input variables on the output of the system. Therefore, non-linear models based on feedforward neural networks (NNARX) are built using the information provided by the linear estimate. The results obtained from the ARX and NNARX models are compared, concluding that NNARX models provide better results than ARX models, but the analysis of ARX models is necessary to obtain guidelines in the choice of the best regression vector as input for neural models.
Conventional closed loop speed control of electro-mechanical drives incorporating dc or ac motors requires the use of a speed transducer, typically a tachogenerator or shaft encoder. In order to improve reliability and to eradicate mounting and wiring problems it is often desirable to eliminate such transducers from the system whilst still retaining good control of speed.
A number of speed estimator schemes have previously been proposed and developed including model-based schemes which require a knowledge of appropriate machine equations and parameters and estimator/observerbased schemes which are also computationally intensive. This paper examines the shortcomings of these approaches and demonstrates the need for methods which do not require detailed, ‘a priori’ drive modelling information and which are not computationally demanding. The paper presents two artificial-intelligence-based solutions for the development of speed estimators for use in electro-mechanical drives and which meet the above constraints. The first is a fuzzy logic-based system processing information from a simple microphone pick-up; the alternative is a neural network-based system which processes voltage and current information.
Analysis of acoustic noise generated by electrical machines can be used for both condition monitoring and control purposes. The paper describes the development of a fuzzy-logic-based system and presents experimental results to demonstrate that such a system is capable of extracting speed information from the noise signal which is of a quality suitable for control purposes.
Details of the structure, design and development of recursive ANN-based speed estimators for both dc and ac induction motor drives is also given. Experimental results obtained from real-time DSP implementations of speed controlled drives validate the performance of the speed estimators and demonstrate that such estimators can be used to give a control performance comparable to that obtainable from conventional closed loop control schemes. It is also shown that ANN-based estimators have a small computational overhead and require only easily obtainable drive response data for their development.
Based on 42 GPS velocity vectors existing in NW Iran, we estimated new velocity vectors for some arbitrary positions in study area by using two activation functions of Back Propagation Artificial Neural Networks (BPANN): “Sigmoid Function” and “Continuous Log-Sigmoid Function”. The results show sigmoid function’s RMSEs in E and N components are ±1 mm and ±1.5 mm respectively which is less than Continuous Log-Sigmoid Function RMSE. So it seems BPANN method could be considered as a good alternative to estimate geodetic velocity field relative to other classical estimation methods.
Two MATLAB packages have been implemented: the Neural Network Based System Identification toolbox (NNSYSID) and the Neural Network Based Control System Design Toolkit (NNCTRL). The NNSYSID toolbox has been developed to assist identification of nonlinear dynamic systems and it offers the possibility to work with a number of different nonlinear model structures based on neural networks. The NNCTRL toolkit is an add-on to the NNSYSID toolbox and contains tools for design and simulation of control systems based on neural networks. This paper gives an overview of the contents of NNSYSID and NNCTRL.
Surrogate models are used to approximate complex problems in order to reduce the final cost of the design process. This study has evaluated the potential for employing surrogate modelling methods in turbo-machinery component design optimization. Specifically four types of surrogate models are assessed and compared, namely: neural networks, Radial Basis Function (RBF) Networks, polynomial models and Kriging models. Guidelines and automated setting procedures are proposed to set the surrogate models, which are applied to two turbo-machinery application case studies.
As parsimonious and flexible universal approximator, the one hidden layer perceptron can be used for non linear prediction, An application is described in the framework of Air-Fuel Ratio (AFR) control in spark-ignition engines, a critical point to satisfy pollutant emission legislation, AFR control depends essentially on the prediction of the air mass to be admitted in cylinder. The building of an air mass predictive neural network is described and its performances are evaluated, Compared to classical solutions based on static mappings, the neural predictor allows for reduction of AFR excursions on rapid torque transients
This article presents a comparison of Artificial Neural Networks and NeuroFuzzy Systems applied for modeling and controlling a real system. The main objective is to control the temperature inside of a ceramics kiln. The details of all system components are described. The steps taken to arrive at the direct and inverse models using the two architectures: Adaptive Neuro Fuzzy Inference System and Feedfoward Neural Networks are described and compared. Finally, real time control results using Internal Model Control strategy are presented.
A nonlinear control law is applied to an experimental pH neutralization process. A partitioned model is used to capture the dynamic behavior of the process, consisting of a linear ARX model and a nonlinear neural network model. When integrated into the internal model control scheme, the resulting controller is shown to have better performance compared to the linear controller for the pH neutralization process.
Many current industry branches use hybrid approaches to solve complex applicationproblems.Overthelastdecades,differenttoolsforthesimulationofsuchhybridsystems(e.g. Hysdel and YAMLIP) as well as the identification of hybrid systems (e.g. HIT, MLPand OAF NN) have been developed. The framework presented in this work facilitates theintegration of artificial feed-forward neural networks in the modelling process of hybriddynamical systems (HDS). Additionally, the framework provides a structured languagefor characterising these feed-forward networks itself. Therefore, an interdisciplinaryexchange in the field of neural networks and its integration into hybrid dynamicalsystems is enabled. Focusing on hybrid systems with autonomous events, two differentapproaches, namely the artificial hybrid model and the artificial hybrid dynamics, areintroduced. Challenges of the modelling process of HDS are reflected and advantagesas well as disadvantages are discussed. The case study includes two common examplesof HDS and analyses the simulation results and examines limitations of the modellingframework.
https://authors.elsevier.com/a/1dF7c,dqti14En
This work analyses the influence of water vapor on the atmospheric transmission loss of solar radiation between heliostats and the receiver of solar power tower plants. To this purpose, an atmospheric transmission code (MODTRAN) is used to generate values of direct normal irradiance (DNI) reaching the mirror and the receiver under different geometries (including sun position, tower height, and mirror-to-receiver slant range) and atmospheric conditions related to water vapor and aerosols. These variables are then used as inputs to an artificial neural network (ANN), which is trained to calculate the corresponding DNI attenuation. Two different aerosol scenarios are simulated: an ideal aerosol-free atmosphere, and a widely different one corresponding to semi-hazy conditions. The developed ANN model is then able to provide the DNI attenuation over a wide range of the input variables considered here, with root mean square differences of only 0.8%. The transmission loss due to water vapor is found to decrease with sun elevation. This is explained by the saturation effect in the incident irradiance at the mirror. The simplicity and accuracy of the algorithm are its great strengths, allowing its anticipated inclusion into the actual energy simulation codes currently used for solar tower plant design.
Proton exchange membrane fuel cells are considered as one of the most promising power sources for transportation in the near future. Proper and robust thermal management is a key issue in fuel cell applications. This paper mainly studies the sensor fault detection and isolation and fault tolerant control for the thermal management of proton exchange membrane fuel cell systems. The thermal model of fuel cell is established and analyzed by structure analysis, and the residual generator of sensor faults are designed by Dulmage Mendelsohn decomposition to rearrange the bi-adjacency matrix of the thermal model. A sliding-mode-based active fault tolerant control strategy is proposed for the thermal management of fuel cell systems. The effectivenesses of the proposed fault detection and isolation method and active fault tolerant control strategy are verified on a fuel cell test bench. The experimental results show that the temperature of the proton exchange membrane fuel cell stack can be maintained at the reference value with high accuracy even when sensor fails.
The ionospheric plasma bubbles cause unpredictable changes in the ionospheric electron density. These variations in the ionospheric layer can cause a phenomenon known as the ionospheric scintillation. Ionospheric scintillation could affect the phase and amplitude of the radio signals traveling through this medium. This phenomenon occurs frequently around the magnetic equator and in low latitudes, mid as well as high latitude regions. ionospheric scintillation is a very complex phenomenon to be modeled. Patterns of ionospheric scintillation occurrence are depended on spatial and temporal ionospheric variabilities. Neural Network (NN) is a data-dependent method, that its performance improves with the sample size. According to the advantages of NN for large datasets and noisy data, the NN model has been implemented for predicting the occurrences of
amplitude scintillations. In this paper, the GA technique was considered to obtain primary weights in the NN model to identify appropriate S4 values for GUAM GPS station in Guam country (latitude: 144.8683, Longitude:13.5893). The modeling was carried out for the whole month of June 2017, while this model along with ionospheric physical data was used for predicting ionospheric scintillation on the first day of July 2017, the day after the modeling. The designed model has the ability to predict daily ionospheric scintillation with an accuracy of about 78%.
A system for monitoring the life cycle of a road using innovative systems for controlling the working bodies of road machines based on GPS technology is developed. The basic principles of the intellectualization of modern road machines are given. A structural diagram and a mathematical model of the intellectual system of road machines, which provides adaptive optimization of working processes in the aspect of modern trends of BIM technologies in the road sector, are proposed. The prospects for managing the design, construction, and operation of road construction objects are considered.
Composite systems, encompassing soft-computing and conventional processing modules, are attractive approaches to tackle complex applications, especially those requiring adaptability and generalization abilities.
This chapter presents an innovative high-level design methodology that provides a comprehensive framework to achieve the most efficient and effective solutions. The system is first specified at a high-abstraction level by using description languages suited to capture the system behavior and the other non-functional constraints. In the computational paradigm co-design the system description is then partitioned into modules, described with various computational paradigms, and modules are completely configured. The result is an algorithmic description of the whole system, from which conventional hardware/software co-design techniques can derive the final implementation.
ResearchGate has not been able to resolve any references for this publication.