Neural Network Topology Research Articles

State of charge estimation for a group of lithium-ion batteries using long short-term memory neural network

The present paper estimates for the first time the State of Charge (SoC) of a high capacity grid-scale lithium-ion battery storage system used to improve the power profile in a distribution network. The proposed long short-term memory (LSTM) neural network model can overcome the problems associated with the nonlinear battery model and adapt to the complexity and uncertainty of the estimation process. The accuracy of the developed model was compared with results obtained from Feed-Forward Neural Network (FFNN) topology and Deep-Feed-Forward Neural Network (DFFNN) topology under three different time series. The system was trained using real data from the Al-Manara PV power plant. The LSTM with learn-and-adapt-to-train-date properties, as well as the idea of “forget gate,” shows exceptional ability to determine the SoC under various ID data. The LSTM properly calculated the SoC for all three-time models with a maximum standard error (MSE) of less than 0.62%, while the FFNN and DFFNN provided a fair estimate for the SoC with MSEs of 5.37 to 9.22% and 4.03 to 7.37%, respectively. The promising results can lead to excellent monitoring and control of battery management systems.

Economic Forecasting Model Based on Chaos Simulated Annealing Neural Network

In view of the prototype of the traditional chaotic network model, a neural network model with continuously updated chaotic noise is innovatively improved. This model has the advantages of two network models. On this basis, a logic graph is proposed, which can replace the chaotic noise generated by the running process of the model function. Models with these advantages can innovatively solve problems such as constrained optimization with dimensions larger than three, high discreteness, and weak linear relationship between convex and concave. Simulation results can be confirmed that the algorithm of this model is very close to the predicted value. This neural network model is an improved model with strong applicability and can be applied to optimization problems of economic systems or other industrial systems. In order to effectively alleviate the predictive control problem of nonlinear research objects, we propose a control method based on chaotic neural network in this study. Taking the economic model as the nonlinear object, establishing the basic structure of the chaotic neural network model, reconstructing the time and space structure, and obtaining the optimal solution of time continuation and embedding dimension through information entropy and pseudonearest neighbor, then the chaotic properties of nonlinear objects and the topology of chaotic neural networks are determined. In the simulation, test samples and experimental samples are established, and the predicted and true values are compared. The prediction results of the model established by this method in this study prove the effectiveness of the method. The training time of the chaotic neural network will not fall into a local minimum, thereby reducing the training time and ensuring the accuracy of the prediction.

A Data-Driven Model to Forecast Multi-Step Ahead Time Series of Turkish Daily Electricity Load

It is critical to maintain a balance between the supply and the demand for electricity because of its non-storable feature. For power-producing facilities and traders, an electrical load is a piece of fundamental and vital information to have, particularly in terms of production planning, daily operations, and unit obligations, among other things. This study offers a deep learning methodology to model and forecast multistep daily Turkish electricity loads using the data between 5 January 2015, and 26 December 2021. One major reason for the growing popularity of deep learning is the creation of new and creative deep neural network topologies and significant computational advancements. Long Short-Term Memory (LSTM), Gated Recurrent Network, and Convolutional Neural Network are trained and compared to forecast 1 day to 7 days ahead of daily electricity load. Three different performance metrics including coefficient of determination (R2), root mean squared error, and mean absolute error were used to evaluate the performance of the proposed algorithms. The forecasting results on the test set showed that the best performance is achieved by LSTM. The algorithm has an R2 of 0.94 for 1 day ahead forecast, and the metric decreases to 0.73 in 7 days ahead forecast.

Real time obstacle motion prediction using neural network based extended Kalman filter for robot path planning

Navigation in dynamic environments for mobile robots is a difficult problem as it involves estimating the path of moving obstacles. The measured data usually contains a bias and noise in addition to its true value. Based on a stacked denoising autoencoder (SDAE), the enhanced Kalman filter developed in this paper can estimate the obstacle position from any type of noisy input. The extended Kalman filter's ability to predict an error-free path is impacted by the measurement noise covariance matrix employed. The SDAE is a neural network topology based on deep learning that can be used to determine the optimum covariance matrix. Both Adam and stochastic gradient learning algorithms are used to train the neural network. The robot's path is re-planned based on the predicted obstacle path to ensure safe navigation. MATLAB-based numerical simulations are used to demonstrate the utility and superiority of the proposed method over the traditional Kalman filter and Particle filter methodologies. The simulation results show that in the presence of any sort of noise, the proposed technique is exceptionally durable and reliable. The simulation findings also reveal that when it comes to denoising the measured data, the stacked denoising autoencoder with Adam optimizer is more efficient than the stochastic approach. The performance of the developed algorithm is validated in MATLAB simulated environments, and it can be extended for navigation tasks. In terms of computation time and robustness in closely spaced obstacles, simulation experiments demonstrated that the path planning using the proposed algorithm outperforms the hybrid A star, artificial potential field, and decision algorithms.

Smart Road Damage Detection and Warning using Machine Learning

We present a neural network topology, as well as training and prediction algorithms, in this research. To create a safe road environment, we present a deep neural network technique to detect road surface deterioration conditions. For training and testing, we provide an image dataset as input. Various sorts of road anxiety are depicted in the photographs. The suggested approach is compared to a variety of deep learning models from different disciplines. The findings of this study are expected to play a significant role in guaranteeing safe driving in the future by effectively detecting poor road conditions.

The Influence of Genetic Algorithms on Learning Possibilities of Artificial Neural Networks

The presented research study focuses on demonstrating the learning ability of a neural network using a genetic algorithm and finding the most suitable neural network topology for solving a demonstration problem. The network topology is significantly dependent on the level of generalization. More robust topology of a neural network is usually more suitable for particular details in the training set and it loses the ability to abstract general information. Therefore, we often design the network topology by taking into the account the required generalization, rather than the aspect of theoretical calculations. The next part of the article presents research whether a modification of the parameters of the genetic algorithm can achieve optimization and acceleration of the neural network learning process. The function of the neural network and its learning by using the genetic algorithm is demonstrated in a program for solving a computer game. The research focuses mainly on the assessment of the influence of changes in neural networks’ topology and changes in parameters in genetic algorithm on the achieved results and speed of neural network training. The achieved results are statistically presented and compared depending on the network topology and changes in the learning algorithm.

Strength Model of Cemented Filling Body Based on a Neural Network Algorithm

As one of the key measures for comprehensive management of goaf in various mines, filling mining has been recognized by practitioners in recent years due to its functions (e.g., resource utilization of solid waste and thorough goaf treatment). The performance of the filling material is the core challenge of filling mining, and it is influenced by the settling speed, conveying characteristics, and filling body strength. To understand the strength characteristics of a cemented filling body composed of medium-fine tailings, in this study, filling material ratio tests under different content of cement, tailings, and water were conducted. A backpropagation (BP) neural network topology structure was established in this study. The strength after different curing times was used as the output variable to analyze the impact of the cement, tailings, and water content on the filling body. A 3-Hn-3 structural model was employed. When the number of hidden layers Hn was 7, the model achieved the best learning and training effect. The results show that the predicted value, which is close to the measured value (fitting accuracy of 92.43–99.92%; average error of 0.0792–7.5682%), satisfies the engineering requirements. The neural network model can be employed to predict the filling body’s strength and provide a good reference to analyze the change law in the filling body’s strength.

A new empirical model and neural network-based approach for evaluation of isobaric heat capacity of natural gas

Isobaric heat capacity of natural gas is of prominent thermophysical property as it is directly related to thermodynamic energy functions, thereby its trustworthy prediction can open a new window for better establishment of the basis for its theoretical and engineering studies. In the present study, a new accurate and simple empirical correlation as a function of specific gravity (γg), temperature (T), and pressure (P) was developed to rapidly estimate the isobaric specific heat capacity (Cp) of natural gas without using gas composition. Due to the limitations of proposed experimental models and relations derived from gas compressibility factor (Z-factor) equations in the literature to a certain temperature, pressure and specific gravity range, an artificial neural network (ANN) based on back-propagation method was also applied for reliable prediction of Cp of natural gas. 847 sets of data from a diverse range of T, P, and γg were used to develop the neural network architecture and topology. Moreover, Genetic algorithm (GA) and Particle swarm (PS) optimization as population-based stochastic search algorithms were also utilized to optimize the weights and biases of networks to establish the best combination of input variables leading to the minimum Cp. The ANN demonstrated an accurate and promising prediction with a correlation coefficient of 0.99691 and 0.99518 for total dataset and test data, respectively.

Dimensionality reduction method of dynamic networks for evolutionary mechanism of neuronal systems

Understanding the development process of biological neuronal systems is one of the major challenges to explore the formation mechanism of human brain intelligence and biological behaviors, which is of great importance to enlighten the designing of various artificial neural network algorithms as well. Here, with the whole neural connection map of C.elegans, we explore the evolutionary mechanism of its neuronal system based on mean-field theory and network dimension reduction method. Firstly, we use a set of activation equations to capture the neurons’ interaction in the networks, and adopt the dimensionality reduction framework to decouple the nematode’s multi-dimensional neural network into an one-dimensional system. Then we propose two control index after the system is decoupled, which can be used to describe the structure and stable state of the whole neural network, as well as the development of an organism on the basis of birth time and process length. Our theoretical approach of multidimensional system to the one-dimensional system can help distract our attention from the micro-dynamics of a single neuron to the macro-dynamics of the whole network. Our results reveal some factors that influence the evolution of neural system, explore the evolutionary constraints and contribution of these factors at the level of neural networks function and topology. In conclusion, our analytical framework provides an overview of quantitative understanding of the growth process and the evolutionary constraints in the neural system.

Optimization and Power Management of Solar PV-based Integrated Energy System for Distributed Green Hydrogen Production

Photovoltaic-based integrated energy systems act as a possible modern technological solution for clean and affordable green hydrogen production. However, research attempts are required from the scientific community to develop power management, control, optimization algorithms, and new techniques for these integrated energy systems effective and economical operation. The integrated energy system considered in this work consists of a solar photovoltaic, battery, grid and proton exchange membrane (PEM) electrolyzer. PEM electrolyzer with a power rating of 100 kW is modelled as a controlled current sink to interfere with the DC bus directly. This work proposes a power management and control algorithm for the photovoltaic-based integrated energy system considering different parameters like availability of solar photovoltaic power, DC link voltage, battery state of charge (SOC), tariff and availability of grid power. Effective power-sharing among the different power sources increases system reliability and stability. Optimization is performed for optimal sizing and costing to achieve economical operation such that these photovoltaic-based integrated energy systems become affordable and are encouraged for wide use. The photovoltaic system operates at the maximum power point through a neural network-based control strategy; in addition to this, stacking in neural network topologies is also discussed for improving system accuracy in tracking the reference voltage for MPPT operation. The DC link voltage is controlled by interfacing the battery with the DC link via a bidirectional buck-boost power converter. The contribution of the work is to highlight the role of renewable energy sources for continuous green hydrogen production. Green hydrogen will be a critical driving force for the future transportation system, fuel cell-based power generation, and industries utilizing hydrogen in direct or indirect forms-that will help transition towards a carbon-free society. Optimization and time-domain simulation results indicated the economic and technical feasibility of the proposed photovoltaic-based integrated energy system for green hydrogen production with the best optimal configuration producing hydrogen at the cost of $ 4.806/kg and total net present cost of $749904.

Optimization of an artificial neural network topology using response surface methodology for microbial fuel cell power prediction.

Microbial fuel cells (MFCs) are among the newest bioelectrical devices that have attracted significant attention because they convert biodegradable organic matter to electricity. MFC design can be improved by understanding and predicting the performance of MFC under different conditions and substrate concentrations. However, few mathematical models have been investigated due to problems caused by the high sensitivity of MFC systems. In this research, a multilayer neural network (NN) was used to predict the generated power of a cell with three inputs (concentration, time, and resistance). Response surface methodology with factors including the number of first layer neurons, number of second layer neurons, training epochs, validation check, and training percentage was used to obtain the optimum structure of the network, and mean squared error (MSE). NN had the minimum MSE when the Number of neurons in the first and second hidden layers, the training epochs, validation check, training percentage were 28, 20, 1000, 100, and 70, respectively. This built network had an excellent ability to predict, and was 98%. According to the results, increasing COD concentration increases generated power and system utilization time. In addition, reducing the external resistance up to 100 Ω can lead to more power obtained.

Forecasting performance of wavelet neural networks and other neural network topologies: A comparative study based on financial market data sets

In this study, we analyse the advantageous effects of neural networks in combination with wavelet functions on the performance of financial market predictions. We implement different approaches in multiple experiments and test their predictive abilities with different financial time series. We demonstrate experimentally that both wavelet neural networks and neural networks with data pre-processed by wavelets outperform classical network topologies. However, the precision of conducted forecasts implementing neural network algorithms still propose potential for further refinement and enhancement. Hence, we discuss our findings, comparisons with “buy-and-hold” strategies and ethical considerations critically and elaborate on future prospects.

Hybrid optimized convolutional neural network for efficient classification of ECG signals in healthcare monitoring

For the prediction of cardiovascular diseases, a wireless sensor network platform that employs a wireless sensor network-enabled electrocardiogram telemetry system will involve the following steps: the electrocardiogram signal’s acquisition, the electrocardiogram signal’s processing, and alerting the physician in case of any emergencies. This system will aid the physician in early as well as an accurate analysis of heart diseases. Currently, a wireless sensor network enabled electrocardiogram monitoring system is being developed for analyzing the electrocardiogram signal. In the past, the convolutional neural networks have been employed for resolving diverse artificial intelligence problems with good outcomes. Even so, designing the convolutional neural networks architecture continues to be a cumbersome as well as meticulous procedure that needs the field experts’ involvement. This work has explored the neuro-evolution application to the automatic design of convolutional neural networks topologies and also has devised a novel solution which is based on the Artificial Bee Colony as well as the Grey Wolf Optimizer. Moreover, this work has given the proposal of the hybrid Grey Wolf Optimizer-Artificial Bee Colony algorithm, in which the wolves will adopt the bees’ strategy of sharing information for promotion of their exploration capability whilst continuing with their original strategy of hunting strategy for maintenance of their exploitation capability. The experimental outcomes demonstrate the proposed algorithm’s superior performance over other existing algorithms.

A store-and-forward neural network to solve multicriteria optimal path problem in time-dependent networks

This paper introduces the constrained multi-objective optimal path problem in time-dependent networks. In the existing literatures, the constraints are all imposed on the objective function while the problem constraints are related to the non-objective function. It is the difference that makes the traditional algorithm unable to get a better solution quality. In this light, we propose a store-and-forward neural network (SFNN) that finds the better result. In the design of SFNN, the topology of neural network is the same as that of time-varying network, and each node is designed as store-and-forward neuron. Each neuron transmits information to other neurons by sending signals. The experimental results show that compared with the traditional methods, the accuracy is significantly improved when the calculation time is acceptable.

Adaptive ML-based technique for renewable energy system power forecasting in hybrid PV-Wind farms power conversion systems

Large scale integration of renewable energy system with classical electrical power generation system requires a precise balance to maintain and optimize the supply–demand limitations in power grids operations. For this purpose, accurate forecasting is needed from wind energy conversion systems (WECS) and solar power plants (SPPs). This daunting task has limits with long-short term and precise term forecasting due to the highly random nature of environmental conditions. This paper offers a hybrid variational decomposition model (HVDM) as a revolutionary composite deep learning-based evolutionary technique for accurate power production forecasting in microgrid farms. The objective is to obtain precise short-term forecasting in five steps of development. An improvised dynamic group-based cooperative search (IDGC) mechanism with a IDGC-Radial Basis Function Neural Network (IDGC-RBFNN) is proposed for enhanced accurate short-term power forecasting. For this purpose, meteorological data with time series is utilized. SCADA data provide the values to the system. The improvisation has been made to the metaheuristic algorithm and an enhanced training mechanism is designed for the short term wind forecasting (STWF) problem. The results are compared with two different Neural Network topologies and three heuristic algorithms: particle swarm intelligence (PSO), IDGC, and dynamic group cooperation optimization (DGCO). The 24 h ahead are studied in the experimental simulations. The analysis is made using seasonal behavior for year-round performance analysis. The prediction accuracy achieved by the proposed hybrid model shows greater results. The comparison is made statistically with existing works and literature showing highly effective accuracy at a lower computational burden. Three seasonal results are compared graphically and statistically.

Performance evaluation of neural networks in modeling exhaust gas aftertreatment reactors

A comprehensive evaluation of neural network (NN) methods to model exhaust gas aftertreatment reactors is presented in this study. A combination of non-linear autoregressive model with exogenous inputs (NARX) network and static NN is utilized to capture the dynamic behavior of the state and outlet variables, such as coverage fractions, temperatures, and conversions, of monolith reactors used in the abatement of automobile emissions. The efficacy of this methodology is illustrated by modeling the diesel oxidation catalyst (DOC) and selective catalytic reduction (SCR) reactors of a diesel engine exhaust gas aftertreatment system. The final NN topology is designed by carefully comparing the model performances in the training and testing datasets. Model parsimony quantified by Akaike’s Information Criterion is employed for choosing the number of input and feedback delays in the NARX network. The novelty of this work lies in generating quality training and validation data using a physics-based system level model comprising of the engine and aftertreatment reactor components. The accuracy and generalizability of the NN models are demonstrated by comparing the model predictions with physical model simulation results in five disparate test datasets used in vehicle certification protocols. Overall, the NN models are found to reproduce all the qualitative trends in the various state and output variables while needing minimum computing resources resulting in better than real-time performance. Therefore, with appropriate training, these models can be embedded in the electronic control unit (ECU) and utilized for model-based predictive control systems.

Neural network assisted design of plasmonic nanostructures on superconducting transition-edge-sensors for single photon detectors

Transition edge sensors (TESs) are extremely sensitive thermometers made of superconducting materials operating at their transition temperature, where small variations in temperature give rise to a measurable increase in electrical resistance. Coupled to suitable absorbers, they are used as radiation detectors with very good energy resolution in several experiments. Particularly interesting are the applications that TESs may bring to single photon detection in the visible and infrared regimes. In this work, we propose a method to enhance absorption efficiency at these wavelengths. The operation principle exploits the generation of highly absorbing plasmons on the metallic surface. Following this approach, we report nanostructures featuring theoretical values of absorption reaching 98%, at the telecom design frequency (λ = 1550 nm). The optimization process takes into account the TES requirements in terms of heat capacity, critical temperature and energy resolution leading to a promising design for an operating device. Neural networks were first trained and then used as solvers of the optical properties of the nanostructures. The neural network topology takes the geometrical parameters, the properties of materials and the wavelength of light as input, predicting the absorption spectrum at single wavelength as output. The incorporation of the material properties and the dependence with frequency was crucial to reduce the number of required spectra for training. The results are almost indistinguishable from those calculated with a commonly used numerical method in computational electromagnetism, the finite-difference time-domain algorithm, but up to 106 times faster than the numerical simulation.

Identification of Flux Rope Orientation via Neural Networks

Geomagnetic disturbance forecasting is based on the identification of solar wind structures and accurate determination of their magnetic field orientation. For nowcasting activities, this is currently a tedious and manual process. Focusing on the main driver of geomagnetic disturbances, the twisted internal magnetic field of interplanetary coronal mass ejections (ICMEs), we explore a convolutional neural network’s (CNN) ability to predict the embedded magnetic flux rope’s orientation once it has been identified from in situ solar wind observations. Our work uses CNNs trained with magnetic field vectors from analytical flux rope data. The simulated flux ropes span many possible spacecraft trajectories and flux rope orientations. We train CNNs first with full duration flux ropes and then again with partial duration flux ropes. The former provides us with a baseline of how well CNNs can predict flux rope orientation while the latter provides insights into real-time forecasting by exploring how accuracy is affected by percentage of flux rope observed. The process of casting the physics problem as a machine learning problem is discussed as well as the impacts of different factors on prediction accuracy such as flux rope fluctuations and different neural network topologies. Finally, results from evaluating the trained network against observed ICMEs from Wind during 1995–2015 are presented.

Progressive trends onthe application ofartificial neural networks inanimal sciences - Areview.

Inrecent years, artificial neural networks have become the subject ofintensive research inanumber ofscientific areas. The high performance and operational speed ofneural models open up awide spectrum ofapplications invarious areas oflife sciences. Objectives pursued bymany scientists, who use neural modelling intheir research, focus - among others - onintensifying real-time calculations. This study shows the possibility ofusing Multilayer-Perceptron (MLP) and Radial Basis Function (RBF) models ofartificial neural networks for the future development ofnew methods for animal science. The process should beexplained explicitly tomake the MLP and RBF models more readily accepted bymore researchers. This study describes and recommends certain models aswell asuses forecasting methods, which are represented bythe chosen neural network topologies, inparticular MLP and RBF models for more successful operations inthe field ofanimals sciences.

Artificial neural network as an alternative for peach fruit mass prediction by non-destructive method

The mass and diameter traits of peach fruits are important parameters to define fruit quality, harvest planning, and management of peach production through the estimation of fresh mass. However, currently, fruit mass estimation methods are based on conventional destructive analyzes which, as they are invasive measurements, can take considerable time, in addition to being costly. Thus, this work aimed to use Artificial Neural Networks in the prediction of the fresh mass of peach fruits as an alternative to the destructive method. Fruits of the peach cultivars, 'Tropic Beauty' (G1), 'BRS Kampai' (G2), and 'BRS Rubimel' (G3) were used in the 2017 and 2018 harvests, and the cultivars (G1) and (G2) were used in the 2019 harvest. Fruits were individually measured for fresh mass (FM) (dependent variable), and suture diameter (SD), equatorial diameter (ED), fruit height (FH) (independent variables). Multiple linear regression analysis (R-MLM) and multilayer perceptron artificial neural network (ANN-MLP) were performed, involving the set of dependent variables SD, ED, and FH, all expressed as a function of the independent variable FM. In all scenarios, we observed that ANN_MLP was more effective than R-MLM in predicting FM. The best ANN-MLP was used to predict the mean peach fruit yield of the genotypes and also for their combinations in the 2018 and 2019 harvests. The presence of high values of R2 in the validation sample indicates that the trained network is efficient and has generalization power to predict FM in all situations evaluated. The use of neural networks to predict the fresh mass (FM) of the peach fruit is an alternative to conventional destructive analyzes and the chance to use the same neural network topology in consecutive harvests will allow the collection of data directly in the field, reducing the analysis time and costs.