Explore chapters and articles related to this topic
Biomanufacturing in Microalgae Industry
Published in Pau Loke Show, Wai Siong Chai, Tau Chuan Ling, Microalgae for Environmental Biotechnology, 2023
Leong Wei, Wen Yi Chia, Kit Wayne Chew, Pau Loke Show
Hence, while it is apparent and obvious that statistical methods like response surface methodology would be of use to study and optimize the processes of microalgae cultivation, the introduction of ANN of AI into the game would greatly improve the optimization process while modelling nonlinear systems at the same time without further physical information provided. Moreover, ANN would also be used to conduct studies on the composition of culture medium and to observe how it would affect the growth of microalgae in the culture units (Wu and Shi 2007). For a particular example, for the heterotrophic cultivation of a microalgae species of Chlorella sp., a hybrid neural network model is proposed and programmed for study of the culture. In the proposed model, concentration of glucose in the culture unit of the species would be the input, while the analyzed specific growth rate of Chlorella sp. would be the output parameter of the study (Teng et al. 2020).
Intelligent Monitoring for Network Fault Management
Published in Witold Pedrycz, Athanasios Vasilakos, Computational Intelligence in Telecommunications Networks, 2018
Each MIB has a hybrid neural network—HMM to determine the MIB’s state. For the neural network portion, a feedforward network is trained using the standard backpropagation algorithm to provide the a posteriori state probabilities p(si|θ(t)), where si is the state, θ(t) is the time-series data of the MIB variables at time t and m is the number of states. Given these probabilities, the HMM then provides temporal context by estimating the state probabilities as p(si|θ(t))=αi(t)∑j=1mαi(t),
Short-Term Electricity Price Forecasting
Published in João P. S. Catalão, Electric Power Systems, 2017
As described in the previous section, electricity price is a nonlinear mapping function of many input variables. It is very hard for a single neural network to correctly learn the impact of all of these inputs on electricity price. Thus, enhancing the learning capability of a neural network-based forecasting engine can significantly improve its price forecast performance. A well-designed combination of different neural networks can potentially enhance their learning capability in modeling a complex process. For example, some parallel and cascaded structures for combining neural networks with improved price prediction performance have been proposed in various studies [8,10,44]. Moreover, CNNs, introduced in the previous section, are a cascaded structure of neural networks. However, these structures share the input data among their building blocks. Thus, the extracted knowledge of a block is not really shared with other blocks. In the hybrid neural network (HNN), on the other hand, a knowledge transfer procedure is envisioned from one neural network to another. For instance, the architecture of HNN, including three neural networks, is shown in Figure 4.13 [25]. These neural networks, denoted by NN1, NN2, and NN3 in the figure, have the same MLP structure.
Hybrid neural network controller for uncertain nonlinear discrete-time systems with non-symmetric dead zone and unknown disturbances
Published in International Journal of Control, 2023
Rahul Kumar, Uday Pratap Singh, Arun Bali, Kuldip Raj
Hybrid neural network (HNN). Let us consider a compact set and a continuous function on . For some arbitrary there exists a HNN so that where is the weight vector which is adjustable and initially generated by using PSO, l>1 is the number of nodes in the network, is the input vector and is the Gaussian function used as activation function of the HNN and expressed as where is the centre and is the width of the Gaussian function. The structure of this HNN can be visualised from Figure 4. In view of the approximation/estimation property of HNN (Fei & Wang, 2019), the function can be expressed as where is the vector of optimal weights and be the tolerance value or minimum possible approximation error which is bounded i.e. .
A hybrid neural network model based modelling methodology for the rubber bushing
Published in Vehicle System Dynamics, 2021
Liangcheng Dai, Maoru Chi, Chuanbo Xu, Hongxing Gao, Jianfeng Sun, Xingwen Wu, Shulin Liang
It can be seen that the total output of the hybrid neural network is a composite function of each neuron, which is determined by the weight coefficient and threshold of each neuron. To guarantee the accuracy of the model, the model error is controlled by a mean squared error (MSE) between the calculated output and the target output , as shown in Equation (31). Based on the comparison between the calculated result and the target value, the weight coefficient and threshold of each neuron are updated. This process is executed in the loop until the MSE matches the target value. where is the target output, is the calculated output of neural network and N is the number of trained input-target pairs.
Initial cost forecasting model of mid-rise green office buildings
Published in Journal of Asian Architecture and Building Engineering, 2020
An Evolutionary Fuzzy Hybrid Neural Network (EFHNN) model was introduced in a study by Cheng et al. based on the Artificial Neural Network (ANN) technique (Cheng, Tsai, and Sudjono 2010). In their research, a hybrid neural network (HNN) was integrated with high-order NNs that operated with alternating non-linear and linear and neuron layer connectors. FL was then employed in the HNN to address the uncertainties, and GA was applied on the HNN and FL for the FHNN optimization. The developed model was found to generate accurate and effective early-stage cost estimation in construction projects. Several other models based on NN have been developed in recent years, such as those reported by Aibinu et al. (2015) and Juszczyk (2013).