Evolutionary design of dynamic neural networks for evaporator control

This paper presents a novel neural network (NN) to control an ammonia refrigerant evaporator. Inspired by the latest findings on the biological neuron, a dynamic synaptic unit (DSU) is proposed to enhance the information processing capacity of artificial neurons. Treating the dynamic synaptic activity after the nonlinear somatic activity helps to capture the dynamics demarcated by the Gaussian activation pertaining to the input space. This practice leads to a remarkable reduction in curse of dimensionality. The proposed NN architecture has been compared with two other conventional architectures; one with dynamic neural units (DNUs) and the other with nonlinear static functions as perceptrons. The objective is to control evaporator heat flow rate and secondary fluid outlet temperature while keeping the degree of refrigerant superheat in the range 4–7 K at the evaporator outlet by manipulating refrigerant and evaporator secondary fluid flow rates. The drawbacks of conventional approaches to this problem are discussed, and how the novel method can overcome them are presented. An evolutionary approach is adopted to optimize the parameters of the NN controllers. Then evaporator process model is accomplished as a combination of governing equations and a sub NN resulting in a simple and sufficiently accurate model. The effectiveness of the proposed dynamic NN controller for the evaporator system model is validated using experimental data from the ammonia refrigeration plant.     

An innovative air-conditioning load forecasting model based on RBF neural network and combined residual error correction

Accurate air-conditioning load forecasting is the precondition for the optimal control and energy saving operation of HVAC systems. They have developed many forecasting methods, such as multiple linear regression (MLR), autoregressive integrated moving average (ARIMA), grey model (GM) and artificial neural network (ANN), in the field of air-conditioning load prediction. However, none of them has enough accuracy to satisfy the practical demand. On the basis of these models existed, a novel forecasting method, called ‘RBF neural network (RBFNN) with combined residual error correction’, is developed in this paper. The new model adopts the advanced algorithm of neural network based on radial basis functions for the air-conditioning load forecasting, and uses the combined forecasting model, which is the combination of MLR, ARIMA and GM, to estimate the residual errors and correct the ultimate foresting results. A study case indicates that RBFNN with combined residual error correction has a much better forecasting accuracy than RBFNN itself and RBFNN with single-model correction.     

Loss-efficiency model of single and variable-speed compressors using neural networks

Compressor is the critical component to the performance of a vapor-compression refrigeration system. The loss-efficiency model including the volumetric efficiency and the isentropic efficiency is widely used for representing the compressor performance. A neural network loss-efficiency model is developed to simulate the performance of positive displacement compressors like the reciprocating, screw and scroll compressors. With one more input, frequency, it can be easily extended to the variable speed compressors. The three-layer polynomial perceptron network is developed because the polynomial transfer function is found very effective in training and free of over-learning. The selection of input parameters of neural networks is also found critical to the network prediction accuracy. The proposed neural networks give less than 0.4% standard deviations and ±1.3% maximum deviations against the manufacturer data.     

Liquid density prediction of five different classes of refrigerant systems (HCFCs,HFCs, HFEs,PFAs and PFAAs) using the artificial neural network-group contribution method

In this work, the densities of 48 refrigerant systems from 5 different categories including hydrochlorofluorocarbons (HCFCs), hydrofluorocarbons (HFCs), hydrofluoroethers (HFEs), perfluoroalkanes (PFAs), and perfluoroalkylalkanes (PFAAs) have been studied using a combined method that includes an artificial neural network (ANN) and a simple group contribution method (GCM). A total of 3825 data points of liquid density at several temperatures and pressures have been used to train, validate and test the model. This study shows that the ANN-GCM model represents an excellent alternative to estimate the density of different refrigerant systems with a good accuracy. The average absolute deviations for train, validation, and test sets are 0.18, 0.26, and 0.28, respectively. A comparison between our results and those obtained from some previous methods shows that as well as generality, this model can predict the density of different refrigerants in a better accord with experimental data up to high temperature, high pressure (HTHP) conditions.     

Prediction of the heat transfer rate of a single layer wire-on-tube type heat exchanger using ANFIS

In this paper, we applied an Adaptive Neuro-Fuzzy Inference System (ANFIS) model for prediction of the heat transfer rate of the wire-on-tube type heat exchanger. Limited experimental data was used for training and testing ANFIS configuration with the help of hybrid learning algorithm consisting of backpropagation and least-squares estimation. The predicted values are found to be in good agreement with the actual values from the experiments with mean relative error less than 2.55%. Also, we compared the proposed ANFIS model to an ANN approach. Results show that the ANFIS model has more accuracy in comparison to ANN approach. Therefore, we can use ANFIS model to predict the performances of thermal systems in engineering applications, such as modeling heat exchangers for heat transfer analysis.     

Modified neural network correlation of refrigerant mass flow rates through adiabatic capillary and short tubes: Extension to CO2 transcritical flow

This paper presents a modified dimensionless neural network correlation of refrigerant mass flow rates through adiabatic capillary tubes and short tube orifices. In particular, CO₂ transcritical flow is taken into account. The definition of neural network input and output dimensionless parameters is grounded on the homogeneous equilibrium model and extended to supercritical inlet conditions. 2000 sets of experimental mass flow-rate data of R12, R22, R134a, R404A, R407C, R410A, R600a and CO₂ (R744) in the open literature covering capillary and short tube geometries, subcritical and supercritical inlet conditions are collected for neural network training and testing. The comparison between the trained neural network and experimental data reports 0.65% average and 8.2% standard deviations; 85% data fall into ±10% error band. Particularly for CO₂, the average and standard deviations are −2.5% and 6.0%, respectively. 90% data fall into ±10% error band.     

La flore bacterienne dans un reseau de distribution: Bacteria in distribution systems

Water quality upkeep in distribution systems is a major concern of drinking water suppliers. In the case of extended systems transporting water through several thousand kilometers of pipelines, the problem is not a simple one. This concern was, among others, at the root of the Water Authority for the Paris Suburbs' decision to modify treatment at the Méry-sur-Oise drinking water production plant (Damez, 1982).As can be seen in Fig. 1, the old line was upgraded by the addition of a raw water basin, two more ozonation stages, and activated carbon filtration. Furthermore, pre-chlorination is no longer used at the plant.A study to assess the impact of these changes on the quality of water in the system was launched prior to the switchover, in collaboration with researchers from the Pasteur Institute in Paris. It involved the enumeration and identification of the germs in analysed samples of river water, treated water and distributed water.Bacterial aftergrowth in the distribution system was significant with the old line, the mean value of the total germs at 20°C exceeding the limit value of European standards (100 germs ml⁻¹). With the new line, bacterial aftergrowth has been limited to a great extent (Table 6).Removal of the biodegradable fraction of organics during treatment probably has a fundamental effect on this limitation of the aftergrowth (Brunet et al., 1981; Gaïd, 1981), which is achieved notwithstanding lower post-chlorination dosages. Moreover, chlorine is now in the form of free chlorine (Table 1).As regards organics, TOC removal has been only slightly increased by the new line (Table 7), but the permanganate consumption has been greatly reduced, from 1.5 to 0.5 mg ⁻¹ at the plant outlet.In river water, the germs are mainly Escherichía, Aeromonas hydrophila, Acinetobacter calcoaceticus, Alcaligenes faecalis, Yersinia enterocolitica were isolated in 6% of the samples. At the plant outlet, the samples are mostly uncontaminated in analysis conditions. In the distribution system, the species isolated are Aeromonas hydrophila and Alcaligenes faecalis. Enterobacteriaceae and Pseudomonodaceae also contaminated the samples, but with a very low frequency (5%) (Tables 3 and 4). As regards acid-fast bacteria, many species were isolated in river water (M. fortuitum, M. chelonei, M. terrae and M. lactis). But no species were found in samples of 250 ml at the plant outlet. Still, 39 acid-fast bacteria were isolated from 126 samples of distributed water (Table 5). Most germs belong to Runyon's group 11 but were not identified (Runyon, 1959).The study is still under way. The onset of deterioration of bacteriological quality may be a criterion for selecting the regeneration cycle for activated carbon.     

Electronic expansion valve mass flow rate prediction based on dimensionless correlation and ANN model

Refrigerant mass flow rate through electronic expansion valve (EEV) makes significant sense for refrigeration system intelligent control and energy conservation. Objectives of this study were to present experimental data of R134a mass flow rate through EEV and to develop models for EEV mass flow rate prediction via two approaches: dimensionless correlation based on Buckingham π-theorem and artificial neural network (ANN) model based on dimensionless parameters. The database utilized for model training and test was comprised of our experimental data and data available in open literatures including R22, R407C, R410A and R134a. Compared with three existing dimensionless correlations, the proposed dimensionless correlation and ANN model demonstrated higher accuracy. The proposed dimensionless correlation gave mean relative error (MRE) of 6.60%, relative mean square error of (RMSE) 12.05 kg h⁻¹ and correlation coefficient (R²) of 0.9810. The ANN model with the configuration of 8-6-1 showed MRE, RMSE and R² of 3.97%, 7.59 kg h⁻¹ and 0.9924, respectively.     

Estimation of vapour liquid equilibria for the system carbon dioxide–difluoromethane using artificial neural networks

In this paper, an alternate tool, i.e. the artificial neural network technique has been applied for estimation of vapour liquid equilibria (VLE) for the binary system, carbon dioxide–difluoromethane, which is an attractive alternative to chlorofluorocarbons and hydrochlorofluorocarbons, normally used as refrigerants. The model can satisfactorily estimate the vapour liquid equilibrium pressure and mole fraction carbon dioxide in vapour phase in the temperature range 222.04–343.23 K and in the pressure range 0.105–7.46 MPa. The average absolute error for the system in the estimation of vapour phase mole fraction is 0.0086 and 0.056 MPa for the pressure.     

A model for reasoning about persistence and causation

Reasoning about change requires predicting how long a proposition, having become true, will continue to be so. Lacking perfect knowledge, an agent may be constrained to believe that a proposition persists indefinitely simply because there is no way for the agent to infer a contravening proposition with certainty. In this paper, we describe a model of causal reasoning that accounts for knowledge concerning cause-and-effect relationships and knowledge concerning the tendency for propositions to persist or not as a function of time passing. Our model has a natural encoding in the form of a network representation for probabilistic models. We consider the computational properties of our model by reviewing recent advances in computing the consequences of models encoded in this network representation. Finally, we discuss how our probabilistic model addresses certain classical problems in temporal reasoning (e. g., the frame and qualification problems).