From Neural Networks to Qualitative Models in Environmental Engineering

As an alternative to physical models, artificial neural networks (ANNs) are a valuable forecast tool in environmental sciences. They can be used effectively due to their learning capabilities and their low computational costs. Once all relevant variables of the system are identified and put into the network, it works quickly and accurately. However, one of the major shortcomings of neural networks is that they do not reveal causal relationships between major system components and thus are unable to improve the explicit knowledge of the user. Another problem is due to the fact that reasoning is only done from the inputs to the outputs. In cases where the opposite is requested (i.e., deriving inputs leading to a given output), neural networks can hardly be used. To overcome these problems, we introduce a novel approach for deriving qualitative information out of neural networks. Some of the resulting rules can directly be used by a qualitative simulator for producing possible future scenarios. Because of the explicit representation of knowledge, the rules should be easier to understand and can be used as a starting point for creating models wherever a physical model is not available. Moreover, the resulting rules are well adapted to be used in decision support systems. We illustrate our approach by introducing a network for predicting surface ozone concentrations and show how rules can be derived from the network and how the approach can be naturally extended for use in decision support systems.     

浅层地下水位预测的小波网络模型

针对浅层地下水位时间序列动态变化的非线性和复杂性 ,提出了基于小波分析与人工神经网络相结合的预测方法———小波网络模型。小波网络模型吸取了小波分析的多分辨功能和人工神经网络的非线性逼近能力。实例计算结果表明 ,建议模型不同预见期的拟合和检验精度很高。小波网络模型延长了预见期 ,提高了预报精度。     

A Computerized Feature Selection Method Using Genetic Algorithms to Forecast Freeway Accident Duration Times

Abstract: This study presents a feature selection method that uses genetic algorithms to create two artificial neural network‐based models that provide a sequential forecast of accident duration from the time of accident notification to the accident site clearance. These two models can provide the estimated duration time by plugging in relevant traffic data as soon as an accident is notified. To select data feature, the genetic algorithm is designed to decrease the number of model inputs while preserving the relevant traffic characteristics. Using the proposed feature selection method, the mean absolute percentage error for forecasting accident duration at each time point is mostly under 29%, which indicates that these models have a reasonable forecasting ability. Thanks to this model, travelers and traffic management units can better understand the impact of accidents. This study shows that the proposed models are feasible in the Intelligent Transportation Systems context.     

Adaptive neuro-fuzzy based inferential sensor model for estimating the average air temperature in space heating systems

The heating systems are conventionally controlled by open-loop control systems because of the absence of practical methods for estimating average air temperature in the built environment. An inferential sensor model, based on adaptive neuro-fuzzy inference system modeling, for estimating the average air temperature in multi-zone space heating systems is developed. This modeling technique has the advantage of expert knowledge of fuzzy inference systems (FISs) and learning capability of artificial neural networks (ANNs). A hybrid learning algorithm, which combines the least-square method and the back-propagation algorithm, is used to identify the parameters of the network. This paper describes an adaptive network based inferential sensor that can be used to design closed-loop control for space heating systems. The research aims to improve the overall performance of heating systems, in terms of energy efficiency and thermal comfort. The average air temperature results estimated by using the developed model are strongly in agreement with the experimental results.     

结合神经网络的热泵冷凝器模型优化

本文采用了按照制冷剂状态来划分分布参数模型微元的方法,建立了冷凝器的稳态分布参数模型,并将其计算结果与实验数据做了验证。在此基础上建立了和稳态模型相结合的BP神经网络和RBF神经网络模型,通过神经网络对稳态模型进行修正,有效地提高了冷凝器仿真模型在热泵系统中的仿真精度,并比较了用两种神经网络模型进行校正的差异。     

边坡非线性位移的神经网络-时间序列分析

边坡的变形表现出复杂的非线性演化特征，大量的工程实践表明利用部分实测的边坡位移时间序列来预测未来边坡的位移更为准确。以神经网络和时间序列分析方法为基础，使用零均值化和标准偏差预处理方法，以及规则化能量函数法和贝叶斯规则化方法进行BP神经网络建模，利用BP网络对边坡位移非平稳时序进行趋势项提取，使非平稳监测时序转化为平稳时序以进行常规ARMA时序分析。结合滚动预测方法，建立了适合岩土体位移预测的神经网络-时间序列分析联合模型，以隔河岩水电站进水口边坡变形和水布垭水电站大岩淌滑坡位移为例进行预测分析。研究结果表明：新模型的预测精度高、实时可靠，可应用于实际工程。     

边坡位移预测的RBF神经网络方法

利用边坡实测位移序列来预测边坡未来时间的位移,可以有效地判断边坡的稳定性。由于神经网络可以通过对样本的反复学习来反映边坡复杂的非线性演化关系,其预测效果要优于传统的预测方法。RBF神经网络作为一种性能良好的前馈网络,具有更好的逼近能力和全局最优特性。以边坡位移时间序列为基础,采用RBF神经网络建立边坡位移预测模型,通过最近邻聚类学习算法实现边坡位移预测,具有结构简单、学习速度快、预测精度高的特点,网络的外推能力也较强。通过2个工程实例说明边坡位移预测的RBF神经网络方法的有效性。     

On the energy consumption in residential buildings

A neural network approach is used in the present study for modelling and estimating the energy consumption time series for a residential building in Athens, using as inputs several climatic parameters.The hourly values of the energy consumption, for heating and cooling the building, are estimated for several years using feed forward backpropagation neural networks. Various neural network architectures are designed and trained for the output estimation, which is the building’s energy consumption. The results are tested with extensive sets of non-training measurements and it is found that they correspond well with the actual values.Furthermore, “multi-lag” output predictions of ambient air temperature and total solar radiation are used as inputs to the neural network models for modelling and predicting the future values of energy consumption with sufficient accuracy.     

Wavelet and ANN combination model for prediction of daily suspended sediment load in rivers

In this research, a new wavelet artificial neural network (WANN) model was proposed for daily suspended sediment load (SSL) prediction in rivers. In the developed model, wavelet analysis was linked to an artificial neural network (ANN). For this purpose, daily observed time series of river discharge (Q) and SSL in Yadkin River at Yadkin College, NC station in the USA were decomposed to some sub-time series at different levels by wavelet analysis. Then, these sub-time series were imposed to the ANN technique for SSL time series modeling. To evaluate the model accuracy, the proposed model was compared with ANN, multi linear regression (MLR), and conventional sediment rating curve (SRC) models. The comparison of prediction accuracy of the models illustrated that the WANN was the most accurate model in SSL prediction. Results presented that the WANN model could satisfactorily simulate hysteresis phenomenon, acceptably estimate cumulative SSL, and reasonably predict high SSL values.     

Predicting chaotic time series of PM10 concentration using artificial neural network

Respirable particulate matter (PM10) concentration at one residential site in Delhi, India was predicted using the neural network approach. The concepts of chaotic systems theory were utilized to build the neural network model. The embedding dimension was estimated to provide the inputs to the neural network. The model evaluation results indicated the importance of noise reduction before selecting the embedding dimension of the time series. The selection of a proper embedding dimension is considered to be essential for obtaining reliable predictions. The model’s performance shows the capability of neural networks in modelling the chaotic time series.     

Neural network-based model for prediction of permanent deformation of unbound granular materials

Several available mechanistic-empirical pavement design methods fail to include predictive model for permanent deformation (PD) of unbound granular materials (UGMs), which make these methods more conservative. In addition, there are limited regression models capable of predicting the PD under multi-stress levels, and these models have regression limitations and generally fail to cover the complexity of UGM behaviour. Recent researches are focused on using new methods of computational intelligence systems to address the problems, such as artificial neural network (ANN). In this context, we aim to develop an artificial neural model to predict the PD of UGMs exposed to repeated loads. Extensive repeated load triaxial tests (RLTTs) were conducted on base and subbase materials locally available in Victoria, Australia to investigate the PD properties of the tested materials and to prepare the database of the neural networks. Specimens were prepared over different moisture contents and gradations to cover a wide testing matrix. The ANN model consists of one input layer with five neurons, one hidden layer with twelve neurons, and one output layer with one neuron. The five inputs were the number of load cycles, deviatoric stress, moisture content, coefficient of uniformity, and coefficient of curvature. The sensitivity analysis showed that the most important indicator that impacts PD is the number of load cycles with influence factor of 41%. It shows that the ANN method is rapid and efficient to predict the PD, which could be implemented in the Austroads pavement design method.     

An Implementation of the Mode-Split Traffic-Assignment Method

A software called Optimal Traffic Signal Control System (OTSCS) was developed by us for testing the feasibility of dynamically controlling a traffic signal by finding optimal signal timing to minimize delay at signalized intersections. It also was designed as a research tool to study the learning behavior of artificial neural networks and the properties of heuristic search methods. It consists of a level-of-service evaluation model that is based on an artificial neural network and a heuristic optimization model that interacts with the level-of-service evaluation model. This article discusses the latter model, named the Optimal Traffic Signal Timing Model (OTSTM). The OTSTM was applied to determine optimal signal timing of two-phase traffic signals to evaluate the model's performance. Two search methods were employed: a depth-first search method (an enumeration method) and a direction-search method that the authors developed. It was found that the OTSTM with the direction search resulted in "optimal" signal timings similar to the depth-first search, which would always produce a global optimal timing. Yet the cost of the direction search, as measured by the CPU time of the computer used for analysis, was found to be much less than the cost of obtaining an optimal solution by the depth-first search cases—more than 10 times less. The study showed that once the artificial neural network is properly trained, heuristic optimal signal timing combined with artificial networks can be used as a decision-support tool for dynamic signal control. This article demonstrates how OTSTM can quickly find an optimal signal-timing solution for two-phase traffic signals.     

安全监测中常用的变形预测方法及适用性

总结了目前常用的回归分析法、时序分析法、灰色系统理论、人工神经网络等预测方法的基本理论和特点,并对各种方法的适用性进行了概括和讨论,以期在工程建设中更好的利用监测数据对未来的变形进行预测。     

Selection of Methodology for Neural Network Modeling of Constitutive Hystereses Behavior of Soils

Classic constitutive modeling of geomaterials based on the elasticity and plasticity theories suffers from limitations pertaining to formulation complexity, idealization of behavior, and excessive empirical parameters. This article capitalizes on the modeling capabilities of neural networks as substitutes for the classic approaches. The neural network–based modeling overcomes the difficulties encountered in understanding the underlying microscopic processes governing the material's behavior by redirecting the efforts into learning the cause-effect relations from behavioral examples. Several methodologies are presented and cross-compared for effectiveness in approximating a theoretical hysteresis model resembling stress-strain behavior. The most effective methodology was used in modeling the constitutive behavior of an experimentally tested soil and produced models that simulated the real behavior of the soil with high accuracy. Although these models are empirical, they are retrainable and thus, unlike classic constitutive modeling techniques, can be revised and generalized easily when new data become available.     

人工神经元网络在动测桩基承载力中的应用

本文将人工智能领域的分支人工神经元网络方法应用在动测桩领域中,使得大量专家的经验和动测桩基承载力技术有机地结合在一起,在确定桩基承载力时发挥特有的作用。     

Data Fusion of Fixed Detector and Probe Vehicle Data for Incident Detection

An important feature of many advanced traveler information systems (ATIS) is real-time information about incidents on the street network. This paper describes a system for automatically detecting incidents for such an ATIS developed using artificial neural networks and statistical prediction methods. The system monitors traffic conditions using two types of data: inductive loop detectors (ILDs) and vehicle probes. For both neural network and statistical methods, incident detection is accomplished using two approaches: by processing traffic input data directly and by processing the output of specialized algorithms that detect incidents using information from each data source. Analysis data generated from a simulation of a typical suburban signalized major arterial street are used. Different model configurations are examined and tested to identify the input variables and methods that are the best predictors of incident occurrence. The neural network approaches consistently perform at least as well as the discriminant analysis models, especially when results are adjusted to avoid false alarms.     

Peak shear strength prediction for discontinuities between two different rock types using a neural network approach

The peak shear strength of discontinuities between two different rock types is essential to evaluate the stability of a rock slope with interlayered rocks. However, current research has paid little attention to shear strength parameters of discontinuities with different joint wall compressive strength (DDJCS). In this paper, a neural network methodology was used to predict the peak shear strength of DDJCS considering the effect of joint wall strength combination, normal stress and joint roughness. The database was developed by laboratory direct shear tests on artificial joint specimens with seven different joint wall strength combinations, four designed joint surface topographies and six types of normal stresses. A part of the experimental data was used to train a back-propagation neural network model with a single-hidden layer. The remaining experimental data was used to validate the trained neural network model. The best geometry of the neural network model was determined by the trial-and-error method. For the same data, multivariate regression analysis was also conducted to predict the peak shear strength of DDJCS. Prediction precision of the neural network model and multivariate regression model was evaluated by comparing the predicted peak shear strength of DDJCS with experimental data. The results showed that the capability of the developed neural network model was strong and better than the multivariate regression model. Finally, the established neural network model was applied in the stability evaluation of a typical rock slope with DDJCS as the critical surface in the Badong formation of China.

     

Development of Recurrent Neural Network Considering Temporal‐Spatial Input Dynamics for Freeway Travel Time Modeling

Abstract: The artificial neural network (ANN) is one advance approach to freeway travel time prediction. Various studies using different inputs have come to no consensus on the effects of input selections. In addition, very little discussion has been made on the temporal–spatial aspect of the ANN travel time prediction process. In this study, we employ an ANN ensemble technique to analyze the effects of various input settings on the ANN prediction performances. Volume, occupancy, and speed are used as inputs to predict travel times. The predictions are then compared against the travel times collected from the toll collection system in Houston. The results show speed or occupancy measured at the segment of interest may be used as sole input to produce acceptable predictions, but all three variables together tend to yield the best prediction results. The inclusion of inputs from both upstream and downstream segments is statistically better than using only the inputs from current segment. It also appears that the magnitude of prevailing segment travel time can be used as a guideline to set up temporal input delays for better prediction accuracies. The evaluation of spatiotemporal input interactions reveals that past information on downstream and current segments is useful in improving prediction accuracy whereas past inputs from the upstream location do not provide as much constructive information. Finally, a variant of the state‐space model (SSNN), namely time‐delayed state‐space neural network (TDSSNN), is proposed and compared against other popular ANN models. The comparison shows that the TDSSNN outperforms other networks and remains very comparable with the SSNN. Future research is needed to analyze TDSSNN's ability in corridor prediction settings.     

Prediction of fire resistance of concrete filled tubular steel columns using neural networks

A functional relationship between the fire resistance of a concrete filled steel column and the parameters which cause the fire resistance is represented using an artificial neural network. Experimental data obtained from previous laboratory fire tests are used for training the neural network model. The model predicted values are compared with actual test results. The results indicate that the model can predict the fire resistance with adequate accuracy required for practical design purpose. The developed neutral network can be used to predict the fire resistance of similar columns under fire by observing various factors influencing the resistance such as: (a) structural factors, (b) material factors, and (c) loading conditions. The structural engineer is required to provide the magnitude of these influencing factors as inputs to the neural network and the network will predict the fire resistance, based on the combined effect of these factors. This system can be used by structural engineers to predict the resistance of fire in similar concrete filled steel columns without conducting costly fire tests, by using the known parameters such as column dimensions, column height, and loading conditions.     

ANN-based empirical modelling of pile behaviour under static compressive loading

Artificial neural networks have been widely used over the past two decades to successfully develop empirical models for a variety of geotechnical problems. In this paper, an empirical model based on the product-unit neural network (PUNN) is developed to predict the load-deformation behaviour of piles based SPT values of the supporting soil. Other parameters used as inputs include particle grading, pile geometry, method of installation as well as the elastic modulus of the pile material. The model is trained using full-scale pile loading tests data retrieved from FHWA deep foundations database. From the results obtained, it is observed that the proposed model gives a better simulation of pile load-deformation curves compared to the Fleming’s hyperbolic model and t-z approach.