Original Artificial Neural Network Research Articles

Detonation Cell Size Prediction Using Artificial Neural Networks (ANNs) for Hydrogen/Hydrocarbon/Ammonia/Nitrous Oxide Mixtures

In this work, a previously developed three-feature Artificial Neural Network (ANN) model with dimensional inputs is directly applied to predict the cell size of hydrocarbon/ammonia/nitrous oxide mixtures and compare these to experimental data. This model uses as inputs three ZND parameters (MCJ, ΔI, and σ˙max), which are mainly calculated using Konnov’s and Mével’s mechanisms. A similar prediction is obtained with the two mechanisms for the biogas–O2, H2–O2, H2–N2O, and NH3–O2 mixtures, indicating that the model is not only limited to Konnov’s chemical kinetic mechanism which was used for its training. The overall good agreement between the ANN predictions and the actual experimental values for the aforementioned mixtures, which are not used in the original training of the ANN model, is promising and shows its potential for application and extension to other mixtures and initial conditions, provided that the chemical kinetic parameters describing the ideal reaction zone structure could be computed. The model is then used to compare experimental cell size data from two detonation tube facilities, and also different chemical kinetic mechanisms for NH3-N2O mixtures. In the end, the original ANN model is expanded with the inclusion of additional cell size data, showing a slightly lower mean error for the predicted cell sizes if the data for the mixtures considered in this study are taken into account for the training of the new ANN model.

A universal ANN-to-SNN framework for achieving high accuracy and low latency deep Spiking Neural Networks

Spiking Neural Networks (SNNs) have become one of the most prominent next-generation computational models owing to their biological plausibility, low power consumption, and the potential for neuromorphic hardware implementation. Among the various methods for obtaining available SNNs, converting Artificial Neural Networks (ANNs) into SNNs is the most cost-effective approach. The early challenges in ANN-to-SNN conversion work revolved around the susceptibility of converted SNNs to conversion errors. Some recent endeavors have attempted to mitigate these conversion errors by altering the original ANNs. Despite their ability to enhance the accuracy of SNNs, these methods lack generality and cannot be directly applied to convert the majority of existing ANNs. In this paper, we present a framework named DNISNM for converting ANN to SNN, with the aim of addressing conversion errors arising from differences in the discreteness and asynchrony of network transmission between ANN and SNN. The DNISNM consists of two mechanisms, Data-based Neuronal Initialization (DNI) and Signed Neuron with Memory (SNM), designed to respectively address errors stemming from discreteness and asynchrony disparities. This framework requires no additional modifications to the original ANN and can result in SNNs with improved accuracy performance, simultaneously ensuring universality, high precision, and low inference latency. We verify it experimentally on challenging object recognition datasets, including CIFAR10, CIFAR100, and ImageNet-1k. Experimental results show that the SNN converted by our framework has very high accuracy even at extremely low latency.

A Grain Size Profile Prediction Method Based on Combined Model of Extreme Gradient Boosting and Artificial Neural Network and Its Application in Sand Control Design

Summary The grain size distribution along the well depth is of great significance for the prediction of the physical properties and the staged sand control design of the unconsolidated or weakly consolidated sandstone reservoir. In this paper, a new method for predicting the formation median grain size profile based on the combination model of extreme gradient boosting (XGBoost) and artificial neural network (ANN) is proposed. The machine learning algorithm and weighted combination model are applied to the prediction and analysis of reservoir grain size. The prediction model is improved from two aspects: First, the feature engineering of the XGBoost-ANN model is constructed by using the data of multiple sampling points on the logging curve. Second, the prediction accuracy is improved by increasing the dimension of the prediction model, that is, the XGBoost and ANN single-prediction models are weighted by the error reciprocal method and a combined prediction model containing multidimensional information is established. The research results show that compared with the single-point mapping model, the prediction accuracy of the multipoint mapping model considering the vertical geological continuity of the reservoir is higher than that of the single-point prediction and the coefficient of determination in the testing set can be improved up to 14.5%. The influence of different weighting methods on prediction performance is studied, and the prediction performance of original XGBoost, ANN, and XGBoost-ANN combined models is compared. The combined prediction model has a higher prediction accuracy than the single XGBoost and ANN models with the same number of sampling points and the coefficient of determination can be improved by up to 16.5%. The prediction accuracy and generalization ability of the XGBoost-ANN combined model are evaluated comprehensively. The combined model is used to design layered sand control of a well in an adjacent block, and good results have been achieved in production practice. This study provides a new method with high accuracy and efficiency for the prediction of unconsolidated sand median grain size profile.

Spiking Deep Residual Networks.

Spiking neural networks (SNNs) have received significant attention for their biological plausibility. SNNs theoretically have at least the same computational power as traditional artificial neural networks (ANNs). They possess the potential of achieving energy-efficient machine intelligence while keeping comparable performance to ANNs. However, it is still a big challenge to train a very deep SNN. In this brief, we propose an efficient approach to build deep SNNs. Residual network (ResNet) is considered a state-of-the-art and fundamental model among convolutional neural networks (CNNs). We employ the idea of converting a trained ResNet to a network of spiking neurons named spiking ResNet (S-ResNet). We propose a residual conversion model that appropriately scales continuous-valued activations in ANNs to match the firing rates in SNNs and a compensation mechanism to reduce the error caused by discretization. Experimental results demonstrate that our proposed method achieves state-of-the-art performance on CIFAR-10, CIFAR-100, and ImageNet 2012 with low latency. This work is the first time to build an asynchronous SNN deeper than 100 layers, with comparable performance to its original ANN.

Artificial neural network prediction and grey relational grade optimisation of friction stir processing

The study predicts friction stir process (FSP) parameters using an artificial neural network (ANN) model. A design of experiment (DoE) approach is used to conduct experiments on FSP and to evaluate the best operating parameters of FSP. ANN uses 30 neurons, 1 hidden layer, and an input and output layer in the Matlab® environment. ANN predicts the output parameters with R values of 0.999, 0.995, and 0.992 for training, validation, and test datasets, respectively, while the overall R-value is 0.997. GRA is used to optimize and rank the parameters of the processes and revealed that the rotational tool speed should be 1180 rpm, traverse feed rate 38 mm/min, and tool tilt angle 1° for best results. The optimized values obtained are 380 MPa, 3.8 µm, 138 HV, and 14% for tensile strength, grain size, microhardness, and elongation, respectively. According to the analysis of variance (ANOVA) and grey relational analysis, the order of influencing parameters on output factors was rotational speed (80%), followed by transverse feed (19%) and tool tilt angle (1%). An ANN is further used to predict using the two most significant parameters—rotational speed and traverse feed. The modified ANN has nearly the same R values as the original ANN. Thus modified ANN may be used for prediction.

Classification for Covid-19 Diseases Based on Ensembled Models

After Covid19 became a worldwide health issue, rapid diagnosis based on clinical symptoms from many diseases with similar symptoms to Covid19 became important to slow down the spread of the pandemic. This study attempts to find ways to classify and diagnose diseases with the help of computer technology quickly and accurately. In this study, the author developed an ensembled machine learning model to categorize four diseases using information on their distinct clinical signs. The authors used a Support Vector Machine (SVM) and Artificial Neural Network (ANN)-based ensemble model. To improve the accuracy of the classification result, this model adds the strong classifier SVM to the result of an intermediate hidden layer of the fundamental ANN deep learning model. The result of the study shows that the integrated model's prediction performance is significantly better than that of the original ANN model after the support of another strong classification algorithm. In conclusion, the effectiveness of the proposed method was proved for classifying the symptoms of patients with allergies, colds, flu, and Covid-19 in this study.

Privacy-Preserving Outsourced Artificial Neural Network Training for Secure Image Classification

Artificial neural network (ANN) is powerful in the artificial intelligence field and has been successfully applied to interpret complex image data in the real world. Since the majority of images are commonly known as private with the information intended to be used by the owner, such as handwritten characters and face, the private constraints form a major obstacle in developing high-precision image classifiers which require access to a large amount of image data belonging to multiple users. State-of-the-art privacy-preserving ANN schemes often use full homomorphic encryption which result in a substantial overhead of computation and data traffic for the data owners, and are restricted to approximation models by low-degree polynomials which lead to a large accuracy loss of the trained model compared to the original ANN model in the plain domain. Consequently, it is still a huge challenge to train an ANN model in the encrypted-domain. To mitigate this problem, we propose a privacy-preserving ANN system for secure constructing image classifiers, named IPPNN, where the server is able to train an ANN-based classifier on the combined image data of all data owners without being able to observe any images using primitives, such as randomization and functional encryption. Our system achieves faster training time and supports lossless training. Moreover, IPPNN removes the need for multiple communications among data owners and servers. We analyze the security of the protocol and perform experiments on a large scale image recognition task. The results show that the IPPNN is feasible to use in practice while achieving high accuracy.

Prediction of ignition delay times of Jet A-1/hydrogen fuel mixture using machine learning

To control the global warming trends, carbon footprint of all the human activities needs to be restricted, including the aviation industry. Mixing hydrogen with commercial kerosene jet fuels appears as a promising alternative fuel to reduce the carbon dioxide emissions of aviation engines. The addition of hydrogen can significantly impact the auto-ignition process of aviation fuels, which is a key ingredient of engine reliability. However, accurate calculations or measurements of ignition delay times (IDTs) over a wide range of pressures, temperatures and fuel blending ratios are complicated and time-consuming. To achieve real-time prediction of ignition delay time for hydrogen-blended jet fuels under various operating conditions, machine learning methods are introduced to build a data-driven proxy model in this work. First, the ignition delay times of Jet A-1/hydrogen fuel mixture are simulated using the well-known HyChem combustion reaction mechanism under different pressures, temperatures, equivalence ratios and blending molar ratios of hydrogen. After some validation against experimental results, an artificial neural network (ANN) model is trained using the database of ignition delay times. Furthermore, a sub-ANN is nested to the original ANN model as an improvement on certain local conditions. The results show that the single original ANN model leads to a large local relative error for IDT <103μs, generally referring to high temperatures and pressures conditions. With the help of the nested sub-ANN approach, the improved model achieves a significantly better accuracy for very fast ignition. Compared with other machine learning models such as random forest, the nested sub-ANN model is more efficient to predict the ignition delay times of Jet A-1/hydrogen fuel mixture, still mitigating computational cost. The sensitivity analysis shows how the nested sub-ANN model is less sensitive to the uncertainties of the input parameters than the original ANN model.

Axial Compression Prediction and GUI Design for CCFST Column Using Machine Learning and Shapley Additive Explanation

Axial bearing capacity is the key index of circular concrete-filled steel tubes (CCFST). A hybrid PSO-ANN model consisting of an artificial neural network (ANN) optimized with particle swarm algorithm (PSO) was proposed to reliably and accurately predict the axial bearing capacity in this paper. The predictive performance of the model was evaluated and compared with the EC4 code and original ANN based on a dataset of 227 experiments, and a graphical user interface (GUI) was developed to achieve the automatic output of the results. The influence of each design parameter on the bearing capacity was analyzed and quantified using the Shapley additive explanation (SHAP) method and sensitivity analysis. The results show that the prediction performance of the PSO-ANN model is superior, and can be recommended as a candidate for the prediction of axial compression bearing capacity of the CCFST column in terms of performance indices. Shapley additive explanation-based parameter analysis indicated that the diameter and thickness of the steel tube are the most two important parameters to the bearing capacity; in particular, the fluctuation of the diameter under the stochastic environment leads to the variation of the axial compression bearing capacity beyond the diameter itself.

Building and training a deep spiking neural network for ECG classification

The electrocardiogram (ECG) reflects the electrical activity of the heart, and is one the most widely used biophysical signals that evaluate heart-related conditions. With years of experiences, medical professionals are able to identify and classify various ECG patterns. However, manually classifying ECG signals is prone to errors and takes considerable amount of time and effort, and thus people start to explore computational models for ECG classification. In recent years, deep artificial neural networks (ANNs) have gained increasing popularity in many fields for their outstanding performances. Traditional ANNs consist of computational units which are inspired from biological neurons but ignore the neural signal transmission details. Spiking neural networks (SNNs), on the other hand, are based on impulse neurons that more closely mimic biological neurons, and thus have a great potential to achieve similar performance with much less power. Nevertheless, SNNs have not become prevalent, and one of the primary reasons is that training SNNs especially the ones with deep structures remains a challenge. In this paper, we aim to propose an efficient way to build and train a deep SNN for ECG classification by constructing a counterpart structure of a deep ANN, transferring the trained parameters, and replacing the activation functions with leaky integrate-and-fire (LIF) neurons. The results show that the accuracy of the deep SNN even exceeds the original ANN. In addition, we compare and discuss the effects of different ANN activation functions on the SNN performance.

MlGeNN: accelerating SNN inference using GPU-enabled neural networks

Abstract In this paper we present mlGeNN—a Python library for the conversion of artificial neural networks (ANNs) specified in Keras to spiking neural networks (SNNs). SNNs are simulated using GeNN with extensions to efficiently support convolutional connectivity and batching. We evaluate converted SNNs on CIFAR-10 and ImageNet classification tasks and compare the performance to both the original ANNs and other SNN simulators. We find that performing inference using a VGG-16 model, trained on the CIFAR-10 dataset, is 2.5× faster than BindsNet and, when using a ResNet-20 model trained on CIFAR-10 with FewSpike ANN to SNN conversion, mlGeNN is only a little over 2× slower than TensorFlow.

Deep learning models for quality estimation of gas tungsten arc welding

Gas tungsten arc (GTA) welding is characterized by the use of non-consumable electrodes and shielding gas for effective welding of various metals. Welding quality can be evaluated by a visual test, but it has the disadvantage that accuracy is limited by the experience of the inspector. The quality of GTA welding can be predicted better through deep learning models that use machine vision, image processing, and welding data such as voltage, current, and feeding speed. In this study, artificial neural network (ANN) models and convolutional neural network (CNN) models are developed for estimating welding quality using v-groove GTA welding experiments. The GTA welding system comprises welding equipment and a vision system. After completing the welding experiments, 21 features and 1 label are collected from the welding data. The observations are then normalized, shuffled, and divided into training data and test data. ANN and CNN models are developed with varied architectures, then verified by calculating root mean squared error, mean absolute error, and accuracy. ed ANN models are developed by applying the relative weight and product measure methods to the original ANN models. The abstract ANN models show similar or better accuracy with fewer features.

Improved artificial neural networks for 3D body posture and lumbosacral moment predictions during manual material handling activities

Body posture measurement approaches, required in biomechanical models to assess risk of musculoskeletal injuries, are usually costly and/or impractical for use in real workplaces. Therefore, we recently developed three artificial neural networks (ANNs), based on measured posture data on several individuals, to predict whole body 3D posture (coordinates of 15 markers located on body’s main joints), segmental orientations (Euler angles of 14 body segments), and lumbosacral (L5-S1) moments during static manual material handling (MMH) activities (ANNPosture, ANNAngle, and ANNMoment, respectively). These ANNs require worker's body height, body weight (only for ANNMoment), hand-load 3D position, and its mass as inputs to accurately predict 3D marker coordinates (RMSE = 7.0 cm), segmental orientations (RMSE = 29.9°) and L5-S1 moments (RMSE = 16.5 N.m) for various static MMH activities. The current work aims to further improve the accuracy of these ANNs by performing outlier elimination and data normalization (as effective tools to improve the accuracy of ANNs) as well as by introducing participant’s knee flexion angle (i.e., lifting technique: stoop, semi-squat, and full-squat) and body weight as new inputs into these ANNs. Results indicate that the RMSE of the new ANNPosture, ANNAngle, and ANNMoment reduced by, respectively, ∼43%, 10%, and 29% (from 7.0 cm, 29.9°, and 16.5 Nm in the original ANNs to, respectively, 4.0 cm, 27.0°, and 11.8 Nm). Such significant improvements in the predictive power of our ANNs further confirm their effectiveness as alternative posture-prediction approaches requiring minimal in vivo data collection in real workplaces.

Comparison of Artificial and Spiking Neural Networks on Digital Hardware.

Despite the success of Deep Neural Networks—a type of Artificial Neural Network (ANN)—in problem domains such as image recognition and speech processing, the energy and processing demands during both training and deployment are growing at an unsustainable rate in the push for greater accuracy. There is a temptation to look for radical new approaches to these applications, and one such approach is the notion that replacing the abstract neuron used in most deep networks with a more biologically-plausible spiking neuron might lead to savings in both energy and resource cost. The most common spiking networks use rate-coded neurons for which a simple translation from a pre-trained ANN to an equivalent spike-based network (SNN) is readily achievable. But does the spike-based network offer an improvement of energy efficiency over the original deep network? In this work, we consider the digital implementations of the core steps in an ANN and the equivalent steps in a rate-coded spiking neural network. We establish a simple method of assessing the relative advantages of rate-based spike encoding over a conventional ANN model. Assuming identical underlying silicon technology we show that most rate-coded spiking network implementations will not be more energy or resource efficient than the original ANN, concluding that more imaginative uses of spikes are required to displace conventional ANNs as the dominant computing framework for neural computation.

Towards a comprehensive optimization of engine efficiency and emissions by coupling artificial neural network (ANN) with genetic algorithm (GA)

In response to the stringent emission regulations, artificial neural network (ANN) coupled with genetic algorithm (GA) is employed to optimize a novel internal combustion engine strategy named direct dual fuel stratification (DDFS). An enhanced ANN model is introduced to improve the accuracy and stability of predictions. Compared to the conventional computational fluid dynamics (CFD)-GA optimization method, the ANN-GA method can identify a better solution to achieve higher fuel efficiency and lower nitrogen oxide (NOx) emissions with the lower computational time. This is attributed to the lower cost of ANN calculation, which allows ANN-GA to introduce larger population to seek optimal solutions. Through combining the required new training data with the previous ones, the original ANN model can be updated to adapt to a wider parameter range. Thus, ANN-GA can readily deal with the optimization problems with variable parameters and objectives. When more re-optimizations are required, ANN-GA can save the computational time over 75% than CFD-GA owing to the data-driven nature of ANN-GA by fully utilizing the available data. Overall, the ANN-GA method shows the superiority in accuracy, efficiency, expansibility, and flexibility for DDFS strategy optimization. It is promising to integrate ANN with optimization algorithm for further improvements of engine performance.

Nonlinear Equalization Based on Artificial Neural Network in DML-Based OFDM Transmission Systems

This article reports the application of an equalizer based on an artificial neural network (ANN), in the form of nonlinear waveform regression, to mitigate nonlinear impairments in directly modulated laser (DML)-based orthogonal frequency-division multiplexing (OFDM) optical transmission. Experiments involving transmission over 0-200 km demonstrate that using an ANN with one hidden layer can greatly reduce nonlinear distortion. The proposed scheme outperformed a Volterra nonlinear equalizer at transmission distances exceeding 25 km. Using a 10G-class DML, the proposed scheme achieved the following data rates: 39.2 Gbps at 100 km (an improvement of 59%) and 33.5 Gbps at 150 km (an improvement of 57%). We also modified the cost function of the ANN during the training procedure to overcome the poor signal-to-noise ratio of the original ANN at low frequencies. This resulted in >30-Gbps transmission over 0-200 km.

Artificial Neural Network for Prediction of Full-Scale Seepage Flow Rate at the Equity Silver Mine

The Equity Silver mine site, British Columbia, Canada, is known for its management and treatment of acid rock drainage. The accurate prediction of full-scale seepage flow rate in field conditions is critical for mine site management, flood control, and contaminant treatment strategies. In this study, a machine learning methodology based on artificial neural network (ANN) is proposed to correlate measured seepage flow rate with the weather monitoring data collected from the field of Equity Silver mine. Comparing with traditional numerical and water balance approaches, the ANN approach generally requires much fewer investigations on hydrogeological mechanisms or definitions of parameters and equations. The measured seepage flow rates from the Bessemer Dump and the weather monitoring data during 1998–2017 are used for training the proposed ANN. The calculated results from the trained ANN show good agreement with the measurements. In addition, the proposed ANN approach is further refined by including time tags of the monitoring data in the input layer, which enable the ANN to investigate the effects of long-term weather trends, climate change, and dynamics inside waste rock piles on seepage flow. The prediction results for the following year (2017–2018) from the refined ANN and the original ANN are compared, which shows the improvement of the refined ANN approach.

An Integrative Remote Sensing Application of Stacked Autoencoder for Atmospheric Correction and Cyanobacteria Estimation Using Hyperspectral Imagery

Hyperspectral image sensing can be used to effectively detect the distribution of harmful cyanobacteria. To accomplish this, physical- and/or model-based simulations have been conducted to perform an atmospheric correction (AC) and an estimation of pigments, including phycocyanin (PC) and chlorophyll-a (Chl-a), in cyanobacteria. However, such simulations were undesirable in certain cases, due to the difficulty of representing dynamically changing aerosol and water vapor in the atmosphere and the optical complexity of inland water. Thus, this study was focused on the development of a deep neural network model for AC and cyanobacteria estimation, without considering the physical formulation. The stacked autoencoder (SAE) network was adopted for the feature extraction and dimensionality reduction of hyperspectral imagery. The artificial neural network (ANN) and support vector regression (SVR) were sequentially applied to achieve AC and estimate cyanobacteria concentrations (i.e., SAE-ANN and SAE-SVR). Further, the ANN and SVR models without SAE were compared with SAE-ANN and SAE-SVR models for the performance evaluations. In terms of AC performance, both SAE-ANN and SAE-SVR displayed reasonable accuracy with the Nash–Sutcliffe efficiency (NSE) &gt; 0.7. For PC and Chl-a estimation, the SAE-ANN model showed the best performance, by yielding NSE values &gt; 0.79 and &gt; 0.77, respectively. SAE, with fine tuning operators, improved the accuracy of the original ANN and SVR estimations, in terms of both AC and cyanobacteria estimation. This is primarily attributed to the high-level feature extraction of SAE, which can represent the spatial features of cyanobacteria. Therefore, this study demonstrated that the deep neural network has a strong potential to realize an integrative remote sensing application.

Performance Prediction and Interpretation of a Refuse Plastic Fuel Fired Boiler

In order to cater to the energy requirement in the form of steam at a reasonable cost, the process industries are relying on the waste incineration plants by engaging themselves through industry symbiosis. However, before the establishment of industrial symbiosis, it is very crucial to monitor and predict the operational performance of the boiler used in the waste incineration plants. The existing works focus on using Artificial Neural Networks (ANNs) for prediction of the performance of the boiler in terms of pressure, temperature, and mass flow rate of steam using the input parameters viz. feed water temperature, feed water pressure, incinerator exit temperature and conveyor speed. However, the problem with this approach is that shallow ANNs cannot model the complex mathematical non-linear relationships so precisely. In addition, ANNs are not interpretable which makes stakeholders apprehensive to use these networks in production. In this paper, we address these drawbacks of ANNs by modeling the complex relationship governing the boiler performance by using a set of machine learning and deep learning models. Also, the research paper introduces multiple techniques like feature importance, Partial Dependence Plots(PDP) plots etc. which interpret the reason behind the model's output to make it more reliable for the stakeholders. It has been empirically shown that the new Machine Learning and Deep Learning models performed better than the ANNs for predicting the boiler performance. The Random Forest model made a Mean Absolute Percentage Error (MAPE) of 1.12 and LSTMs had a MAPE of 1.14 in the prediction of steam temperature C <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">o</sup> which is a significant improvement in comparison to the original ANN model which had a MAPE of 6.93. In the case of the predictions for steam pressure kgf /cm <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> the MAPE for the Random Forest model and LSTM was 5.54 and 4.21 respectively as compared to ANNs MAPE of 1.49. Similarly for steam mass flow rate(t/h), the MAPE was improved to 15.6 and 9.63 by Random Forest Model and LSTM respectively, which was originally 18.77 for ANN based model. These results clearly show that LSTM based models outperformed ANNs and Random Forests in terms of prediction accuracy.

Application of Artificial Neural Network in Tunnel Engineering: A Systematic Review

Due to the lack of living space and the increase in population, there has been a construction boom in the underground space to improve the quality of human life. Tunnel engineering plays a vital role in the development of underground space. In addition to traditional methods, some intelligent methods such as artificial neural networks (ANNs) have been applied to various problems in the tunnel domain in recent years. This paper systematically reviews the application of ANNs from different aspects of tunnel engineering. It reveals that the backpropagation algorithm (BPA) and Levenberg-Marquardt algorithm (LMA) are the most widely used. Due to the limitations of some original models, some scholars use optimization algorithms such as particle swarm optimization (PSO) and genetic algorithm (GA) to optimize the original ANNs to obtain better prediction results. A comparison between the ANN-based methods and methods like statistical methods is conducted. Finally, the following conclusions can be drawn: (1) The recommended ratio of the training set and test set is 3:1; (2) The advantage of optimized ANNs is not apparent when the optimization algorithm varies. Additionally, the performance of ANNs is always better than that of statistical methods.