PM2.5 concentration prediction using hidden semi-Markov model-based times series data mining

In this paper, a novel framework and methodology based on hidden semi-Markov models (HSMMs) for high PM_2.5 concentration value prediction is presented. Due to lack of explicit time structure and its short-term memory of past history, a standard hidden Markov model (HMM) has limited power in modeling the temporal structures of the prediction problems. To overcome the limitations of HMMs in prediction, we develop the HSMMs by adding the temporal structures into the HMMs and use them to predict the concentration levels of PM_2.5. As a model-driven statistical learning method, HSMM assumes that both data and a mathematical model are available. In contrast to other data-driven statistical prediction models such as neural networks, a mathematical functional mapping between the parameters and the selected input variables can be established in HSMMs. In the proposed framework, states of HSMMs are used to represent the PM_2.5 concentration levels. The model parameters are estimated through modified forward–backward training algorithm. The re-estimation formulae for model parameters are derived. The trained HSMMs can be used to predict high PM_2.5 concentration levels. The validation of the proposed framework and methodology is carried out in real world applications: prediction of high PM_2.5 concentrations at O’Hare airport in Chicago. The results show that the HSMMs provide accurate predictions of high PM_2.5 concentration levels for the next 24 h.     

Performance of radial basis and LM-feed forward artificial neural networks for predicting daily watershed runoff

This study investigated the effects of upstream stations’ flow records on the performance of artificial neural network (ANN) models for predicting daily watershed runoff. As a comparison, a multiple linear regression (MLR) analysis was also examined using various statistical indices. Five streamflow measuring stations on the Cahaba River, Alabama, were selected as case studies. Two different ANN models, multi layer feed forward neural network using Levenberg–Marquardt learning algorithm (LMFF) and radial basis function (RBF), were introduced in this paper. These models were then used to forecast one day ahead streamflows. The correlation analysis was applied for determining the architecture of each ANN model in terms of input variables. Several statistical criteria (RMSE, MAE and coefficient of correlation) were used to check the model accuracy in comparison with the observed data by means of K-fold cross validation method. Additionally, residual analysis was applied for the model results. The comparison results revealed that using upstream records could significantly increase the accuracy of ANN and MLR models in predicting daily stream flows (by around 30%). The comparison of the prediction accuracy of both ANN models (LMFF and RBF) and linear regression method indicated that the ANN approaches were more accurate than the MLR in predicting streamflow dynamics. The LMFF model was able to improve the average of root mean square error (RMSE_ave) and average of mean absolute percentage error (MAPE_ave) values of the multiple linear regression forecasts by about 18% and 21%, respectively. In spite of the fact that the RBF model acted better for predicting the highest range of flow rate (flood events, RMSE_ave/RBF = 26.8 m³/s vs. RMSE_ave/LMFF = 40.2 m³/s), in general, the results suggested that the LMFF method was somehow superior to the RBF method in predicting watershed runoff (RMSE/LMFF = 18.8 m³/s vs. RMSE/RBF = 19.2 m³/s). Eventually, statistical differences between measured and predicted medians were evaluated using Mann-Whitney test, and differences in variances were evaluated using the Levene's test.     

Artificial neural network-based model for predicting VO2max from a submaximal exercise test

The goal of this study is to develop an accurate artificial neural network (ANN)-based model to predict maximal oxygen uptake (VO₂max) of fit adults from a single stage submaximal treadmill jogging test. Participants (81 males and 45 females), aged from 17 to 40 years, successfully completed a maximal graded exercise test (GXT) to determine VO₂max. The variables; gender, age, body mass, steady-state heart rate and jogging speed are used to build the ANN prediction model. Using 10-fold cross validation on the dataset, the average values of standard error of estimate (SEE), Pearson’s correlation coefficient (r) and multiple correlation coefficient (R) of the model are calculated as 1.80 ml kg^?1 min^?1, 0.95 and 0.93, respectively. Compared with the results of the other prediction models in literature that were developed using Multiple Linear Regression Analysis, the reported values of SEE, r and R in this study are considerably more accurate.     

A 24-h forecast of ozone peaks and exceedance levels using neural classifiers and weather predictions

A neural network combined to a neural classifier is used in a real time forecasting of hourly maximum ozone in the centre of France, in an urban atmosphere. This neural model is based on the MultiLayer Perceptron (MLP) structure. The inputs of the statistical network are model output statistics of the weather predictions from the French National Weather Service. These predicted meteorological parameters are very easily available through an air quality network. The lead time used in this forecasting is (t + 24) h. Efforts are related to a regularisation method which is based on a Bayesian Information Criterion-like and to the determination of a confidence interval of forecasting. We offer a statistical validation between various statistical models and a deterministic chemistry-transport model. In this experiment, with the final neural network, the ozone peaks are fairly well predicted (in terms of global fit), with an Agreement Index = 92%, the Mean Absolute Error = the Root Mean Square Error = 15 μg m⁻³ and the Mean Bias Error = 5 μg m⁻³, where the European threshold of the hourly ozone is 180 μg m⁻³.To improve the performance of this exceedance forecasting, instead of the previous model, we use a neural classifier with a sigmoid function in the output layer. The output of the network ranges from [0,1] and can be interpreted as the probability of exceedance of the threshold. This model is compared to a classical logistic regression. With this neural classifier, the Success Index of forecasting is 78% whereas it is from 65% to 72% with the classical MLPs. During the validation phase, in the Summer of 2003, six ozone peaks above the threshold were detected. They actually were seven.Finally, the model called NEUROZONE is now used in real time. New data will be introduced in the training data each year, at the end of September. The network will be re-trained and new regression parameters estimated. So, one of the main difficulties in the training phase – namely the low frequency of ozone peaks above the threshold in this region – will be solved.     

Prediction of the strength of mineral admixture concrete using multivariable regression analysis and an artificial neural network

This study applies multiple regression analysis and an artificial neural network in estimating the compressive strength of concrete that contains various amounts of blast furnace slag and fly ash, based on the properties of the additives (blast furnace slag and fly ash in this case) and values obtained by non-destructive testing rebound number and ultrasonic pulse velocity for 28 different concrete mixtures (M_control and M₁–M₂₇) at different curing times (3, 7, 28, 90, and 180 days). The results obtained using the two methods are then compared and discussed. The results reveal that although multiple regression analysis was more accurate than artificial neural network in predicting the compressive strength using values obtained from non-destructive testing, the artificial neural network models performed better than did multiple regression analysis models. The application of an artificial neural network to the prediction of the compressive strength in admixture concrete of various curing times shows great potential in terms of inverse problems, and it is suitable for calculating nonlinear functional relationships, for which classical methods cannot be applied.     

ANN based simulation and experimental verification of analytical four- and five-parameters models of PV modules

In this article, artificial neural network (ANN) is adopted to predict photovoltaic (PV) panel behaviors under realistic weather conditions. ANN results are compared with analytical four and five parameter models of PV module. The inputs of the models are the daily total irradiation, air temperature and module voltage, while the outputs are the current and power generated by the panel. Analytical models of PV modules, based on the manufacturer datasheet values, are simulated through Matlab/Simulink environment. Multilayer perceptron is used to predict the operating current and power of the PV module. The best network configuration to predict panel current had a 3–7–4–1 topology. So, this two hidden layer topology was selected as the best model for predicting panel current with similar conditions. Results obtained from the PV module simulation and the optimal ANN model has been validated experimentally. Results showed that ANN model provide a better prediction of the current and power of the PV module than the analytical models. The coefficient of determination (R²), mean square error (MSE) and the mean absolute percentage error (MAPE) values for the optimal ANN model were 0.971, 0.002 and 0.107, respectively. A comparative study among ANN and analytical models was also carried out. Among the analytical models, the five-parameter model, with MAPE = 0.112, MSE = 0.0026 and R² = 0.919, gave better prediction than the four-parameter model (with MAPE = 0.152, MSE = 0.0052 and R² = 0.905). Overall, the 3–7–4–1 ANN model outperformed four-parameter model, and was marginally better than the five-parameter model.     

Prediction of artificial soil’s unconfined compression strength test using statistical analyses and artificial neural networks

Laboratory prediction of the unconfined compression strength (UCS) of cohesive soils is important to determine the shear strength properties. However, this study presents the application of different methods simple–multiple analysis and artificial neural networks for the prediction of the UCS from basic soil properties. Regression analysis and artificial neural networks prediction indicated that there exist acceptable correlations between soil properties and unconfined compression strength. Besides, artificial neural networks showed a higher performance than traditional statistical models for predicting UCS. Regression analysis and artificial neural network prediction indicated strong correlations (R² = 0.71–0.97) between basic soil properties and UCS. It has been shown that the correlation equations obtained by regression analyses are found to be reliable in practical situations.     

Short-term Maharashtra state electrical power load prediction with special emphasis on seasonal changes using a novel focused time lagged recurrent neural network based on time delay neural network model

In this paper, the parameter-wise optimization training process is implemented to achieve an optimal configuration of focused time lagged recurrent neural network (FTLRNN) models by embedding the gamma, laguarre, and multi-channel tapped delay line memory structure. The aim is to examine the prediction ability of the proposed models in order to predict one-day-ahead electric power load simultaneously as usual to oppose 1–24 h forecast in sequel with a special emphasis on seasonal changes over a year. An improved delta-bar-delta algorithm is used to accelerate the training of neural networks and to improve the stability of the convergence.Experimental results indicate that the FTLRNN with time delay neural network (TDNN) clearly outperformed the gamma and laguarre based short-term memory structure in various performance metrics such as mean square error (MSE), normalized MSE, correlation coefficient (r) and mean absolute percentage error (MAPE) during evaluation process. Empirical results show that the proposed dynamic NN model consistently performs well on daily, weekly, and monthly average basis in terms of prediction accuracy. It is noticed from the literature review that an optimally configured FTLRNN with multi-channel tapped delay line memory structure is not currently available to solve short-term electrical power load prediction. The proposed method gives acceptable errors in all seasons, months and on daily basis. The average prediction error on three weeks is obtained as low as 1.67%.     

Prediction of cutting temperature in orthogonal machining of AISI 316L using artificial neural network

In this study, an approach based on artificial neural network (ANN) was proposed to predict the experimental cutting temperatures generated in orthogonal turning of AISI 316L stainless steel. Experimental and numerical analyses of the cutting forces were carried out to numerically obtain the cutting temperature. For this purpose, cutting tests were conducted using coated (TiCN + Al₂O₃ + TiN and Al₂O₃) and uncoated cemented carbide inserts. The Deform-2D programme was used for numerical modelling and the Johnson–Cook (J–C) material model was used. The numerical cutting forces for the coated and uncoated tools were compared with the experimental results. On the other hand, the cutting temperature value for each cutting tool was numerically obtained. The artificial neural network model was used to predict numerical cutting temperatures by means of the numerical cutting forces. The best results in predicting the cutting temperature were obtained using the network architecture with a hidden layer which has seven neurons and LM learning algorithm. Finally, the experimental cutting temperatures were predicted by entering the experimental cutting forces into a formula obtained from the artificial neural networks. Statistical results (R², RMSE, MEP) were quite satisfactory. This demonstrates that the established ANN model is a powerful one for predicting the experimental cutting temperatures.     

Support vector regression and multilayer feed forward neural networks for non-exercise prediction of VO2max

The purpose of this study is to develop non-exercise (N-Ex) VO₂max prediction models by using support vector regression (SVR) and multilayer feed forward neural networks (MFFNN). VO₂max values of 100 subjects (50 males and 50 females) are measured using a maximal graded exercise test. The variables; gender, age, body mass index (BMI), perceived functional ability (PFA) to walk, jog or run given distances and current physical activity rating (PA-R) are used to build two N-Ex prediction models. Using 10-fold cross validation on the dataset, standard error of estimates (SEE) and multiple correlation coefficients (R) of both models are calculated. The MFFNN-based model yields lower SEE (3.23 ml kg^?1 min^?1) whereas the SVR-based model yields higher R (0.93). Compared with the results of the other N-Ex prediction models in literature that are developed using multiple linear regression analysis, the reported values of SEE and R in this study are considerably more accurate. Therefore, the results suggest that SVR-based and MFFNN-based N-Ex prediction models can be valid predictors of VO₂max for heterogeneous samples.     

Human lower extremity joint moment prediction: A wavelet neural network approach

Joint moment is one of the most important factors in human gait analysis. It can be calculated using multi body dynamics but might not be straight forward. This study had two main purposes; firstly, to develop a generic multi-dimensional wavelet neural network (WNN) as a real-time surrogate model to calculate lower extremity joint moments and compare with those determined by multi body dynamics approach, secondly, to compare the calculation accuracy of WNN with feed forward artificial neural network (FFANN) as a traditional intelligent predictive structure in biomechanics.To aim these purposes, data of four patients walked with three different conditions were obtained from the literature. A total of 10 inputs including eight electromyography (EMG) signals and two ground reaction force (GRF) components were determined as the most informative inputs for the WNN based on the mutual information technique. Prediction ability of the network was tested at two different levels of inter-subject generalization. The WNN predictions were validated against outputs from multi body dynamics method in terms of normalized root mean square error (NRMSE (%)) and cross correlation coefficient (ρ).Results showed that WNN can predict joint moments to a high level of accuracy (NRMSE < 10%, ρ > 0.94) compared to FFANN (NRMSE < 16%, ρ > 0.89). A generic WNN could also calculate joint moments much faster and easier than multi body dynamics approach based on GRFs and EMG signals which released the necessity of motion capture. It is therefore indicated that the WNN can be a surrogate model for real-time gait biomechanics evaluation.     

Thermodynamic study of the cerium–cadmium system

Cadmium vapor pressures were determined over Ce–Cd samples by an isopiestic method. The measurements were carried out in the temperature range from 690 to 1080 K and over a composition range of 48–85 at% Cd. From the vapor pressures thermodynamic activities of Cd were derived for all samples at their respective sample temperatures, and partial molar enthalpies of Cd were obtained from the temperature dependence of the activities. With these partial molar enthalpies the Cd activities were converted to a common temperature of 823 K. By means of a Gibbs–Duhem integration Ce activities were calculated, using a corresponding literature value for the two-phase field (CeCd₁₁+L) as integration constant. Finally integral Gibbs energies were calculated for the composition range 48–100 at% Cd with a minimum value of −37 kJ g-atom⁻¹ at 823 K in the phase CeCd. Phase boundaries of the intermetallic compounds CeCd, CeCd₂, Ce₁₃Cd₅₈, and CeCd₁₁ were estimated from the vapor pressure measurements and from SEM analyses.     

Neural networks to predict earthquakes in Chile

A new earthquake prediction system is presented in this work. This method, based on the application of artificial neural networks, has been used to predict earthquakes in Chile, one of the countries with larger seismic activity. The input values are related to the b-value, the Bath's law, and the Omori–Utsu's law, parameters that are strongly correlated with seismicity, as shown in solid previous works. Two kind of prediction are provided in this study: The probability that an earthquake of magnitude larger than a threshold value happens, and the probability that an earthquake of a limited magnitude interval might occur, both during the next five days in the areas analyzed. For the four Chile's seismic regions examined, with epicenters placed on meshes with dimensions varying from 0.5° × 0.5° to 1° × 1°, a prototype of neuronal network is presented. The prototypes predict an earthquake every time the probability of an earthquake of magnitude larger than a threshold is sufficiently high. The threshold values have been adjusted with the aim of obtaining as few false positives as possible. The accuracy of the method has been assessed in retrospective experiments by means of statistical tests and compared with well-known machine learning classifiers. The high success rate achieved supports the suitability of applying soft computing in this field and poses new challenges to be addressed.     

Boosting a fast neural network for supervised land cover classification

It is demonstrated that the use of an ensemble of neural networks for routine land cover classification of multispectral satellite data can lead to a significant improvement in classification accuracy. Specifically, the AdaBoost.M1 algorithm is applied to a sequence of three-layer, feed-forward neural networks. In order to overcome the drawback of long training time for each network in the ensemble, the networks are trained with an efficient Kalman filter algorithm. On the basis of statistical hypothesis tests, classification performance on multispectral imagery is compared with that of maximum likelihood and support vector machine classifiers. Good generalization accuracies are obtained with computation times of the order of 1 h or less. The algorithms involved are described in detail and a software implementation in the ENVI/IDL image analysis environment is provided.     

Wind farm power prediction based on wavelet decomposition and chaotic time series

In this paper, a prediction model is proposed for wind farm power forecasting by combining the wavelet transform, chaotic time series and GM(1, 1) method. The wavelet transform is used to decompose wind farm power into several detail parts associated with high frequencies and an approximate part associated with low frequencies. The characteristic of each high frequencies signal is identified, if it is chaotic time series then use weighted one-rank local-region method to predict it. If not, use GM(1, 1) model to predict it. And the GM(1, 1) model is also used to predict the approximate part of the low frequencies. In the end, the final forecasted result for wind farm power is obtained by summing the predicted results of all extracted high frequencies and the approximate part. According to the predicted results, the proposed method can improve the prediction accuracy of the wind farm power.     

Prediction of California bearing ratio (CBR) of fine grained soils by AI methods

Advances in field of artificial intelligence (AI) offers opportunities of utilizing new algorithms and models that enable researchers to solve the most complex systems. As in other engineering fields, AI methods have widely been used in geotechnical engineering. Unlikely, there seems quite insufficient number of research related to the use of AI methods for the estimation of California bearing ratio (CBR). There were actually some attempts to develop prediction models for CBR, but most of these models were essentially statistical correlations. Nevertheless, many of these statistical correlation equations generally produce unsatisfactory CBR values. However, this paper is likely one of the very first research which aims to investigate the applicability of AI methods for prediction of CBR. In this context, artificial neural network (ANN) and gene expression programming (GEP) were applied for the prediction of CBR of fine grained soils from Southeast Anatolia Region/Turkey. Using CBR test data of fine grained soils, some proper models are successfully developed. The results have shown that the both ANN and GEP are found to be able to learn the relation between CBR and basic soil properties. Additionally, sensitivity analysis is performed and it is found that maximum dry unit weight (γ_d) is the most effective parameter on CBR among the others such as plasticity index (PI), optimum moisture content (w_opt), sand content (S), clay + silt content (C + S), liquid limit (LL) and gravel content (G) respectively.     

Artificial neural networks for earthquake prediction using time series magnitude data or Seismic Electric Signals

The aim of this study is to evaluate the performance of artificial neural networks in predicting earthquakes occurring in the region of Greece with the use of different types of input data. More specifically, two different case studies are considered: the first concerns the prediction of the earthquake magnitude (M) of the following day and the second the prediction of the magnitude of the impending seismic event following the occurrence of pre-seismic signals, the so-called Seismic Electric Signals (SES), which are believed to occur prior to an earthquake, as well as the time lag between the SES and the seismic event itself. The neural network developed for the first case study used only time series magnitude data as input with the output being the magnitude of the following day. The resulting accuracy rate was 80.55% for all seismic events, but only 58.02% for the major seismic events (M ? 5.2 on the Richter scale). Our second case study for earthquake prediction uses SES as input data to the neural networks developed. This case study is separated in two parts with the differentiating element being the way of constructing the missing SES. In the first part, where the missing SES were constructed randomly for all the seismic events, the resulting accuracy rates for the magnitude of upcoming seismic events were just over 60%. In the second part, where the missing SES were constructed for the major seismic events (M ? 5.0 on the Richter scale) only by the use of neural networks reversely, the resulting accuracy rate by predicting only the magnitude was 84.01%, and by predicting both the magnitude and time lag was 83.56% for the magnitude and 92.96% for the time lag. Based on the results we conclude that, when the neural networks are trained by using the appropriate data they are able to generalise and predict unknown seismic events relatively accurately.     

Application of feedforward neural network in the study of dissociated gas flow along the porous wall

This paper concerns the use of feedforward neural networks (FNN) for predicting the nondimensional velocity of the gas that flows along a porous wall. The numerical solution of partial differential equations that govern the fluid flow is applied for training and testing the FNN. The equations were solved using finite differences method by writing a FORTRAN code. The Levenberg–Marquardt algorithm is used to train the neural network. The optimal FNN architecture was determined. The FNN predicted values are in accordance with the values obtained by the finite difference method (FDM). The performance of the neural network model was assessed through the correlation coefficient (r), mean absolute error (MAE) and mean square error (MSE). The respective values of r, MAE and MSE for the testing data are 0.9999, 0.0025 and 1.9998 · 10^?5.     

Investigating data preprocessing methods for circuit complexity models

Preprocessing the data is an important step while creating neural network (NN) applications because this step usually has a significant effect on the prediction performance of the model. This paper compares different data processing strategies for NNs for prediction of Boolean function complexity (BFC). We compare NNs’ predictive capabilities with (1) no preprocessing (2) scaling the values in different curves based on every curve’s own peak and then normalizing to [0, 1] range (3) applying z-score to values in all curves and then normalizing to [0, 1] range, and (4) logarithmically scaling all curves and then normalizing to [0, 1] range. The efficiency of these methods was measured by comparing RMS errors in NN-made BFC predictions for numerous ISCAS benchmark circuits. Logarithmic preprocessing method resulted in the best prediction statistics as compared to other techniques.     

A genetic algorithm to refine input data selection for air temperature prediction using artificial neural networks

The accurate prediction of air temperature is important in many areas of decision-making including agricultural management, transportation and energy management. Previous research has focused on the development of artificial neural network (ANN) models to predict air temperature from one to twelve hours in advance. The inputs to these models included a constant duration of prior data with a fixed resolution for all environmental variables for all prediction horizons. The overall goal of this research was to develop more accurate ANN models that could predict air temperature for each prediction horizon. The specific objective was to determine if the ANN model accuracy could be improved by applying a genetic algorithm (GA) for each prediction horizon to determine the preferred duration and resolution of input prior data for each environmental variable. The ANN models created based on this GA based approach provided smaller errors than the models created based on the existing constant duration and fixed data resolution approach for all twelve prediction horizons. Except for a few cases, the GA generally included a longer duration for prior air temperature data and shorter durations for other environmental variables. The mean absolute errors (MAEs) for the evaluation input patterns of the one-, four-, eight-, and twelve-hour prediction models that were based on this GA approach were 0.564 °C, 1.264 °C, 1.766 °C and 2.018 °C, respectively. These MAEs were improvements of 3.98%, 4.59%, 2.55% and 1.70% compared to the models that were created based on the existing approach for the same corresponding prediction horizons. Thus, the GA based approach to determine the duration and resolution of prior input data resulted in more accurate ANN models than the existing ones for air temperature prediction. Future work could examine the effects of various GA and fitness evaluation parameters that were part of the approach used in this study.