Comment on hess-2022-365

<strong class="journal-contentHeaderColor">Abstract.</strong> Hydrological models are widely used to characterise, understand and manage hydrosystems. Data-driven models are of particular interest in karst environments given the complexity and heterogeneity of these systems. There is a multitude of data-driven modelling approaches, which can make it difficult for a manager or researcher to choose. We therefore conducted a comparison of two data-driven modelling approaches: artificial neural networks (ANN) and reservoir models. We investigate five karst systems in the Mediterranean and Alpine regions with different characteristics in terms of climatic conditions, hydrogeological properties and data availability. We compare the results of ANN and reservoir modelling approaches using several performance criteria over different hydrological periods. The results show that both ANN and reservoir models can accurately simulate karst spring discharge, but also that they have different advantages and drawbacks: (i) ANN models are very flexible regarding the format and amount of input data, (ii) reservoir models can provide good results even with short calibration periods, and (iii) ANN models seem robust for reproducing high-flow conditions while reservoir models are superior for reproducing low-flow conditions. However, both modelling approaches struggle to reproduce extreme events (droughts, floods), which is a known problem in hydrological modelling. For research purposes, ANN models have shown to be useful to identify recharge areas and delineate catchment, based on insights into the input data. Reservoir models are adapted to understand the hydrological functioning of a system, by studying model structure and parameters.

As a major component of the hydrologic cycle, rainfall runoff plays a key role in water resources management and sustainable development. Conceptual models of the rainfall-runoff process are governed by parameters that can rarely be directly determined for use in distributed models, but should be either inferred through good judgment or calibrated against the historical record. Artificial neural network (ANN) models require comparatively fewer such parameters, but their accuracy needs to be checked. This paper compares a Hydrologic Engineering Centre-Hydrologic Modeling System (HEC-HMS) conceptual model and an ANN model based on the conjugate gradient method for streamflow prediction. Daily precipitation, temperature, and streamflow data of the Upper Indus River for a period of 20 years (1985–2004) are used as input for calibrating in the case of the HEC-HMS, and for training in case of the ANN. Ten years of data (2005–2014) are used to validate the HEC-HMS model and test the ANN. The performance of the models is assessed using different statistical indicators such as the Nash–Sutcliffe efficiency (NSE), root mean square error (RMSE), mean bias error (MBE), and the coefficient of determination (R2). The results show good predictions for streamflow in the case of both HEC-HMS and ANN models. A parametric study is conducted using Monte Carlo analysis and finds that the most important parameters for HEC-HMS models are the storage coefficient and the time of concentration; while for ANN models, input combinations are the most important. This study investigates the sensitivity of these parameters, which can be used to determine preliminary estimation ranges of their values for future modeling. Finally, evaluating the impact of the simulated streamflow’s accuracy on the flow duration curve shows that the curve is significantly affected by any streamflow simulation inaccuracy.

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

The generalization ability of artificial neural networks (ANNs) and M5 model tree (M5Tree) in modeling reference evapotranspiration (ET 0 ) is investigated in this study. Daily climatic data, average temperature, solar radiation, wind speed, and relative humidity from six different stations operated by California Irrigation Management Information System (CIMIS) located in two different regions of the USA were used in the applications. King-City Oasis Rd., Arroyo Seco, and Salinas North stations are located in San Joaquin region, and San Luis Obispo, Santa Monica, and Santa Barbara stations are located in the Southern region. In the first part of the study, the ANN and M5Tree models were used for estimating ET 0 of six stations and results were compared with the empirical methods. The ANN and M5Tree models were found to be better than the empirical models. In the second part of the study, the ANN and M5Tree models obtained from one station were tested using the data from the other two stations for each region. ANN models performed better than the CIMIS Penman, Hargreaves, Ritchie, and Turc models in two stations while the M5Tree models generally showed better accuracy than the corresponding empirical models in all stations. In the third part of the study, the ANN and M5Tree models were calibrated using three stations located in San Joaquin region and tested using the data from the other three stations located in the Southern region. Four-input ANN and M5Tree models performed better than the CIMIS Penman in only one station while the two-input ANN models were found to be better than the Hargreaves, Ritchie, and Turc models in two stations.

The present study was carried out using artificial neural network (ANN) model for predicting the sodium hazardness, i.e., sodium adsorption ratio (SAR), percent sodium (%Na) residual, Kelly’s ratio (KR), and residual sodium carbonate (RSC) in the groundwater of the Pratapgarh district of Southern Rajasthan, India. This study focuses on verifying the suitability of water for irrigational purpose, wherein more groundwater decline coupled with water quality problems compared to the other areas are observed. The southern part of the Rajasthan State is more populated as compared to the rest of the parts. The southern part of the Rajasthan is more populated as compared to the rest of the Rajasthan, which leads to the industrialization, urbanization, and evolutionary changes in the agricultural production in the southern region. Therefore, it is necessary to propose innovative methods for analyzing and predicting the water quality (WQ) for agricultural use. The study aims to develop an optimized artificial neural network (ANN) model to predict the sodium hazardness of groundwater for irrigation purposes. The ANN model was developed using ‘nntool’ in MATLAB software. The ANN model was trained and validated for ten years (2010–2020) of water quality data. An L-M 3-layer back propagation technique was adopted in ANN architecture to develop a reliable and accurate model for predicting the suitability of groundwater for irrigation. Furthermore, statistical performance indicators, such as RMSE, IA, R, and MBE, were used to check the consistency of ANN prediction results. The developed ANN model, i.e., ANN4 (3-12-1), ANN4 (4-15-1), ANN1 (4-5-1), and ANN4 (3-12-1), were found best suited for SAR, %Na, RSC, and KR water quality indicators for the Pratapgarh district. The performance analysis of the developed model (3-12-1) led to a correlation coefficient = 1, IA = 1, RMS = 0.14, and MBE = 0.0050. Hence, the proposed model provides a satisfactory match to the empirically generated datasets in the observed wells. This development of water quality modeling using an ANN model may help to useful for the planning of sustainable management and groundwater resources with crop suitability plans as per water quality.

Monthly river flow forecasting using artificial neural network and support vector regression models coupled with wavelet transform

BackgroundThis study aimed to develop an artificial neural network (ANN) model for predicting synchronous organ-specific metastasis in lung cancer (LC) patients.MethodsA total of 62,151 patients who diagnosed as LC without data missing between 2010 and 2015 were identified from Surveillance, Epidemiology, and End Results (SEER) program. The ANN model was trained and tested on an 75/25 split of the dataset. The receiver operating characteristic (ROC) curves, area under the curve (AUC) and sensitivity were used to evaluate and compare the ANN model with the random forest model.ResultsFor distant metastasis in the whole cohort, the ANN model had metrics AUC = 0.759, accuracy = 0.669, sensitivity = 0.906, and specificity = 0.613, which was better than the random forest model. For organ-specific metastasis in the cohort with distant metastasis, the sensitivity in bone metastasis, brain metastasis and liver metastasis were 0.913, 0.906 and 0.925, respectively. The most important variable was separate tumor nodules with 100% importance. The second important variable was visceral pleural invasion for distant metastasis, while histology for organ-specific metastasis.ConclusionsOur study developed a “two-step” ANN model for predicting synchronous organ-specific metastasis in LC patients. This ANN model may provide clinicians with more personalized clinical decisions, contribute to rationalize metastasis screening, and reduce the burden on patients and the health care system.

In this research work, we developed a mathematical model of a digital currency market, involving daily closing price as a function of time. We proposed the Artificial Neural Network (ANN) model. We observed that our ANN model was able to predict the daily closing price of Bitcoin and also make six weeks forecast to a reasonable degree of accuracy. We equally observe that the time dependent ANN model can actually give digital currency traders and investors a clue on when to trade off their digital assets with minimum risk. We therefore, recommend that ANN model should be incorporated into digital currency trading platforms as a signal tool to enable digital currency traders take more informed and less risky trading decisions. From our findings, we would advise traders who wish to employ ANN model to consider a smaller time frame say a few weeks’ time interval for their predictions. We observed also that ANN models have limitations when it comes to manual computation or implementation in Microsoft Excel, especially when dealing with very large input values. This is because of the saturation characteristic of our ANN inner layer activation function (viz; tanh function) which can lead to identical output values for different input values, making it difficult to replicate the ANN model's behavior. Furthermore, ANN models often involve complex interactions between multiple neurons, layers, and activation functions, which can be challenging to replicate manually.

Purpose – This paper aims to predict the needle penetration force (NPF) in denim fabrics using the artificial neural network (ANN) and multiple linear regression (MLR) models based on the effects of various sewing parameters. Design/methodology/approach – In order to design the ANN and MLR models, four parameters including fabric weight, number of fabric layers, weave pattern, and sewing needle size are taken into account as the input parameters and NPF as the output parameter. According to these parameters, 140 samples of data were resulted. Each sample was tested five times. From these 140 data (input-output data pairs), 112 were used for training the ANN and MLR models and 28 samples were used to test the performance of ANN and MLR. Also, the NPF was measured on the Instron tensile tester to simulate sewing process. Findings – The results indicated that the NPF in denim fabrics can be well predicted in terms of sewing parameters by using ANN and MLR models, in which the ANN model exhibits greater performance than MLR (RANN=0.989&gt; RMLR=0.901). Research limitations/implications – The NPF measurement method is limited at low speed. Originality/value – Using the ANN model for forecasting NPF in denim fabrics can help the garment manufactures to produce high-quality denim products and improve the sewing process through reducing seam damage. The NPF could be also measured in the cycle loading conditions similar to sewing machine process by using a special designed tools mounted on the Instron tensile tester.

The purpose of this study was to validate the use of artificial neural network (ANN) models for predicting quality of life (QOL) after breast cancer surgery and to compare the predictive capability of ANNs with that of linear regression (LR) models. The European Organization for Research and Treatment of Cancer Quality of Life Questionnaire and its supplementary breast cancer measure were completed by 402 breast cancer patients at baseline and at 2 years postoperatively. The accuracy of the system models were evaluated in terms of mean square error (MSE) and mean absolute percentage error (MAPE). A global sensitivity analysis was also performed to assess the relative significance of input parameters in the system model and to rank the variables in order of importance. Compared to the LR model, the ANN model generally had smaller MSE and MAPE values in both the training and testing datasets. Most ANN models had MAPE values ranging from 4.70 to 19.96 %, and most had high prediction accuracy. The ANN model also outperformed the LR model in terms of prediction accuracy. According to global sensitivity analysis, pre-operative functional status was the best predictor of QOL after surgery. Compared with the conventional LR model, the ANN model in the study was more accurate for predicting patient-reported QOL and had higher overall performance indices. Further refinements are expected to obtain sufficient performance improvements for its routine use in clinical practice as an adjunctive decision-making tool.

The power capacity of solar photovoltaics (PVs) in Korea has grown dramatically in recent years, and an accurate estimation of solar resources is crucial for the efficient management of these solar PV systems. Since the number of solar irradiance measurement sites is insufficient for Korea, satellite images can be useful sources for estimating solar irradiance over a wide area of Korea. In this study, an artificial neural network (ANN) model was constructed to calculate hourly global horizontal solar irradiance (GHI) from Korea Communication, Ocean and Meteorological Satellite (COMS) Meteorological Imager (MI) images. Solar position variables and five COMS MI channels were used as inputs for the ANN model. The basic ANN model was determined to have a window size of five for the input satellite images and two hidden layers, with 30 nodes on each hidden layer. After these ANN parameters were determined, the temporal and spatial applicability of the ANN model for solar irradiance mapping was validated. The final ANN ensemble model, which calculated the hourly GHI from 10 independent ANN models, exhibited a correlation coefficient (R) of 0.975 and root mean square error (RMSE) of 54.44 W/m² (12.93%), which were better results than for other remote-sensing based works for Korea. Finally, GHI maps for Korea were generated using the final ANN ensemble model. This COMS-based ANN model can contribute to the efficient estimation of solar resources and the improvement of the operational efficiency of solar PV systems for Korea.

ABSTRACTStatistical methods such as Poisson distribution, negative binomial regression (NB), and zero inflated negative binomial regression (ZINB) have generally been used in road safety studies to establish the complex relationships between variables. Over the last few years, the artificial neural networks (ANN) model has also been used. The ANN model does not have any prior limitations such as the equality condition of mean and variance observed in Poisson regression. However, though the ANN model has been used in the analysis of different accident types, to the best of our knowledge, no study has used the ANN model for establishing this relationship with truck accident data on divided multilane interurban roads. In this study, the road sections D750/07–D750/15 in Ankara–Aksaray–Eregli, Turkey, were considered and truck accident data from 2008 to 2011 were analyzed using NB and ANN. The analysis show that the ANN model has lower errors and higher R2 values than NB and performs slightly better than NB for predicting the number of trucks involved in accidents. Based on a comparison of performances the study concludes, that ANN could be used as an alternative model for analyzing truck accident data on divided multilane interurban roads in Turkey.

According to the thought of intelligent forecasting and hybrid forecasting, an Intelligent Hybrid (IH) model for short-term traffic flow forecasting was presented. The IH model had three sub-models: History Mean (HM) model, Artificial Neural Network (ANN) model and the Fuzzy Combination (FC) model. By means of the good static stabilization character of HM method, the HM model predicted the traffic flow by the Single Exponential Smoothing method based on the historical traffic data. Otherwise, the ANN model was a 1.5-layer feed-forward neural network built by some common S-function neurons. Because of the strong dynamic nonlinear mapping ability of ANN, the ANN model can estimate the actual traffic flow in a very precise and satisfactory sense. The FC model mixed the two individual forecasting results by fuzzy logic and its output was regarded as the final forecasting of the traffic flow. Factual application results show that the IH model, which takes advantage of the unique strength of the HM model and the ANN model, can produce more precise forecasting than that of two individual models. Thus, the IH model can be an efficient method to the short-term traffic flow forecasting.

Nitric acid production plants emit small amounts of nitrogen oxides (NOx) to the environment. As the regulatory authorities demand the reduction of the resulting air pollution, existing plants are looking for economical ways to comply with this demand. Several Artificial Neural Networks (ANN) models were trained from several months of operating plant data to predict the NOx concentration in the tail gas, and their total amount emitted the environment. The training of the ANN model was done by the Guterman-Boger algorithm set that generates a non-random initial connection weights, suggests a small number of hidden neurons, avoids, and escapes from, local minima encountered during the training. The ANN models gave small errors, 0.6% relative error on the NOx concentration prediction and 0.006 kg/hour on daily emission in the 20-45 kg NOx/hour range. Knowledge extraction from the trained ANN models revealed the underlying relationships between the plant operating variables and the NOx emission rate, especially the beneficial effect of cooling the absorbed gas and reticulating liquids in the absorption towers. Clustering the data by the patterns of the hidden neurons outputs of auto-associative ANN models of the same data revealed interesting insights.

Data-driven simulation models can be used for optimal control and decision making in building systems only if they are developed accurately and reliably. An unsolved issue pertains to the data collected from BEMS that generally contain errors owing to sensors, malfunctioning systems, and other unknown reasons. These anomalies generally occur in a multidimensional space and cannot be easily detected. Unfortunately, these anomalies can prevent the simulation model from being accurate, reliable, and scalable. This paper presents an automatic anomaly detection method using a support vector data description (SVDD). The data obtained from a compression chiller in a real office building were classified into raw and filtered datasets based on SVDD. Two artificial neural network (ANN) models were developed (ANNraw and ANNSVDD). Based on the comparison between the two ANNs, the ANNSVDD is found to be better than ANNraw in terms of the model reliability and reproducibility. It is also interesting that the model accuracy differences between the two ANN models are marginal. However, the accuracy of the two ANN models can be improved as long as they are tested against filtered data by SVDD. The improvements in accuracy signify the importance of the elimination of these anomalies.