Performance Of Data-driven Models Research Articles

Retrieval of surface solar irradiance from satellite imagery using machine learning: pitfalls and perspectives

Abstract. Knowledge of the spatial and temporal characteristics of solar surface irradiance (SSI) is critical in many domains. While meteorological ground stations can provide accurate measurements of SSI locally, they are sparsely distributed worldwide. SSI estimations derived from satellite imagery are thus crucial to gain a finer understanding of the solar resource. Inferring SSI from satellite images is, however, not straightforward, and it has been the focus of many researchers in the past 30 to 40 years. For long, the emphasis has been on models grounded in physical laws with, in some cases, simple statistical parametrizations. Recently, new satellite SSI retrieval methods have been emerging, which directly infer the SSI from the satellite images using machine learning. Although only a few such works have been published, their practical efficiency has already been questioned. The objective of this paper is to better understand the potential and the pitfalls of this new family of methods. To do so, simple multi-layer-perceptron (MLP) models are constructed with different training datasets of satellite-based radiance measurements from Meteosat Second Generation (MSG) with collocated SSI ground measurements from Météo-France. The performance of the models is evaluated on a test dataset independent from the training set in both space and time and compared to that of a state-of-the-art physical retrieval model from the Copernicus Atmosphere Monitoring Service (CAMS). We found that the data-driven model's performance is very dependent on the training set. Provided the training set is sufficiently large and similar enough to the test set, even a simple MLP has a root mean square error (RMSE) that is 19 % lower than CAMS and outperforms the physical retrieval model at 96 % of the test stations. On the other hand, in certain configurations, the data-driven model can dramatically underperform even in stations located close to the training set: when geographical separation was enforced between the training and test set, the MLP-based model exhibited an RMSE that was 50 % to 100 % higher than that of CAMS in several locations.

Self-optimization of training dataset improves forecasting of cyanobacterial bloom by machine learning

Data-driven model (DDM) prediction of aquatic ecological responses, such as cyanobacterial harmful algal blooms (CyanoHABs), is critically influenced by the choice of training dataset. However, a systematic method to choose the optimal training dataset considering data history has not yet been developed. Providing a comprehensive procedure with self-based optimal training dataset-selecting algorithm would self-improve the DDM performance. In this study, a novel algorithm was developed to self-generate possible training dataset candidates from the available input and output variable data and self-choose the optimal training dataset that maximizes CyanoHAB forecasting performance. Nine years of meteorological and water quality data (input) and CyanoHAB data (output) from a site on the Nakdong River, South Korea, were acquired and pretreated via an automated process. An artificial neural network (ANN) was chosen from among the DDM candidates by first-cut training and validation using the entire collected dataset. Optimal training datasets for the ANN were self-selected from among the possible self-generated training datasets by systematically simulating the performance in response to 46 periods and 40 sizes (number of data elements) of the generated training datasets. The best-performing models were screened to identify the candidate models. The best performance corresponded to 6–7 years of training data (∼18 % lower error) for forecasting 1–28 d ahead (1–28 d of forecasting lead time (FLT)). After the hyperparameters of the screened model candidates were fine-tuned, the best-performing model (7 years of data with 14 d FLT) was self-determined by comparing the forecasts with unseen CyanoHAB events. The self-determined model could reasonably predict CyanoHABs occurring in Korean waters (cyanobacteria cells/mL ≥ 1000). Thus, our proposed method of self-optimizing the training dataset effectively improved the predictive accuracy and operational efficiency of the DDM prediction of CyanoHAB.

Unraveling the morphological complexity of two-dimensional macromolecules

Summary2D macromolecules, such as graphene and graphene oxide, possess a rich spectrum of conformational phases. However, their morphological classification has only been discussed by visual inspection, where the physics of deformation and surface contact cannot be resolved. We employ machine learning methods to address this problem by exploring samples generated by molecular simulations. Features such as metric changes, curvature, conformational anisotropy and surface contact are extracted. Unsupervised learning classifies the morphologies into the quasi-flat, folded, crumpled phases and interphases using geometrical and topological labels or the principal features of the 2D energy map. The results are fed into subsequent supervised learning for phase characterization. The performance of data-driven models is improved notably by integrating the physics of geometrical deformation and topological contact. The classification and feature extraction characterize the microstructures of their condensed phases and the molecular processes of adsorption and transport, comprehending the processing-microstructures-performance relation in applications.

Downhole data correction for data-driven rate of penetration prediction modeling

In recent years, machine learning has been adopted in the Oil and Gas industry as a promising technology for solutions to the most demanding problems like downhole parameters estimations and incidents detection. A big amount of available data makes this technology an attractive option for solving a wide variety of drilling problems, as well as a reliable candidate for performing big-data analysis and interpretation. Nevertheless, this approach may cause, in some cases, that petroleum engineering concepts are disregarded in favor of more data-intensive approaches. This study aims to evaluate the impact of drilling data measurement correction on data-driven model performance. In our study, besides using the standard data processing technologies, like gap filling, outlier removal, noise reduction etc., the physics-based drilling models are also implemented for data quality improvement and data correction in consideration of the measurement physics, rarely mentioned in most of publications. In our case study, recurrent neural networks (RNN) that are able to capture temporal natures of a signal are employed for the rate of penetration (ROP) estimation with an adjustable predictive window. The results show that the RNN model produces the best results when using the drilling data recovered through analytical methods. Moreover, the comprehensive data-driven model evaluation and engineering interpretation are conducted to facilitate better understanding of the data-driven models and their applications.

Background Noise Removal Technique using Deep Learning Segmentation Network without Segmentation Map

Audio data acquired at industrial manufacturing sites often include unexpected background noise. Since the performance of data-driven models can be worse by background noise. Therefore, it is important to get rid of unwanted background noise. There are two main techniques for noise canceling in a traditional manner. One is Active Noise Canceling (ANC), which generates an inverted phase of the sound that we want to remove. The other is Passive Noise Canceling (PNC), which physically blocks the noise. However, these methods require large device size and expensive cost. Thus, we propose a deep learning-based noise canceling method. This technique was developed using audio imaging technique and deep learning segmentation network. However, the proposed model only needs the information on whether the audio contains noise or not. In other words, unlike the general segmentation technique, a pixel-wise ground truth segmentation map is not required for this method. We demonstrate to evaluate the separation using pump sound of MIMII dataset, which is open-source dataset.

Experimental Evaluation of Solar Radiation and Solar Efficacy Models and Performance of Data-Driven Models

Abstract Weather data are major input for building energy usage predictions. However, solar-radiation-related historical and real-time weather data are unavailable or incomplete in many locations. ...

Data pre-processing effect on ANN-based prediction intervals construction of the evaporation process at different climate regions in Iran

The point predictions of stochastic processes, such as evaporation, by data-driven methods such as Artificial Neural Network (ANN), are associated with uncertainties. Furthermore, the performance of data-driven models, as well as their uncertainty, are dependent on the quality and quantity of the used data. The main aim of this research was the Uncertainty Quantifying (UQ) of the ANN-based evaporation predictions by Prediction Intervals (PIs) analysis using the data from three stations in Iran (i.e. Tabriz, Urmia, and Ardabil). In this way, data pre-processing methods i.e. Wavelet-based De-noising (WD), training with Jitted Data (JD) and also their combination i.e. Hybrid Wavelet De-noising and Jitted Data (HWDJD) were applied to examine their effects on the estimated values of PIs. The Lower-Upper-Bound Estimation (LUBE) method as the direct NN-based PI construction was utilized for estimating the PI values. Since the efficiency of any ANN model and consequently, the robustness of the uncertainty analysis is sensitive to the correct selection of input variables, the first order Partial Derivation (PaD) sensitivity analysis method was also used to select dominant inputs among all potential input variables. The results indicated that the LUBE technique could provide acceptable results in estimating the uncertainty bounds, whereas the uncertainty quantity would be affected by the used data pre-processing methods. The reduction of uncertainty effect via data pre-processing methods was significant in modeling Urmia and Ardabil stations. Results showed that the reduction of PI bandwidths via HWDJD, WD, and JD methods were up to 30%, 21%, and 9%, respectively. It means that to reduce the ANN-based modeling uncertainty due to the uncertainty involved in the input data set, at first, the contaminant noise of the available data should be eliminated from the time series, then artificially generated time series (JD) that mimic the smoothed time series pattern can be generated and used in the training process.

Compressive strength Prediction recycled aggregate incorporated concrete using Adaptive Neuro-Fuzzy System and Multiple Linear Regression

Compressive strength of concrete, renowned as one of the most substantial mechanical properties of concrete and key factors for the quality assurance of concrete. In the present study, two different data-driven models, i.e., Adaptive Neuro-Fuzzy Inference System (ANFIS), and Multiple Linear Regression (MLR) were used to predict the 28 days compressive strength of recycled aggregate concrete (RAC). 16 different input parameters, including both dimensional and non-dimensional parameters, were used for predicting the 28 days compressive strength of concrete. The present study established that estimation of 28 days compressive strength of recycled aggregate concrete was performed better by ANFIS in comparison to MLR. Besides, the performance of data-driven models with and without the non-dimensional parameters is explored. It was observed that the data-driven models show better accuracy when the non-dimensional parameters were used as additional input parameters. Furthermore, the effect of each non-dimensional parameter on the performance of each data-driven model is investigated and 28 days compressive strength of concrete is examined.

Predicting strength of recycled aggregate concrete using Artificial Neural Network, Adaptive Neuro-Fuzzy Inference System and Multiple Linear Regression

Abstract Compressive strength of concrete, recognized as one of the most significant mechanical properties of concrete, is identified as one of the most essential factors for the quality assurance of concrete. In the current study, three different data-driven models, i.e., Artificial Neural Network (ANN), Adaptive Neuro-Fuzzy Inference System (ANFIS), and Multiple Linear Regression (MLR) were used to predict the 28 days compressive strength of recycled aggregate concrete (RAC). Recycled aggregate is the current need of the hour owing to its environmental pleasant aspect of re-using the wastes due to construction. 14 different input parameters, including both dimensional and non-dimensional parameters, were used in this study for predicting the 28 days compressive strength of concrete. The present study concluded that estimation of 28 days compressive strength of recycled aggregate concrete was performed better by ANN and ANFIS in comparison to MLR. In other words, comparing the test step of all the three models, it can be concluded that the MLR model is better to be utilized for preliminary mix design of concrete, and ANN and ANFIS models are suggested to be used in the mix design optimization and in the case of higher accuracy necessities. In addition, the performance of data-driven models with and without the non-dimensional parameters is explored. It was observed that the data-driven models show better accuracy when the non-dimensional parameters were used as additional input parameters. Furthermore, the effect of each non-dimensional parameter on the performance of each data-driven model is investigated. Finally, the effect of number of input parameters on 28 days compressive strength of concrete is examined.

Conjunction of ANN and threshold based wavelet de-noising approach for forecasting suspended sediment load

Information on suspended sediment load (SSL) is fundamental for numerous water resources management and environmental protection projects. This phenomenon has the inherent complexity due to a large number of vague parameters and existence of both spatial variability of the basin characteristics and temporal climatic patterns. This complexity turns into a barrier to get accurate prediction by conventional linear methods. On the other hand, the extent of the noise on hydrological data reduces the performance of data-driven models like Artificial Neural Networks (ANNs). Although ANNs could capture the complex nonlinear relationship between input and output parameters, being data-driven method positioned it in a state of need to preprocessed data. In this paper, the application of ANN approach focusing on wavelet- based denoising method for modeling daily streamflow-sediment relationship was proposed. The daily streamflow and SSL data observed at outlet of the Potomac River in USA were used as the case study. Achieving this purpose, Daubechies (db) was used as mother wavelet to decompose both streamflow and sediment time series into detailed and approximation subseries. Decomposition at level ten via db3 and at level eight via db5 were examined for streamflow and SSL time series, respectively. At first, the appropriate input combination with raw data to estimate current SSL was determined and re-imposed to ANN with denoised data. The comparison of results reveals that in term of determination coefficient, the obtained result by denoised data was improved up to 23.2% with raged to use noisy, raw data and this exhibits that denoised data can be employed productively in ANN-based daily SSL forecasting.

Influence of lag time on event-based rainfall–runoff modeling using the data driven approach

This study investigated the effect of lag time on the performance of data-driven models, specifically the adaptive network-based fuzzy inference system (ANFIS), in event-based rainfall–runoff modeling. Rainfall and runoff data for a catchment in Singapore were chosen for this study. For the purpose of this study, lag time was determined from cross-correlation analysis of the rainfall and runoff time series. Rainfall antecedents were the only inputs of the models and direct runoff was the desired output. An ANFIS model with three sub-models defined based on three different ranges of lag times was developed. The performance of the sub-models was compared with previously developed ANFIS models and the physically-based Storm Water Management Model (SWMM). The ANFIS sub-models gave significantly superior results in terms of the RMSE, r2, CE and the prediction of the peak discharge, compared to other ANFIS models where the lag time was not considered. In addition, the ANFIS sub-models provided results that were comparable with results from SWMM. It is thus concluded that the lag time plays an important role in the selection of events for training and testing of data-driven models in event-based rainfall–runoff modeling.