A Machine Learning based Closed-Loop Efficiency Optimization Strategy for Data-Driven Optimization Methods

Lean premixed combustion is the state-of-the-art technology to achieve ultra low NOx emissions in stationary gas turbines. However, lean premixed flames are susceptible to thermoacoustic instabilities, lean blowout, and flashback. The design of such a combustion system is thus always related to the balancing between the levels of emissions and flame stability. Data-driven optimization methods and the adaptation of models through artificial intelligence have experienced a surge in development in the past years. The goal of this study is to show the potential of these methods for gas turbine burner development. A special pilot burner that features 61 different positions of fuel injection, manufactured by means of selective laser melting is used to modify the gas mixture close to the flame anchoring position. Each of the injector lines is equipped with an individual valve, such that the distribution of fuel-air mixture can be modified variously. Installed into an industrial MGT6000 swirl combustor, a data-driven optimization method is used to find an optimal subset of injection locations by automated experiments. The method uses a surrogate model that is based on Gaussian processes regression. It is adopted for experimental optimization, keeping measurement efforts to a minimum. The optimizer controls the fuel valves and uses live measurements to find a distribution that generates minimal NOx emissions while ensuring flame stability. The solutions found by the optimization scheme are analyzed and advantages and limitations of the approach are discussed.

Accurate prediction of endpoint carbon at the dynamic control stage in the converter is crucial for achieving smelting targets. Currently, there are two main methods for converter endpoint prediction: the data-driven method and the mechanism-based method. Data-driven methods exhibit high accuracy but are vulnerable to data quality variations and lack interpretability. Mechanism-based methods provide great interpretability but face challenges in precisely identifying key parameters in the mechanism formula. Inspired by the design concept of physics-informed neural networks (PINNs), an integrated data-driven and mechanism-based method for endpoint carbon prediction in BOF (dmPINNs, data-driven and mechanism-based physics-informed neural networks) is proposed, which has four parts: feature extraction, mechanism-based calculation, data-driven prediction, and integrated prediction. We identify key parameters of the mechanism formula through the neural network to obtain the specified formula for each heat and supervise the training process of the neural network through the mechanism formula to ensure interpretability. Experimental results show that, within the ±0.012% error range, the hit rate of endpoint carbon content using dmPINNs improved by 5.23% compared with the traditional data-driven method and has greater robustness with the supervision of the mechanism formula.

An integrated method for critical clearing time prediction based on a model-driven and ensemble cost-sensitive data-driven scheme

Image set annotation is an important task in the supervised training of the deep neural network. Manual and data-driven dataset annotation methods are commonly used approaches. Both of them have shortcomings, especially in the case of the dataset requiring professional knowledge, which leads to high cost with manual annotation methods and poorly diversified annotation samples with data-driven annotation methods. Although the recommendation annotation method based on cosine similarity using deep neural network features takes advantages of manual annotation and data-driven method, there are still problems such as low accuracy and click-through rate. In order to improve the recommendation accuracy and click-through rate, we propose a confusion graph recommendation annotation method, which builds a confusion graph based on the Largest Margin Nearest Neighbor (LMNN) distance among deep neural network features, to recommend the most confusing images to annotators. In this paper, we made ablation studies on the self-built child face dataset in terms of Precision, mAP (mean Average Precision), and CTR (click-through-rate). The experimental results show that the proposed method achieves superior performance, compared with the cosine similarity recommendation annotation method and the manual annotation method.

A hybrid remaining useful life prognostic method for proton exchange membrane fuel cell

A systematic review of supply and demand side optimal load scheduling in a smart grid environment

Modern power systems are pushed to operate close to their secure limits due to increased system complexity and high penetration of renewable energy sources (RES). Conventional stability assessment (SA) and control methods face significant adaptiveness and robustness issues under highly uncertain and variable modern power systems. In recent years, with advanced real-time system monitoring and communication techniques like phasor measurement unit (PMU) devices, huge volumes of network-wide time-synchronized measurements can be collected and utilized. Under this new reality, data-driven methods have been identified as powerful tools for online and real-time stability assessment and control. This Ph.D. research aims to develop data-analytic methodologies online and real-time assess the system stability status and provide intelligent control actions when the power system is subject to a severe contingency. Specifically, this research covers a series of new methods in pre-fault SA, post-fault real-time SA and power system emergency control. Special efforts are devoted to enhancing the accuracy and robustness of data-driven methods under some extreme scenarios and practical problems, and a confidence-aware data-driven control scheme to manage the potential prediction error better. In data-driven SA, the point and consecutive data missing at both offline and online stages are studied, and two missing data-robust methods are developed, respectively for pre-fault SA and post-fault SA. For the pre-fault method, to enhance the robustness of SA against incomplete PMU measurements, a hybrid ensemble learning-based online SA method is proposed with feature-validation-based missing data estimators. Considering the possibility that some features may have a low correlation with other measurements, a feature validation process detects the potentially inaccurately estimated features to ensure the model performance and data restoration quality. Further, a novel spatial-temporal recurrent imputation network, SRIN, is proposed for time-series data missing at both sensor and PMU levels. The proposed SRIN method jointly learns the imputation for missing values and prediction output through an integrated multi-task recurrent network model, where the imputed values and the SA output can be simultaneously and effectively updated during backpropagation. In this way, the proposed method can be effectively trained by incompletely recorded datasets, which is immensely valuable for practical applications without high-quality historical data. Moreover, by learning shared representations or features across tasks, multi-task models can better exploit the information present in the data and improve the overall performance of both tasks. In addition, considering the time-varying and location-varying missing data patterns, the imputations of SRIN are derived by an intelligent combination of history-based and feature-based estimations through a trainable weighting component. Thereby, the proposed method is highly adaptive to various data missing scenarios, including both spatial and temporal consecutive missing patterns. In emergency control, data-driven methods for event-based load shedding (ELS) and response-based load shedding (RLS) are developed. Specifically, a hierarchical data-driven method is proposed for real-time prediction of ELS to mitigate fault-induced delayed voltage recovery (FIDVR). First, a novel adaptive trajectory sensitivity analysis is developed to optimize the ELS solutions offline and construct the ELS database. Then, by hierarchically modelling the ELS problem as a multi-output classification subproblem and a regression subproblem, the critical load buses mainly responsible for the delayed voltage recovery can be identified, and the over-shedding buses can be effectively reduced. Moreover, based on an improved weighted kernel extreme learning machine (WKELM) method, the direct mapping between system pre-fault operating conditions and corresponding control variables is constructed. Meanwhile, the established weighted learning scheme can extract the decision-making knowledge from the highly sparse and imbalanced dataset and emphasize on the rare yet critical cases. Therefore, the risk of insufficient load shedding is reduced, and better ELS effectiveness can be achieved. Subsequently, a probabilistic confidence-aware method for real-time RLS is developed to manage the potential prediction error better. The method is based on a two-stage process where a scalable Gaussian process (SGP) method determines the average load shedding amount at the first stage. The upper bound of the prediction interval is considered supplementary in the second stage. In parallel, if an RLS case is identified as unconfident, the conservative rule-based load shedding is taken as a backup plan rather than adopting LS estimations that cannot be trusted. Guided by the probabilistic information, the potential prediction error can be fully considered, and the confidence information can be incorporated into the control process, thereby more accurate and reliable load shedding actions can be taken. Leveraging confident predictions-based 2-stage RLS process, the excessive load loss can be reduced while ensuring high control effectiveness. All the proposed methods have been well verified on simulation experiments over standard power system models.

The development of heavy oil reservoirs in China is of great significance to safeguard national energy security, but great challenges are faced due to the complex and heterogeneous reservoir properties. Inter-well connectivity analysis is critical to enhancing the development performance, as it is a good way to interpret fluid flow and provides a theoretical basis for injection-production optimization. Data-driven deep learning methods have been widely used in reservoir development and can be employed to develop surrogate models of injection and production and to infer inter-well connectivity. In this study, the model performance of a recurrent neural network (RNN) and its four variants were evaluated and compared in a temporal production prediction. The comparison results showed that bidirectional gated recurrent unit (Bi-GRU) is the optimal algorithm with the highest accuracy of 0.94. A surrogate model was established to simulate the inter-well connectivity of steam-assisted gravity drainage (SAGD) in the research area by utilizing the Bi-GRU algorithm. A global sensitivity analysis method, Fourier amplitude sensitivity testing (FAST), was introduced and combined with the surrogate model to explain the influence of the input variables on the output variables by quantitatively calculating the sensitivity of each variable. Quantitative results for the inter-well connectivity of SAGD were derived from the sensitivity analysis of the proposed method, which was effectively applied to typical linear patterns and five-spot patterns. Inter-well connectivity varied from 0.1 to 0.58 in test applications, and mutual corroboration with previous geological knowledge can further determine the distribution of the interlayer in the reservoir. The workflow proposed in this study provides a new direction for analyzing and inferring the inter-well connectivity of SAGD in Northeast China heavy oil reservoirs.

Design of pump-jet propulsor based on data-driven optimization method

Currently, the continuous change prediction of PM2.5 concentration is an air pollution research hotspot. Combining physical methods and deep learning models to divide the pollution process of PM2.5 into effective multiple types is necessary to achieve a reliable prediction of the PM2.5 value. Therefore, a candlestick chart sample generator was designed to generate the candlestick chart from the online PM2.5 continuous monitoring data of the Guilin monitoring station site. After these generated candlestick charts were analyzed through the Gaussian diffusion model, it was found that the characteristics of the physical transmission process of PM2.5 pollutants can be reflected. Based on a set three-day period, using the time linear convolution method, 2188 sets of candlestick chart data were obtained from the 2013–2018 PM2.5 concentration data. There existed 16 categories generated by unsupervised classification that met the established classification judgment standards. After the statistical analysis, it was found that the accuracy rate of the change trend of these classifications reached 99.68% during the next period. Using the candlestick chart data as the training dataset, the Visual Geometry Group (VGG) model, an improved convolutional neural network model, was used for the classification. The experimental results showed that the overall accuracy (OA) value of the candlestick chart combination classification was 96.19%, and the Kappa coefficient was 0.960. IN the VGG model, the overall accuracy was improved by 1.93%, on average, compared with the support vector machines (SVM), LeNet, and AlexNet models. According to the experimental results, using the VGG classification method to classify continuous pollution data in the form of candlestick charts can more comprehensively retain the characteristics of the physical pollution process and provide a classification basis for accurately predicting PM2.5 values. At the same time, the statistical feasibility of this method has been proved.

As the growth of Industry 4.0, online fault detection plays a crucial role in ensuring the manufacturing quality. Generally, the fault detection methods can be classified into model-based and data-driven methods. There are advantages/disadvantages between two methods. In this study, we integrated both methods in order to develop an efficient fault detection method for non-Gaussian industrial processes. The data-driven method, independent component analysis (ICA) is used to extract non-Gaussian information and dimensionality reduction. Meanwhile, the model-based method, generalized likelihood ratio (GLR) test is adopted as the charting statistic. The proposed ICA-GLR method has advantages of 1) detecting a wide range of process changes, 2) estimating the change points and 3) needless prior parameters to be specified by practitioner. The efficiency of the proposed ICA-GLR fault detection method will be verified via implementing one simulated non-Gaussian process and two real manufacturing processes: Tennessee Eastman process and semiconductor manufacturing process. Results demonstrate that the proposed ICA-GLR method has superior fault detectability when compared to traditional methods, such as principal component analysis and ICA.

Examination of load-deformation characteristics of long-span bridges in harsh natural environments based on real-time updating artificial neural network

An interpretable data-driven method for degradation prediction of proton exchange membrane fuel cells based on temporal fusion transformer and covariates

Mild cognitive impairment (MCI) is an important stage in Alzheimer's disease (AD) research, focusing on early pathogenic factors and mechanisms. Examining MCI patient subtypes and identifying their cognitive and neuropathological patterns as the disease progresses can enhance our understanding of the heterogeneous disease progression in the early stages of AD. However, few studies have thoroughly analyzed the subtypes of MCI, such as the cortical atrophy, and disease development characteristics of each subtype. In this study, 396 individuals with MCI, 228 cognitive normal (CN) participants, and 192 AD patients were selected from ADNI database, and a semi-supervised mixture expert algorithm (MOE) with multiple classification boundaries was constructed to define AD subtypes. Moreover, the subtypes of MCI were obtained by using the multivariate linear boundary mapping of support vector machine (SVM). Then, the gray matter atrophy regions and severity of each MCI subtype were analyzed and the features of each subtype in demography, pathology, cognition, and disease progression were explored combining the longitudinal data collected for 2 years and analyzed important factors that cause conversion of MCI were analyzed. Three MCI subtypes were defined by MOE algorithm, and the three subtypes exhibited their own features in cortical atrophy. Nearly one-third of patients diagnosed with MCI have almost no significant difference in cerebral cortex from the normal aging population, and their conversion rate to AD are the lowest. The subtype characterized by severe atrophy in temporal lobe and frontal lobe have a faster decline rate in many cognitive manifestations than the subtype featured with diffuse atrophy in the whole cortex. APOE ε4 is an important factor that cause the conversion of MCI to AD. It was proved through the data-driven method that MCI collected by ADNI baseline presented different subtype features. The characteristics and disease development trajectories among subtypes can help to improve the prediction of clinical progress in the future and also provide necessary clues to solve the classification accuracy of MCI.

Enthalpies of formation and reaction are important thermodynamic properties that have a crucial impact on the outcome of chemical transformations. Here we implement the calculation of enthalpies of formation with a general-purpose ANI-1ccx neural network atomistic potential. We demonstrate on a wide range of benchmark sets that both ANI-1ccx and our other general-purpose data-driven method AIQM1 approach the coveted chemical accuracy of 1 kcal/mol with the speed of semiempirical quantum mechanical methods (AIQM1) or faster (ANI-1ccx). It is remarkably achieved without specifically training the machine learning parts of ANI-1ccx or AIQM1 on formation enthalpies. Importantly, we show that these data-driven methods provide statistical means for uncertainty quantification of their predictions, which we use to detect and eliminate outliers and revise reference experimental data. Uncertainty quantification may also help in the systematic improvement of such data-driven methods.