Trained Long Short-term Memory Research Articles

Joint Image Processing with Learning-Driven Data Representation and Model Behavior for Non-Intrusive Anemia Diagnosis in Pediatric Patients

Anemia diagnosis is crucial for pediatric patients due to its impact on growth and development. Traditional methods, like blood tests, are effective but pose challenges, such as discomfort, infection risk, and frequent monitoring difficulties, underscoring the need for non-intrusive diagnostic methods. In light of this, this study proposes a novel method that combines image processing with learning-driven data representation and model behavior for non-intrusive anemia diagnosis in pediatric patients. The contributions of this study are threefold. First, it uses an image-processing pipeline to extract 181 features from 13 categories, with a feature-selection process identifying the most crucial data for learning. Second, a deep multilayered network based on long short-term memory (LSTM) is utilized to train a model for classifying images into anemic and non-anemic cases, where hyperparameters are optimized using Bayesian approaches. Third, the trained LSTM model is integrated as a layer into a learning model developed based on recurrent expansion rules, forming a part of a new deep network called a recurrent expansion network (RexNet). RexNet is designed to learn data representations akin to traditional deep-learning methods while also understanding the interaction between dependent and independent variables. The proposed approach is applied to three public datasets, namely conjunctival eye images, palmar images, and fingernail images of children aged up to 6 years. RexNet achieves an overall evaluation of 99.83 ± 0.02% across all classification metrics, demonstrating significant improvements in diagnostic results and generalization compared to LSTM networks and existing methods. This highlights RexNet’s potential as a promising alternative to traditional blood-based methods for non-intrusive anemia diagnosis.

Recurrent neural networks as virtual cavity pressure and temperature sensors in high-pressure die casting

High-pressure die casting (HPDC) is a permanent mold-based production technology that facilitates the casting of near net shape components from nonferrous alloys. The pressure and temperature conditions within the cavity impact the cast product quality during and after the conclusion of the die filling process. Die surface cavity sensors can deliver information describing the conditions at the die-casting interface. They are associated with high costs and limited service lifetimes below the achievable total cycle count of the die inserts and therefore ill-suited for industrial use cases. In this work, the suitability of long short-term memory (LSTM) recurrent neural networks (RNN) for substituting physical cavity temperature and pressure sensors virtually after the production ramp-up or at the end of the sensor service life is investigated. Training LSTMs with data of 233 casting cycles with different process parameters provides networks which are then applied to 99 further cycles. The prediction accuracy is investigated for different time interval lengths in the solidification and cooling phase. For longer time intervals, the cavity pressure prediction deteriorates, potentially due to a highly individual and hardly ascertainable buildup of casting distortion and internal stresses. Overall, however, the accuracy of the developed LSTMs is excellent for the cavity temperatures and good for the cavity pressures.

Corporate bankruptcy prediction based on the Adaboost algorithm for optimisation of long and short-term memory networks

In this study, a Long Short-Term Memory (LSTM) network was used to model and predict time series data, providing an innovative approach to corporate bankruptcy prediction. Subsequently, the predictions and real labels of the LSTM models were used to train the Adaboost algorithm to improve the accuracy and robustness of the models. Eventually, multiple trained LSTM models are combined into a more robust integrated model by adjusting the weights. It is worth mentioning that the integrated model achieves 95.57% prediction accuracy on the training set and 94.39% prediction accuracy on the test set, which indicates that the model has good prediction effect and generalisation ability. The method proposed in this study is of great significance, firstly, by combining LSTM and Adaboost algorithm, we not only improve the accurate prediction ability of corporate bankruptcy, but also enhance the ability of identifying anomalies. Second, by combining multiple LSTM models and adjusting the weights to form a more powerful integrated model, we effectively improve the overall prediction performance. This approach can provide financial institutions, investors, and government regulators with a more reliable and accurate tool for assessing corporate bankruptcy risk, which can help identify potential risks and take appropriate measures in a timely manner.

Converter Capacitor Temperature Estimation Based on Continued Training LSTM under Variable Load Conditions.

Capacitors are crucial components in power electronic converters, responsible for harmonic elimination, energy buffering, and voltage stabilization. However, they are also the most susceptible to damage due to their operational environment. Accurate temperature estimation of capacitors is essential for monitoring their condition and ensuring the reliability of the converter system. This paper presents a novel method for estimating the core temperature of capacitors using a long short-term memory (LSTM) algorithm. The approach incorporates a continued training mechanism to adapt to variable load conditions in converters. Experimental results demonstrate the proposed method's high accuracy and robustness, making it suitable for real-time capacitor temperature monitoring in practical applications.

Comparing ChatGPT And LSTM In Predicting Changes In Quarterly Financial Metrics

In the financial industry, the ability to predict financial metrics accurately and in a timely manner can significantly impact investment decisions, risk management, and strategic planning. In recent years, machine learning has emerged as a powerful tool for such predictions. This study aims to explore the heretofore underexplored predictive potential of ChatGPT by predicting positive or negative changes in quarterly financial metrics, such as revenue and sales numbers, using textual data from social media. The performance of ChatGPT is compared against Long Short-Term Memory (LSTM) neural network models developed as part of this research. The methodology involves preprocessing large datasets from Twitter concerning major companies such as Amazon, Google, and Tesla, training LSTM models, and prompt engineering for ChatGPT-based predictions. Initial findings indicate that LSTM models can predict quarterly financial metric changes with up to 87% accuracy, significantly outperforming ChatGPT, which achieves a maximum accuracy of 67%. Therefore, at the current time, ChatGPT cannot be considered a reliable predictive tool for changes in quarterly financial metrics.

Fast fluid–structure interaction simulation method based on deep learning flow field modeling

The rapid acquisition of high-fidelity flow field information is of great significance for engineering applications such as multi-field coupling. Current research in flow field modeling predominantly focuses on low Reynolds numbers and periodic flows, exhibiting weak generalization capabilities and notable issues with temporal inferring error accumulation. Therefore, we establish a reduced order model (ROM) based on Convolutional Auto-Encoder (CAE) and Long Short-Term Memory (LSTM) neural network and propose an unsteady flow field modeling method for the airfoil with a high Reynolds number and strong nonlinear characteristics. The attention mechanism and weak physical constraints are integrated into the model architecture to improve the modeling accuracy. A broadband excitation training strategy is proposed to overcome the error accumulation problem of long-term inferring. With only a small amount of latent codes, the relative error of the flow field reconstructed by CAE can be less than 5‰. By training LSTM with broadband excitation signals, stable dynamic evolution can be achieved in the time dimension. CAE-LSTM can accurately predict the forced response and complex limit cycle behavior of the airfoil in a wide range of amplitude and frequency under subsonic/transonic conditions. The relative errors of predicted variables and integral force are less than 1%. The fluid–structure interaction framework is built by coupling the ROM and motion equations of the structure. CAE-LSTM predicts the time series response of pitch displacement and moment coefficient at different reduced frequencies, which is in good agreement with computational fluid dynamics, and the simulation time savings exceed one order of magnitude.

Tool Wear Prediction Based on LSTM and Deep Residual Network

To improve the accuracy and efficiency of tool wear predictions, this study proposes a tool wear prediction model called LSTM_ResNet which is based on the long short-term memory (LSTM) network and the Residual Network (ResNet). The model utilizes LSTM layers for processing, where the first block and loop blocks serve as the core modules of the deep residual network. The model employs a series of methods including convolution, batch normalization (BN), and Rectified Linear Unit (ReLU) to enhance the model’s expression and prediction capabilities. The performance of the LSTM_ResNet model was evaluated using experimental data from the PHM2010 datasets and two different depths (64 and 128 layers), training both LSTM_ResNet models for 200 epochs. The 64-layer model’s root mean square error (RMSE) values are 3.36, 4.35, and 3.59, and the mean absolute error (MAE) values are 2.42, 2.85, and 2.21; using 128 layers, the RMSE values are 3.66, 3.99, and 3.77, and the MAE values are 2.49, 2.73, and 3.01. The results indicate that the 64-layer LSTM has smaller average errors, suggesting that compared to other common network structures, the LSTM_ResNet network has a higher performance. This research provides an effective solution for tool wear prediction and helps to improve the technical level of tool wear prediction in China.

Displacement estimation for a high-rise building during Super Typhoon Mangkhut based on field measurements and machine learning

Knowledge of displacement responses of high-rise buildings under harsh wind excitations is essential for their wind-resistant structural design. This paper develops a machine learning model named long short-term memory (LSTM) to estimate the displacements of a 420-m-high building during Super Typhoon Mangkhut based on available field measurements. The developed model is trained and validated using the field measurements on the building during typhoon events, and the performance of the model is assessed against several evaluation criteria. Then, the trained LSTM model is employed to estimate the displacements of the skyscraper during Mangkhut. The accuracy of the estimated displacements is validated in time and frequency domains. Moreover, the background and resonant components of the estimated displacements during the extreme windstorm are analyzed. This paper aims to provide valuable reference for the wind-resistant design of high-rise buildings in tropical cyclone-prone regions.

Optimizing Internet of Things Fog Computing: Through Lyapunov-Based Long Short-Term Memory Particle Swarm Optimization Algorithm for Energy Consumption Optimization.

In the era of continuous development in Internet of Things (IoT) technology, smart services are penetrating various facets of societal life, leading to a growing demand for interconnected devices. Many contemporary devices are no longer mere data producers but also consumers of data. As a result, massive amounts of data are transmitted to the cloud, but the latency generated in edge-to-cloud communication is unacceptable for many tasks. In response to this, this paper introduces a novel contribution-a layered computing network built on the principles of fog computing, accompanied by a newly devised algorithm designed to optimize user tasks and allocate computing resources within rechargeable networks. The proposed algorithm, a synergy of Lyapunov-based, dynamic Long Short-Term Memory (LSTM) networks, and Particle Swarm Optimization (PSO), allows for predictive task allocation. The fog servers dynamically train LSTM networks to effectively forecast the data features of user tasks, facilitating proper unload decisions based on task priorities. In response to the challenge of slower hardware upgrades in edge devices compared to user demands, the algorithm optimizes the utilization of low-power devices and addresses performance limitations. Additionally, this paper considers the unique characteristics of rechargeable networks, where computing nodes acquire energy through charging. Utilizing Lyapunov functions for dynamic resource control enables nodes with abundant resources to maximize their potential, significantly reducing energy consumption and enhancing overall performance. The simulation results demonstrate that our algorithm surpasses traditional methods in terms of energy efficiency and resource allocation optimization. Despite the limitations of prediction accuracy in Fog Servers (FS), the proposed results significantly promote overall performance. The proposed approach improves the efficiency and the user experience of Internet of Things systems in terms of latency and energy consumption.

A methodology for stress-strain behavior characterization and mixture optimization of recycled aggregate concrete based on machine learning

The characterization of stress-strain behaviors of recycled aggregate concrete (RAC) excluding physical properties of recycled aggregates (RAs) may result in an inaccurate prediction of mechanical responses in practical applications. In this study, a data-driven model using a refined long short-term memory (LSTM) network is established based on the Bayesian optimization algorithm, with the motivation to accurately predict the uniaxial compressive stress-strain behaviors of RAC, including the stress-strain relation, elastic modulus, peak stress, and the peak strain. Training and testing of the proposed model require the integration of the mixture content and the fundamental physical properties of RAs with the stress-strain relation of ordinary concrete featured by prominent sequential attributes. Upon a dataset containing 100 experimental samples from independent studies, covering a wide range of RA substitution rates, the superior prediction capability of the proposed LSTM network is demonstrated in comparison with the analytical results of three empirical mechanics-driven models. Finally, the trained LSTM network is further employed to optimize the mixture for RAC using the Bayesian optimization technique innovatively, to achieve a balance between the mechanical performance and requirement to the quality of RAs. This study constitutes a beneficial addition to the understanding of the RAs-dependent stress-strain behaviors of RAC from the perspective of the data paradigm and provides a practical tool for the optimal design of RAC materials and structures.

Validation set sampling strategies for predictive process monitoring

Previous studies investigating the efficacy of long short-term memory (LSTM) recurrent neural networks in predictive process monitoring and their ability to capture the underlying process structure have raised concerns about their limited ability to generalize to unseen behavior. Event logs often fail to capture the full spectrum of behavior permitted by the underlying processes. To overcome these challenges, this study introduces innovative validation set sampling strategies based on control-flow variant-based resampling. These strategies have undergone extensive evaluation to assess their impact on hyperparameter selection and early stopping, resulting in notable enhancements to the generalization capabilities of trained LSTM models. In addition, this study expands the experimental framework to enable accurate interpretation of underlying process models and provide valuable insights. By conducting experiments with event logs representing process models of varying complexities, this research elucidates the effectiveness of the proposed validation strategies. Furthermore, the extended framework facilitates investigations into the influence of event log completeness on the learning quality of predictive process models. The novel validation set sampling strategies proposed in this study facilitate the development of more effective and reliable predictive process models, ultimately bolstering generalization capabilities and improving the understanding of underlying process dynamics.

Traffic Flow Forecast using Time Series Analysis based on Machine Learning Algorithm

Time series forecasting plays a critical role in predicting future trends and making informed decisions across various domains. In this research paper, we present a multi-model approach to improve the accuracy and robustness of time series forecasting. Our methodology involves integrating two powerful models, Long Short-Term Memory (LSTM) and Seasonal Autoregressive Integrated Moving Average (SARIMA), to capture complex temporal patterns and seasonality in the data. We leverage real-world traffic data from different junctions and employ a comprehensive pipeline for data preprocessing, including data selection, resampling, and sequence generation. Our study focuses on four junctions, and for each junction, we individually train LSTM models, followed by SARIMA models for comparison. The LSTM models are designed to capture long-term dependencies in the time series data, while the SARIMA models are tailored to handle seasonality and autocorrelation. We implement a custom sequence generation process and split the dataset into training and testing sets for model evaluation. The LSTM models are trained using Pytorch, optimizing them for accurate short-term predictions. Our results show that the combination of LSTM and SARIMA models yields superior forecasting performance compared to using each model individually. We present a detailed analysis of the forecasted results, including Root Mean Square Error (RMSE) calculations and visualizations to demonstrate the effectiveness of our multi-model approach. This research contributes to the field of time series forecasting by showcasing the benefits of combining deep learning and classical statistical methods. The proposed approach provides a flexible and robust framework applicable to various time series prediction tasks, offering improved accuracy and reliability in forecasting future trends.

Memristive Neural Networks for Predicting Seizure Activity.

The biological part of the investigation. Young healthy outbred CD1 mice were used in our study. They were divided into two groups: control (n=6) and the group with induced chronic epileptiform activity (n=6). Local field potentials (LFP) were recorded from the hippocampus and medial entorhinal cortex of the mice of both groups to register neuronal activity. The LFP recordings were used for deep ANN training. Epileptiform activity in mice was modeled by intraperitoneal injection of pilocarpine (280 mg/kg). LFP were recorded in the awake mice a month after the induction of epileptiform activity.Mathematical part of the investigation. A deep long short-term memory (LSTM) ANN capable of predicting biological signals of neuronal activity in mice has been developed. The ANN implementation is based on memristive devices, which are described by the equations of the redox processes running in the memristive thin metal-oxide-metal films, e.g., Au/ZrO2(Y)/TiN/Ti and Au/SiO2(Y)/TiN/Ti. In order to train the developed ANN to predict epileptiform activity, a supervised learning algorithm was used, which allowed us to adjust the network parameters and train LSTM on the described recordings of neuronal activity. After training on the LFP recordings from the hippocampus and medial entorhinal cortex of the mice with chronic epileptiform activity, the proposed deep ANN has demonstrated high values of evaluation metric (root-mean-square error, RMSE) and successfully predicted epileptiform activity shortly before its occurrence (40 ms). The results of the numerical experiments have shown that the RMSE value of 0.019 was reached, which indicates the efficacy of proposed approach. The accuracy of epileptiform activity prediction 40 ms before its occurrence is a significant result and shows the potential of the developed neural network architecture. The proposed deep ANN can be used to predict pathological neuronal activity including epileptic seizure (focal) activity in mice before its actual occurrence. Besides, it can be applied for building a long-term prognosis of the disease course based on the LFP data. Thus, the proposed ANN based on memristive devices represents a novel approach to the prediction and analysis of pathological neuronal activity possessing a potential for improving the diagnosis and prognostication of epileptic seizures and other diseases associated with neuronal activity.

Accurate detection of vehicle, pedestrian, cyclist and wheelchair from roadside light detection and ranging sensors

Accurate detection plays a critical role in improving the safety situation of vulnerable road users. This study extends infrastructure-based LiDAR application to all three major vulnerable road user groups including pedestrians, cyclists, and wheelchair users. Two critical problems for accurate detection of small-sized road users are scanning angle variability and feature fluctuation. To address these issues, a feature-based classification method combined with prior LiDAR trajectory information is developed. Effective dimension-related features are proposed and five classifiers including artificial neural network (ANN), random forest (RF), adaptive boosting (AdaBoost), random under-sampling boosting (RUSBoost), and long short-term memory (LSTM) are tested with a novel feature engineering process. A total of seven features are selected from the point cloud of clusters for vehicle/pedestrian/cyclist/wheelchair classification. By updating these significant features based on prior information of the entire trajectory, the performance of road user classification (imbalanced datasets) has been significantly improved. Experimental study is conducted to examine the recall rate, F1-score, and AUC of vehicles, pedestrians, cyclists, and wheelchairs before and after integration with prior trajectory information. The result shows the trained AdaBoost, RUSBoost, and LSTM classifiers with prior trajectory information can achieve recall/F1-score/AUC: (1) Low traffic volumes – vehicles (100%/99.96%/99.96%), pedestrians (99.96%/99.96%/99.97%), cyclists (99.74%/99.45%/99.67%), and wheelchairs (99.22%/99.68%/99.01%) and (2) Moderate traffic volumes – vehicles (99.39%/99.44%/99.69%), pedestrians (98.33%/97.99%/98.64%), and cyclists (95.41%/94.29%/94.40%), using 32-laser LiDAR sensors (10 Hz).

Predict the priority of end-users’ maintenance requests and the required technical staff through LSTM and Bi-LSTM recurrent neural networks

PurposeThis paper aims to develop predictive methods, based on recurrent neural networks, useful to support facility managers in building maintenance tasks, by collecting information coming from a computerized maintenance management system (CMMS).Design/methodology/approachThis study applies data-driven and text-mining approaches to a CMMS data set comprising more than 14,500 end-users’ requests for corrective maintenance actions, collected over 14 months. Unidirectional long short-term memory (LSTM) and bidirectional LSTM (Bi-LSTM) recurrent neural networks are trained to predict the priority of each maintenance request and the related technical staff assignment. The data set is also used to depict an overview of corrective maintenance needs and related performances and to verify the most relevant elements in the building and how the current facility management (FM) relates to the requests.FindingsThe study shows that LSTM and Bi-LSTM recurrent neural networks can properly recognize the words contained in the requests, thus correctly and automatically assigning the priority and predicting the technical staff to assign for each end-user’s maintenance request. The obtained global accuracy is very high, reaching 93.3% for priority identification and 96.7% for technical staff assignment. Results also show the main critical building elements for maintenance requests and the related intervention timings.Research limitations/implicationsThis work shows that LSTM and Bi-LSTM recurrent neural networks can automate the assignment process of end-users’ maintenance requests if trained with historical CMMS data. Results are promising; however, the trained LSTM and Bi-LSTM RNN can be applied only to different hospitals adopting similar categorization.Practical implicationsThe data-driven and text-mining approaches can be integrated into the CMMS to support corrective maintenance management by facilities management contractors, i.e. to properly and timely identify the actions to be carried out and the technical staff to assign.Social implicationsThe improvement of the maintenance of the health-care system is a key component of improving health service delivery. This work shows how to reduce health-care service interruptions due to maintenance needs through machine learning methods.Originality/valueThis study develops original methods and tools easily integrable into IT workflow systems (i.e. CMMS) in the FM field.

Using Time-Series Generative Adversarial Networks to Synthesize Sensing Data for Pest Incidence Forecasting on Sustainable Agriculture

A sufficient amount of data is crucial for high-performance and accurate trend prediction. However, it is difficult and time-consuming to collect agricultural data over long periods of time; the consequence of such difficulty is datasets that are characterized by missing data. In this study we use a time-series generative adversarial network (TimeGAN) to synthesize multivariate agricultural sensing data and train RNN (Recurrent Neural Network), LSTM (Long Short-Term Memory), and GRU (Gated Recurrent Unit) neural network prediction models on the original and generated data to predict future pest populations. After our experiment, the data generated using TimeGAN and the original data have the smallest EC value in the GRU model, which is 9.86. The results show that the generative model effectively synthesizes multivariate agricultural sensing data and can be used to make up for the lack of actual data. The pest prediction model trained on synthetic data using time-series data generation yields results that are similar to that of the model trained on actual data. Accurate prediction of pest populations would represent a breakthrough in allowing for accurate and timely pest control.

Application of LSTM Approach for Predicting the Fission Swelling Behavior within a CERCER Composite Fuel

Irradiation-induced swelling plays a key role in determining fuel performance. Due to their high cost and time demands, experimental research methods are ineffective. Knowledge-based multiscale simulations are also constrained by the loss of trustworthy theoretical underpinnings. This work presents a new trial of integrating knowledge-based finite element analysis (FEA) with a data-driven deep learning framework, to predict the hydrostatic-pressure–temperature dependent fission swelling behavior within a CERCER composite fuel. We employed the long short-term memory (LSTM) deep learning network to mimic the history-dependent behaviors. Training of the LSTM is achieved by processing the sequential order of the inputs to do the forecasting; the input features are fission rate, fission density, temperature, and hydrostatic pressure. We performed the model training based on a leveraged dataset of 8000 combinations of a wide range of input states and state evaluations that were generated by high-fidelity simulations. When replicating the swelling plots, the trained LSTM deep learning model exhibits outstanding prediction effectiveness. For various input variables, the model successfully pinpoints when recrystallization first occurs. The preliminary study for model interpretation suggests providing quantified insights into how those features affect solid and gaseous portions of swelling. The study demonstrates the efficacy of combining data-driven and knowledge-based modeling techniques to assess irradiation-induced fuel performance and enhance future design.

Prediction of transonic wing buffet pressure based on deep learning

In the present study, a deep learning approach based on a long short-term memory (LSTM) neural network is applied for the prediction of transonic wing buffet pressure. In particular, fluctuations in surface pressure over a certain time period as measured by a piezoresistive pressure sensor, are considered. As a test case, the generic XRF-1 aircraft configuration developed by Airbus is used. The XRF-1 configuration has been investigated at different transonic buffet conditions in the European Transonic Wind tunnel (ETW). During the ETW test campaign, sensor data has been obtained at different local span—and chordwise positions on the lower and upper surface of the wing and the horizontal tail plane. For the training of the neural network, a buffet flow condition with a fixed angle of attack alpha and a fixed sensor position on the upper wing surface is considered. Subsequent, the trained network is applied towards different angles of attack and sensor positions considering the flow condition applied for training the network. As a final step, the trained LSTM neural network is used for the prediction of pressure data at a flow condition different from the flow condition considered for training. By comparing the results of the wind tunnel experiment with the results obtained by the neural network, a good agreement is indicated.

How can we benefit from regime information to make more effective use of long short-term memory (LSTM) runoff models?

Abstract. To date, long short-term memory (LSTM) networks have been successfully applied to a key problem in hydrology: the prediction of runoff. Unlike traditional conceptual models, LSTM models are built on concepts that avoid the need for our knowledge of hydrology to be formally encoded into the model. The question, then, is how we can still make use of our domain knowledge and traditional practices, not to build the LSTM models themselves, as we do for conceptual models, but to use them more effectively. In the present paper, we adopt this approach, investigating how we can use information concerning the hydrologic characteristics of catchments for LSTM runoff models. In this first application of LSTM in a French context, we use 361 gauged catchments with very diverse hydrologic conditions from across France. The catchments have long time series of at least 30 years. Our main directions for investigation include (a) the relationship between LSTM performance and the length of the LSTM input sequence within different hydrologic regimes, (b) the importance of the hydrologic homogeneity of catchments when training LSTMs on a group of catchments, and (c) the interconnected influence of the local tuning of the two important LSTM hyperparameters, namely the length of the input sequence and the hidden unit size, on the performance of group-trained LSTMs. We present a classification built on three indices taken from the runoff, precipitation, and temperature regimes. We use this classification as our measure of homogeneity: catchments within the same regime are assumed to be hydrologically homogeneous. We train LSTMs on individual catchments (local-level training), on catchments within the same regime (regime-level training), and on the entire sample (national-level training). We benchmark local LSTMs using the GR4J conceptual model, which is able to represent the water gains/losses in a catchment. We show that LSTM performance has the highest sensitivity to the length of the input sequence in the Uniform and Nival regimes, where the dominant hydrologic process of the regime has clear long-term dynamics; thus, long input sequences should be chosen in these cases. In other regimes, this level of sensitivity is not found. Moreover, in some regimes, almost no sensitivity is observed. Therefore, the size of the input sequence in these regimes does not need to be large. Overall, our homogeneous regime-level training slightly outperforms our heterogeneous national-level training. This shows that the same level of data adequacy with respect to the complexity of representation(s) to be learned is achieved in both levels of training. We do not, however, exclude a potential role of the regime-informed property of our national LSTMs, which use previous classification variables as static attributes. Last but not least, we demonstrate that the local selection of the two important LSTM hyperparameters (the length of the input sequence and the hidden unit size) combined with national-level training can lead to the best runoff prediction performance.

Physics-informed long short-term memory networks for response prediction of a wind-excited flexible structure

Slender and flexible infrastructures such as sign supports, cantilever traffic signal structures and high mast lighting towers are sensitive to wind force and were reported to have fatigue-related issues due to the large-amplitude vibrations throughout thier life. Simulating wind-induced structural response can be an important step to evaluate their fatigue life and reliability. However, wind simulations are usually quite complicated. A comprehensive wind force model was usually developed by conducting multiple wind tunnel tests. However, due to the high cost of wind tunnel tests and the limitation of a wind tunnel, aerodynamic and aeroelastic coefficients were usually extracted only at certain wind speeds and wind directions. Interpolation or extrapolation methods were commonly used when coefficients were not available, which makes the simulation result questionable. In this study, a methodology was proposed to simulate wind-induced structural response with lower costs. The proposed method uses monitoring data in the field to develop long short-term memory (LSTM) networks. In training LSTM networks, only the monitoring data in regular wind condition was used. However, the trained LSTM network can still predict the wind-induced response in high and extreme wind conditions observed during the monitoring of the structure. The proposed method can be useful when simulating wind-induced structural response in a wide range of wind speeds and can be widely used on other structures suspected of having fatigue damage due to wind-induced vibrations.