Long Short-term Memory Method Research Articles

Real-Time Multipath Mitigation in Multi-GNSS Short Baseline Positioning via CNN-LSTM Method

Multipath is the main systematic error of the Global Navigation Satellite System (GNSS) short baseline positioning. Multipath cannot be eliminated by the double-differenced technique and is difficult to parameterize, which severely restrict the high-precision GNSS positioning application. Based on the spatiotemporal repeatability of multipath, the sidereal filtering in coordinate-domain (SF-CD), the sidereal filtering in observation-domain (SF-OD), and the multipath hemispherical map (MHM) can be used to mitigate the multipath in real-time. However, the multipath model with large matrix for multi-GNSS multipath mitigation is difficult to achieve lightweight calculation and the SF-CD cannot be applied to mitigate the multi-GNSS multipath. In this paper, we propose a new multipath mitigation strategy in the coordinate-domain that shakes off the formation mechanism of multipath, a CNN (convolutional neural network)-LSTM (long short-term memory) method is used to mine the deep multipath features in GNSS coordinate series. Furthermore, multipath will be mitigated in real-time by constantly predicting the value of the next epoch. The experimental results show that the CNN-LSTM effectively mitigates the multi-GNSS multipath. The method can reduce the average RMS (root-mean square) of multi-GNSS positioning errors in the east, north, and vertical directions by 62.3%, 70.8%, and 66.0%. Moreover, comparing with the SF-CD, SF-OD, and MHM, CNN-LSTM can more effectively mitigate the effects of the GPS multipath, and the ability of multipath mitigation is almost not affected over time.

An Effective Machine Learning Scheme to Analyze and Predict the Concentration of Persistent Pollutants in the Great Lakes

Persistent organic pollutants (POPs) are highly toxic and difficult to degrade in the natural ecology, which has a severe negative impact on the ecological environment. Quantifying changes in the concentrations of persistent organic pollutants in the Great Lakes is challenging work. Machine learning (ML) methods are potent predictors that have recently achieved impressive performance on time series tasks. ARIMA model, Linear Regression methods, XGBoost algorithm, and Long Short-Term Memory (LSTM) are commonly used for estimating time-series changes. Traditionally Mean Squared Error (MSE) and Root Mean Squared Error (RMSE) have been standard criteria to measure the error between the actual value and predicted value; however, Euclidean distance (ED) cannot effectively calculate the similarity between two-time series. We proposed an alternative criterion called Penalty Dynamic Time Wrapping (Penalty-DTW) based on Dynamic Time Wrapping (DTW). It can accurately measure the difference between the actual value and the predicted value. We study the benefits of Penalty-DTW vs. ED under the above ML algorithms. Further, considering the machine learning algorithm’s uncertainty, we proposed combining LSTM and deep ensemble methods to quantify algorithms uncertainty and make a confident prediction. We find improved LSTM model outperformed other predictive power models by comparing pollutant concentration prediction. The prediction results show that the concentration of pollutants has a stable downward trend in recent years. Simultaneously, we found that pollutants’ concentration correlates with seasons, which positively guides environmental pollution control in the Great Lakes.

Comparing the Performance of Deep Learning Methods to Predict Companies’ Financial Failure

One of the most crucial problems in the field of business is financial forecasting. Many companies are interested in forecasting their incoming financial status in order to adapt to the current financial and business environment to avoid bankruptcy. In this work, due to the effectiveness of Deep Learning methods with respect to classification tasks, we compare the performance of three well-known Deep Learning methods (Long-Short Term Memory, Deep Belief Network and Multilayer Perceptron model of 6 layers) with three bagging ensemble classifiers (Random Forest, Support Vector Machine and K-Nearest Neighbor) and two boosting ensemble classifiers (Adaptive Boosting and Extreme Gradient Boosting) in companies' financial failure prediction. Because of the inherent nature of the problem addressed, three extremely imbalanced datasets of Spanish, Taiwanese and Polish companies' data have been considered in this study. Thus, five oversampling balancing techniques, two hybrid balancing techniques (oversampling-undersampling) and one clustering-based balancing technique have been applied to avoid data inconsistency problem. Considering the real financial data complexity level and type, the results show that the Multilayer Perceptron model of 6 layers, in conjunction with SMOTE-ENN balancing method, yielded the best performance according to the accuracy, recall and type II error metrics. In addition, Long-Short Term Memory and ensemble methods obtained also very good results, outperforming several classifiers used in previous studies with the same datasets.

Prediction of Geological Parameters during Tunneling by Time Series Analysis on In Situ Data.

A tunnel boring machine (TBM) is a type of heavy load equipment that is widely used in underground tunnel construction. The geological conditions in the tunneling process are decisive factors that directly affect the control of construction equipment. Because TBM tunneling always takes place underground, the acquisition of geological information has become a key issue in this field. This study focused on the internal relationships between the sequential nature of tunnel in situ data and the continuous interaction between equipment and geology and introduced the long short-term memory (LSTM) time series neural network method for processing in situ data. A method for predicting the geological parameters in advance based on TBM real-time state monitoring data is proposed. The proposed method was applied to a tunnel project in China, and the R2 of the prediction results for five geological parameters are all higher than 0.98. The performance of the LSTM was compared with that of an artificial neural network (ANN). The prediction accuracy of the LSTM was significantly higher compared with that of the ANN, and the generalization and robustness of LSTM are also better than those of ANN, which indicates that the proposed LSTM method could extract the sequence properties of the in situ data. The rule of equipment-geology interaction was reflected by increasing the memory structure of the model through the introduction of the “gate” concept, and the accurate prediction of geological parameters during tunneling was realized. Additionally, the influence of time window and distance of prediction on the model is discussed. The proposed method provides a new approach toward obtaining geological information during TBM construction and also provides a certain reference for the effective analysis of the in situ data with sequence properties.

A new hybrid model for wind speed forecasting combining long short-term memory neural network, decomposition methods and grey wolf optimizer

Reliable and accurate wind speed forecasting (WSF) is fundamental for efficient exploitation of wind power. In particular, high accuracy short-term WSF (ST-WSF) has a significant impact on the efficiency of wind power generation systems. Due to the non-stationarity and stochasticity of the wind speed (WS), a single model is often not sufficient in practice for the accurate estimation of the WS. Hybrid models are being proposed to overcome the limitations of single models and increase the WS forecasting performance. In this paper, a new hybrid WSF model is developed based on long short-term memory (LSTM) network and decomposition methods with grey wolf optimizer (GWO). In the pre-processing stage, the missing data is filled by the weighted moving average (WMA) method, the WS time series (WSTS) data are smoothed by WMA filtering and the smoothed data are used as model input after Z-score normalization. The forecasting model is formed by the combination of a single model, a decomposition method and an advanced optimization algorithm. Successively, the hybrid WSF model is developed by combining the LSTM and decomposition methods, and optimizing the intrinsic mode function (IMF) estimated outputs with a grey wolf optimizer (GWO). The developed non-linear hybrid model is utilized on the data collected from five wind farms in the Marmara region, Turkey. The obtained experimental results indicate that the proposed combined model can capture non-linear characteristics of WSTS, achieving better forecasting performance than single forecasting models, in terms of accuracy.

Application of LSTM approach for modelling stress–strain behaviour of soil

This paper presents a new trial to reproduce soil stress–strain behaviour by adapting a long short-term memory (LSTM) deep learning method. LSTM is an approach that employs time sequence data to predict future occurrences, and it can be used to consider the stress history of soil behaviour. The proposed LSTM method includes the following three steps: data preparation, architecture determination, and optimisation. The capacity of the adapted LSTM method is compared with that of feedforward and feedback neural networks using a new numerical benchmark dataset. The performance of the proposed LSTM method is verified through a dataset collected from laboratory tests. The results indicate that the LSTM deep-learning method outperforms the feed forward and feedback neural networks based on both accuracy and the convergence rate when reproducing the soil’s stress–strain behaviour. One new phenomenon referred to as “bias at low stress levels”, which was not noticed before, is first discovered and discussed for all neural network-based methods.

Detection of electricity theft using data processing and LSTM method in distribution systems

Electricity theft is a big problem faced by all energy distribution services and continues to rising. Therefore, studies on electricity theft detection techniques have increased in recent years. Unsuitable calibration and illegal calibration of energy meters during production may cause non-technical losses. Non-technical losses have been a major concern for the resulting security risks and the immeasurable loss of income. In most of the meter tampered locations, damaged meter terminals and/or illegal applications cannot be distinguishable during checking. In fact, electric distribution companies will never be able to eliminate electricity theft. But it is possible to take measure to detect, prevent and reduce it. In this paper, we developed by using deep learning methods on real daily electricity consumption data (Electricity consumption dataset of State Grid Corporation of China). Data reduction has been made by developing a new method to make the dataset more usable and to extract meaningful results. A Long Short-Term Memory (LSTM) based deep learning method has been developed for the dataset to be able to recognize the actual daily electricity consumption data of 2016. In order to evaluate the performance of the proposed method, the accuracy, prediction and recall metric was used by considering the five cross-fold technique. Performance of the proposed methods were found to be better than previously reported results.

Machine learning for demand forecasting in the physical internet: a case study of agricultural products in Thailand

Supply chains are complex, stochastic systems. Nowadays, logistics managers face two main problems: increasingly diverse and variable customer demand that is difficult to predict. Classical forecasting methods implemented in many business units have limitations with the fluctuating demand and the complexity of fully connected supply chains. Machine Learning methods have been proposed to improve prediction. In this paper, a Long Short-Term Memory (LSTM) is proposed for demand forecasting in a physical internet supply chain network. A hybrid genetic algorithm and scatter search are proposed to automate tuning of the LSTM hyperparameters. To assess the performance of the proposed method, a real-case study on agricultural products in a supply chain in Thailand was considered. Accuracy and coefficient of determination were the key performance indicators used to compare the performance of the proposed method with other supervised learnings: ARIMAX, Support Vector Regression, and Multiple Linear Regression. The results prove the better forecasting efficiency of the LSTM method with continuous fluctuating demand, whereas the others offer greater performance with less varied demand. The performance of hybrid metaheuristics is higher than with trial-and-error. Finally, the results of forecasting model are effective in transportation and holding costs in the distribution process of the Physical Internet.

Intent Detection Problem Solving via Automatic DNN Hyperparameter Optimization

Accurate intent detection-based chatbots are usually trained on larger datasets that are not available for some languages. Seeking the most accurate models, three English benchmark datasets that were human-translated into four morphologically complex languages (i.e., Estonian, Latvian, Lithuanian, Russian) were used. Two types of word embeddings (fastText and BERT), three types of deep neural network (DNN) classifiers (convolutional neural network (CNN); long short-term memory method (LSTM), and bidirectional LSTM (BiLSTM)), different DNN architectures (shallower and deeper), and various DNN hyperparameter values were investigated. DNN architecture and hyperparameter values were optimized automatically using the Bayesian method and random search. On three datasets of 2/5/8 intents for English, Estonian, Latvian, Lithuanian, and Russian languages, accuracies of 0.991/0.890/0.712, 0.972/0.890/0.644, 1.000/0.890/0.644, 0.981/0.872/0.712, and 0.972/0.881/0.661 were achieved, respectively. The BERT multilingual vectorization with the CNN classifier was proven to be a good choice for all datasets for all languages. Moreover, in the majority of models, the same set of optimal hyperparameter values was determined. The results obtained in this research were also compared with the previously reported values (where hyperparameter values of DNN models were selected by an expert). This comparison revealed that automatically optimized models are competitive or even more accurate when created with larger training datasets.

An Enhanced Stacked LSTM Method With No Random Initialization for Malware Threat Hunting in Safety and Time-Critical Systems

Malware detection is an increasingly important operational focus in cyber security, particularly, given the fast pace of such threats (e.g., new malware variants introduced every day). In recent years, there has been increased interest in exploring the use of machine learning techniques in automating and enhancing the effectiveness of malware detection and analysis. In this paper, we present a deep recurrent neural network solution as a stacked long short-term memory (LSTM) with a pre-training as a regularization method to avoid random network initialization. In our proposed approach, we use global and short dependencies of the inputs. With pre-training, we avoid random initialization and are able to improve the accuracy and robustness of malware threat hunting. The proposed method speeds up the convergence, in comparison to the stacked LSTM, by reducing the length of malware OpCode or bytecode sequences. Hence, the complexity of our final method is reduced. This leads to better accuracy , higher Mattews Correlation Coefficients (MCC) , and Area Under the Curve (AUC) in comparison to a standard LSTM with similar detection time. Our proposed method can be applied in real-time malware threat hunting, particularly, for safety critical systems, such as electronic health or Internet of Battlefield / Military of Things, where poor convergence of the model could lead to catastrophic consequences. We evaluate the effectiveness of our proposed method on Windows, Ransomware, Internet of Things (IoT), and Android malware datasets using both static and dynamic analysis. For the IoT malware detection, we also present a comparative summary of the performance on an IoT-specific dataset of our proposed method and the standard stacked LSTM method. More specifically, our proposed method achieves an accuracy of 99.1% in detecting IoT malware samples, with AUC of 0.985 and MCC of 0.95; thus, outperforming standard LSTM-based methods in these key metrics.

An Enhanced LSTM Method to Improve the Accuracy of the Business Process Prediction

Prediction of the process behavior plays a key role in business process management. This research benefits from recent development in the field of deep learning to predict the next event in business processes. The proposed method uses Long Short-Term Memory (LSTM) as a promising architecture of recurrent neural networks. This architecture is implemented using a number of configurations with the aim of investigating how each of them affects the performance of the prediction models. In order to build and evaluate our prediction models, we used two publicly available datasets (BPI 2012 and BPI 2017). After developing 300 prediction models, the results indicated that the proposed method outperforms the state-of-the-art methods in terms of precision. The best result in terms of Accuracy (0.907) was achieved through “one-hidden” layer LSTM architecture and by using “Big” configuration in the absence of “feedback”.

De-identifying free text of Japanese electronic health records

BackgroundRecently, more electronic data sources are becoming available in the healthcare domain. Electronic health records (EHRs), with their vast amounts of potentially available data, can greatly improve healthcare. Although EHR de-identification is necessary to protect personal information, automatic de-identification of Japanese language EHRs has not been studied sufficiently. This study was conducted to raise de-identification performance for Japanese EHRs through classic machine learning, deep learning, and rule-based methods, depending on the dataset.ResultsUsing three datasets, we implemented de-identification systems for Japanese EHRs and compared the de-identification performances found for rule-based, Conditional Random Fields (CRF), and Long-Short Term Memory (LSTM)-based methods. Gold standard tags for de-identification are annotated manually for age, hospital, person, sex, and time. We used different combinations of our datasets to train and evaluate our three methods. Our best F1-scores were 84.23, 68.19, and 81.67 points, respectively, for evaluations of the MedNLP dataset, a dummy EHR dataset that was virtually written by a medical doctor, and a Pathology Report dataset. Our LSTM-based method was the best performing, except for the MedNLP dataset. The rule-based method was best for the MedNLP dataset. The LSTM-based method achieved a good score of 83.07 points for this MedNLP dataset, which differs by 1.16 points from the best score obtained using the rule-based method. Results suggest that LSTM adapted well to different characteristics of our datasets. Our LSTM-based method performed better than our CRF-based method, yielding a 7.41 point F1-score, when applied to our Pathology Report dataset. This report is the first of study applying this LSTM-based method to any de-identification task of a Japanese EHR.ConclusionsOur LSTM-based machine learning method was able to extract named entities to be de-identified with better performance, in general, than that of our rule-based methods. However, machine learning methods are inadequate for processing expressions with low occurrence. Our future work will specifically examine the combination of LSTM and rule-based methods to achieve better performance.Our currently achieved level of performance is sufficiently higher than that of publicly available Japanese de-identification tools. Therefore, our system will be applied to actual de-identification tasks in hospitals.

Restoration of Missing Pressures in a Gas Well Using Recurrent Neural Networks with Long Short-Term Memory Cells

This study proposes a data-driven method based on recurrent neural networks (RNNs) with long short-term memory (LSTM) cells for restoring missing pressure data from a gas production well. Pressure data recorded by gauges installed at the bottom hole and wellhead of a production well often contain abnormal or missing values as a result of gauge malfunctions, noise, outliers, and operational instability. RNNs employing LSTM cells to prevent long-term memory loss have been widely used to predict time series data. In this study, an RNN with the LSTM method was used to restore abnormal or missing wellhead and bottom-hole pressures in three intervals within a production sequence of more than eight years in duration. The pressure restoration was performed using various input features for RNNs with LSTM models based on the characteristics of the available data. It was carried out through three sequential processes and the results were acceptable with a mean absolute percentage error no more than 5.18%. The reliability of the proposed method was verified through a comparison with the results of a physical model.

Prediction Model of Suspension Density in the Dense Medium Separation System Based on LSTM

In the dense medium separation system of coal preparation plant, the fluctuation of raw coal ash and lag of suspension density adjustment often causes the instability of product quality. To solve this problem, this study established a suspension density prediction model for the dense medium separation system based on Long Short-Term Memory (LSTM). First, the historical data in the dense medium separation system of a coal preparation plant were collected and preprocessed. Moving average and cubic exponential smoothing methods were used to replace abnormal data and to fill in the missing data, respectively. Second, a LSTM network was used to construct the density prediction model, and the optimal number of time steps, hidden layers, and nodes was determined. Finally, the model was employed on a testing set for prediction, and a Back-Propagation (BP) network without a time series was used for comparison. Root Mean Squared Error (RMSE) were the minimum when the number of the hidden layers, nodes, and time steps was 6, 12, and 5, respectively. In this case, the RMSE and Mean Absolute Percent Error (MAPE) of the LSTM method were 0.009 and 0.007, respectively, while those of the BP method were 0.019 and 0.015, respectively. Therefore, the model established using LSTM can be used to accurately predict the suspension density of the dense medium separation system.

Software defect prediction via LSTM

Software quality plays an important role in the software lifecycle. Traditional software defect prediction approaches mainly focused on using hand-crafted features to detect defects. However, like human languages, programming languages contain rich semantic and structural information, and the cause of defective code is closely related to its context. Failing to catch this significant information, the performance of traditional approaches is far from satisfactory. In this study, the authors leveraged a long short-term memory (LSTM) network to automatically learn the semantic and contextual features from the source code. Specifically, they first extract the program's Abstract Syntax Trees (ASTs), which is made up of AST nodes, and then evaluate what and how much information they can preserve for several node types. They traverse the AST of each file and fed them into the LSTM network to automatically the semantic and contextual features of the program, which is then used to determine whether the file is defective. Experimental results on several opensource projects showed that the proposed LSTM method is superior to the state-of-the-art methods.

Probabilistic Data-Driven Prognostic Methodology for Proton Exchange Membrane Fuel Cells


 
 
 Hydrogen fuel cells, particularly proton exchange membrane fuel cells (PEMFC), are promising, robust, clean energy sources. However, their high cost and short lifespan under dynamic loads impedes their widespread usage. Accurate and real-time prognostics, especially remaining useful time (RUL) estimation, can help ameliorate the commercial viability of PEMFCs. Data-driven methods are increasingly considered for RUL estimation. This paper looks at two such methods – Gaussian Process Regression (GPR) and Long-Short Term Memory (LSTM) Networks and assesses them in terms of accuracy and suitability for real-time applications when tested against the IEEE PHM Challenge 2014 data set. Gaussian Process Regression is a non-parametric kernel method. LSTM, on the other hand, is a recurrent neural network based architecture that is effective at detecting both long term and short term trends in time series predictions. For the cases investigated here, the results derived using LSTM are more accurate, especially since they effectively capture long term trends. However, GPR assigns a probability to its prediction - a desirable aspect in a real-time setting so that corrective action can be applied appropriately. The paper then proposes the use of a variant of these methods - Gaussian Process-Long Short Term Memory Network (GP-LSTM) as an alternative that combines the higher accuracies of the LSTM method and the probabilistic output from GPR. The results attained using GP-LSTM are close in accuracy to the LSTM results and have a probability associated with them, making them suitable for real-time applications. The effectiveness of GP-LSTM is further proven using a dynamic data set and strategies are suggested to appropriately apply GP-LSTM to real-world scenarios.
 
 

Monitoring of volcanic ash cloud from heterogeneous data using feature fusion and convolutional neural networks–long short-term memory

Heterogeneous data have become a key issue restricting the monitoring accuracy of volcanic ash cloud and rapid application of remote sensing. In view of the characteristics of classification and heterogeneous data in volcanic ash cloud monitoring, a monitoring method of volcanic ash cloud using feature fusion, convolutional neural networks (CNN) and long short-term memory (LSTM) (FF–CNN–LSTM) is presented in this paper. In this method, firstly, the target features in image are identified by CNN and the time sequence information in text is extracted by LSTM. Secondly, the mapping relationship between text information and image features is learned by fusing image target and text high-level features, and then the identification and diffusion of volcanic ash cloud from heterogeneous data containing only image and text were performed. Finally, the presented FF–CNN–LSTM method is tested by the simulation experiment and the true heterogeneous data of Enta volcanic ash cloud case. The experimental results show that compared with the single CNN, LSTM and the simple combination between CNN and LSTM (CNN–LSTM), the presented FF–CNN–LSTM method in this paper can be fitted with less training steps and has high accuracy and low loss rate; the obtained distribution of volcanic ash cloud is clear and intuitive and has the characteristics with fewer model parameters, simple calculation and high accuracy. It also shows the feasibility and effectiveness of the presented method in volcanic ash cloud monitoring to some extent. The results reveal that the presented method can potentially contribute the monitoring of volcanic ash cloud and disaster prevention and mitigation.

직류 나노그리드에서 신재생에너지 출력예측기반의 전기자동차 최적 스케줄링 기법

This paper focuses on optimal operation of a DC nanogrid that includes multiple electric vehicles (EV) and photovoltaic (PV) generators. The operator of DC nanogrid optimizes EV charging and discharging schedule to maximize economic benefits by charging more EVs and utilizing more electric power generated by PV while limiting the power imports from the grid. This paper employs the sophisticated forecasting algorithms for solar PV output and load demand in DC nanogrid using RNN(Recurrent Neural Network)-LSTM(Long Short-Term Memory) method. Then, we find the optimal EV charging and discharging schedule with mixed-integer linear programming (MILP) considering various operating concerns for EV owners and the DC nanogrid operator. The details on the overall operation algorithm is described by the definition of objective function as well as designs on equality and inequality constraints. The performance of the proposed method is verified by simulation studies considering multiple EVs and various PV outputs.

FasTag: Automatic text classification of unstructured medical narratives.

Unstructured clinical narratives are continuously being recorded as part of delivery of care in electronic health records, and dedicated tagging staff spend considerable effort manually assigning clinical codes for billing purposes. Despite these efforts, however, label availability and accuracy are both suboptimal. In this retrospective study, we aimed to automate the assignment of top-level International Classification of Diseases version 9 (ICD-9) codes to clinical records from human and veterinary data stores using minimal manual labor and feature curation. Automating top-level annotations could in turn enable rapid cohort identification, especially in a veterinary setting. To this end, we trained long short-term memory (LSTM) recurrent neural networks (RNNs) on 52,722 human and 89,591 veterinary records. We investigated the accuracy of both separate-domain and combined-domain models and probed model portability. We established relevant baseline classification performances by training Decision Trees (DT) and Random Forests (RF). We also investigated whether transforming the data using MetaMap Lite, a clinical natural language processing tool, affected classification performance. We showed that the LSTM-RNNs accurately classify veterinary and human text narratives into top-level categories with an average weighted macro F1 score of 0.74 and 0.68 respectively. In the “neoplasia” category, the model trained on veterinary data had a high validation accuracy in veterinary data and moderate accuracy in human data, with F1 scores of 0.91 and 0.70 respectively. Our LSTM method scored slightly higher than that of the DT and RF models. The use of LSTM-RNN models represents a scalable structure that could prove useful in cohort identification for comparative oncology studies. Digitization of human and veterinary health information will continue to be a reality, particularly in the form of unstructured narratives. Our approach is a step forward for these two domains to learn from and inform one another.

Analyzing Crude Oil Prices under the Impact of COVID-19 by Using LSTARGARCHLSTM

Under the influence of the COVID-19 pandemic and the concurrent oil conflict between Russia and Saudi Arabia, oil prices have exhibited unusual and sudden changes. For this reason, the volatilities of the West Texas Intermediate (WTI), Brent and Dubai crude daily oil price data between 29 May 2006 and 31 March 2020 are analysed. Firstly, the presence of chaotic and nonlinear behaviour in the oil prices during the pandemic and the concurrent conflict is investigated by using the Shanon Entropy and Lyapunov exponent tests. The tests show that the oil prices exhibit chaotic behavior. Additionally, the current paper proposes a new hybrid modelling technique derived from the LSTARGARCH (Logistic Smooth Transition Autoregressive Generalised Autoregressive Conditional Heteroskedasticity) model and LSTM (long-short term memory) method to analyse the volatility of oil prices. In the proposed LSTARGARCHLSTM method, GARCH modelling is applied to the crude oil prices in two regimes, where regime transitions are governed with an LSTAR-type smooth transition in both the conditional mean and the conditional variance. Separating the data into two regimes allows the efficient LSTM forecaster to adapt to and exploit the different statistical characteristics and ARCH and GARCH effects in each of the two regimes and yield better prediction performance over the case of its application to all the data. A comparison of our proposed method with the GARCH and LSTARGARCH methods for crude oil price data reveals that our proposed method achieves improved forecasting performance over the others in terms of RMSE (Root Mean Square Error) and MAE (Mean Absolute Error) in the face of the chaotic structure of oil prices.