Search

• A prototype of a machine-learned global dB/dt model is introduced using a multivariate LSTM network. • The LSTM model catches geomagnetic disturbances dB/dt ≥ 100 nT/min relatively well. • This model can be further developed to provide an advanced warning of GICs that typically comes with a significantly large dB/dt. Geomagnetically Induced currents (GICs) pose a potentially catastrophic threat to the safety and security of large-scale power distribution systems. Using a recurrent neural network, we predict high (≥60°) magnetic latitude (MLAT) and night side (06–18) magnetic local time (MLT) geomagnetic field disturbances (surface dB s /dt and vertical dB z /dt), the trigger of GICs, during a geomagnetic storms occurred on March 09, 2012. We use the high latitude geomagnetic field data from SuperMAG and the solar wind particle and magnetic field data from NASA-OMNIWeb. We divide the high latitude night side of Earth's globe into multiple sectors and obtain a time series of maximum geomagnetic disturbances observed in each sector. We use the Long Short Term Memory (LSTM) network technique sector by sector on both, the solar wind data and the sectored geomagnetic field disturbance (10° geomagnetic latitude, 1-hour MLT) time series obtained from SuperMAG. Each sector prediction is plotted on a polar map to reveal high latitude night side geomagnetic disturbance patterns. This high latitude nigh side model has great potential. With solar wind input at a Lagrangian-1 point, and geomagnetic field data from SuperMAG, it can predict the peak magnetic disturbance at any given sector for couple of hours in advance. Our machine learning techniques provide the best set of solar wind input that gives the best disturbance prediction. This paper presents a prototype of the global geomagnetic disturbance model with low-resolution spatial grids and a subset of potential solar wind input parameters. Once matured, the global magnetic disturbance model will be provided to the public. [ABSTRACT FROM AUTHOR]

This paper presents a comprehensive study on the application of machine learning techniques in the prediction of respiratory rate via time-series-based statistical and machine learning methods using several physiological signals. Two different models, ARIMA and LSTM, were developed. The LSTM model showed a stronger capacity for learning and capturing complicated patterns in the data compared to the ARIMA model. The findings imply that LSTM models, by incorporating many variables, have the ability to provide predictions that are more accurate, particularly in situations where respiratory rate values vary significantly. [ABSTRACT FROM AUTHOR]

With the fast growth of online content and social media platforms, the amount of user- generated data available in the form of product evaluations, social media postings, and online debates has rapidly expanded. It has become essential for organizations to comprehend client attitudes, make wise decisions, and improve their goods and services by analysing the sentiment indicated in this data. This paper aims to develop a sentiment analysis system for product reviews using URLs of particular item as data sources. It utilizes web scraping techniques to collect textual information from web pages. Natural Language Processing (NLP) techniques are applied for data pre-processing, including tokenization and stop word removal. Machine learning models like Support Vector Classifier (SVC), Random Forest, KNN, Logistic Regression, Naive Bayes and Deep learning models like LSTM and GRU are used for sentiment analysis. Performance measures such as recall, accuracy, and precision are employed to evaluate the models. High scores were consistently achieved by the LSTM and GRU models across all measures, demonstrating their potency in successfully capturing sequential relationships. With strong F-1 ratings of 91% and 90%, respectively, Long Short Term Memory (LSTM) and Gated Recurrent Units (GRU) regularly beat other models, demonstrating their outstanding precision-recall balance. Conversely, K-Nearest Neighbor (KNN) presents a significant trade-off, offering a blazing-fast training time of 0.01 seconds but demanding an extensive 102.34 seconds for prediction, resulting in a lower F-1 score of 47%. Support Vector Machine (SVM) and Logistic Regression achieve competitive F-1 scores of 85% and 84%, respectively, with relatively slow training and prediction times, striking a balance between performance and efficiency. Naive Bayes and Random Forest, while respectable in accuracy, fall slightly behind with F-1 scores of 71% and 80%, respectively, highlighting the significance of taking into account both performance and computing efficiency when choosing the best approach. Through this research, valuable insights are gained into the effectiveness and accuracy of different algorithms in sentiment analysis. The findings contribute to the understanding of sentiment analysis techniques and offer guidance for selecting the most suitable algorithm for URL-based sentiment analysis applications. [ABSTRACT FROM AUTHOR]

Machine learning is poised to revolutionize medicine with algorithms that spot cardiac arrhythmia. An automated diagnostic approach can boost the efficacy of diagnosing life-threatening arrhythmia disorders in routine medical procedures. In this paper, we propose a deep learning network CLINet for ECG signal classification. Our network uses convolution, LSTM and involution layers to bring their unique advantages together. For both convolution and involution layers, we use multiple, large size kernels for multi-scale representation learning. CLINet does not require complicated pre-processing and can handle electrocardiograms of any length. Our network achieves 99.90% accuracy on the ICCAD dataset and 99.94% accuracy on the MIT-BIH dataset. With only 297 K parameters, our model can be easily embedded in smart wearable devices. The source code of CLINet is available at https://github.com/CandleLabAI/CLINet-ECG-Classification-2024. [ABSTRACT FROM AUTHOR]

Product quality is a vital aspect in the operation of manufacturing systems, manufacturers need to implement at least one quality assurance method to assure the desired quality. The most recent approach for quality assurance is named Zero Defect Manufacturing (ZDM). The scope of the current paper is the implementation of ZDM approach in the automotive industry and specifically for the spot-welding process. Using a machine learning method that is utilizing linear regression and LSTM and consuming data from the production such as sensors and other engineering data to predict the quality of future spot welds. Using this quality prediction there is a root cause analysis to identify why the spot weld will fail and the appropriate prevention actions are proposed. The next step is the training and validation of the machine learning model and the calculation of the accuracy of the model. Once the accuracy of the model is validated a series of simulations, using a dynamic scheduling tool, are performed in order to calculate a series of KPIs to evaluate the impact of the proposed method to the production. [ABSTRACT FROM AUTHOR]

This paper proposes an innovative approach towards real-time earthquake forecasting in Chile by integrating intelligent technologies and machine learning methods. Earthquakes pose significant risks to communities and infrastructure in Chile, making accurate and timely forecasting crucial for disaster preparedness and mitigation. Traditional forecasting methods have limitations in providing real-time insights into seismic activity. In contrast, intelligent technologies such as artificial intelligence (AI) and machine learning offer promising avenues for enhancing prediction accuracy and speed. A new earthquake forecasting method using a modified clustering approach LMSCAN(Local Maxima-based Spatio-Cluster Analysis Network) and enhanced neural network time series analysis LSTM-IC (Long Short Term Memory Inverse Correlation) is presented in this paper. This neural network-based technology has been utilized to forecast earthquakes in the Chile region, one of the countries with the largest seismic activity. This study explores the integration of intelligent systems with machine learning algorithms to analyze seismic data and predict earthquake occurrences in Chile. By leveraging historical seismic data, sensor networks, and advanced predictive models, our approach aims to provide timely warnings and insights into seismic events, thereby improving disaster response and resilience. The proposed framework holds the potential to revolutionize earthquake forecasting by enabling real-time monitoring and proactive measures to safeguard communities and infrastructure in Chile and beyond. The remarkable 95 % accuracy achieved by this model is a testament to its exceptional learning process, which sets it apart from other models. Its ability to learn and adapt to new data is unparalleled, allowing it to forecast incredibly precisely and produce highly reliable results. [ABSTRACT FROM AUTHOR]

Failure probability estimation of hysteretic systems subjected to random process excitation appears in many practical engineering applications, such as structures subjected to earthquake and vibration control. This task is computationally challenging due to the need for repeated solutions of a high-resolution model. This cost grows significantly with the stochastic dimension — characterized by the number of random variables, in this case, arising from the discretization of the random excitation. To address this issue, a novel non-intrusive reduced order model (ROM) is proposed in this paper. Accordingly, the complexity due to spatial discretization is first reduced using proper orthogonal decomposition. However, interpolation among response time histories in the resulting reduced solution space is computationally infeasible due to large stochastic dimensionality. An existing ROM developed by the authors is now improved by adopting a long short-term memory (LSTM) network for this interpolation. Furthermore, this adoption required two additional modifications in the mathematical structure of the ROM related to data compression and nonlinear coupling. The proposed methodology can be seamlessly integrated with black-box nonlinear solvers. Through detailed numerical studies on a beam on Winkler foundation and a multi-storied building frame subjected to stationary and non-stationary excitations, the proposed ROM is found to be accurate, efficient, and robust with respect to the frequency band of interest. The ROM has shown reduction of two orders of magnitude in the computational time. Furthermore, the proposed ROM is inherently parallelizable and holds promise to be scaled to other types of nonlinear systems with large stochastic dimensionality. • New non-intrusive ROM for nonlinear stochastic dynamical system. • It enhances a POD based non-intrusive ROM with LSTM-based surrogate. • Addressing large stochastic dimensionality due to random process excitation. • Bilinear hysteric nonlinearity is considered. • Achieved significant speed-up and robustness. [ABSTRACT FROM AUTHOR]

This paper proposes a new method for predicting downlink traffic volume in mobile networks, aiming to minimize overprovisioning while meeting specified service-level agreement (SLA) violation rates. We introduce a multivariate and multi-step prediction approach and compare four machine learning (ML) architectures: long short-term memory (LSTM), convolutional neural network (CNN), transformer, and light gradient-boosting machine (LightGBM). Our models predict up to 24 steps ahead and are evaluated under both single-step and multi-step conditions. Additionally, we propose parametric loss functions to adhere to SLA violation rate constraints. Our results emphasize the importance of using parametric loss functions to meet SLA constraints. We discovered that LSTM when paired with our custom multivariate feature sets, outperforms the transformer architecture in short-term forecasting up to 4 h ahead. For these short-term predictions, we demonstrate that methods based on domain knowledge, like our custom feature sets combined with simpler models such as LSTM, surpass more complex models like transformers. However, for long-term forecasting (8 to 24 h ahead), transformers outperform all other models. [ABSTRACT FROM AUTHOR]

• The VMD decomposition processing module can improve prediction accuracy. • VMD-CBiLSTM can improve the accuracy of drought prediction in arid and semi-arid regions. • The VMD-CBiLSTM prediction model has good stability. Drought can cause great harm and loss, so accurate and efficient drought prediction has certain research significance. In previous studies, a single machine learning model is often used to predict the factors related to drought, such as precipitation and temperature, to indirectly explain the influence degree of drought, but the overall prediction accuracy is low, which can not fully and effectively predict the nonlinear and non-stationary drought characteristic information. In this paper, the variational modal decomposition model (VMD) is used to decompose the meteorological drought time series, and the convolution neural network (CNN) and the improved bidirectional long short-term memory neural network (BiLSTM) are combined to construct the meteorological drought hybrid prediction model (VMD-CBiLSTM). The research results show that using VMD-CBiLSTM model to forecast the monthly daily evapotranspiration deficit index (DEDI) of three weather stations, compared with the results predicted by VMD-LSTM, VMD-BiLSTM and VMD-ARIMA models, the average prediction accuracy R2 is increased by 0.076, 0.034 and 0.328 respectively, and the average RMSE is decreased by 0.178, 0.094 and 1.373 respectively. Compared with single model, VMD-CBiLSTM can not only reduce the uncertainty of meteorological drought prediction caused by climate model heterogeneity, but also improve the accuracy of drought prediction in arid and semi-arid regions, which can provide reference for coping with drought occurrence and drought early warning in advance. [ABSTRACT FROM AUTHOR]

Due to their dynamic properties such as irregular sampling rate and high-frequency sampling, Continuous Time Series (CTS) are found in many applications. Since CTS with irregular sampling rate are difficult to model with standard Recurrent Neural Networks (RNNs), RNNs have been generalised to have continuous-time hidden dynamics defined by a Neural Ordinary Differential Equation (Neural ODE), leading to the ODE-RNN model. Another approach that provides a better modelling is that of the Latent ODE model, which constructs a continuous-time model where a latent state is defined at all times. The Latent ODE model uses a standard RNN as the encoder and a Neural ODE as the decoder. However, since the RNN encoder leads to difficulties with missing data and ill-defined latent variables, a Latent ODE-RNN model has recently been proposed that uses a ODE-RNN model as the encoder instead. Both the Latent ODE and Latent ODE-RNN models are difficult to train due to the vanishing and exploding gradients problem. To overcome this problem, the main contribution of this paper is to propose and illustrate a new model based on a new Latent ODE using an ODE-LSTM (Long Short-Term Memory) network as an encoder - the Latent ODE-LSTM model. To limit the growth of the gradients, the Norm Gradient Clipping strategy was embedded on the Latent ODE-LSTM model. The performance evaluation of the new Latent ODE-LSTM (with and without Norm Gradient Clipping) for modelling CTS with regular and irregular sampling rates is then demonstrated. Numerical experiments show that the new Latent ODE-LSTM performs better than Latent ODE-RNNs and can avoid the vanishing and exploding gradients during training. Code implementations developed in this work are available at github.com/CeciliaCoelho/LatentODELSTM. [ABSTRACT FROM AUTHOR]

This paper focuses on the hypothesis of optimizing time series predictions using fractal interpolation techniques. In general, the accuracy of machine learning model predictions is closely related to the quality and quantitative aspects of the data used, following the principle of garbage-in, garbage-out. In order to quantitatively and qualitatively augment datasets, one of the most prevalent concerns of data scientists is to generate synthetic data, which should follow as closely as possible the actual pattern of the original data. This study proposes three different data augmentation strategies based on fractal interpolation, namely the Closest Hurst Strategy , Closest Values Strategy and Formula Strategy. To validate the strategies, we used four public datasets from the literature, as well as a private dataset obtained from meteorological records in the city of Braşov, Romania. The prediction results obtained with the LSTM model using the presented interpolation strategies showed a significant accuracy improvement compared to the raw datasets, thus providing a possible answer to practical problems in the field of remote sensing and sensor sensitivity. Moreover, our methodologies answer some optimization-related open questions for the fractal interpolation step using Optuna framework. [ABSTRACT FROM AUTHOR]

With the significant increase in photovoltaic (PV) electricity generation, more attention has been given to PV power forecasting. Indeed, accurate forecasting allows power grid operators to better schedule and dispatch their assets, such as energy storage systems and reserve. In this paper, a hybrid deep learning model and a convolutional neural network with memory is proposed, to provide intraday (2 h) solar irradiance forecasts using sequentially-collected images from public webcams. The performance of the proposed model is compared to those of a standard time-series forecast models, a linear regression as well as state-of-the-art neural networks. All models are trained and tested using images collected from two webcams on EPFL's campus for just over a year. The results show that the proposed method outperforms all other models and matches the state-of-the-art methodology while providing simplicity of implementation and efficient computation. • A novel method for estimating global horizontal irradiance is proposed. • The method uses images public webcams instead of all-sky cameras. • The method combines images from different cameras to improve forecasts' accuracy. • The method can be extended to include other types of data (meteorological data). • This simple method matches state-of-the-art methods. [ABSTRACT FROM AUTHOR]

With the increasing impact of climate change and relative sea level rise, low-lying coastal communities face growing risks from recurrent nuisance flooding and storm tides. Thus, timely and reliable predictions of coastal water levels are critical to resilience in vulnerable coastal areas. Over the past decade, there has been increasing interest in utilizing machine learning (ML) based models for emulation and prediction of coastal water levels. However, flood advisory systems still rely on running computationally demanding hydrodynamic models. To alleviate the computational burden, these physics-based models are either run at small scales with high resolution or at large scales with low resolution. While ML-based models are very fast, they face challenges in terms of ensuring reliability and ability to capture any surge levels. In this paper, we develop a deep neural network for spatiotemporal prediction of water levels in coastal areas of the Chesapeake Bay in the U.S. Our model relies on data from numerical weather prediction models as the atmospheric input and astronomical tide levels, while its outputs are time series of predicted water levels at several tide gauge locations across the Chesapeake Bay. We utilized a CNN-LSTM setting as the architecture of the model. The CNN part extracts the features from a sequence of gridded wind fields and fuses its output to several independent LSTM units. The LSTM units concatenate the atmospheric features with respective astronomical tide levels and produce water level time series. The novel contribution of the present work is in spatiotemporality and in prioritization of the physical relationships in the model to maintain a high analogy to hydrodynamic modeling, either in the network architecture or in the selection of predictors and predictands. The results show that this setting yields a strong performance in predicting coastal water levels that cause flooding from minor to major levels. We also show that the model stands up successfully to the rigorous comparison with a high-fidelity ADCIRC model, yielding mean RMSE and correlation coefficient of 14.3 cm and 0.94, respectively, in two extreme cases, versus 12.30 cm and 0.96 for the ADCIRC model. The results highlight the practical feasibility of employing fast yet inexpensive data-driven models for resilient coastal management. • A deep learning model is developed to predict a wide range of coastal water levels using wind and tides as input. • CNN layers extract features from the wind field and fuse them to independent LSTM units, creating spatiotemporality. • The model compares well with the data for different storm scenarios. • The model is comparable to or surpasses a high-fidelity hydrodynamic model while taking a fraction of a second to run. [ABSTRACT FROM AUTHOR]

Due to the influx of advancements in technology and the increased simplicity of communication through emails, there has been a severe threat to the global economy and security due to upsurge in volume of unsolicited During the training of models, high-dimensional and redundant datasets may reduce the classification results of the model due to high memory costs and high computation. An important data processing technique is feature selection which helps in selecting relevant features and subsets of information from the dataset. Therefore, choosing efficient feature selection techniques is very important for the best performance of classification of a model. Moreover, most of the research has been performed using traditional machine learning techniques, which are not enough to deal with the huge amount of data and its variations. Also, spammers are becoming smarter with technological advancement. Therefore, there is a need for hybrid techniques consisting of deep learning and conventional algorithms to cope with these problems. We have proposed a novel scheme in this paper for email spam detection, which will result in an improved feature selection approach from the original dataset and increase the accuracy of the classifier as well. The literature has been studied to explore the efficient machine learning models that have been applied by different researchers for email spam detection and feature selection to acquire the best results. Our method, GWO-BERT, has given remarkable results with deep learning techniques such as CNN, biLSTM and LSTM. We have compared our models with RF and LSTM and used dataset: "Lingspam," which is a publicly available dataset. With different experiments, our technique, GWO-BERT, obtained 99.14% accuracy, which is almost equal to 100 percent. [ABSTRACT FROM AUTHOR]

Accurate traffic congestion estimation and prediction are critical building blocks for smart trip planning and rerouting decisions in transportation systems. Over the decades, there have been many studies focusing on traffic congestion estimation and prediction with different statistical approaches (e.g., Markov chain) and machine learning models (e.g., clustering, Bayesian networks, and artificial neural networks). However, there is a lack of a unified framework to address the mechanisms of different models and integrate the advantages of different methods through combinations. This paper introduces the FD-Markov-LSTM model, a hybrid interpretable approach that combines the fundamental diagram (FD), Markov chain, and long short-term memory (LSTM). The aim is to estimate and predict traffic states by integrating statistical data in both congested and uncongested scenarios. The FD-Markov-LSTM model leverages the FD to identify hierarchical traffic states and utilizes the Markov process to capture the probabilistic transitions between these states. We employ the LSTM model to further capture the residual time series produced by the Markov chain model (assuming a memoryless property) to enhance the estimation and prediction performance. The proposed model's accuracy in estimating and predicting traffic flow is evaluated using empirical data from three case studies conducted in Beijing and Los Angeles. The results highlight a significant improvement in accuracy compared to classical benchmark models such as the Markov model, ARIMA model, k-Nearest Neighbor model, Random Forest model, and LSTM. Specifically, the FD-Markov-LSTM model achieves reductions of over 39% in mean absolute error, 35% in root mean squared error, and 7.4% in mean absolute percentage error. These results clearly demonstrate that the FD-Markov-LSTM model outperforms the benchmark models, enabling more precise predictions of traffic flow. [ABSTRACT FROM AUTHOR]

With a recently defined AutoGCOP framework, the design of local search algorithms has been defined as the composition of elementary algorithmic components. The effective compositions of the best algorithms thus retain useful knowledge of effective algorithm design. This paper investigates machine learning to learn and extract useful knowledge in effective algorithmic compositions. The process of forecasting algorithmic components in the design of effective local search algorithms is defined as a sequence classification task, and solved by a long short-term memory (LSTM) neural network to systematically analyse algorithmic compositions. Compared with other learning models, the results reveal the superior prediction performance of the proposed LSTM. Further analysis identifies some key features of algorithmic compositions and confirms their effectiveness for improving the prediction, thus supporting effective automated algorithm design. • The design of local search algorithms is defined as a sequence classification task. • LSTM is applied to forecast algorithmic components for automated composition. • LSTM has a better classification performance as compared with other classifiers. • Key features for sequence classification are identified to support algorithm design. [ABSTRACT FROM AUTHOR]

Machine learning models are widely used for identifying malicious software. However, existing models suffer from issues such as imprecise polysemous representations and a lack of contextual semantic representations, leading to the failure to recognize certain types of malicious software. In this paper, we propose a semantic-based intelligent malware detection model called SeMalBERT for identifying malicious software in Windows systems. Specifically, the model utilizes the API function sequences of malicious software as the learned features. Firstly, BERT is applied to accomplish word representation tasks and extract semantic information from the sequences. Secondly, a hybrid discriminator based on Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) is used to explore the chaining relationships between functions. Lastly, an external attention mechanism is added after the LSTM to enable the model to better focus on key information within the text. Experimental results demonstrate that SeMalBERT outperforms existing malware detection techniques in terms of accuracy, F1 score, and loss function value on a general dataset. [ABSTRACT FROM AUTHOR]

Edge computing has gained widespread adoption for time-sensitive applications by offloading a portion of IoT system workloads from the cloud to edge nodes. However, the limited resources of IoT edge devices hinder service deployment, making auto-scaling crucial for improving resource utilization in response to dynamic workloads. Recent solutions aim to make auto-scaling proactive by predicting future workloads and overcoming the limitations of reactive approaches. These proactive solutions often rely on time-series data analysis and machine learning techniques, especially Long Short-Term Memory (LSTM), thanks to its accuracy and prediction speed. However, existing auto-scaling solutions often suffer from oscillation issues, even when using a cooling-down strategy. Consequently, the efficiency of proactive auto-scaling depends on the prediction model accuracy and the degree of oscillation in the scaling actions. This paper proposes a novel approach to improve prediction accuracy and deal with oscillation issues. Our approach involves an automatic featurization phase that extracts features from time-series workload data, improving the prediction's accuracy. These extracted features also serve as a grid for controlling oscillation in generated scaling actions. Our experimental results demonstrate the effectiveness of our approach in improving prediction accuracy, mitigating oscillation phenomena, and enhancing the overall auto-scaling performance. [ABSTRACT FROM AUTHOR]

The proliferation of fake news has raised concerns regarding its detection, posing a significant challenge. Motivated by the ongoing discussion on the sustainability of machine learning algorithms, this paper discusses the usefulness of data reduction for fake news detection. This is accomplished by using the fuzzy transform (or F-transform for short), which has already been proven effective, in the literature, to reduce the training time. A Long Short Term Memory architecture is then employed for classification to determine the authenticity of the news. From the formal perspective, we discuss in general the role of the F-transform in a learning system. Regarding the numerical experiments, we use five publicly available datasets, trying different compression ratios and different types of F-transform, assessing accuracy, F1 score, training time and energy consumption with and without F-transform. Although the F-transform is a lossy compression technique, the results show a negligible variation in accuracy and F1-score when comparing results with and without F-transform (i.e. 1%–3% in most cases and around 10% in one case). This seems to be congruent with the theoretical achievement. Furthermore, the approach yields substantial training time and energy savings, with over 50% reduction in energy consumption. • Deep architecture based on LSTM to detect fake news. • Fuzzy transform to reduce data cardinality and improve sustainability. • Formal discussion on compression in a generic learning system (new property). • Numerical experiments on five publicly available datasets. • Good accuracy, substantial training time and energy savings (over 50% reduction). [ABSTRACT FROM AUTHOR]

Forecasting cryptocurrency volatility can help investors make better-informed investment decisions in order to minimize risks and maximize potential profits. Accurate forecasting of cryptocurrency price fluctuations is crucial for effective portfolio management and contributes to the stability of the financial system by identifying potential threats and developing risk management strategies. The objective of this paper is to provide a comprehensive study of statistical and machine learning methods for predicting daily and weekly volatility of the following four cryptocurrencies: Bitcoin, Ethereum, Litecoin, and Monero. Several models and forecasting methods are compared in terms of their forecasting accuracy, i.e., HAR (heterogeneous autoregressive), ARFIMA (autoregressive fractionally integrated moving average), GARCH (generalized autoregressive conditional heteroscedasticity), LASSO (least absolute shrinkage and selection operator), RR (ridge regression), SVR (support vector regression), MLP (multilayer perceptron), FNM (fuzzy neighbourhood model), RF (random forest), and LSTM (long short-term memory). The realized variance calculated from intraday returns is used as the input variable for the models. In order to assess the predictive power of the models considered, the model confidence set (MCS) procedure is applied. Our experimental results demonstrate that there is no single best method for forecasting volatility of each cryptocurrency, and different models may perform better depending on the specific cryptocurrency, choice of the error metric and forecast horizon. For daily forecasts, the method that is always found in a set of best models is linear SVR, while for weekly forecasts, there are two such methods, namely FNM and RR. Furthermore, we show that simple linear models such as HAR and ridge regression, perform not worse than more complex models like LSTM and RF. The research provides a useful reference point for the development of more sophisticated models. • Comprehensive study of the forecasting methods for cryptocurrency volatility. • 12 popular methods compared including HAR, GARCH, LASSO, SVR, MLP, RF, LSTM. • No single best method for each cryptocurrency. • Different models perform better depending on error metric and forecast horizon. • Simple linear models can perform as well as more complex ML models. [ABSTRACT FROM AUTHOR]