A Bayesian non-stationary heteroskedastic time series model for multivariate critical care data.

We propose a multivariate GARCH model for non-stationary health time series by modifying the observation-level variance of the standard state space model. The proposed model provides an intuitive and novel way of dealing with heteroskedastic data using the conditional nature of state-space models. We follow the Bayesian paradigm to perform the inference procedure. In particular, we use Markov chain Monte Carlo methods to obtain samples from the resultant posterior distribution. We use the forward filtering backward sampling algorithm to efficiently obtain samples from the posterior distribution of the latent state. The proposed model also handles missing data in a fully Bayesian fashion. We validate our model on synthetic data and analyze a data set obtained from an intensive care unit in a Montreal hospital and the MIMIC dataset. We further show that our proposed models offer better performance, in terms of WAIC than standard state space models. The proposed model provides a new way to model multivariate heteroskedastic non-stationary time series data. Model comparison can then be easily performed using the WAIC.

Event prediction by estimating continuously the completion of a single temporal pattern's instances.

OBJECTIVE: Develop a new method for continuous prediction that utilizes a single temporal pattern ending with an event of interest and its multiple instances detected in the temporal data. METHODS: Use temporal abstraction to transform time series, instantaneous events, and time intervals into a uniform representation using symbolic time intervals (STIs). Introduce a new approach to event prediction using a single time intervals-related pattern (TIRP), which can learn models to predict whether and when an event of interest will occur, based on multiple instances of a pattern that end with the event. RESULTS: The proposed methods achieved an average improvement of 5% AUROC over LSTM-FCN, the best-performed baseline model, out of the evaluated baseline models (RawXGB, Resnet, LSTM-FCN, and ROCKET) that were applied to real-life datasets. CONCLUSION: The proposed methods for predicting events continuously have the potential to be used in a wide range of real-world and real-time applications in diverse domains with heterogeneous multivariate temporal data. For example, it could be used to predict panic attacks early using wearable devices or to predict complications early in intensive care unit patients.

Anomaly Detection for Asynchronous Multivariate Time Series of Nuclear Power Plants Using a Temporal-Spatial Transformer.

Industrial process monitoring is a critical application of multivariate time-series (MTS) anomaly detection, especially crucial for safety-critical systems such as nuclear power plants (NPPs). However, some current data-driven process monitoring approaches may not fully capitalize on the temporal-spatial correlations inherent in operational MTS data. Particularly, asynchronous time-lagged correlations may exist among variables in actual NPPs, which further complicates this challenge. In this work, a reconstruction-based MTS anomaly detection approach based on a temporal-spatial transformer is proposed. It employs a two-stage temporal-spatial attention mechanism combined with a multi-scale strategy to learn the dependencies within normal operational data at various scales, thereby facilitating the extraction of temporal-spatial correlations from asynchronous MTS. Experiments on simulated datasets and real NPP datasets demonstrate that the proposed model possesses stronger feature learning capabilities, as evidenced by its improved performance in signal reconstruction and anomaly detection for asynchronous MTS data. Moreover, the proposed TS-Trans model enables earlier detection of anomalous events, which holds significant importance for enhancing operational safety and reducing potential losses in NPPs.

Multiview Spatial-Temporal Meta-Learning for Multivariate Time Series Forecasting.

Multivariate time series modeling has been essential in sensor-based data mining tasks. However, capturing complex dynamics caused by intra-variable (temporal) and inter-variable (spatial) relationships while simultaneously taking into account evolving data distributions is a non-trivial task, which faces accumulated computational overhead and multiple temporal patterns or distribution modes. Most existing methods focus on the former direction without adaptive task-specific learning ability. To this end, we developed a holistic spatial-temporal meta-learning probabilistic inference framework, entitled ST-MeLaPI, for the efficient and versatile learning of complex dynamics. Specifically, first, a multivariate relationship recognition module is utilized to learn task-specific inter-variable dependencies. Then, a multiview meta-learning and probabilistic inference strategy was designed to learn shared parameters while enabling the fast and flexible learning of task-specific parameters for different batches. At the core are spatial dependency-oriented and temporal pattern-oriented meta-learning approximate probabilistic inference modules, which can quickly adapt to changing environments via stochastic neurons at each timestamp. Finally, a gated aggregation scheme is leveraged to realize appropriate information selection for the generative style prediction. We benchmarked our approach against state-of-the-art methods with real-world data. The experimental results demonstrate the superiority of our approach over the baselines.

Predicting the Posture of High-Rise Building Machines Based on Multivariate Time Series Neural Network Models.

High-rise building machines (HBMs) play a critical role in the successful construction of super-high skyscrapers, providing essential support and ensuring safety. The HBM's climbing system relies on a jacking mechanism consisting of several independent jacking cylinders. A reliable control system is imperative to maintain the smooth posture of the construction steel platform (SP) under the action of the jacking mechanism. Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Temporal Convolutional Network (TCN) are three multivariate time series (MTS) neural network models that are used in this study to predict the posture of HBMs. The models take pressure and stroke measurements from the jacking cylinders as inputs, and their outputs determine the levelness of the SP and the posture of the HBM at various climbing stages. The development and training of these neural networks are based on historical on-site data, with the predictions subjected to thorough comparative analysis. The proposed LSTM and GRU prediction models have similar performances in the prediction process of HBM posture, with medians R2 of 0.903 and 0.871, respectively. However, the median MAE of the GRU prediction model is more petite at 0.4, which exhibits stronger robustness. Additionally, sensitivity analysis showed that the change in the levelness of the position of the SP portion of the HBM exhibited high sensitivity to the stroke and pressure of the jacking cylinder, which clarified the position of the cylinder for adjusting the posture of the HBM. The results show that the MTS neural network-based prediction model can change the HBM posture and improve work stability by adjusting the jacking cylinder pressure value of the HBM.

Adaptive prediction for effluent quality of wastewater treatment plant: Improvement with a dual-stage attention-based LSTM network.

The accurate effluent prediction plays a crucial role in providing early warning for abnormal effluent and achieving the adjustment of feedforward control parameters during wastewater treatment. This study applied a dual-staged attention mechanism based on long short-term memory network (DA-LSTM) to improve the accuracy of effluent quality prediction. The results showed that input attention (IA) and temporal attention (TA) significantly enhanced the prediction performance of LSTM. Specially, IA could adaptively adjust feature weights to enhance the robustness against input noise, with R2 increased by 13.18%. To promote its long-term memory ability, TA was used to increase the memory span from 96 h to 168 h. Compared to a single LSTM model, the DA-LSTM model showed an improvement in prediction accuracy by 5.10%, 2.11%, 14.47% for COD, TP, and TN. Additionally, DA-LSTM demonstrated excellent generalization performance in new scenarios, with the R2 values for COD, TP, and TN increasing by 22.67%, 20.06%, and 17.14% respectively, while the MAPE values decreased by 56.46%, 63.08%, and 42.79%. In conclusion, the DA-LSTM model demonstrated excellent prediction performance and generalization ability due to its advantages of feature-adaptive weighting and long-term memory focusing. This has forward-looking significance for achieving efficient early warning of abnormal operating conditions and timely management of control parameters.

Group Surrogate Data Generating Models and similarity quantification of multivariate time-series: A resting-state fMRI study.

Advancements in non-invasive brain analysis through novel approaches such as big data analytics and in silico simulation are essential for explaining brain function and associated pathologies. In this study, we extend the vector auto-regressive surrogate technique from a single multivariate time-series to group data using a novel Group Surrogate Data Generating Model (GSDGM). This methodology allowed us to generate biologically plausible human brain dynamics representative of a large human resting-state (rs-fMRI) dataset obtained from the Human Connectome Project. Simultaneously, we defined a novel similarity measure, termed the Multivariate Time-series Ensemble Similarity Score (MTESS). MTESS showed high accuracy and f-measure in subject identification, and it can directly compare the similarity between two multivariate time-series. We used MTESS to analyze both human and marmoset rs-fMRI data. Our results showed similarity differences between cortical and subcortical regions. We also conducted MTESS and state transition analysis between single and group surrogate techniques, and confirmed that a group surrogate approach can generate plausible group centroid multivariate time-series. Finally, we used GSDGM and MTESS for the fingerprint analysis of human rs-fMRI data, successfully distinguishing normal and outlier sessions. These new techniques will be useful for clinical applications and in silico simulation.

Benefit-aware early prediction of health outcomes on multivariate EEG time series.

Given a cardiac-arrest patient being monitored in the ICU (intensive care unit) for brain activity, how can we predict their health outcomes as early as possible? Early decision-making is critical in many applications, e.g. monitoring patients may assist in early intervention and improved care. On the other hand, early prediction on EEG data poses several challenges: (i) earliness-accuracy trade-off; observing more data often increases accuracy but sacrifices earliness, (ii) large-scale (for training) and streaming (online decision-making) data processing, and (iii) multi-variate (due to multiple electrodes) and multi-length (due to varying length of stay of patients) time series. Motivated by this real-world application, we present BeneFitter that infuses the incurred savings from an early prediction as well as the cost from misclassification into a unified domain-specific target called benefit. Unifying these two quantities allows us to directly estimate a single target (i.e. benefit), and importantly, (a) is efficient and fast, with training time linear in the number of input sequences, and can operate in real-time for decision-making, (b) can handle multi-variate and variable-length time-series, suitable for patient data, and (c) is effective, providing up to 2× time-savings with equal or better accuracy as compared to competitors.

A self-supervised learning-based approach to clustering multivariate time-series data with missing values (SLAC-Time): An application to TBI phenotyping.

Self-supervised learning approaches provide a promising direction for clustering multivariate time-series data. However, real-world time-series data often include missing values, and the existing approaches require imputing missing values before clustering, which may cause extensive computations and noise and result in invalid interpretations. To address these challenges, we present a Self-supervised Learning-based Approach to Clustering multivariate Time-series data with missing values (SLAC-Time). SLAC-Time is a Transformer-based clustering method that uses time-series forecasting as a proxy task for leveraging unlabeled data and learning more robust time-series representations. This method jointly learns the neural network parameters and the cluster assignments of the learned representations. It iteratively clusters the learned representations with the K-means method and then utilizes the subsequent cluster assignments as pseudo-labels to update the model parameters. To evaluate our proposed approach, we applied it to clustering and phenotyping Traumatic Brain Injury (TBI) patients in the Transforming Research and Clinical Knowledge in Traumatic Brain Injury (TRACK-TBI) study. Clinical data associated with TBI patients are often measured over time and represented as time-series variables characterized by missing values and irregular time intervals. Our experiments demonstrate that SLAC-Time outperforms the baseline K-means clustering algorithm in terms of silhouette coefficient, Calinski Harabasz index, Dunn index, and Davies Bouldin index. We identified three TBI phenotypes that are distinct from one another in terms of clinically significant variables as well as clinical outcomes, including the Extended Glasgow Outcome Scale (GOSE) score, Intensive Care Unit (ICU) length of stay, and mortality rate. The experiments show that the TBI phenotypes identified by SLAC-Time can be potentially used for developing targeted clinical trials and therapeutic strategies.

A novel deep learning ensemble model based on two-stage feature selection and intelligent optimization for water quality prediction.

Accurate prediction of effluent total nitrogen (E-TN) can assist in feed-forward control of wastewater treatment plants (WWTPs) to ensure effluent compliance with standards while reducing energy consumption. However, multivariate time series prediction of E-TN is a challenge due to the complex nonlinearity of WWTPs. This paper proposes a novel prediction framework that combines a two-stage feature selection model, the Golden Jackal Optimization (GJO) algorithm, and a hybrid deep learning model, CNN-LSTM-TCN (CLT), aiming to effectively capture the nonlinear relationships of multivariate time series in WWTPs. Specifically, convolutional neural network (CNN), long short-term memory (LSTM), and temporal convolutional network (TCN) combined to build a hybrid deep learning model CNN-LSTM-TCN (CLT). A two-stage feature selection method is utilized to determine the optimal feature subset to reduce the complexity and improve the accuracy of the prediction model, and then, the feature subset is input into the CLT. The hyperparameters of the CLT are optimized using GJO to further improve the prediction performance. Experiments indicate that the two-stage feature selection model learns the optimal feature subset to predict best, and the GJO-CLT achieves the best performance for different backtracking windows and prediction steps. These results demonstrate that the prediction system excels in the task of multivariate water quality time series prediction of WWTPs.

Anomaly prediction of CT equipment based on IoMT data.

BACKGROUND: Large-scale medical equipment, which is extensively implemented in medical services, is of vital importance for diagnosis but vulnerable to various anomalies and failures. Most hospitals that conduct regular maintenance have been suffering from medical equipment-related incidents for years. Currently, the Internet of Medical Things (IoMT) has emerged as a crucial tool in monitoring the real-time status of the medical equipment. In this paper, we develop an IoMT system of Computed Tomography (CT) equipment in the West China Hospital, Sichuan University and collected the system status time-series data. Novel multivariate time-series classification models and frameworks are proposed to predict the anomalies of CT equipment. The important features that are closely related to the equipment anomalies are identified with the model. METHODS: We extracted the real-time CT equipment status time-series data of 11 equipment between May 19, 2020 and May 19, 2021 from the IoMT, which includes the equipment oil temperature, anode voltage, etc. The arcs are identified as labels of anomalies due to their relationship with decreased imaging quality and CT equipment failures. To improve prediction accuracy, the statistics and transformations of the raw historical time-series data segment in the sliding time window are used to construct new features. Due to the particularity of time-series data, two frameworks are proposed for splitting the training and test sets. Then the Decision Tree, Support Vector Machine, Logistic Regression, Naive Bayesian, and K-Nearest Neighbor classification models are used to classify the system status. We also compare our model to state-of-the-art models. RESULTS: The results show that the anomaly prediction accuracy and recall of our method are 79% and 77%, respectively. The oil temperature and anode voltage are identified as the decisive features that may lead to anomalies. The proposed model outperforms the others when predicting the anomalies of the CT equipment based on our dataset. CONCLUSIONS: The proposed method could predict the state of CT equipment and be used as a reference for practical maintenance, where unexpected anomalies of medical equipment could be reduced. It also brings new insights into how to handle non-uniform and imbalanced time series data in practical cases.

Unsupervised Anomaly Detection for IoT-Based Multivariate Time Series: Existing Solutions, Performance Analysis and Future Directions.

The recent wave of digitalization is characterized by the widespread deployment of sensors in many different environments, e.g., multi-sensor systems represent a critical enabling technology towards full autonomy in industrial scenarios. Sensors usually produce vast amounts of unlabeled data in the form of multivariate time series that may capture normal conditions or anomalies. Multivariate Time Series Anomaly Detection (MTSAD), i.e., the ability to identify normal or irregular operative conditions of a system through the analysis of data from multiple sensors, is crucial in many fields. However, MTSAD is challenging due to the need for simultaneous analysis of temporal (intra-sensor) patterns and spatial (inter-sensor) dependencies. Unfortunately, labeling massive amounts of data is practically impossible in many real-world situations of interest (e.g., the reference ground truth may not be available or the amount of data may exceed labeling capabilities); therefore, robust unsupervised MTSAD is desirable. Recently, advanced techniques in machine learning and signal processing, including deep learning methods, have been developed for unsupervised MTSAD. In this article, we provide an extensive review of the current state of the art with a theoretical background about multivariate time-series anomaly detection. A detailed numerical evaluation of 13 promising algorithms on two publicly available multivariate time-series datasets is presented, with advantages and shortcomings highlighted.

DCFF-MTAD: A Multivariate Time-Series Anomaly Detection Model Based on Dual-Channel Feature Fusion.

The detection of anomalies in multivariate time-series data is becoming increasingly important in the automated and continuous monitoring of complex systems and devices due to the rapid increase in data volume and dimension. To address this challenge, we present a multivariate time-series anomaly detection model based on a dual-channel feature extraction module. The module focuses on the spatial and time features of the multivariate data using spatial short-time Fourier transform (STFT) and a graph attention network, respectively. The two features are then fused to significantly improve the model's anomaly detection performance. In addition, the model incorporates the Huber loss function to enhance its robustness. A comparative study of the proposed model with existing state-of-the-art ones was presented to prove the effectiveness of the proposed model on three public datasets. Furthermore, by using in shield tunneling applications, we verify the effectiveness and practicality of the model.

Masked Graph Neural Networks for Unsupervised Anomaly Detection in Multivariate Time Series.

Anomaly detection has been widely used in grid operation and maintenance, machine fault detection, and so on. In these applications, the multivariate time-series data from multiple sensors with latent relationships are always high-dimensional, which makes multivariate time-series anomaly detection particularly challenging. In existing unsupervised anomaly detection methods for multivariate time series, it is difficult to capture the complex associations among multiple sensors. Graph neural networks (GNNs) can model complex relations in the form of a graph, but the observed time-series data from multiple sensors lack explicit graph structures. GNNs cannot automatically learn the complex correlations in the multivariate time-series data or make good use of the latent relationships among time-series data. In this paper, we propose a new method-masked graph neural networks for unsupervised anomaly detection (MGUAD). MGUAD can learn the structure of the unobserved causality among sensors to detect anomalies. To robustly learn the temporal context from adjacent time points of time-series data from the same sensor, MGUAD randomly masks some points of the time-series data from the sensor and reconstructs the masked time points. Similarly, to robustly learn the graph-level context from adjacent nodes or edges in the relation graph of multivariate time series, MGUAD masks some nodes or edges in the graph under the framework of a GNN. Comprehensive experiments are conducted on three public datasets. According to the experimental findings, MGUAD outperforms state-of-the-art anomaly detection methods.

Application of 1D ResNet for Multivariate Fault Detection on Semiconductor Manufacturing Equipment.

Amid the ongoing emphasis on reducing manufacturing costs and enhancing productivity, one of the crucial objectives when manufacturing is to maintain process tools in optimal operating conditions. With advancements in sensing technologies, large amounts of data are collected during manufacturing processes, and the challenge today is to utilize these massive data efficiently. Some of these data are used for fault detection and classification (FDC) to evaluate the general condition of production machinery. The distinctive characteristics of semiconductor manufacturing, such as interdependent parameters, fluctuating behaviors over time, and frequently changing operating conditions, pose a major challenge in identifying defective wafers during the manufacturing process. To address this challenge, a multivariate fault detection method based on a 1D ResNet algorithm is introduced in this study. The aim is to identify anomalous wafers by analyzing the raw time-series data collected from multiple sensors throughout the semiconductor manufacturing process. To achieve this objective, a set of features is chosen from specified tools in the process chain to characterize the status of the wafers. Tests on the available data confirm that the gradient vanishing problem faced by very deep networks starts to occur with the plain 1D Convolutional Neural Network (CNN)-based method when the size of the network is deeper than 11 layers. To address this, a 1D Residual Network (ResNet)-based method is used. The experimental results show that the proposed method works more effectively and accurately compared to techniques using a plain 1D CNN and can thus be used for detecting abnormal wafers in the semiconductor manufacturing industry.

Dynamic Correlation Adjacency-Matrix-Based Graph Neural Networks for Traffic Flow Prediction.

Modeling complex spatial and temporal dependencies in multivariate time series data is crucial for traffic forecasting. Graph convolutional networks have proved to be effective in predicting multivariate time series. Although a predefined graph structure can help the model converge to good results quickly, it also limits the further improvement of the model due to its stationary state. In addition, current methods may not converge on some datasets due to the graph structure of these datasets being difficult to learn. Motivated by this, we propose a novel model named Dynamic Correlation Graph Convolutional Network (DCGCN) in this paper. The model can construct adjacency matrices from input data using a correlation coefficient; thus, dynamic correlation graph convolution is used for capturing spatial dependencies. Meanwhile, gated temporal convolution is used for modeling temporal dependencies. Finally, we performed extensive experiments to evaluate the performance of our proposed method against ten existing well-recognized baseline methods using two original and four public datasets.

Anomaly Detection and Repairing for Improving Air Quality Monitoring.

Clean air in cities improves our health and overall quality of life and helps fight climate change and preserve our environment. High-resolution measures of pollutants' concentrations can support the identification of urban areas with poor air quality and raise citizens' awareness while encouraging more sustainable behaviors. Recent advances in Internet of Things (IoT) technology have led to extensive use of low-cost air quality sensors for hyper-local air quality monitoring. As a result, public administrations and citizens increasingly rely on information obtained from sensors to make decisions in their daily lives and mitigate pollution effects. Unfortunately, in most sensing applications, sensors are known to be error-prone. Thanks to Artificial Intelligence (AI) technologies, it is possible to devise computationally efficient methods that can automatically pinpoint anomalies in those data streams in real time. In order to enhance the reliability of air quality sensing applications, we believe that it is highly important to set up a data-cleaning process. In this work, we propose AIrSense, a novel AI-based framework for obtaining reliable pollutant concentrations from raw data collected by a network of low-cost sensors. It enacts an anomaly detection and repairing procedure on raw measurements before applying the calibration model, which converts raw measurements to concentration measurements of gasses. There are very few studies of anomaly detection in raw air quality sensor data (millivolts). Our approach is the first that proposes to detect and repair anomalies in raw data before they are calibrated by considering the temporal sequence of the measurements and the correlations between different sensor features. If at least some previous measurements are available and not anomalous, it trains a model and uses the prediction to repair the observations; otherwise, it exploits the previous observation. Firstly, a majority voting system based on three different algorithms detects anomalies in raw data. Then, anomalies are repaired to avoid missing values in the measurement time series. In the end, the calibration model provides the pollutant concentrations. Experiments conducted on a real dataset of 12,000 observations produced by 12 low-cost sensors demonstrated the importance of the data-cleaning process in improving calibration algorithms' performances.

Industrial Transfer Learning for Multivariate Time Series Segmentation: A Case Study on Hydraulic Pump Testing Cycles.

Industrial data scarcity is one of the largest factors holding back the widespread use of machine learning in manufacturing. To overcome this problem, the concept of transfer learning was developed and has received much attention in recent industrial research. This paper focuses on the problem of time series segmentation and presents the first in-depth research on transfer learning for deep learning-based time series segmentation on the industrial use case of end-of-line pump testing. In particular, we investigate whether the performance of deep learning models can be increased by pretraining the network with data from other domains. Three different scenarios are analyzed: source and target data being closely related, source and target data being distantly related, and source and target data being non-related. The results demonstrate that transfer learning can enhance the performance of time series segmentation models with respect to accuracy and training speed. The benefit can be most clearly seen in scenarios where source and training data are closely related and the number of target training data samples is lowest. However, in the scenario of non-related datasets, cases of negative transfer learning were observed as well. Thus, the research emphasizes the potential, but also the challenges, of industrial transfer learning.

A zero-inflated endemic-epidemic model with an application to measles time series in Germany.

Count data with an excess of zeros are often encountered when modeling infectious disease occurrence. The degree of zero inflation can vary over time due to nonepidemic periods as well as by age group or region. A well-established approach to analyze multivariate incidence time series is the endemic-epidemic modeling framework, also known as the HHH approach. However, it assumes Poisson or negative binomial distributions and is thus not tailored to surveillance data with excess zeros. Here, we propose a multivariate zero-inflated endemic-epidemic model with random effects that extends HHH. Parameters of both the zero-inflation probability and the HHH part of this mixture model can be estimated jointly and efficiently via (penalized) maximum likelihood inference using analytical derivatives. We found proper convergence and good coverage of confidence intervals in simulation studies. An application to measles counts in the 16 German states, 2005-2018, showed that zero inflation is more pronounced in the Eastern states characterized by a higher vaccination coverage. Probabilistic forecasts of measles cases improved when accounting for zero inflation. We anticipate zero-inflated HHH models to be a useful extension also for other applications and provide an implementation in an R package.

Improving performance of deep learning predictive models for COVID-19 by incorporating environmental parameters.

The Coronavirus disease 2019 (COVID-19) pandemic has severely crippled the economy on a global scale. Effective and accurate forecasting models are essential for proper management and preparedness of the healthcare system and resources, eventually aiding in preventing the rapid spread of the disease. With the intention to provide better forecasting tools for the management of the pandemic, the current research work analyzes the effect of the inclusion of environmental parameters in the forecasting of daily COVID-19 cases. Three univariate variants of the long short-term memory (LSTM) model (basic/vanilla, stacked, and bi-directional) were employed for the prediction of daily cases in 9 cities across 3 countries with varying climatic zones (tropical, sub-tropical, and frigid), namely India (New Delhi and Nagpur), USA (Yuma and Los Angeles) and Sweden (Stockholm, Skane, Uppsala and Vastra Gotaland). The results were compared to a basic multivariate LSTM model with environmental parameters (temperature (T) and relative humidity (RH)) as additional inputs. Periods with no or minimal lockdown were chosen specifically in these cities to observe the uninhibited spread of COVID-19 and explore its dependence on daily environmental parameters. The multivariate LSTM model showed the best overall performance; the mean absolute percentage error (MAPE) showed an average of 64% improvement from other univariate models upon the inclusion of the above environmental parameters. Correlation with temperature was generally positive for the cold regions and negative for the warm regions. RH showed mixed correlations, most likely driven by its temperature dependence and effect of allied local factors. The results suggest that the inclusion of environmental parameters could significantly improve the performance of LSTMs for predicting daily cases of COVID-19, although other positive and negative confounding factors can affect the forecasting power.