DESIGN AND DEVELOPMENT OF TIME SERIES FORECASTING MODELS FOR COVID-19 PREDICTION

Abstract

The Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) trig- gered the COVID-19 pandemic, which became a global health crisis with severe impacts on humanity. The first wave of COVID-19 was reported in most countries at the beginning of 2020, and the World Health Organization (WHO) declared it a pandemic on March 11, 2020. During the early phase of the pandemic, most countries relied on non-pharmaceutical interventions, such as bans on in- ternational travel, mandatory face masks, quarantine protocols, contact tracing, and complete lockdowns, to curb the spread of the virus. Governments faced numerous challenges, including educating the public about the pandemic, man- aging resources, ensuring medical facilities, and addressing economic impacts. Therefore, identifying future cases and predicting the spread of the virus became critical for healthcare systems to take proactive measures and minimize casual- ties. Consequently, predictive analysis of pandemics emerged as a vital research area, aiding healthcare services and governments in planning and controlling the spread of COVID-19. The highly contagious SARS-CoV-2 virus spread rapidly and evolved into numerous mutants and variants, leading to second and third waves of infections across many countries. Over time, vaccines were developed to mitigate the impact of COVID-19, adding further complexity to the dynamics of COVID-19 time series data. This has underscored the importance of developing models capable of handling highly dynamic and non-stationary data for accurate time series forecasting associated with multiple waves of the pandemic driven by successive mutations of the virus. In this thesis various aspects of the COVID-19 pandemic related to time series forecasting and modeling are analyzed using state-of-the-art methods and novel forecasting models. Initially, the well-known forecasting models namely, Autoregressive Integrated Moving Average (ARIMA), Facebook Prophet (FB- Prophet), Exponential smoothing models (ETS), Artificial Neural Network (ANN), and Long Short-Term Memory (LSTM) are evaluated and compared using diverse datasets spanning different timelines of the pandemic. LSTM outperformed all the other compared models for the time series forecasting of the COVID-19 cases. Fuzzy time series (FTS) models are particularly effective at handling uncertain and imprecise time series data, particularly when the underlying patterns are nonlinear. Evolutionary optimization has proven to be a powerful approach for hyperparameter tuning, achieving strong results in complex problem-solving with minimal computational expense. There are three main hyperparameters of a FTS model:- i) number of intervals, ii) length of intervals, iii) fuzzy order. Therefore, two new algorithms based on FTS with hyperparameter optimization using PSO are proposed, namely, nested-FTS-PSO, and exhaustive-search-FTS- PSO. The forecasting results of the algorithms are compared with ARIMA, v FB-Prophet, FTS, and FTS-PSO models using COVID-19 cases from 10 highly affected countries. Exhaustive-search-FTS-PSO algorithm outperformed all the other compared models. The integration of fuzzy techniques with deep learning has emerged as a promising research area, as it enhances both interpretability and explainability of deep learning based systems. Consequently, the development of a hybrid fuzzy time series forecasting model that combines FTS, deep learning, and swarm intelligence is envisioned to achieve more accurate time series forecasting of dynamic, non-stationary data pertaining to multiple waves of the pandemic caused by successive mutations of the virus. Deep learning models, specifically, stacked- LSTM, bidirectional-LSTM, convolution-LSTM, attention-LSTM, attention-bi- LSTM are integrated with FTS and PSO. The hybrid models are compared with the ETS, ARIMA, ANN, LSTM, and FTSF-PSO models. Hybrid of FTS, PSO, and attention-bi-LSTM outperformed all the other compared models on the USA and India COVID-19 datasets. Compartmental epidemiological models are among the most traditional and widely used approaches to represent the progression of an epidemic. Advance- ments in Artificial Intelligence (AI) hold significant potential to aid in combating pandemics. Fully leveraging the capabilities of AI, epidemiological modeling, and optimization techniques in an integrated forecasting solution is crucial for predicting the impact of a pandemic. Therefore, Susceptible-Infected-Recovered- Deceased (SIRD) epidemiological model integrating with PSO and deep learning (stacked-LSTM) is proposed to model the evolution of the COVID-19 pandemic in India, UK and the USA. Time-varying model parameters are used to deal with multiple waves of the COVID-19. The proposed hybrid model outperformed the stacked-LSTM and hybrid of SIRD and PSO. Further, a novel epidemiological compartmental model, which provides realistic projections of epidemic spread based on viral characteristics, is proposed, that incorporates time-varying hyper- parameters and deep learning models. The COVID-19 time series data is highly dynamic in nature due to rapidly changing transmission rates and government policy measures such as lockdowns and vaccination campaigns. Recognizing the numerous factors influencing epidemic spread, a 10-compartmental epidemiological model is presented, incorporating restriction policies, multi-dose vaccinations, and vaccine efficacy. COVID-19 case studies of the USA and India are carried out for demonstrating the efficacy of the proposed approach. The proposed approach outperformed ETS, ARIMA, ANN, LSTM, and SEIRD models in the performance evaluation. Future work may enhance the forecasting accuracy of the proposed models by integrating advanced optimization algorithms and deep learning techniques. Time series forecasting models have the potential to drive changes in the society by enhancing decision-making processes and optimizing resource allocation across various industries. From this perspective, the models introduced in this thesis can be utilized across various fields, including economics, healthcare, and public policy, which can lead to significant positive societal outcomes.