Application of Hybrid Methods Combining Recurrent Neural Networks and Modified Metaheuristics for Time-Series Forecasting

Time-series forecasting is becoming very important in many different industries, including cybersecurity, healthcare, energy, finance and cloud computing. This thesis explores the use of advanced metaheuristic algorithms together with recurrent neural networks (RNNs) to produce hybrid models enhancing the accuracy, stability, and generalizing ability of forecasting. We address problems including nonlinear behavior, seasonality, and noise in time-series data by utilizing the benefits of both Recurrent Neural Networks (RNNs) and metaheuristics.
This work uses an enhanced Moth Flame Optimizer (MFO) to improve the optimization of logistic regression models, which boosts the accuracy of spam email classification. Furthermore, the efficiency of hybrid models in capturing intricate seasonal trends is demonstrated by the use of an Enhanced Sine Cosine Algorithm (ESCA) in conjunction with Long Short-Term Memory (LSTM) networks for wind energy ahead-of-time prediction. With a modified Reptile Search Algorithm, Convolutional Neural Networks (CNNs) and Extreme Gradient Boosting (XGBoost) show potential in hybrid techniques in enhancing the security of IoT networks in the field of cybersecurity.
Moreover, the practical benefits in the field of healthcare are shown by the identification of seizures in electroencephalogram (EEG) data using metaheuristic-optimized recurrent neural networks (RNNs). The study on improving the performance of LSTM networks utilizing the Enhanced Harris Hawks Optimization Algorithm for crude oil price prediction underlines the need to integrate different methodologies in financial modelling. Using modified metaheuristic-optimized LSTM networks for earthquake magnitude prediction emphasizes the several applications of these hybrid approaches in preventing disasters.
Additionally, this study addresses the challenges associated with cloud instance pricing, which are increased by the complex nature of budgeting due to fluctuating demand and changing costs. A modified optimizer is introduced and tested on cloud instance data using multi-headed recurrent architectures, specifically LSTM and Gated Recurrent Unit (GRU) networks. The GRU model, enhanced by this technique, attained remarkable forecasting IV precision, validated through statistical analysis and revealed by using artificial intelligence techniques to clarify feature importance.
The thesis significantly improves the current degree of knowledge in time-series data prediction, resulting in new possibilities for the application of hybrid methods across various areas. By addressing the limitations of previous approaches and using the powers of RNNs and metaheuristics, this work helps predictive modelling and its useful applications to advance.