A hybrid approach to time series forecasting

This study presents a hybrid framework that combines deep learning and metaheuristic methods for the analysis of time series with different characteristic patterns. The hybrid framework integrates Particle Swarm Optimization (PSO) and Long Short-Term Memory (LSTM) methods. It involves the adaptive use of the stochastic gradient descent optimization algorithm used in traditional Long Short Term Memory models with the learning rate Particle Swarm Optimization meta-heuristic. This approach provides an efficient future forecasting approach for Wikipedia web traffic time series of different languages. LSTM is one of the DL methods used to model contextual dependencies in time series. PSO, on the other hand, is a metaheuristic optimization method widely used for time series analysis. This method provides an optimization process to find the best parameter values for future predictions. For the LSTM model, hyperparameters such as the number of units, learning rate, and number of iterations are important, while for the PSO algorithm, parameters like individuality factor, social factor, swarm weight factor, number of particles, and number of iterations are crucial. Properly selecting these hyperparameter values is of vital importance to ensure the model fits the dataset and prevent issues such as overfitting or underfitting. In this study, performance evaluation was conducted using different LSTM architectures and optimization methods for time series analysis. The methods were applied to web traffic time series data. Firstly, experiments were carried out using the Unidirectional LSTM model. Subsequently, Bidirectional LSTM and Multilayer LSTM models were also evaluated. Upon examining the results, it was observed that the Unidirectional LSTM model provided the lowest Mean Squared Error (MSE) values. In the experiments with the LSTM model, it was observed that the Stochastic Gradient Descent (SGD) optimization algorithm yielded good results in high iterations. However, to achieve better results in low iterations, the PSO-LSTM hybrid method was employed by adjusting the learning rate. The PSO algorithm was used for hyperparameter optimization to determine the optimal learning rate. The obtained results demonstrate that the analysis performed using PSO and LSTM outperformed other traditional methods in terms of performance. Especially in terms of MSE error metrics, it has been observed that the hybrid PSO-LSTM method achieves lower error values. This finding indicates that the hybrid approach provides a more effective combination for modeling time series data and making future predictions. This study makes a significant contribution to the evaluation of different LSTM architectures and optimization methods for time series analysis. The approach of trying different hyperparameter values can lead to more reliable and better-performing results for both LSTM and PSO. By doing so, a more comprehensive research can be conducted to improve the overall performance of the model and obtain the most suitable model for the dataset.