Publication Date: 2024/01/03
Abstract: As one-third of the world's energy consumption, crude oil is vital to the global economy, yet because of its volatility and complexity, it is still difficult to estimate its price.Although machine learning models might enhance forecasts, they are not impervious to unanticipated shocks, geopolitical events, and uncertainty in the world economy. Few studies have employed hybrid models to increase prediction accuracy, despite a large body of research on machine- learning models' potential to improve forecasting. The current approach for predicting petroleum prices forecasts for a short time horizon (10 days) by ignoring outside data that could enhance the prediction performance. Though useful for many in the oil and gas sector, short-term petroleum price forecasting has some limitations and challenges of its own, including limited accuracy, volatility and uncertainty, and a potential inability to fully account for the unpredictable effects of government policies on petroleum prices. In the oil and gas sector, medium- to long-term predictions may offer more consistent and trustworthy direction for strategic planning. This degree of selective attention is also lacking in current skip connection-based forecasting algorithms. When the network is producing predictions, attention mechanisms enable it to choose focus on distinct segments of the input sequence. As a result, the skip connection increases the model's computational complexity, necessitates a large amount of memory, adds noise and redundancy, and needs to be carefully designed and tuned to fit the network architecture and data domain. In order to increase the accuracy of petroleum price predictions, this study suggests combining the benefits of long short-term memory (LSTM), CNN, and attention connection. The proposed model outperformed the classical skip base CNN-LSTM algorithm, which came in second place with an MAE and RMSE of 0.0231 and 0.0297, and skip base CNN- GRU, which achieved the highest MAE and RMSE of 0.0236 and 0.0318, respectively, according to experimental results on MATLAB 2022a. The proposed model also achieved the lowest MAE and RMSE values of 0.0175 and 0.0199.
Keywords: Attention; Deep Learning; Convolutional Neural Network; Long Short-Term Memory; Machine Learning; Petroleum Price and Forecasting.
DOI: https://doi.org/10.5281/zenodo.10453152
PDF: https://ijirst.demo4.arinfotech.co/assets/upload/files/IJISRT23DEC820.pdf
REFERENCES