DEEP LEARNING MODELS FOR NATURAL GAS DEMAND FORECASTING: A COMPARATIVE STUDY OF MLP, CNN AND LSTM
| Autor: | Artemis Aidoni, Konstantinos Kofidis, Catalina Lucia Cocianu, Lazar Avram |
|---|---|
| Jazyk: | angličtina |
| Rok vydání: | 2023 |
| Předmět: | |
| Zdroj: | Romanian Journal of Petroleum & Gas Technology, Vol 4, Iss 1, Pp 133-148 (2023) |
| Druh dokumentu: | article |
| ISSN: | 2734-5319 2972-0370 |
| DOI: | 10.51865/JPGT.2023.01.12 |
| Popis: | This study aims to investigate the use of various deep learning techniques to predict future residential natural gas consumption in Italy, with a particular emphasis on the correlation between gas consumption and temperature. Four models were evaluated, including Multi-layer Perceptron (MLP), Convolutional Neural Network (CNN), Simple Long-Short Term Memory (LSTM), and Stack-LSTM, with the latter chosen due to its two-layer LSTM and potential to improve forecasting accuracy. Feature scaling was conducted with the MinMaxScaler method to ensure uniform values among variables. Statistical analysis was performed using Mean Squared Error (MSE), Mean Absolute Error (MAE), and R-squared accuracy metrics, with ANOVA tests and boxplots, used to visualize the distribution of accuracy metrics across test and full datasets. Results implied that the CNN and Stack-LSTM models were more effective in accurately predicting the target variable compared to the other models, as indicated by MSE and R-squared scores, as well as graphical comparisons of actual and predicted values. Finally, the research recommends the utilization of supplementary features in future research to increase the precision of forecasts. |
| Databáze: | Directory of Open Access Journals |
| Externí odkaz: |
|