Assessment of deep learning and classical statistical methods on forecasting hourly natural gas demand at multiple sites in Spain
| Autor: | Javier Segovia, Aniqa Rehman, Paul R. Anderson, Junjie Zhu |
|---|---|
| Rok vydání: | 2022 |
| Předmět: |
Energy demand
business.industry Mechanical Engineering Deep learning Building and Construction Demand forecasting Missing data Pollution Industrial and Manufacturing Engineering General Energy Natural gas Statistics Environmental science Artificial intelligence Electrical and Electronic Engineering business Extreme value theory Civil and Structural Engineering |
| Zdroj: | Energy. 244:122562 |
| ISSN: | 0360-5442 |
| DOI: | 10.1016/j.energy.2021.122562 |
| Popis: | Prediction of natural gas demand can help to better manage energy demand and supply and recent developments of deep learning methods make it possible to improve forecast performance. This study examined the feasibility of hourly natural gas demand forecast, and compared statistical and deep learning methods to evaluate their prediction performance at five different sites in Spain. Hourly forecast can achieve an adjusted R2 ∼0.99 and MAPE lower to 2.7%. SMLR yields high prediction accuracy (MAPE: 3%–10%) in four sites, but suffers issues of missing data (5%–17%) and has relatively more extreme predictions (653 observations; ±100% away from the values). MLP has less amount of extreme predictions (517 observations) with a similar accuracy (MAPE: 3%–11%), but still suffers issue of missing data (11%–28%). LSTMs also achieves good prediction accuracy (MAPE: 3%–13%) and is able to manage most extreme values. Other methods are generally less promising but can be site-specific. Understanding their distinctive characteristics can help decision-makers to rule better decisions. Exploration of future forecasts based on LSTMs shows promising (adjusted R2: 0.90–0.99; MAPE: 11%–32%) in near future ( |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|
Full Text from ScienceDirect