|
Forecasting electricity consumption in the United Kingdom |
|---|---|
| รหัสดีโอไอ | |
| Title | Forecasting electricity consumption in the United Kingdom |
| Creator | Yanisa Peerabenjakul |
| Contributor | Jirachai Buddhakulsomsiri, Advisor |
| Publisher | Thammasat University |
| Publication Year | 2568 |
| Keyword | Forecasting, Time series, Regression, Machine learning, Neural network |
| Abstract | Producing highly precise predictions for immediate power grid requirements is essential for maintaining stable utility operations and facilitating strategic energy management. To address this need, this research projects 30 minute load intervals across the United Kingdom by analyzing historical usage logs from the year 2023. Given the collection frequency, this temporal dataset yields exactly 48 distinct measurement points within every 24 hour cycle. The main objective is to compare the forecasting performance of classical time-series models, a machine-learning model, and an Ensemble forecasting approach under the same experimental framework. Four forecasting models were developed and evaluated: SARIMA, Holt–Winters, Artificial Neural Network (ANN), and Ensemble model. The SARIMA model was applied to capture autoregressive and seasonal patterns, while the Holt–Winters model was used to model level, trend, and seasonal components. The ANN model used lagged electricity demand variables as inputs to capture nonlinear demand patterns. The Ensemble model combined the forecasts from SARIMA, Holt–Winters, and ANN using a simple averaging method. Model performance was evaluated using MAE, RMSE, MAPE, and MASE. The results show that the ANN model achieved the best forecasting performance on the testing dataset, with MAE of 344.479 MW, RMSE of 454.746 MW, MAPE of 1.06%, and MASE of 0.2081. The Ensemble model ranked second, with a testing MAPE of 6.84%, outperforming SARIMA and Holt–Winters but not ANN. The findings indicate that ANN was the most suitable model for this study, while Ensemble forecasting can improve traditional statistical models but may require weighted or advanced combination methods to achieve better accuracy. |
