Short term net load forecasting using computational intelligence techniques
Net load forecasting has become increasingly complex due to the high penetration of renewables and evolving data subject to so-called concept drift, i.e., sudden and large changes in the energy flow pattern. Accurate net load forecasting is essential to prevent unexpected imbalances across all volta...
Uloženo v:
| Vydáno v: | 7th E-Mobility Power System Integration Symposium (EMOB 2023) Ročník 2023; s. 139 - 145 |
|---|---|
| Hlavní autoři: | , , , , |
| Médium: | Konferenční příspěvek |
| Jazyk: | angličtina |
| Vydáno: |
The Institution of Engineering and Technology
2023
|
| On-line přístup: | Získat plný text |
| Tagy: |
Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
|
| Shrnutí: | Net load forecasting has become increasingly complex due to the high penetration of renewables and evolving data subject to so-called concept drift, i.e., sudden and large changes in the energy flow pattern. Accurate net load forecasting is essential to prevent unexpected imbalances across all voltage levels of the electricity grid, as well as to promote the stability and reliability of the power system. Therefore, the use of accurate forecasting techniques is essential to manage and optimize the use of available resources at the TSO-DSO interface. This paper evaluates several forecasting methods, including an adaptive random forest method based on incremental learning incorporating a drift detector, a method based on a recurrent neural network using long short-term memory (LSTM), and a method based on an ensemble of models including decision trees (DT), support vector machines (SVM), extreme gradient boosting algorithms (XGBoost), and Lasso regressions. The experiment was conducted using the net load data collected at the TSO-DSO interface in Portugal, where concept drift can be observed, possibly due to increasing integration of distributed energy resources behind the meter. The study examined two scenarios. In the first scenario, the models were trained using a large training set that included significant drifts, while in the second scenario, the models were trained prior to the occurrence of the drifts. The results showed that the approach using the adaptive model is more robust to the concept drift and performs better compared to the other traditional methods, especially in scenarios where there are significant changes in the net load patterns over time. |
|---|---|
| DOI: | 10.1049/icp.2023.2696 |