 |
International Society of Science and Applied Technologies |
| Machine Learning with Functional Outputs: A Framework for Power System Reliability Forecasting | ||||
| Author |
Yu Zhou
|
|||
| Co-Author(s) |
Sen-hui Wang; Xiang-xia Li
|
|||
| Abstract | Load forecasting is a key task in power system scheduling and reliability analysis. Compared with load values at individual time points, the target daily load curve provides a more complete description of the continuous variation and overall pattern of electricity demand over a day. This paper proposes a Functional Long Short-Term Memory Network (FLSTM) framework for daily load curve forecasting with functional response. The framework represents functional outputs through basis function expansion and designs a unified feature extraction and fusion structure for historical load sequences, target-day temperature, and the day-of-week one-hot vector. We further compare two functional-output training strategies: coefficient-space training and function-space training. Experimental results based on load and temperature data from the DUQ region of the PJM electricity market show that function-space training outperforms coefficient-space training under all three input scenarios. Compared with the discrete-output baseline and the functional linear baseline, FLSTM achieves lower seasonal-average MAPE and RMSE, demonstrating its effectiveness in improving the forecasting accuracy and reconstruction quality of daily load curves.
|
|||
| Keywords | Functional data analysis, functional output, load curve forecasting, Functional Long Short-Term Memory Network | |||
| Article #: RQD2026-247 | ||||
Proceedings of 31st ISSAT International Conference on Reliability & Quality in Design |