1. Introduction
Grid operators require accurate day-ahead demand forecasts for unit commitment and reserve planning, and forecasting error directly translates into either costly reserve over-procurement or reliability risk, motivating continued refinement of hybrid statistical-and-deep-learning approaches.
2. Methodology
Three years of hourly demand data from a regional distribution utility was used to train a hybrid forecasting pipeline in which Prophet first decomposed the series into trend, weekly and holiday-effect components, with the residual series then modelled by a two-layer LSTM network, and forecasts reconstructed by summing the Prophet trend-seasonal component with the LSTM residual prediction.
3. Results
The hybrid LSTM-Prophet model achieved a MAPE of 2.9 percent for 24-hour-ahead forecasts on the held-out test year, compared with 3.8 percent for a standalone LSTM and 4.6 percent for standalone Prophet, with the largest relative improvement observed around public holidays where Prophet calendar effects corrected LSTM under-forecasting.
4. Conclusion
Decomposing calendar-driven seasonality explicitly before applying sequence modelling improves short-term electricity demand forecast accuracy over either method alone. Future work will incorporate temperature forecast uncertainty into the pipeline.
References
[1] Taylor S. J. and Letham B., Forecasting at scale (Prophet), The American Statistician, 2018. [2] Hochreiter S. and Schmidhuber J., Long short-term memory, Neural Computation, 1997.