Unearthing Details of Time Series of Load: A Dual-scale Input Structured LSTM Approach.

Ultrashort term load forecasting for fine-grained residential load has been a challenge due to the uncertainty of residential electricity consumption behaviors. It is a good idea to mine more power consumption details for precise forecasting. Long short-term memory network (LSTM) can do well in characterizing general load shapes referring to daily life routine, but is limited in characterizing peaks and valleys referring to detailed knowledge lying in event-driven activities. To unearth more details, a new input structure named event window (EW) is designed based on reliable analysis that explains a symbiotic relationship between randomness and determinacy of household consumptive activities. Combining an EW structured LSTM to the conventional LSTM, this paper proposes a dual-scale input structured LSTM ultrashort term load forecaster (DILF). The DILF shows satisfactory increased performance in forecasting accuracy in the experiment with an increase in the proportion of training data carrying details led by the synthetic input. More understanding is delivered via cases analyses on the two original networks, i.e., conventional LSTM and pure EW structured LSTM. Finally, applicability of the dual-scale input structure to other neural networks is verified in the experiment.