Comparison of Two Modified Deterministic LSTM Models with a Probabilistic LSTM Model for a Day-Ahead Forecasting of Electricity Demands

This paper presents two simple modifications of a standard deterministic LSTM method, which are specifically implemented for the probabilistic day-ahead forecasting of electricity demands. Both modifications feature a two-stage model, where a standard deterministic LSTM method is built in the first stage, and then used in the second stage in a hindcasting-based evaluation of the model uncertainty (first model modification) and model parameter uncertainty (second model modification). The estimated uncertainty ranges from both modified models are added to the deterministic predictions of the first-stage LSTM model, to provide probabilistic demand forecasting. The two presented modified deterministic LSTM models are discussed in detail and compared with a single-stage probabilistic LSTM forecasting model, to evaluate their performance and suitability for a day-ahead probabilistic electricity demand forecasting. The obtained result confirm that the second modification and probabilistic LSTM model are best in capturing ranges of the uncertainties in demand variations.