Assessment of Load Forecasting Uncertainties by Deterministic and Probabilistic LSTM Methods with Meteorological Data as Additional Inputs

This paper evaluates two approaches for assessing and quantifying uncertainties in electricity demand forecasting (load forecasting), which are both based on a long short-term memory (LSTM) method. The first is a "meteorological index (MI)" method, which modifies conventional deterministic (point-based) LSTM approach by using meteorological data as additional inputs to obtain ranges of corresponding demand variations from hindcasted residual distributions. The second approach is "probabilistic neural network (PNN)", which adds an additional normal distribution layer at the output of the LSTM model to estimate ranges of demand variations. Using autoregressive model as a reference (based on only demand data), impact of different meteorological parameters is in both approaches evaluated for different input data series (temperature, solar irradiance and wind speed). Obtained results show that ambient temperature has the greatest impact on demand variations and although the PNN methods have better overall performance, the MI-based methods allow to quantify uncertainties due to specific meteorological parameter(s), i.e., to provide "explainable" uncertainty forecasts by assessing "parameter uncertainties".