Very short term load forecasting of residential electricity consumption using the Markov-chain mixture distribution (MCM) model

This study utilizes the Markov-chain mixture distribution model (MCM) for very short term load forecasting of residential electricity consumption. The model is used to forecast one step ahead half hour resolution residential electricity consumption data from Australia. The results are compared with Quantile Regression (QR) and Persistence Ensemble (PeEn) as advanced and simple benchmark models. The results were compared in terms of reliability, reliability mean absolute error (rMAE), prediction interval normalized average width (PINAW) and normalized continuous ranked probability score (nCRPS). For 10 steps conditioning for QR and PeEn, the MCM results were on par with QR, and superior to PeEn. As a sensitivity analysis, simulations were performed where the number of data points for conditioning QR and PeEn was varied and compared to the MCM output, which is based on only one data point for conditioning. It was shown that in terms of nCRPS and rMAE the QR results converged towards the MCM results for lower number of conditioning points included in QR. The nCRPS of PeEn never reached the superior MCM and QR results, but in rMAE, for number of conditioning points above 24, PeEn was the most reliable. Based on the sparse complexity design of MCM, high computational speed and competitive performance, it is suggested as a candidate for benchmark model in probabilistic forecasting of electricity consumption.

Automated summary

This study utilized the Markov-chain mixture distribution model (MCM) for very short term load forecasting of residential electricity consumption and compared the results with Quantile Regression (QR) and Persistence Ensemble (PeEn). The MCM showed competitive performance and reliability compared to the benchmark models QR and PeEn, indicating its potential as a candidate for probabilistic forecasting of electricity consumption.