Multivariate dynamic regression: modeling and forecasting for intraday electricity load

Abstract

This paper introduces electricity load curve models for short-term forecasting purposes. A broad class of multivariate dynamic regression models is proposed to model hourly electricity load. Alternative forecasting models, special cases of our general model, include separate time series regressions for each hour and week day. All the models developed include components that represent trends, seasons at different levels (yearly, weekly, etc.), dummies to take into account weekends/holidays and other special days, and short-term dynamics and weather regression effects, discussing the necessity of nonlinx ear functions for cooling effects. Our developments explore the facilities of dynamic linear models such as the use of discount factors, subjective intervention, variance learning and smoothing/filtering. The elicitation of the load curve is considered in the context of subjective intervention analysis, which is especially useful to take into account the adjustments for daylight savings time. The theme of combination of probabilistic forecasting is also briefly addressed.variance learning, is a novelty, as far as we know, in the area of electricity load forecasting. All the models developed include components that represent trends, seasons at different levels (yearly, weekly etc.), dummies to take into account weekends/holidays and other special days, and short-term dynamics and weather regression effects, including nonlinear functions for cooling effects. A comparison between multivariate and univariate models shows that they mainly differ in computational cost. From the point of view of the distribution network operator, this is relevant information, especially because the univariate models takes less time to be fitted.Special care is taken with the role played by weekends/holidays and their effect on the estimated coefficients. In particular, we explore the similarities between the shapes of load curves occurring on the day just before and just after weekends/holidays. Our developments explore the facilities of dynamic linear models such as the use of discount factors, subjective intervention, variance learning and smoothing. The elicitation of the load curve is considered in the context of subjective intervention analysis, which is especially useful to take into account the adjustments for daylight savings time.More precisely, we model a suitable transformation of the electricity load at hour and day t as a linear combination of the transformed load at the same hour on the previous day and appropriate functions of the prevailing temperatures on day t . The coefficients depend on the pair . ; t / in a continuous nonlinear manner. The paper ends by briefly addressing the theme of combination of probabilistic forecasting.The remainder of the paper is organized as follows. A review of alternative models is presented in Section 2. In Section 3, we explore the data set analyzed. Some stylized facts always presented in the load electricity data are also reviewed. Our application to Brazilian southeast hourly electrici...

Time series analysis and forecasting of China’s energy production during Covid-19: statistical models vs machine learning models

“…For traditional time-series models such as ARIMA, covariates can only be added disguised by error-corrected regression, i.e., dynamic regression models, also known as ARIMAX; however, these models have been shown to have limited improvement in model performance in energy forecasting, and in some cases are even weaker than ordinary ARIMA models (Dehghani et al 2017;Migon and Alves 2013).…”

Time series analysis and forecasting of China’s energy production during Covid-19: statistical models vs machine learning models

“…For traditional time-series models such as ARIMA, covariates can only be added disguised by error-corrected regression, i.e., dynamic regression models, also known as ARIMAX; however, these models have been shown to have limited improvement in model performance for energy forecasting, and in some cases are even weaker than ordinary ARIMA models (Dehghani et al 2017;Migon and Alves 2013).…”

On‐Line Warning System for Pipe Burst Using Bayesian Dynamic Linear Models

“…This approach has been used in engineering, epidemiology, economics, to mention a few examples. Examples for applications of DLM are diverse: modeling electric load curves (Migon & Alves, 2013), hydrological forecasting (Ciupak et al., 2015; Ravines et al., 2008) and forecasting of epidemiologic time series for public health surveillance (Nobre et al., 2001). See Schmidt and Lopes (2019) for a recent review.…”