Page - 68 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

2. Methodology Literature does not show any consensus on the best approach to electricity demand forecasting. The range of different approaches used recently includes classical time series models [3,6,8,10,12,17,25,26,27], and machine intelligence framework [5,15]. Although within each one of these categories the sophistication of the applied techniques can be qualified as high, in this chapter, I exclusively consider the most basic time series models i.e. naive and exponential smoothing as benchmarks and ARMAX for the inclusion of explanatory variables. 2.1. Benchmarks methods The simplest benchmark method in forecasting exercises is often known as naïve method, assuming that the forecasted observation is the last real observation available, and is also the simplest benchmark method used in this work. Additionally, the so called Exponential smoothing method is based on the idea of separating the time-series trend from its random disturbance, that is, it ‘‘smoothes” series behavior. It is important to remark that in this method the models are usually constructed based on empirical reasoning. The Winter’s method is one of several exponential smoothing methods that can analyze seasonal time series directly. A very interesting survey for this method can be found in [11]. Examples of this method’s application to electric load forecasting problem can be found in [23,27,13]. 2.2. Multiple regression & time series models A basic conventional structure decomposes the observed load into four components: the normal load, the weather sensitive part, special events, and a random component. Assuming a conventional aggregated energy demand relationship [3,4], a log-linear model can be analytically expressed as: lnCt= pt+st+CSDt+CWEAt+ut (1) Where tC denotes the electricity consumption on day t; pt is the trend and st (part of) the deterministic pattern; CSDt represents special days; CWEAt refers to the meteorological variables, and ut is the disturbance term. The diagnostic of the transitory dynamics displayed by ut term is performed using the ARMA structure. A plot of the autocorrelation function and partial autocorrelation function and some conventional tests, like the Augmented Dickey Fuller test, are used to decide whether a data series is stationary or not. Thus, an ARMAX (p, q, b) model for the electricity load can be also represented as: tttt εuqθXb Cp )()(ln)( (2) Where )(p , tXb)( and )(qθ are the lag polynomials for the natural logarithm of the electricity demand (C), the exogenous variables matrix (X) (which is formed by the variables p, s, CSD, and CWEA) , the moving average term (u) and ε is white noise. M.BakhatandJ.RosselloNadal / ImprovingDailyElectricityLoadsForecasting68