Short-Term Load Foresting Using Combination of Linear and Non-Linear Models

Numerous short-term load forecasting models are available in the literature. However, the improvement in forecast accuracy using the combination models has yet to be analyzed on a daily rolling basis for a very long test period. In this paper, the characteristics of a combination of the Seasonal Autoregressive Integrated Moving Average (SARIMA) – a linear model and Radial Basis Function networks (RBFN) – a non-linear model have been studied in two different modeling frameworks, namely single series (SS) and variable segmented series (VSS). The hourly load data from the Ontario Electricity Market (OEM) and the Iberian Electricity Market (MIBEL) are used for the analysis. This dataset spans 12 years for OEM and one year for MIBEL. The impact on prediction accuracy by the size of training data and the combining individual forecasts has been studied for both markets. To achieve the empirical objective, a large number of models(1,447,740 in number) are estimated to produce load forecasts on a daily rolling basis. The forecast performance has been compared with the other models proposed in the literature. Among the linear models, for all window sizes of training data, the forecast accuracy of the combination model is better than the model selected with the minimum Akaike information criterion (AIC) and Bayesian information criterion (BIC) in both frameworks. Moreover, the ensemble of RBFN and linear models produces the best forecast. The results pinpointed that the proposed model’s precision and stability are higher than the earlier forecasting models proposed for both markets. The novelty in the model is that only a single hourly time series is used for forecasting, and there is no need for other explanatory variables.