Probabilistic long-term electricity demand forecasting in South Africa

Abstract

Electricity demand forecasts inevitably differ from the actual electricity demand, hence the task of a forecaster is to make some educated guesses of future electricity demand and to effectively communicate them. Forecasts come with uncertainties whose quantification is as important as the forecasts themselves. Probabilistic modelling accounts for uncertainties in estimation, prediction or in forecasting. Uncertainties in future electricity demand emanate from changing weather conditions, market penetration of renewable sources of electricity, power saving appliances, electric vehicles, escalating cost of electricity, transition of economic sectors away from reliance on energy-intensive towards a more diverse range of services-oriented activity or vice versa and unpredictable economic growth. Short-term electricity demand forecasts are used to ensure system stability. They are used for planning the day-to-day running of the electricity generation system. Medium-term forecasts are used for maintenance planning. They are used to schedule maintenance in such a way that electricity demand is met during maintenance period. Long-term forecasts are used for capital planning. They assist in determining whether the current generation infrastructure will still generate enough electricity to meet future demand. Literature has shown that a universally best electricity demand forecasting technique is non-existent, it can therefore be argued that a good scientific approach to electricity demand forecasting will produce good forecasts. This study involves forecasting electricity demand for long-term planning purposes using both frequentist and Bayesian modelling approaches. Long-term forecasts for hourly electricity demand from 2007 to 2023 are done with in-sample forecasts from 2007 to 2012 and out-of-sample forecasts from 2013 to 2023 and 2013 to 2015 is used to validate the models. Quantile regression (QR) in the frequestist modeling paradigm is used to forecast hourly electricity demand at various percentiles of the distribution of electricity demand. The findings are that the future distributions of hourly electricity demand and peak daily demand would be more likely to shift towards lower demand over the years until 2023. Bayesian structural time series (BSTS) in the Bayesian modeling paradigm is used to forecast hourly distribution of electricity demand. Accurate trend specification is important in long-term forecasting, otherwise erroneous forecasts could be obtained, especially in South Africa where it is difficult to determine if the demand trend would continue a downward trajectory, stabilise or would revert to an upward trajectory. The findings are that future South African hourly demand until 2023 are less likely to exceed the highest historical hourly demand of 36 826kW. The South African electricity demand from Eskom are more likely to maintain the downward trend until 2023. The developed generalised additive mixed quantile averaging (GAMMQV) model indicates that demand for electricity is unlikely to exceed the highest (36 826 kW) hourly electricity demand until 2023. The probability of electricity exceeding 36 826kW was below 0.15 from 2013 to 2023. The GAMMQV model gave better point forecasts than BSTS and QR.