Abstract
The science of load-forecasting (LF) requires that electric power utilities estimate load profile in advance as electricity needs to be generated when there is a demand for it. Typically, power demand varies according to economic sectors, therefore, power utilities must apply LF to avoid underestimation or overestimation of load – which also can increase operating expenditure (OPEX). However, unanticipated global outbreaks such as the 2019 coronavirus disease (COVID-19) pandemic have demonstrated that the normal LF is inadequate during such phenomena and is incapable of predicting the required load for all energy sectors. The pandemic caused severe dramatic changes to major sectors of the economy, especially the energy sector. As a consequence, governments across the world have implemented several restrictions to curb the spread of the virus, some of which include lockdowns, with lockdown levels having different impacts on power consumption. These actions have resulted in normal LF patterns becoming inadequate and insufficient for each lockdown level leading to power utilities experiencing huge financial losses around this period.
This study investigates the impact of COVID-19 on commercial areas, and how to classify corresponding load profiles for LF under different COVID-19 lockdowns in South Africa, to mitigate future losses on power utilities during lockdowns. The load profile datasets used for this investigation were obtained from a local municipal utility provider City Power, Johannesburg, measured in kWh at 30-minute interval for a period of two years (2019 and 2020). This pandemic problem was firstly analyzed statistically, by analyzing 2019 and 2020 datasets, for a commercial area of interest. Due to their modelling suitability, two statistical distributions – normal and lognormal distributions – are employed to analyze the attributes of peak and average loads using daily, monthly, and annual datasets. The performance of each model was investigated by using error functions that seek to quantify the best distribution. The results demonstrated that these two distribution models performed better with the 2019 dataset, as compared to the 2020 dataset.
The study goes further to investigate and validate the 2020 results using another set of data from a different commercial area with dissimilar geographic and physical characteristics. The investigation revealed that power consumption is generally higher in 2019, compared to 2020. By employing dataset similarity indices, datasets from these two areas are found exhibit high similarities, especially during episodes of intense lockdowns in 2020. Additionally, it was observed that reduction in power consumption at these commercial areas is dependent on the implemented COVID-19 lockdown level. Therefore, the 2020 dataset was further re-classified according to different lockdown levels (levels 1 to 5) during the COVID-19 first wave, starting from March to November 2020. Time series classification tools were employed on these datasets to investigate the underlying characteristic behaviour and unique features using K-nearest neighbour (KNN) algorithm and dynamic time warping (DTW) for each lockdown level.