Browsing by Author "Weyermuller, Evan"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Adaptive neuro-fuzzy inference systems and data clustering for short-term load forecasting in a capacity constrained market
    (Stellenbosch : Stellenbosch University, 2021-12) Weyermuller, Evan; Vermeulen, Johan; Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering.
    ENGLISH ABSTRACT: Accurate short-term load forecasting of electrical power systems is of critical importance when it comes to accurately managing and controlling a modern grid. It allows for the effective allocation and dispatch of generation capacity in order to meet demand in the most efficient and cost-effective manner. This problem is even more pertinent in South Africa as there are severe generation capacity constraints, which demand the need for crippling load reduction measures like load shedding. An adaptive neuro-fuzzy inference system (ANFIS) model for day-ahead short-term load forecasting (STLF) was designed. The model consisted of 24 separate hourly sub-models, accepting the month, day of the week, and the load value from 24 hours prior as inputs. Holidays were input as day type 8 to distinguish them from normal weekdays. This model was then applied to the South African national load profile dataset and was able to achieve a mean absolute percentage error (MAPE) of 1.866% for the ‘normal’ years of 2005–2007, and 1.424% for the capacity-limited years of 2008–2019. These accuracy values compare favourably to other models in literature and the accuracy actually improves in the later years during the energy crisis. This model is able to accurately predict all daily, weekly, and yearly cycles in the demand profiles, while remaining minimalistic enough to be trained on a consumer-grade computer. A clustered ANFIS model was also developed. The daily load profiles were optimally clustered into two distinct shapes: one primarily weekdays, and the other primarily weekends and holidays. A random forest model was designed to predict the day-ahead cluster index and achieved 99.36% accuracy. Separate ANFIS models per cluster were then trained using similar inputs to the model above. The clustered model was, however, unable to achieve any better accuracy and was very erratic and unstable in its forecast. Therefore, it can be concluded that the unclustered model is superior in both accuracy and simplicity.