Forecasting Electricity Price Spikes: A Comparative Analysis of GAM and SVM

Abstract

The Australian National Electricity Market (NEM) exhibits high price volatility, with prices determined every five minutes. Sudden and unpredictable price spikes create substantial uncertainty and financial risk for market participants, including generators, retailers, and households. Forecasting these rare events is challenging because non-spike periods can dominate the data, leading to the issue of data imbalance. This can bias traditional models toward the majority class, reducing their ability to accurately forecast the minority class (price spikes). This thesis addresses this challenge by adapting the Generalised Additive Model (GAM) and Support Vector Machine (SVM) with class weights to improve the prediction of rare spikes. Moreover, model performance is evaluated using imbalance-aware metrics such as the F1-score, G-mean, and the precision–recall area under the curve (PR-AUC). Results indicate that the GAM outperforms the SVM in predicting spikes while maintaining low false-positive rates. Both models balance performance across spike and non-spike classes and are robust to variations in weighting schemes and hyperparameters. These findings demonstrate the potential of GAM and SVM for electricity price spike forecasting (EPSF) in Australia, enabling generators to optimise production, retailers to maximise profits, and households to reduce costs.