Negative Price Forecasting in Australian Energy Markets using gradient-boosted Machines: Predictive and Probabilistic Analysis

With the integration of distributed energy resources such as roof-top solar panels and wind turbines into the grid, power generation can surpass demand-generation and thus, giving rise to the negative pricing, especially in the summer months. In this regard, a scientific case study is conducted in this paper to analyse and predict the increasing instances of negative energy prices against demand-generation in Australian energy markets (AEMs) using real-time energy data from the Hornsdale power reserve, South Australia. A robust machine learning method, Light gradient boosting machine (LightGBM) is utilised to detect and predict negative prices at different quantiles to quantity the outliers in the pricing data. The implementation results demonstrate that predicting the prices at different quantiles can tackle outliers (negative prices) effectively with the help of extracted upper and lower bounds using quantile regression-based approach. The case study is further extended to learn the complex statistical relationships between different data features using Naive-Bayes Tree Augmented (NB-TAN) algorithm considering 'price' as the dependent feature against the independent features such as demand-generation, battery charging/discharging, and frequency control ancillary services.