A1041
Title: New Zealand electricity price forecasting: An analysis of statistical and machine learning models with feature selection
Authors: Gaurav Kapoor - Auckland University of Technology (New Zealand) [presenting]
Nuttanan Wichitaksorn - Auckland University of Technology (New Zealand)
Abstract: An empirical comparison is presented for statistical and machine learning models for daily electricity price forecasting in the New Zealand electricity market. The effectiveness of GARCH and SV models and their t-distribution variants when paired with feature selection techniques, including LASSO, mutual information, and recursive feature elimination, are demonstrated. A key aspect of the study is the inclusion of a diverse set of explanatory variables in all models. These models are compared against a range of popular machine learning models, including LSTM, GRU, XGBoost, LEAR, and a four-layer DNN, where the latter two are considered benchmarks. The results reveal that GARCH and SV models, particularly their $t$ variants, perform exceptionally well when paired with feature selection techniques and explanatory variables. In most scenarios considered, these models outperform machine learning models when coupled with LASSO feature selection. This contribution provides a comprehensive evaluation of the performance of different models and feature selection techniques for electricity price forecasting in the New Zealand electricity market. The best-performing model improves the symmetric mean absolute percentage error (sMAPE) and means absolute scaled error (MASE) by 2\% to 3\% over the LEAR benchmark model, highlighting the practical relevance of the findings.
