A study on forecasting system marginal price using a hybrid model based on UMAP dimensionality reduction

This study aims to enhance the prediction accuracy of the System Marginal Price (SMP) in the electricity market, which faces growing uncertainty due to the expansion of renewable energy, fuel price volatility, and rising demand.Various exogenous variables-including weather, oil prices, exchange rates, economic indicators, and electricity demand-were used.To address multicollinearity and high dimensionality, dimensionality reduction techniques like PCA and UMAP were applied.After dimensionality reduction, the study compared traditional time series models (ARIMA, ARIMAX), machine learning models (SVR, Random Forest, XGBoost, CatBoost), deep learning models (RNN, LSTM, CNN), and a hybrid CNN-SVR model.All models used the same external variables and a 7-day lag of SMP data.Results showed that the CNN-SVR model achieved the best performance, especially with UMAP-reduced data (4D).In contrast, LSTM and CNN performance dropped with UMAP 6D, likely due to loss of temporal structure.PCA improved results for traditional models like ARIMA and SVR.This research contributes to improved SMP forecasting by systematically comparing models and reduction techniques.Future work may explore online learning, Transformer models, and temporal-aware dimensionality reduction.