Forecasting the clearing price in the day-ahead spot market using eXtreme Gradient Boosting

Day-ahead prediction of electricity market price is a key for market participants to make a bidding strategy. While numerous methods for day-ahead market price (DMP) forecasting have been developed and applied, those lack interpretable models for DMP forecasting. This study provides a new hybrid model for DMP forecasting, centered on eXtreme Gradient Boosting (XGBoost). The suggested proposal uses neural networks (NNs) with entity embedding (EE) to preprocess categorical feature data and applying Shapley Additive exPlanations (SHAP) to select optimal features. In addition, the weights of NNs and SHAP are interpretability methods. The embedding weights from NNs can reveal the distribution of categorical features in the multi-dimensional space by t-distributed Stochastic Neighbor Embedding (t-SNE). SHAP can explain the impact of input features on the forecast result by assessing the importance of global and local features. The proposed model has been proven to be high accuracy, interpretability and outperforms other models on the test set accuracy, including support vector machine, artificial neural works, classification and regression trees, and random forest). Furthermore, the study found that those load/capacity ratios, the historical clearing prices, penetration of renewable energy, month, hour, weekdays, and weekends have a substantial impact on the prediction of the DMP.