Abstract:The precise forecasting of electricity prices for the following day holds significant importance in enhancing the revenue generated by market participants. Addressing the challenge posed by the increased presence of new energy sources in spot market transactions，a novel approach for predicting day-ahead electricity prices using light gradient boosting machine （LightGBM）is introduced. The method incorporates similar wave clustering and self-attention （SA）techniques. Initially，fuzzy C-means（FCM）clustering is applied to categorize fluctuation scenarios based on electricity price data characteristics，resulting in distinct datasets for each scenario. Subsequently，Kendall correlation coefficient is utilized to analyze the correlation of features influencing electricity prices across different fluctuation scenarios， facilitating the selection of input features. The SA-LightGBM prediction model is then constructed，tailored to the specific fluctuation scenario and input characteristics. Finally，the proposed algorithm’s effectiveness and dependability are assessed using historical data sourced from a provincial power trading platform in China.