Similar-Scenario Retrieval and Intelligent Hyperparameter Optimization for Day-Ahead Electricity Price Forecasting

INTRODUCTION: Day-ahead electricity prices exhibit pronounced volatility, nonlinearity, and temporal dependence. Training models on all available historical observations may introduce irrelevant information and reduce forecasting reliability. OBJECTIVES: This study proposes a hybrid forecasting framework that integrates same-period similar-scenario retrieval with intelligent hyperparameter optimization to improve the forecasting performance of day-ahead market-clearing prices. By refining the composition of the training dataset and optimizing the network configuration, the proposed framework enhances the model’s ability to learn the underlying dynamics of historical electricity prices. METHODS: First, similarity measures are used to evaluate historical electricity price sequences from corresponding periods. Historical observations that are highly relevant to the target forecasting period are then selected to construct a tailored training dataset. Subsequently, a hybrid CNN–BiLSTM–AM forecasting model is developed. Within this architecture, the convolutional neural network (CNN) extracts local patterns and salient features from the input data, while the bidirectional long short-term memory network (BiLSTM) captures long-term temporal dependencies in both forward and backward directions. The attention mechanism (AM) adaptively assigns weights to the learned representations, thereby emphasizing features that contribute more substantially to the final prediction. The dung beetle optimization (DBO) algorithm is further employed to optimize the key hyperparameters of the model. Finally, the effectiveness of the proposed method is evaluated using day-ahead market-clearing price data from the Miaojiaba Hydropower Station in the Bailong River Basin, Gansu Province, China. RESULTS: The empirical results demonstrate that the similarity-scenario-based sample reconstruction strategy effectively enhances the model’s ability to capture historical electricity price patterns. Taking a representative market price of 100 CNY/MWh as the reference, the proposed method reduces the standard deviation of forecasting errors by approximately 26 CNY/MWh compared with a conventional BiLSTM model without historical similar-sequence selection, thereby helping to mitigate potential revenue losses caused by electricity price forecasting deviations. CONCLUSION: The proposed framework effectively combines targeted sample selection, spatiotemporal feature learning, attention-based weighting, and hyperparameter optimization. It provides a reliable tool for day-ahead bidding decisions and supports efficient electricity resource allocation.