Abstract: As the degree of coupling between electricity markets continues to deepen, the traditional feature set limited to a single market is not enough to support the demand for high-precision forecasting. Moreover, the model forecasting performance is sensitive to the choice of calibration window, while traditional electricity price forecasting only uses a time-fixed dataset, and the "black box" structure of the prediction model leads to low reliability of the prediction results in engineering applications. Aiming at the above problems, this paper proposes an interpretable two-layer day-ahead electricity price prediction framework considering the integration of calibration windows and coupled market features. The inner framework is the optimal prediction based on the improved adaptive noise complete ensemble empirical mode decomposition (ICEEMDAN), which first decomposes the original electricity price sequence, then applies the LEAR, LSTNet, CNN-LSTM, ARIMA and KELM models to predict the subsequence and selects the optimal prediction algorithm. The outer framework is the calibration window integration prediction based on Bayesian model average (BMA), and assigns weights to predictions under different calibration window length datasets and integrates them to obtain the predicted electricity price. Finally, the interpretable method SHAP is used to analyze how the coupled market features affect the predicted electricity price. In this paper, the superiority of the proposed algorithm and the effectiveness of the calibration window integration scheme are proved by an example analysis of the Nordic electricity market data set.