CNN-TFT explained by SHAP with multi-head attention weights for time series forecasting