SCINet for Mitigating Distributional Bias in Electric Coal Inventory Prediction

Accurately predicting coal storage in power plants is crucial for optimizing coal storage management and ensuring power supply security. The nonlinearity and non-stationarity of coal storage sequences pose challenges to coal inventory prediction. In response to this issue, Sampling Convolution and Interaction Networks (SCINet) combined with an Attention mechanism was used to extract multi-scale nonlinear temporal features and correlations between variables for prediction. An improved Reversible Instance Normalization (RevIn) was introduced during data preprocessing to help to mitigate the effects of the non-stationarity of the coal storage sequence on the prediction model. Finally, by utilizing the relationship between coal stored, consumed, and incoming, the prediction results were adjusted to yield the coal storage prediction sequence. Experimental results showed an average improvement of 4.98%, 5.56%, and 4.29% in terms of single-day, 7-day, and 10-day coal inventory prediction accuracy, compared to before the improvement.