To solve the problem of unbalanced dataset clustering, this paper proposes an unbalanced data clustering method based on adaptive competitive learning. By optimizing competitive learning, new centroids are added adaptively to update the number of subclass centroids and integrate the structural features in the dataset. The two metrics of compactness and divisibility are combined to calculate the subclass merging difficulty coefficient to obtain the final clustering results. And the MCCL algorithm is tested for accuracy on the dataset characterized by imbalance, and the short-term power load forecasting step based on the combined MCCL-BILSTM model is designed. The PCA and K-means clustering algorithms are utilized to screen out the similar days for power load forecasting in response to the realistic demand of uncertain scenarios as well as multiple influencing factors of the combined power load. The MCCL-BILSTM model is utilized to forecast the electricity load for the four categories of similar days. In the comparison of short-term power load forecasting, the forecast curves of LSTM model, RNN model, BP model and GRNN model for similar days can roughly reflect the trend of the peak power load, but the MCCL-BILSTM model using clustering to screen the similar days is able to show a more superior forecasting effect.
