Abstract: To address the strong random fluctuations in load sequences, the large variations in electricity consumption patterns, and the limitations of a single forecasting model in adapting to commercial and industrial users, this study proposes a short-term load forecasting method that integrates adaptive deep clustering with a Fuzzy C-Means–Variational Autoencoder (FCM-VAE) framework. First, we construct an autoencoder based on the Time-Series Mixer (TSMixer) to extract latent -space features from load sequences, and incorporate a dynamic micro-cluster elimination mechanism to achieve adaptive clustering of electricity consumption patterns. Subsequently, separate BiLSTM-based sub-models are trained for each identified load cluster. Finally, the prior membership probabilities of samples with respect to each cluster center are computed using fuzzy C-means clustering, and the conditional likelihood distributions are reconstructed via the variational autoencoder. On this basis, dynamic fusion weights are generated through Bayesian inference, enabling adaptive weighting of the predictions from multiple sub-models. Validation using load data from industrial and commercial users in Shanghai shows that the TSMixer-based adaptive clustering model outperforms peer methods. The proposed model attains a MAPE of 7.3467%, an RMSE of 24.4 kW, and an R² of 0.9153, achieving an 8.28% relative reduction in MAPE over the unweighted fusion baseline, with further gains over both the no-clustering model and the static FCM weighting model. These results confirm that deep clustering combined with dynamic probabilistic fusion substantially improves short-term load forecasting accuracy.