Daily load curve clustering is an important basis for identifying user electricity consumption patterns and group load forecasting. To address the insufficient extraction of peak-valley and multi-time-scale fluctuation features, as well as the weak discriminability of the embedding space in existing deep embedded clustering methods, a multi-scale feature attention contrastive learning deep embedded clustering (MFA-CL-DEC) model was constructed. Multi-scale convolutional attention and soft-assignment contrastive constraints were integrated to improve feature representation and separability. Local and global features of daily load curves were extracted by multi-scale convolution. Key time-period information was highlighted by a temporal attention mechanism. Clustering centers were initialized by K-means++. The target distribution was constructed based on soft assignment probabilities, and the clustering loss was optimized. Positive and negative sample pairs were dynamically constructed according to clustering soft assignment results. A contrastive loss was introduced to enhance intra-cluster compactness and inter-cluster separation. Experiments were conducted on the Irish public load dataset. The results show that the proposed method outperforms the comparison models in terms of DBI, CHI, and SC. Ablation experiments and visualization results verify the effectiveness of multi-scale feature extraction and soft-assignment contrastive constraints. The clustering results are further used for group load forecasting. The prediction errors are lower than those of the non-clustering, random-grouping, and DCEC-1D grouping forecasting. The proposed method provides effective support for user electricity consumption pattern identification and group load forecasting.