In order to address the current issues of insufficient extraction of deep load features, low decomposition accuracy, and high training costs in the load decomposition model, a multi-scale feature fusion model was proposed. The model was composed of two parts: the load decomposition subnetwork and the load recognition subnetwork, both of which were employed with convolutional blocks composed of one-dimensional convolution and batch normalization for the initial extraction of load features. Subsequently, a pyramid pooling module was incorporated to precisely extract deep load features from multiple dimensions and fused them with the initial feature extraction part. Network parameters and training costs were significantly reduced by the pyramid pooling module. At the same time, in contrast to previous models with attention mechanisms, a multi-head self-attention mechanism was incorporated by the network. Different segments of load features were focused on by each attention head, achieving the selection of crucial load characteristics from multiple perspectives and further enhancing the performance of load disaggregation. Finally, experiments on the UK-DALE and REDD datasets show that the proposed model outperforms four benchmark models in both load disaggregation performance and appliance operation state recognition ability.