Non-intrusive load disaggregation technology offers significant potential for power resource optimization in smart grids. However, several challenges are faced by existing methods. Long-term temporal dependencies are inadequately modeled. Diverse appliance features are captured imprecisely. Predictions between power and state are also inconsistent. A multi-task non-intrusive load disaggregation architecture was proposed to address these limitations. The architecture was based on a transformer and a multi-scale attention convolutional network. Two parallel branches were contained within the architecture. These branches were designated as power disaggregation and state recognition. A transformer structure was used in the power disaggregation branch. A multi-head self-attention mechanism was employed by this structure to capture long-range dependencies and dynamic patterns in the load sequence. In the state recognition branch, a dedicated convolutional neural network was designed. Multi-scale convolutional modules and a channel-spatial attention mechanism were utilized by this network to extract key features of appliance switching events. Inter-task consistency was enhanced through a specific fusion mechanism. Element-wise multiplication was performed between predicted power values and state probabilities. The power output was directly constrained by the use of state information. As a result, appliances were identified in off-states. In these states, power false positives were significantly suppressed. The proposed model was evaluated on the public REDD (reference energy disaggregation dataset) and UK-DALE (UK domestic appliance-level electricity) datasets. Superior performance over existing mainstream methods is achieved. This performance is demonstrated across key evaluation metrics. These metrics are MAE (mean absolute error), SAE (signal aggregate error), and the F₁-score.