In multiple appliance load monitoring, variations in learning difficulty and numerical scale across target appliances can create imbalances in network parameter optimization, resulting in degraded performance for certain appliances. To this end, a tailored dynamic multi-target loss function is designed to adaptively assign rational weights for target appliances at each epoch, mitigating model bias toward specific appliances. Specifically, a global percentage error metric is employed to evaluate each appliance's performance on a unified scale, allowing dynamic weight adjustment to balance parameter optimization across appliances. This enables the proposed method to build mapping relationships and learn correlations across multiple target appliances, even in the presence of substantial differences in their usage patterns. Furthermore, a transformer-structure monitor is designed to integrate multimodal signals, combining raw data series with multi-step differential signals. This improves the model's learning capability to capture pattern changes in target appliances while enhancing robustness against anomalies.