基于时空特征编码与动态分类器扩展的非侵入式未知负荷增量识别OA

In the dual context of promoting carbon neutrality globally and addressing the continuous growth of household electricity consumption,refined management of household energy is crucial.However,non-invasive load monitoring technology still faces severe challenges in dealing with unknown new load identification tasks,especially in scenarios where data is scarce and features are dynamically changing.Traditional methods often often suffer from insufficient generalization ability and catastrophic forgetting problems.Therefore,this paper proposes an unknown load increment recognition method based on spatiotemporal feature encoding and dynamic classifier extension fusion.By focusing on high robustness steady-state electrical features and integrating V-I trajectory spatial features with various temporal information,an RGB three channel spatiotemporal feature image is constructed for enhancing feature discrimination.Subsequently,transfer learning was performed on the ResNet34 backbone network and global average pooling layer based on adaptation and reconstruction to characterize known loads.On this basis,a few sample class incremental learning paradigm with dynamic classifier extension mechanism is introduced,which enables fast learning and recognition of new loads with only a small number of new samples(10-shots),and shortens the update time of new loads from hours to minutes in traditional"complete retraining",while effectively suppressing catastrophic forgetting.The experimental results show that the known load recognition performance is excellent on public datasets,and in the three unknown load incremental learning scenarios of self built datasets,the F1 score of this method is on average 1～3 percentage points higher than mainstream algorithms.The ablation experiment further verifies the synergistic effect of spatiotemporal feature encoding and core components in suppressing forgetting.