Abstract:

The DC series arc fault in photovoltaic systems caused by insulation aging or loose wiring is highly prone to electrical fires.Therefore,the arc fault detection devices must be installed in photovoltaic systems.However,the arc fault detection devices easily malfunction due to DC-side high-frequency noise caused by shadow occlusion and inverter startup.A novel arc fault detection method is proposed based on attention weight screening of arc features by combining the attention mechanism with the 1d convolutional neural network.By visualizing the contribution weight of arc features,the critical feature bands of 8~18 kHz and 28~38 kHz are extracted,and the interference arc features bands in the 8~23 kHz frequency band are removed.It has been verified that the arc fault detection model trained with the key arc features can successfully avoid the false activation caused by shadow occlusion and inverter startup,and the an arc detection accuracy of 99.33%is ultimately achieved.

Key words: series arc fault, convolutional neural network, attention mechanism, arc fault recognition, photovoltaic system