In this paper, we propose a subband adaptive neural network based in-loop filter in VVC using cyclic discrete wavelet transform (DWT), named DWT-SALF. DWT-SALF takes advantages of subband adaptive learning based on DWT in the neural network-based in-loop filter (NNLF). Compared to the convolutional neural network (CNN), transformer is effective in capturing low-frequency features but has limited ability of constructing high-frequency representations. Thus, DWT-SALF uses transformer to handle the low-frequency subband, while utilizing CNN to treat the high-frequency subbands. We further enhance the high-frequency subbands with the guidance of the processed low-frequency subband. To increase the network depth and receptive field without increasing parameters, we adopt cyclic DWT that is cyclically used twice in the basic block and its affiliated branches of high and low frequency. Experimental results show that DWT-SALF achieves significant BD-rate gains of {-8.10 \(\%\) (Y), -21.19 \(\%\) (U), -22.28 \(\%\) (V)} over the VTM-11.0_NNVC-2.0 anchor under all intra (AI) configuration.