During　the　process　of　real-time　monitoring,　low　sampling　rate　make　it　difficult　to　construct　a　prediction　model　of　variable　due　to　the　lack　of　data.　Transfer　learning　addresses　the　dilemma　of　lacking　sufficiently　labeled　data　for　training　neural　networks　by　leveraging　relevant　labeled　data　for　knowledge　transfer,　which　can　significantly　improve　the　prediction　accuracy　of　the　low　sampling　rate　variable　by　utilizing　high　sampling　rate　variables.　When　the　feature　spaces　of　low　and　high　sampling　rate　indicators　do　not　coincide,　it　constitutes　a　special　case　of　transfer　learning　known　as　Heterogeneous　Transfer　Learning　(HTL).　One　classical　way　of　HTL　is　fine-tuning　the　pre-trained　source　model.　Nevertheless,　they　mainly　focus　on　optimizing　the　weights　of　pre-trained　models　and　ignore　the　mismatch　of　structure.　Therefore,　in　this　paper,　Domain　Perceptive-Pruning　and　Fine-tuning　(DP-PF)　is　proposed　for　HTL　to　simultaneously　tune　the　structure　and　weights　of　the　source　pre-trained　model　and　improve　its　adaptability　to　the　target　task.　Specifically,　DP-PF　proposes　target-perceptive　pruning　that　removes　unimportant　layers　from　the　source　pre-trained　model　based　on　the　target　pre-trained　model　to　tune　structure,　importance-perceptive　fine-tuning　with　adaptive　learning　rate　based　on　layer　importance　to　tune　the　weights,　and　source-perceptive　regularizing　to　mitigate　catastrophic　forgetting　of　original　knowledge　contained　in　the　source　model.　Experiments　are　constructed　based　on　the　wastewater　treatment　process　and　air　quality　prediction.　The　R-2　predicted　by　DP-PF　is　at　least　12%　higher　than　that　of　other　compared　methods.　The　excellence　of　DP-PF　in　accurate　cross-domain　prediction　proves　the　effectiveness　of　the　proposed　method.