The　risks　associated　with　particulate　matter　on　human　body　are　not　limited　to　outdoor　environments　only.　Research　showed　that　outdoor　particulate　matter　can　infiltrate　inside　spaces　through　various　processes,　such　as　window　and　door　diffusion,　as　well　as　wall　penetration.　Majority　of　the　neural　network　based　existing　methods　for　predicting　indoor　air　quality　are　opaque　models　lacking　clear　physical　interpretability.　These　models　overlook　the　optimization　of　input　samples　as　well　as　the　seasonal　spatial　and　temporal　distribution　of　the　concentration　of　particulate　matter.　As　a　result,　the　accuracy　of　predictions　is　compromised,　and　the　computational　load　of　the　models　is　also　increased.　This　paper　presents　a　novel　model　for　predicting　the　concentration　of　indoor　particulate　matter,　called　the　Warped　K-means　-　Osmotic　Diffusion　Mechanism　-　LASSO　Attention　Temporal　Convolutional　Network　(WKM-ODM-LATCN).　Firstly,　redundant　variables　are　eliminated　using　the　LASSO　regression　algorithm.　Then,　sub-datasets　are　created　with　distinct　seasonal　characteristics　using　the　warped　K-means　algorithm.　Models　of　osmotic　diffusion　mechanism　are　constructed　for　each　season　based　on　the　principle　of　particulate　mass　balance.　Finally,　based　on　the　optimized　dataset　and　predicted　spatio-temporal　information　about　the　mechanism　model　as　the　input　values　of　the　multidimensional　attention　temporal　convolutional　network,　the　concentrations　of　indoor　particulate　matter　in　the　future　time　periods　are　finally　predicted.　The　obtained　results　demonstrate　that　WKM-ODM-LATCN　has　superior　prediction　accuracy　in　comparison　to　earlier　approaches,　hence　confirming　the　practicality　and　dependability　of　the　proposed　approach.　This　study　aims　to　effectively　include　spatial　and　temporal　environmental　data　in　order　to　accurately　estimate　the　concentration　of　indoor　particulate　matter.　Additionally,　it　demonstrates　strong　physical　interpretability.　©　2025　Elsevier　Ltd