The　sulfur　dioxide　(SO2)　concentration　of　flue　gas　in　metal　smelting　industry　is　high　and　difficult　to　recycle.　The　commonly　used　SO2　desulfurization　treatment　is　acid　production　from　flue　gas,　however,　it　is　difficult　to　directly　measure　the　conversion　rate　of　SO2.　In　this　paper,　we　present　a　soft　sensor　method　based　on　Long　Short-Term　Memory　(LSTM)　which　integrates　the　attention　mechanism　for　predicting　SO2　conversion　rate　to　tackle　such　problem.　Considering　that　the　change　of　SO2　conversion　rate　is　affected　by　many　external　factors,　the　attention　mechanism　is　used　to　quickly　and　accurately　predict　the　change　results　by　considering　the　system　data　of　several　past　moments.　The　proposed　attention　mechanism　uses　LSTM　units　to　encode　the　hidden　state　of　the　look　back　time　data,　obtains　different　attention　weights,　and　then　decodes　and　predicts　SO2　conversion　rate　according　to　the　hidden　state.　The　experimental　results　indicate　that　LSTM　model　with　attention　mechanism　has　lower　training　cost　compared　with　LSTM　model.　The　training　accuracy　and　soft　sensor　accuracy　are　also　improved　owing　to　the　attention　mechanism.　It　is　instructive　for　SO2　conversion　rate　soft　sensor　in　acid　production　from　flue　gas.