Fault　prediction　is　a　crucial　strategy　for　ensuring　the　safety　and　efficiency　of　batch　processes,　and　it　aims　to　predict　the　degradation　trend　and　isolate　the　fault.　Operating　mistakes　or　external　environment　changes　can　lead　to　the　degradation　of　batch　processes,　eventually　resulting　in　faults.　The　point　at　which　degradation　begins　is　known　as　the　time　to　start　prediction　(TSP).　Determining　the　TSP　provides　more　reliable　degradation　information　for　fault　prediction　and　enhances　prediction　accuracy.　Nevertheless,　it　is　difficult　to　obtain　TSP　in　batch　processes　because　of　some　complex　data　features,　such　as　multi‐phase　and　strong　dynamics,　which　result　in　a　significant　loss　of　fault　prediction.　To　address　these　challenges,　a　fault　prediction　method　is　proposed　based　on　the　global–local　percentile　method　applied　to　Gaussian　error　linear　unit‐long　short‐term　memory‐encoder‐decoder　(GELU‐LSTM‐encoder‐ED).　Firstly,　squared　prediction　error　(SPE)　is　normalized　to　eliminate　differences　in　data　features.　Secondly,　a　global–local　percentile　method　is　proposed　to　determine　the　TSP.　The　period　is　divided　based　on　the　threshold　fluctuation　to　derive　the　model　input.　Then　the　global　percentile　model　is　employed　to　identify　the　interval　to　which　the　TSP　belongs.　Subsequently,　the　local　percentile　model　is　employed　to　analyze　this　interval　and　identify　the　TSP.　Thirdly,　the　LSTM　based　on　the　encoder‐decoder　model,　enhanced　with　the　GELU,　is　utilized　to　improve　the　performance　of　fault　prediction.　Finally,　two　fault　datasets　on　the　penicillin　fermentation　process　and　Escherichia　coli　fermentation　process　are　employed　to　demonstrate　the　performance　of　the　presented　method　compared　with　other　shared　methods.