There　are　complex　mapping　relationships　among　different　mill　load　parameters　and　multi-scale　frequency　spectra　of　ball　mill＇s　mechanical　vibration　and　acoustic　signals.　Aim　at　to　construct　an　effective　and　meaningful　soft　sensor　model,　how　to　select　interesting　input　variables　of　each　local-scale　frequency　spectrum　and　how　to　fuse　these　different　multi-scale　ones　jointly,　is　still　an　un-solved　open　issue.　A　new　method　based　on　LASSO　(Least　Absolute　Shrinkage　and　Selection　Operator)　and　SEN　(Selective　Ensemble)　algorithm　for　mill　load　parameter　forecasting　(MLPF)　is　proposed　in　this　paper.　Candidate　submodels　are　constructed　with　LASSO　based　on　each　local-scale　frequency　spectrum,　and　the　valued　input　features　are　selected　at　the　same　time.　These　ensemble　sub-models　are　selected　and　combined　based　on　SEN　approach　by　using　branch　&　bound　(BB)　and　adaptive　weighting　fusion　(AWF)　algorithms.　Regularization　coefficients　of　all　candidate　sub-models　are　selected　together　to　ensure　diversities　among　these　selected　ensemble　sub-models.　Multi-scale　frequency　spectra　data　of　mechanical　vibration　and　acoustic　signals　based　on　a　laboratory　scale　ball　mill　are　used　to　validate　the　proposed　method.　©　2017　IEEE.