Short-term　load　forecasting　for　buildings　is　crucial　for　effective　demand　management.　Traditional　approaches　that　treat　data　uniformly　for　training　and　testing　often　fail　to　account　for　internal　pattern　shifts,　potentially　compromising　predictive　accuracy.　This　paper　proposes　a　novel　load　forecasting　method　for　industrial　buildings,　incorporating　a　dynamic　weighting　strategy　that　acknowledges　these　internal　data　patterns.　We　introduce　an　adaptive　hybrid　clustering　algorithm　that　integrates　ordering　points　with　K　nearest　neighbors.　This　algorithm　categorizes　daily　load　curves　into　distinct　patterns,　training　specialized　models　for　each　category.　For　each　target　day　in　the　test　set,　the　K　nearest　neighbor　algorithm　determines　the　appropriate　clustering　category.　To　account　for　variations　in　load　data,　we　construct　separate　input　datasets　based　on　temporal　features　and　clustering　types,　which　are　then　used　for　model　training.　Finally,　dynamic　weights,　informed　by　historical　error　metrics,　optimize　the　integration　of　these　two　data　construction　methods,　enhancing　the　precision　of　the　predictions.　Experimental　validation　in　an　industrial　setting　demonstrates　the　superiority　of　our　method,　which　improves　the　mean　absolute　scaled　error　(MASE)　by　7.94%　and　0.88%,　respectively,　compared　to　methods　that　solely　rely　on　temporal　features　or　cluster　types.　©　2025　Elsevier　Ltd