The　electric　power　industry　is　the　most　important　basic　energy　industry　in　the　development　of　the　national　economy.　The　operation　control　and　dispatch　of　the　electric　power　system　is　of　great　significance　in　ensuring　the　planning　of　the　electric　power　system,　industrial　development,　economic　operation　and　environmental　protection.　Short-term　power　load　forecasting　plays　an　important　role　in　power　system　operation　control　and　dispatch.　This　paper　proposes　a　combined　power　load　forecasting　model　based　on　similar　day　selection　and　improved　LSTM.　First,　a　combined　similar　day　screening　model　based　on　grey　relational　analysis　and　cosine　similarity　is　constructed,　which　makes　up　for　the　shortcomings　of　the　single　selection　method.　The　training　set　and　test　set　ensure　the　quality　of　the　input　data　of　the　model;　then　use　the　single-feature　time　series　training　to　increase　the　dropout　layer　of　the　LSTM　model　for　power　load　prediction,　reduce　the　resource　consumption　of　training　and　prediction,　and　effectively　alleviate　the　occurrence　of　overfitting.,　to　improve　the　robustness　of　the　model.　The　prediction　results　of　the　two　examples　basically　coincide　with　the　real　value　in　trend,　which　confirms　that　the　similar　daily　screening　and　improved　LSTM　combined　prediction　model　constructed　in　this　paper　can　provide　reliable　support　for　power　load　forecasting　and　other　time-series　data　forecasting　applications.　©　2023　IEEE.