Through　accurate　power　load　prediction,　the　smart　grid　can　provide　more　efficient,　reliable　and　environmentally　friendly　power　services　than　the　traditional　power　grid.　In　recent　years,　Recurrent　Neural　Networks　(RNNs)　have　received　more　and　more　attention　in　power　load　prediction　because　traditional　machine　learning　models　cannot　capture　the　time　dependencies　that　often　exist　in　power　load　data.　However,　due　to　the　self-connection　of　the　hidden　layer,　the　vanishing　gradient　problem　is　very　easy　to　occur　with　the　increasing　depth　of　the　simple　RNN,　which　leads　to　the　decline　of　the　prediction　accuracy.　In　order　to　solve　this　problem,　this　paper　adopts　the　recently　developed　RNN　architecture　called　Clock-Work　RNN　(CW-RNN)　on　the　task　of　one　day　ahead　prediction.　The　research　focuses　on　the　construction　strategy　of　recurrent　connection　matrix　related　to　the　hidden　layer　in　CW-RNN.　An　improved　CW-RNN　named　CW-RNN-SCR　adopts　the　strategy　of　connecting　adjacent　modules　into　a　simply　closed　ring,　which　not　only　improves　the　prediction　accuracy　but　also　reduces　the　number　of　parameters　required　for　training.　Experimental　results　show　that　the　improved　CW-RNN　can　achieve　better　performance　than　the　traditional　RNN　architectures.　©　2019　IEEE.