Accurate　electric　load　forecasting　can　prevent　the　waste　of　power　resources　and　plays　a　crucial　role　in　smart　grid.　The　time　series　of　electric　load　collected　by　smart　meters　are　non-linear　and　non-stationary,　which　poses　a　great　challenge　to　the　traditional　forecasting　methods.　In　this　paper,　sparse　representation　model　(SRM)　is　proposed　as　a　novel　approach　to　tackle　this　challenge.　The　main　idea　of　SRM　is　to　obtain　sparse　representation　coefficients　by　the　training　set　and　the　part　of　over-complete　dictionary,　and　the　rest　part　of　over-complete　dictionary　multiplied　with　sparse　representation　coefficients　can　be　used　to　predict　the　future　load　value.　Experimental　results　demonstrate　that　SRM　is　capable　of　forecasting　the　complex　electric　load　time　series　effectively.　It　outperforms　some　popular　machine　learning　methods　such　as　Neural　Network,　SVM,　and　Random　Forest.　©　2019,　Springer　Nature　Switzerland　AG.