Accurate　prediction　of　crude　oil　prices　is　meaningful　for　reducing　firm　risks,　stabilizing　commodity　prices　and　maintaining　national　financial　security.　Wrong　crude　oil　price　forecasts　can　bring　huge　losses　to　governments,　enterprises,　investors　and　even　cause　economic　and　social　instability.　Many　classic　econometrics　and　computational　approaches　show　good　performance　for　the　ordinary　time　series　prediction　tasks,　but　not　satisfactory　in　crude　oil　price　predictions.　They　ignore　the　characteristics　of　non-linearity　and　non-stationarity　of　crude　oil　prices　data,　which　hinder　an　accurate　prediction　and　eventually　lead　to　poor　accuracy　or　the　wrong　result.　Empirical　mode　decomposition　(EMD)　and　ensemble　EMD　(EEMD)　solve　the　problems　of　non-stationary　time　series　forecasting,　but　they　also　generate　new　problems　of　mode　mixing　and　reconstruction　errors.　We　propose　a　hybrid　method　that　is　combination　of　the　complete　ensemble　empirical　mode　decomposition　with　adaptive　noise　(CEEMDAN)　and　multi-layer　gated　recurrent　unit　(ML-GRU)　neural　network　to　solve　the　abovementioned　issues.　This　not　only　deals　with　the　issue　of　mode　mixing　effectively,　but　also　makes　the　reconstruction　error　of　data　close　to　zero.　Multi-layer　GRU　has　an　excellent　ability　of　nonlinear　data-fitting.　The　experimental　results　of　real　WTI　crude　oil　dataset　show　that　the　proposed　approach　perform　better　in　crude　oil　prices　forecasts　than　some　state-of-the-art　models.　©　2020　by　the　authors.　Licensee　MDPI,　Basel,　Switzerland.　This　article　is　an　open　access　article　distributed　under　the　terms　and　conditions　of　the　Creative　Commons　Attribution　(CC　BY)　license　(http://creativecommons.org/licenses/by/4.0/).