To fully explore the timing and weather information in load data and improve the accuracy of power load prediction, a neural network based on variational mode decomposition（VMD）and bi-directional long short-term memory（BiLSTM）is proposed. Multi-dimensional sequential power load forecasting method leverages the strengths of VMD and BiLSTM to improve the accuracy of power load prediction. Firstly, through correlation analysis of multi-dimensional weather information and time sequence information, feature vectors with high correlation are selected as inputs. Meanwhile, VMD is used to decompose the original load data into intrinsic mode functions（IMF）of different frequencies. Then, the IMF and feature vector with high correlation are input to BiLSTM neural network optimized by sparrow search algorithm（SSA）for prediction. Finally, the predicted value of IMF is superimposed to obtain the final predicted value of power load. Load forecasting data set of 2016 electrical mathematical contest in modeling is used as an example to verify. Compared with BiLSTM and VMD-BiLSTM model, VMD-SSA-BiLSTM model can fully mine timing and weather information in data, and improve the prediction accuracy of multidimensional load data.