Abstract: Load forecasting is a key part of many power grid applications and plays an important role. However, the nonlinearity, time-varying characteristics and uncertainty of grid load are the challenging to accurate load prediction, therefore, it plays a vital role to improve prediction accuracy by fully mine potential characteristics of load sequence. It is considered that the position information, trend, periodicity and time information of the load sequence should be fully utilized in feature extraction, and a neural network framework at a deeper level should be constructed for feature mining. For this reason, a load forecasting model based on feature embedding and Transformer framework was proposed, and the proposed model was composed by a feature embedding layer, a Transformer layer and a prediction layer. In the feature embedding layer, firstly, the location information, trend, periodicity and time information of the historical load were embedded into a characteristic vector, and then its output feature vector was blended with meteorology information to obtain the feature vector. The Transform layer accepted the feature vector of historical series and mined the temporal nonlinear dependence hidden in the sequence based on the obtained feature vectors of load sequence. Through the fully connected network the prediction layer implemented the load forecasting. Experimental results show that the forecasting performance of the proposed model is better than that of the comparative model, thus both feasibility and availability of the proposed model are verified.