Abstract: Under the background of new power system construction, the diversification of renewable energy access and electric load type puts forward higher requirements for the accuracy of power load forecasting. However, the existing power load forecasting methods do not consider the negative effects of data noise on the forecasting model, which makes its forecasting accuracy being easily affected by the noise of observation data, and then the methods are difficult to be applied to actual scenarios. This paper firstly combs and studies the load forecasting methods based on deep learning, and introduces transformer algorithm to build a load forecasting model; secondly, a transformer-based robust power load forecasting model is formed by applying probability distribution preprocessing to the power load influencing factors with certain volatility, such as temperature and humidity; thirdly, by comparing and analyzing the transformer basic forecasting model and the transformer based robust forecasting model on the open original data set and the data set with noise, it is verified that the transformer load forecasting model can effectively deal with the data noise, and the transformer load forecasting model considering the data distribution can improve the robustness of the model on the basis of ensuring the forecasting accuracy.