With the development of smart grid, the power consumption of each scenario becomes more diversified, and accurate station load forecasting is the key to ensure that the relevant power sector to develop appropriate maintenance tasks, while providing an important reference for orderly power consumption and economic operation. In order to mine the characteristics of the station load to improve the accuracy of the station load forecasting, a station power load forecasting based on the kernel principal components analysis combined with local preservation projection for dimensionality reduction, K-means clustering algorithm (K-means), and bi-directional long short-term memory network (BiLSTM) is proposed. Firstly, the kernel local preservation projection (KLPP) is used to reduce the dimensionality of multi-featured load data in the station area to extract the main feature information. Secondly, the K-means clustering method is adopted to classify the data with similar features into their respective cluster sets. Finally, for each typical type after clustering, BiLSTM is trained in a targeted way, and the load of a low-voltage station area of a university in China is selected as an example to be compared and analyzed with other classical forecasting methods. The proposed method is more suitable for the actual load direction and effectively improves the prediction effect.