Abstract: To solve the problem of poor performance of conventional time series forecasting methods in long sequence forecasting scenarios, from the perspectives of time resolution dimension reduction and enhancing the mining of long-term dependency features of sequences, two modeling methods for medium-to-long term power forecasting were proposed, achieving power prediction with a span of 10 days and a time resolution of 15 minutes. On the one hand, an improved convolutional neural network gate and recurrent unit (abbr. CNN-GRU) network was proposed to reduce the time scale dimension, which respectively realizes the fusion/reduction of long time series and the forecasting of time series after dimension reduction through CNN module and GRU module. On the other hand, the multi-head attention mechanism based on the Informer network achieves the mining of long-term dependent features of sequences. The example results show that the two methods have different adaptability in different scenarios, and the accuracy/qualification rates on the 10th day can reach 74.21%/73.47% and 71.81%/74.48% respectively, which are significantly improved compared with conventional GRU, CNN and time convolutional network (abbr. TCN) models.