Accurate mid-term power load forecasting is crucial for power dispatch and resource optimization. Addressing the practical need for daily peak/valley load management in power scheduling, medium-term forecasting is studied with daily maximum/minimum load as the prediction granularity. To overcome the error accumulation caused by the decay of coupling relationships between historical loads and multi-dimensional external variables in traditional methods, a deep neural network time-series forecasting approach incorporating a cross multi-head attention (CMA) mechanism is proposed. The model features three innovative designs: first, a dual-branch convolutional neural network (CNN) and long short-term memory (LSTM) network are used to extract the local pattern features of the load sequence and the global temporal correlation of auxiliary variables; second, a cross-multi-head attention layer is designed to establish a dynamic weight mapping between historical load and external variables in future periods; finally, a gated recurrent unit (GRU) achieves adaptive fusion of multi-scale features. Experimental results demonstrate that the model achieves high accuracy and ro-bustness in power load forecasting tasks.