Accurate and reliable electricity load forecasting has provided great opportunities for the general efficient energy management, resource planning and grid operation. Short-term load forecasting is essential for estimating near-future power demand, facilitating efficient scheduling and stable execution of power systems. The technology still poses challenges due to the influences of environmental factors, seasonal and daily patterns as well as social behavior and user habits. In this paper, we propose a novel method called MmSN for short-term load forecasting via a multi-module structural network based on multi-feature fusion. Specifically, the network mainly involves the form of convolutional extraction, bidirectional long short-term memory and attention mechanism, which exquisitely picks the key impact of global feature fusion on the history load output and facilitates the efficient forecasting of the close-to-true value. Experimental results demonstrate that MmSN achieves competitive performance, exhibiting significantly lower errors and higher accuracy compared to other approaches, which greatly indicates the potential in the load forecasting task of electricity resource management.