Accurate prediction of regional electricity sales is crucial for the power sector’s effective energy management and planning. Existing forecasting models largely rely on historical electricity sales data and partially incorporate temperature effects, yet they inadequately consider a broad range of meteorological factors. In response, it introduces a novel forecasting method combining multi- head attention mechanisms with long short-term memory networks（MHAM-LSTM）for regional electricity sales. Initially, key variables are identified and redundant variables are eliminated through correlation analysis. Subsequently, the multi-head attention mechanism is used to focus on the key indicators that have an important impact on electricity sales. Finally, the LSTM network delves into the latent patterns of time-series data to forecast regional electricity sales. Experimental results show that the MHAM-LSTM model surpasses comparative models, including random forest, deep neural networks, long short-term memory networks, temporal convolutional networks, and transformer, in electricity sales forecasting accuracy, demonstrating significant performance advantages. Additionally, the analysis of meteorological factor importance reveals that incorporating multiple meteorological variables, particularly temperature, wind speed, and humidity, plays a crucial role in improving prediction accuracy.