Medium-term load forecasting is influenced by multiple external variables such as temperature, holidays, and weekends. Although long short-term memory (LSTM) networks have shown strong predictive ability in short-term load forecasting, they cannot establish a good correlation between multiple external variables and predicted load in medium -term load forecasting. To address the above issues, parallel LSTM structures and time series N-node tree LSTM (t-N Tree LSTM) structures are proposed. By introducing branch structures and tree structures to construct finer feature granularity, modeling of medium -term load forecasting is achieved. Finally, experiments are conducted on the 2017 global energy forecasting competition dataset GEFCom2017, and the results show that finer feature granularity is beneficial for obtaining higher accuracy prediction results in the medium-term load forecasting process, verifying the effectiveness of the parallel LSTM model and t-N Tree LSTMs model.