Aiming at the low accuracy of long-term power load forecasting in some developed cities due to large load fluctuation, an auto-regressive integral moving average (ARIMA) forecasting method based on data-driven linear clustering (DLC) is proposed. Firstly, linear features are used as clustering criteria to pre-process the annual load data set of large substations in order to prepare for modeling. Then, the optimal auto-regressive integral moving average model is constructed for each sub-sequence to predict its corresponding future load. Finally, the long-term load forecasting results of the power system are obtained by summarizing all the forecasting results of the models. From the results of error analysis and application, both theory and practice verify that the proposed method can reduce the random prediction error while ensuring the accuracy of modeling, thus achieving more stable and accurate load forecasting results of power system.