Aiming at the problems that the existing research methods do not adequately describe the quantitative relationship between time series accumulated temperature effect and power load change, and it is difficult to adapt to the rapidly changing climate environment, this paper proposes a long-term load time series forecasting method based on deep reinforcement learning for reliability assessment and considering the accumulated temperature effect. Firstly, the temperature correction model is learned through the depth deterministic strategy gradient method to quantify the nonlinear relationship between high temperature and load, reflect the impact of high temperature weather on the cumulative rise of load, and modify the climate temperature. Secondly, the other important influencing factors of power load except temperature are screened out, and the influencing factor data are preprocessed. Finally, based on the historical load, the corrected temperature data and the data of other influencing factors, the STL-LSTM-FED model is used to predict the long-term load time series data. The actual data of a province from 2012 to 2014 are selected for example analysis. Experiments show that, compared with the traditional method, the prediction accuracy of the model in extreme high temperature scenarios has been significantly improved, which verifies its adaptability to complex climate conditions. The research results not only improve the climate adaptability of load forecasting, but also provide data support for evaluating the security and stability margin of power grid equipment and reserve capacity configuration under extreme disaster scenarios by accurately capturing the load growth mode caused by temperature accumulation effect, and effectively enhance the prediction ability of power grid reliability.