Abstarct: Abstract Power system operation is significantly affected by short-term load forecasting. While short-term forecasting of the temporal pattern of load is subject to errors due to changing weather conditions and consumer behavior, existing deterministic forecasting methods are limited due to the neglect of uncertainties and random factors. For this reason, it is of great importance to develop methods that provide probabilistic forecasts of load values in addition to providing a deterministic forecast. In this thesis, a deep learning method is proposed for probabilistic forecasting of short-term load. In the first step, an LSTM network with two inputs, urban load and ambient temperature, is used to generate a deterministic forecast. Then, in order to improve the forecast accuracy and take into account uncertainties, the time dependencies of load are modeled as a time series. In the following, a new hybrid method using the SARIMA model is proposed for probabilistic load forecasting. Combining the LSTM network with the SARIMA model allows the proposed method to provide probabilistic load prediction in addition to having higher accuracy in deterministic prediction. The proposed method is validated using a dataset of urban electrical load and ambient temperature in Shahrood city, Iran. Numerical results show that the MAPE index value is less than 2%, indicating the appropriate accuracy of the proposed hybrid method. These results confirm the high potential of using deep learning along with statistical modeling for short-term electrical load prediction.