The invention discloses a load subdivision-based differential user baseline load prediction method, device and system, and the method comprises the steps: obtaining the power consumption load of a user, carrying out the data mining and analysis of the power consumption load of the user, and screening out an incident peak type load; calculating the meteorological sensitivity and the time sensitivity of the peak-attack load; determining a corresponding power consumption mode according to the meteorological sensitivity and the time sensitivity of the peak-attack load, the power consumption mode being a meteorological sensitive power consumption mode, a time sensitive power consumption mode or a meteorological and time insensitive power consumption mode; And user baseline loads under differentpower utilization modes are predicted respectively, and differential user baseline load prediction based on load subdivision is completed. According to the load subdivision-based differential user baseline load prediction method, device and system, on the basis of load subdivision, multi-dimensional power utilization behaviors and influence factors thereof are considered, and user power utilization behavior characteristics are refined, so that the user baseline load prediction precision is further improved.