Abstract: In response to the challenges of assessing the potential for demand response due to the difficulty in observing residential users' internal private information and the lack of historical demand response data, a method for assessing the potential of residential user demand response based on Stackelberg game theory derivation and improved multi-task learning is proposed. First, a Stackelberg game theory derivation model is established between the electricity retail company and the residential user's home energy management systems to address the modeling issues of residential user demand response behavior under incomplete information. This model uses the equilibrium of virtual games to mine the electricity demand price elasticity coefficients from price-volume information and extracts demand response characteristic parameters. Secondly, a multi-task learning model establishes a mapping relationship between electricity usage characteristics and demand response characteristic parameters. A gradient normalization algorithm is employed to address the issues of inconsistent gradient magnitudes and mismatched convergence speeds among subtasks. This facilitates the generalized modeling of demand response characteristics for the residential user group, thereby enabling the assessment of the demand response potential of this group. Finally, the proposed method is validated using residential user data from Northern Ireland, demonstrating its superiority in mining demand response characteristics and assessing demand response potential.