Air conditioning load is a kind of flexible resources with fast response speed and large adjustable capacity. Through the analysis and modeling of the interaction potential of air conditioning load under multiple demand response time states

it can quickly respond to the grid side dispatch with little impact on user comfort

reduce the power demand in peak hours

and alleviate the contradiction between power supply and demand. In order to better reveal the various factors that affect air-conditioning participation in demand response

firstly

based on the thermodynamic model of air-conditioning

air-conditioning load is decomposed into static load and dynamic load. The concept of data-driven is different from the classical model-driven

which uses massive data acquired by collection or simulation. The deep feature relationships are mined and explored

and the problem architecture and solution ideas under the data-driven algorithm are established.Then

the constrained regression method and the data-driven temporal convolutional network-bidirectional gated recurrent unit-attention （TCN-BiGRU-Attention）neural network are used to estimate the adjustable interaction ability under different time states when the air conditioner participates in the demand response. The simulation results show that there is a significant correlation between the setting temperature and other factors and the static load of air conditioning

and there is a huge difference in the interaction ability between air conditioning load and the grid side under different demand response time scales. This method can effectively reduce the total cost of system calls and greatly improve the efficiency of air conditioning load participating in peak load shaving. The validity and accuracy of the method are verified based on the data of real users.