To reduce the negative impact of inner heat gain of buildings caused by holidays on the accuracy of load forecasting in large central air-conditioning systems, an office building complex in Shanghai Expo Park is taken as the research area, a cooling load forecasting method for large central air-conditioning system based on improved Deep Deterministic Policy Gradient is proposed. New date-related feature named Days from Previous Holiday is introduced, the fully connected neural network in Deep Deterministic Policy Gradient structure is replaced by Long Short-Term Memory neural network, and a load forecasting model based on Recurrent Deterministic Policy Gradient is constructed. The experimental results show that the improved prediction model within Days from Previous Holiday can timely capture the load change trend caused by holidays, effectively improve the prediction accuracy which is 0.951 and the error value is 7.08%.