To generate a characteristic model conforming to reality even when past operation result data is not stored.SOLUTION: A chiller plant optimization system 200 generates, as learning data, a plurality of combinations among a plurality of pieces of operation information whose values are made different within a range permitted for operation information representing operating conditions of respective facilities of a chiller plant, a plurality of characteristic models whose values are made different within a range permitted for characteristic models representing physical characteristics of the respective facilities, and a plurality of power consumptions of the respective facilities which a simulator outputs for a plurality of inputs generated from the plurality of pieces of operation information and the plurality of characteristic models. A multilayer neural network model and the learning data are used to learn associations among the plurality of pieces of operation information, the plurality of characteristic models, and the plurality of power consumptions of the respective facilities, and a learning model which outputs one characteristic model for input of a combination between one piece of operation information and one power consumption of the respective facilities is generated. A characteristic model is generated from the learning model and the operation information.SELECTED DRAWING: Figure 2
展开▼