Development of a prediction model tuning method with a dual-structured optimization framework for an entire heating, ventilation and air-conditioning system

Abstract

• the proposed model predicts the behavior of an entire HVAC system. • the proposed method was applied to an actual building. • a prediction model was created with high accuracy and low computational load. • a case study was conducted to clarify required input information. Heating, ventilation, and air-conditioning (HVAC) account for a large proportion of energy consumption. Improving the energy efficiency of HVAC and utilizing increased amounts of renewable energy are effective strategies for achieving decarbonization. While thermal storage is one of the key technologies that solves a renewable energy issue that has the temporal and geographical gaps between supply and demand, operation planning using optimization methods, such as model predictive control, is important owing to the complexity of the system. A prediction model with high accuracy and low computational load is required because the performance of model predictive control typically depends on it. Further, to facilitate its extensive use in common buildings, it is necessary to develop a simple modeling method that requires no expertise. This study proposes a framework based on a dual-structured optimization process as a modeling method to create a prediction model. This method was applied to an actual small-scale office building. It was confirmed that such a model which accurately predicts up to 24 h ahead within approximately 1 s can be created. 1 1 AHP: air source heat pump chiller; ANN: artificial neural network; DB: dry bulb; DNN: deep neural network; GTHP: geothermal heat pump; HVAC: heating, ventilation and air-conditioning; MAE: mean absolute error; MLR: multi linear regression; MPC: model predictive control; m-PSO: particle swarm optimization incorporating mutation method; PSO: particle swarm optimization; ReLU: rectified linear unit; RMSE: root mean square error; TABS: thermally activated building system; WB: wet bulb; a : weight for DNN; b : bias for DNN; c : coefficient for m-PSO, c 1 = c 2 = 1.49618; f DNN : function of deep neural network; m rate : mutation rate for m-PSO; n : number of items/elements; r : random value r ∈ [0,1]; t: time step; t H : prediction horizon; v i : velocity of i th element for m-PSO; w : inertial coefficient for m-PSO, w = 0.7298; x : input data for functions; x gb : global best positions for m-PSO; x pb : personal best positions for m-PSO; y i : observed value of i th element; y ^ i : predicted value of i th element; φ : activation function; U : uniform distribution value in the available range.