According to a report from the United States Energy Information Administration (EIA), about 39% of total US electricity was consumed by residential buildings in 2018 [
1]. With the rapid development of intelligent building control algorithms, including predictive control and adaptive control strategy, it is practical and meaningful to put forward improvements in energy efficiency and energy management. Traditional predictive control algorithms, such as model predictive control (MPC) optimize future control actions at each time interval based on a selected physical system model. A model-based predictive control approach applied in building cooling systems with thermal energy storage was proposed in [
2]. The proposed approach achieved lower electricity cost and better performance by optimizing the operation schedules of the central plant based on predictions of weather conditions and building loads. An application of fuzzy model predictive control (FMPC) in building heating control was presented in [
3]. Disturbances of radiance, ambient temperature, and occupancy were considered in the FMPC-based building climate control system with regard to user comfort and monetary costs of heat supply. However, a widely recognized shortcoming of MPC is that it takes complicated and time-consuming procedures to capture a precise and accurate dynamic model of a physical system. There is always a gap between the optimization model and the realistic physical system. Moreover, traditional MPC is not suitable for applications with high frequency dynamics where the sampling time is measured in milliseconds or microseconds. Online optimization methods were implemented to improve the speed of MPC in [
4]. The improved MPC computed the control actions for a problem with a horizon of 30 time steps in around 5 ms, which allowed MPC to be carried out at 200 Hz, while the algorithm stability required further analysis. Most of the existing model-based predictive methods perform better when the accuracy of modeling is higher. The MPC and rule-based control (RBC) method proposed in [
5] outperformed robust model predictive control (RMPC) with regard to energy-comfort trade-off when the model uncertainty is less than 30% or more than 67%. The accuracy of the building models was found to be critical in the implementation of MPC algorithms [
6]. However, higher costs and more efforts of constructing an accurate dynamics system result from the growing scale of industrial production and the additional sophistication of system structure in recent years. The expectation of an ideal control effect cannot be obtained with the classical model predictive control algorithm which creates an opportunity for model-free control methods [
4]. Different from capturing a physics-based model, model-free control strategy is data-driven which relies on the input/output data of the selected system and does not need an explicit mathematical model. Previous studies [
7,
8,
9] indicated that the neural network-based control strategies were effective in controlling a broad class of nonlinear processes. A data-driven adaptive control method based on the lazy learning (LL) technique for a class of discrete-time nonlinear systems were proved practical in [
10]. The proposed LL-based controller is designed only using the input/output measurement data of the system by means of a pseudo gradient-based dynamic linearization technique. The robustness of the LL-based controller is verified to be enhanced due to the accurate prediction of the desired signal. The position tracking accuracy and processing efficiency of a non-circular cutting-derived Computer Numerical Control (CNC) system could be improved with a compact form dynamic linearization-based model-free adaptive predictive control method [
11]. A neural network-based predictive control method was introduced to solve the tracking problems of a nonlinear system successfully in [
12]. Moreover, multiple authors applied recurrent neural network algorithms in the domain of building energy predictions successfully [
13,
14,
15,
16]. Similar research was carried out on applying model-free methods to assisting system control operations [
17,
18,
19]. Ref. [
20] highlighted that model-free data-based control methods were practical in industrial practice if online calibration were to be performed. The concept of data predictive control (DPC), an alternative approach using control-oriented data-driven models for implementing receding horizon control, was mentioned in [
21]. The proposed control-oriented data-driven model uses two algorithms of a single regression tree and random forest to control system inputs and disturbances. The inputs and disturbances of a selected building model are separated into controllable variables and uncontrollable ones. This separation method is called recursive partitioning, which improves the performance of the system response prediction. Jain et al. [
21] demonstrated that DPC was a practical alternative to MPC in trading off energy consumption against thermal comfort without taking a real-time electricity pricing system into consideration. The primary reason for the modeling choice in [
21] was that complicated models like neural network went through a longer calculation routine and involved more factors, which put up barriers for an engineer to judge whether the operation was correct or not, than interpretable regression trees. However, Ahmad et al. [
22] compared the performance of artificial neural network (ANN) with random forest (RF) for the hourly HVAC energy consumption prediction of a hotel. The results proved that both models had comparable predictive power and nearly equal applicability in building energy applications. Ahmad et al. pointed out that ANN models were good choices to be used as surrogate models instead of detailed dynamic simulation models in real-time control applications. Moreover, ANN performed marginally better than RF with a lower root-mean-square error (RMSE) in general [
22].