Heating, ventilation and cooling (HVAC) is the largest source of residential energy consumption. Occupancy sensors’ data can be used for HVAC control since they indicate the number of people in the building. HVAC/sensor interactions show the essential features of a typical cyber-physical system (CPS). However, there are communication protocol incompatibility issues in the CPS interface between the sensors and the building HVAC server. Through either wired or wireless communication links, the server always needs to understand the communication schedule to receive occupant values from sensors. We will illustrate the background of building energy consumption optimization problem. In this paper, we first propose two hardware-based emulators to investigate the use of wired/wireless communication interfaces for occupancy sensor-based building CPS control. The synchronization scheme between sensors and HVAC server requirements will be discussed. We have built two hardware/software emulation platforms to investigate the sensor/HVAC integration strategies. The first emulator demonstrates the residential building’s energy control by using sensors and Raspberry pi boards to emulate the functions/responses of a static thermostat. In this case, room HVAC temperature settings could be changed in real-time with a high resolution based on the collected sensor data. The second emulator is built to show the energy control in commercial building by transmitting the sensor data and control signals via BACnet in HVAC system. Both emulators discussed above are portable (i.e., all hardware units can be easily taken to a new place) and have extremely low cost. We test the whole system with YABE (Yet Another BACnet Explorer) and WebCTRL. Secondly, power system is facing a rapid transition to a highly interpretable, interactable and intelligent system. Effective simulations based on fast energy data processing algorithms have attracted many attentions due to the massive amount of data generated by the edge sensing devices in the smart grid systems. Machine learning (ML) and deep learning (DL) can be used to improve the performance of energy consumption forecasting. However, substantial computational resources are required for the training and inference of deep neural network. Instead of simply adopting the DL model for the offline processing of aggregated residential load, our platform can perform the online analysis of the load of building energy system with dynamic and stochastic characteristics. Today, edge computing platform that consists of a fine mesh of compute nodes and end devices, has become a promising system to reduce the computation complexity. In this paper, we propose an online, distributed, edge-computing-oriented simulation methodology to analyze the building energy data. A long short-term memory (LSTM) based framework is used for real-time forecasting of the building energy load. A public dataset is used to prove the effectiveness of the simulation model. The results show that the proposed simulation model provides satisfactory online load forecasting performance and has good scalability. At last, In order to reduce the negative influence on the environment and improve sustainability, it is very important to efficiently manage energy consumption. There are some energy prediction methods using sensors, such as collecting data utilizing Internet of Things (IoT) based on widely used smart meters. Some machine learning models including support vector regression are used to predict the energy consumption. However, they are not able to figure out the relationships between time dependency input signals. Therefore, a sequence to sequence convolutional bi-directional long short-term memory (Seq2Seq CNN bi-LSTM) with self-attention model is proposed in this paper in order to predict the building energy consumption. In the framework, the important time series energy information can be extracted from several input variables using CNN model. Because of the long-range temporal dependencies are offered, the Seq2Seq models could provide better accuracy by using two bi-LSTM architectures including encoder and decoder. Meanwhile, bi-LSTMs are also used to capture the pattern of time series data. Specifically, the above information and the trends of time series in two directions including the forward and backward states are used by bi-LSTM layers to make better predictions. Self-attention model can highlight the most relevant input information in the energy prediction by allocating the attention weights. The connection burden can also be alleviated by the attention mechanism. In this case, the self-attention can be used to ignore the irrelevant information and amplify the needed information. We may also apply deep reinforcement learning to the energy optimization problem.