Data play an essential role in the optimal control of smart buildings’ operation, especially in building energy-management for the target of nearly zero buildings. The building monitoring system is in charge of collecting and managing building data. However, device imperfections and failures of the monitoring system are likely to produce low-quality data, such as data loss and inconsistent data, which then seriously affect the control quality of the buildings. This paper proposes a new approach based on Gaussian process regression for data-quality monitoring and sensor network data compensation in smart buildings. The proposed method is proven to effectively detect and compensate for low-quality data thanks to the application of data analysis to the energy management monitoring system of a building model in Viet Nam. The research results provide a good opportunity to improve the efficiency of building energy-management systems and support the development of low-cost smart buildings.

Buildings with solar rooftops have become vital objects in the energy transition in Vietnam. In this context, the demand for research on energy management solutions to use energy efficiently and increase PV energy absorption capacity is rising. In this paper, we present a practical route to developing a low-cost monitoring platform to meet the building energy management in the country. First, our project built a monitoring architecture with high-density wireless sensors in an office building in Vietnam. Next, we discussed the influence of significant obstacles such as technical issues, users, and cost on the resilience and reliability of the monitoring system. Then, we proposed essential solutions for data quality improvement by testing sensors, detecting wireless sensor network errors, and compensating for data losses by embedding machine learning. We found the platform’s potential in developing a rich database of building characteristics and occupants. Finally, we proposed plans exploiting the data to reduce wasted energy in equipment operation, change user behaviors, and increase auto-consumption PV power. The effectiveness of the monitoring platform was an approximate 62% energy reduction in the first year. The results are a cornerstone for implementing advanced research as modeling and real-time optimal control toward nearly zero-energy buildings.

In distributed energy management systems, there is a critical need to provide low-cost, real-time, on-site Global Horizontal Irradiance (GHI) forecasting data. For this need, Tiny Machine Learning (TinyML) techniques on edge devices are a valuable and cost-effective solution. The paper investigates Tiny Transformer, a lightweight transformer version that performs better on time series data and is well-suited for edge devices. To evaluate the model's performance, two general tiny models, including RNN and ANN, are used. The collected data for model development query from an open weather data website, such as temperature, humidity, wind speed, sunshine duration, GHI, and hour of day. On Raspberry Pi and PCs, the results show that Tiny Transformer outperforms other models in terms of size, inference time, and accuracy. The highlight of this paper is the utilization of lookback with threshold error for model degradation analysis and the XAI method, SHAP (Shapley Additive Explanations) technique, to account for the influence of features on the model prediction. The paper shows the feasibility of implementing TinyML models and embedded AI on edge computing for energy prediction.

This study compared Random Forest (RF) and Long Short-Term Memory (LSTM) models for short-term wind speed forecasting. In addition to accuracy, model size, training time and inference latency were assessed. SHAP (SHapley Additive exPlanations) was used to identify key meteorological features and retrain reduced models. The results show that the LSTM achieves higher accuracy at short-term forecasting horizons, whereas the RF model is faster and more lightweight for real-time deployment. SHAP improved model efficiency by facilitating the selection of key input features for the models. It confirms the value for practical wind forecasting applications.