Oh Seung-Taek, Lee You-Bin, Lim Jae-Hyun
Smart Natural Space Research Center, Kongju National University, Cheonan 31080, Republic of Korea.
Department of Computer Science & Engineering, Kongju National University, Cheonan 31080, Republic of Korea.
Sensors (Basel). 2025 Aug 19;25(16):5154. doi: 10.3390/s25165154.
Emerging lighting technology aims to enhance indoor light quality while conserving energy through control systems that integrate with natural light. In related technologies, it is crucial to identify quickly and accurately indoor light environments that are constantly changing due to natural light. Consequently, a large number of sensors must be installed, but installing multiple sensors would cause an increasing data processing load and inconvenience to users' activities. Some have attempted to calculate natural light characteristics, such as solar radiation and color temperature cycles, and implement natural light lighting technology by applying deep learning technology. However, there are only a few cases of using deep learning to analyze indoor illuminance, which is essential for commercializing natural light lighting technology. Research on minimizing the number of sensors is also lacking. This paper proposes a method for generating a detailed indoor illuminance map using deep learning, which calculates the illuminance values of the entire indoor area with a single illuminance sensor. A dataset was constructed by collecting dynamically changing indoor illuminance and the position of the sun, and a single sensor was selected through analysis. Then, a DNN model was built to calculate the illuminance of every region of an indoor space by inputting the illuminance measured by a single sensor and the position of the sun, and it was applied to generate a detailed indoor illuminance map. Research has demonstrated that calculating the illuminance levels across an entire indoor area is feasible. Specifically, on clear days with a color temperature anomaly of about 1%, a detailed illuminance map of the indoor space was created, achieving an average MAE of 2.0 Lux or an MAPE of 2.5%.
新兴照明技术旨在通过与自然光集成的控制系统提高室内光质量，同时节约能源。在相关技术中，快速准确地识别因自然光而不断变化的室内光环境至关重要。因此，必须安装大量传感器，但安装多个传感器会导致数据处理负荷增加，并给用户活动带来不便。一些人试图计算自然光特性，如太阳辐射和色温周期，并通过应用深度学习技术来实现自然光照明技术。然而，利用深度学习分析室内照度的案例很少，而这对于自然光照明技术的商业化至关重要。目前也缺乏关于减少传感器数量的研究。本文提出了一种利用深度学习生成详细室内照度图的方法，该方法用一个照度传感器计算整个室内区域的照度值。通过收集动态变化的室内照度和太阳位置构建了一个数据集，并通过分析选择了一个传感器。然后，建立了一个深度神经网络（DNN）模型，通过输入单个传感器测量的照度和太阳位置来计算室内空间每个区域的照度，并将其应用于生成详细的室内照度图。研究表明，计算整个室内区域的照度水平是可行的。具体而言，在色温异常约为1%的晴天，创建了室内空间的详细照度图，平均平均绝对误差（MAE）为2.0勒克斯，平均绝对百分比误差（MAPE）为2.5%。
