Recurrent Neural Network Architectures Comparison in Time-Series Binary Classification on IoT-Based Smart Lighting Control

Recurrent neural network (RNN) is suitable for sequential cases because they have feedback in their structure to recognize a dynamic behavior and a sequential pattern in time-series data. However, in a binary classification problem on internet of things (IoT)-based smart lighting control, there are several optional applications of RNNs with different architectures that are applicable to the problem. This paper proposes a comparison of several RNN architectures in a case study of IoT-based smart lighting control. An IoT implementation, including passive infrared (PIR) Sensors, NodeMCU, light-emitting diode (LED) lighting, relays, and Raspberry Pi, produces a dataset of smart lighting control. We implement four types of RNN architectures including stacked long short term memory (LSTM), stacked LSTM-decision tree (LSTM-DT), simple RNN, and RNN-DT. Results show that stacked LSTM and simple RNN produce the best performance out of the four architectures with 95% accuracy. However, simple RNN has a significantly lower delay compared to other methods, with a mean ±standard deviation of (109.8±65.8) ms. We discuss several comparisons of the obtained results with state-of-the-art research.