A Dynamic Power Features Selection Method for Multi-appliance Recognition on Cloud-Based Smart Grid

Most current studies concerning appliance recognition focus on single appliance recognition, but for general home users, it is universal to simultaneously switch on and off multiple electric appliances. Therefore, this study discusses the recognition of a multi-appliance load, and aims to establish recognition sample data, while reducing the packet transmission quantity and computation complexity of the cloud server. This study proposes a dynamic power features selection method for multi-appliance recognition, which uses the electricity information detected by the smart meter to evaluate current operating condition and rate of change in order to dynamically determine the transmission interval time. As power features and electrical waveforms are not completely identical, the recognition architecture proposed in this study is divided approximately into two stages. The first stage of prediction is implemented by the Factorial Hidden Markov Model (FHMM), and in addition to doping out the presently probable load operation combinations, as well as their probabilities, the key point is to obtain all values after power feature standardization of each combination. The larger the value, the better the load condition combination represents the power feature. These values are ordered, and a specific percentage of the power features are selected to estimate the error of the recognition sample data, which is combined with the probability of the first stage as the final forecast result. According to the experimental results, in multi-load conditions, the first 25% of the power features have the maximum recognition of 83.75%. In the case of a single load, the first 75% of the power features have the maximum recognition of 94.59%.