	US 11,868,869 B1
	Non-intrusive load monitoring method and device based on temporal attention mechanism
Gang Huang, Hangzhou (CN); Wei Hua, Hangzhou (CN); and Yongfu Li, Hangzhou (CN)
Assigned to ZHEJIANG LAB, Hangzhou (CN)
Filed by ZHEJIANG LAB, Hangzhou (CN)
Filed on Jun. 28, 2023, as Appl. No. 18/215,784.
Claims priority of application No. 202211118482.1 (CN), filed on Sep. 15, 2022.
Int. Cl. G06N 3/04 (2023.01); G06N 3/049 (2023.01)

CPC G06N 3/049 (2013.01)

6 Claims

1. A non-intrusive load monitoring method based on temporal attention mechanism, comprising the following steps:

step 1: obtaining a total load data, an equipment load data, and corresponding sampling time of a building during a certain period of time;

step 2: integrating the total load data and the equipment load data with the corresponding sampling time to obtain an enhanced total load data and an enhanced equipment load data;

step 3: using a sliding window method to segment the enhanced total load data and the enhanced equipment load data, and constructing a deep learning training dataset;

step 4: constructing a neural network model based on a deep learning training framework and training the model using the training dataset obtained in step 3;

wherein, the step 1 is specially as follows:

obtaining a load power consumption of the building during a certain period of time, sampling time T=[t₀, t₁, . . . , t_n], and the total load data comprises a total load active power P_t_₀_:t_{_n}, a reactive power Q_t_₀_:t_{_n}, and an apparent power S_t_₀_:t_{_n}, the equipment load data comprises an active power P_t_₀_:t_{_n}ⁱ, a reactive power Q_t_₀_:t_{_n}ⁱ; and an apparent power S_t_₀_:t_{_n}ⁱof each equipment during the above sampling time, where i is the equipment number;

wherein, the step 2 is specially as follows:

firstly, converting the sampling time T=[t₀, t₁, . . . , t_n] to the corresponding number of weeks, that is T_d=[d₀, d₁, . . . , d_n], wherein, d_nrepresents the number of weeks the sampling time t_nbelongs to;

then, based on the attention mechanism, the total load data and the equipment load data are fused separately with T_d, as shown below:

F=attn(F,T_d)

among them, F can represent the total load active power P_t_₀_:t_{_n}, the reactive power Q_t_₀_:t_{_n}and the apparent power S_t_₀_:t_{_n}, as well as the equipment load active power P_t_₀_:t_{_n}ⁱ, the reactive power Q_t_₀_:t_{_n}ⁱand the apparent power S_t_₀_:t_{_n}ⁱ; F represents the load data after fusion and enhancement with time information T_d; obtaining the enhanced total load data P_t_₀_:t_{_n}, Q_t_₀_:t_{_n}and S_t_₀_:t_{_n}, and the enhanced equipment load data P_t_₀_:t_{_n}ⁱ, Q_t_₀_:t_{_n}ⁱand S_t_₀_:t_{_n}ⁱ; attn( ) represents the attention mechanism function; and

wherein, after the model training is completed, it is applied to load monitoring in other time periods, specifically, the method described in step 2 is used to enhance the active, reactive, and apparent power data of the building for any length of time period w, the trained neural network model is input, and the model output is the equipment's power consumption in the building during this time period.