Non-Intrusive Electric Load Identification Algorithm for Optimizing Convolutional Neural Network Hyper-Parameters

Anjun Zhao; Xiao Zhao; Jing Jing; Jiangtao Xi; Pufang Cui

doi:10.3788/LOP212374

Laser & Optoelectronics Progress, Volume. 60, Issue 2, 0210001(2023)

Non-Intrusive Electric Load Identification Algorithm for Optimizing Convolutional Neural Network Hyper-Parameters

Anjun Zhao¹, Xiao Zhao¹, Jing Jing^2、*, Jiangtao Xi¹, and Pufang Cui¹

¹College of Information and Control Engineering, Xi'an University of Architecture and Technology, Xi'an 710055, Shaanxi, China

²China Northwest Architecture Design and Research Institute, Xi'an 710018, Shaanxi, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(12)

Fig. 1. Schematic diagrams of VI trajectories of refrigerator and compact fluorescent lamp. (a) (b) VI trajectories; (c) (d) pixelated images

Download full size

Fig. 2. Structure diagram of CNN model

Download full size

Fig. 3. PSO-CNN recognition model training flow chart

Download full size

Fig. 4. Comparison of recognition accuracy

Download full size

Fig. 5. Confusion matrix comparison of recognition results of each algorithm. (a) (b) CNN confusion matrix; (c) (d) PSO-CNN confusion matrix

Download full size

Fig. 6. Comparison of feature extraction of each layer between CNN and PSO-CNN. (a) CFL; (b) refrigerator; (c) hairdryer; (d) fan

Download full size

Fig. 7. Model classification performance comparison. (a) F_macro based on PLAID dataset; (b) F_macro based on WHITED dataset

Download full size

Table 1. Hyper-parameters to be optimized and their range

View table

Table 1. Hyper-parameters to be optimized and their range

Layer	Hyper-parameter	Dynamicrange
Conv	Number of convolution kernels	30-150
	Convolution kernel size	2-7
	Convolution kernel stride	1-4
Pooling	Pooling kernel size	2-7
Pooling	Pooling kernel stride	1-4
Dropout	Dropout probability	0-1

Table 2. Parameters definition of particles in PSO

View table

Table 2. Parameters definition of particles in PSO

Particle parameter	Parameters to be optimized
a1	The number of convolution kernels of convolution layer 1
a2	The convolution kernels size of the convolution layer 1
a3	The convolution kernel stride of convolution layer 1
a4	The pooling kernel size of pooling layer 1
a5	The pooling kernel stride of pooling layer 1
a6	The number of convolution kernels of convolution layer 2
a7	The convolution kernels size of the convolution layer 2
a8	The convolution kernels stride of convolution layer 2
a9	The pooling kernel size of pooling layer 2
a10	The pooling kernel stride of pooling layer 2
a11	Dropout probability

Table 3. Comparison of recognition accuracy of algorithms
View table
Table 3. Comparison of recognition accuracy of algorithms
Model Recognition accuracy
PLAID WHITED
PSO-CNN 93.20 91.69
CNN 87.33 82.17
Reference［3］ 79.40 72.01
Reference［5］ 80.70 72.98
Reference［14］ 82.40 73.01

Table 4. Comparison of hyper-parameter information between CNN and PSO-CNN

View table

Table 4. Comparison of hyper-parameter information between CNN and PSO-CNN

Layers	Parameter	PLAID		WHIED
Layers	Parameter	CNN	PSO-CNN	CNN	PSO-CNN
Con1	Number	32	50	32	64
	Size	5×5	5×5	5×5	5×5
	Step	1	2	1	2
Pool1	Size	2×2	4×4	2×2	2×2
Pool1	Step	1	1	1	2
Con2	Number	64	80	64	120
	Size	5×5	3×3	5×5	3×3
	Step	1	2	1	2
Pool2	Size	2×2	2×2	2×2	4×4
Pool2	Step	1	1	1	2
Dropout	Probability	0.3	0.5	0.3	0.4

Table 5. Comparison of number of convolutional layer parameters, receptive field, and running time between CNN and PSO-CNN
View table
Table 5. Comparison of number of convolutional layer parameters, receptive field, and running time between CNN and PSO-CNN
Dataset Model Number of parameters of convolutions Model receptive field Running time /h
Con1 Con2 Sum
PLAID CNN 832 1664 2496 11 0.17
PSO-CNN 1300 800 2100 19 0.25
WHITED CNN 832 1664 2496 11 0.15
PSO-CNN 1664 1200 2864 39 0.32

Tools

Get Citation

Copy Citation Text

Anjun Zhao, Xiao Zhao, Jing Jing, Jiangtao Xi, Pufang Cui. Non-Intrusive Electric Load Identification Algorithm for Optimizing Convolutional Neural Network Hyper-Parameters[J]. Laser & Optoelectronics Progress, 2023, 60(2): 0210001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites