Experiments were carried out on the measured data of two homes that verify the effectiveness of the proposed method. The measured data within 7 days were collected from two independent homes using acquisition devices. Indeed, the electrical appliances of the home were used as usual during the acquisition process, so the home appliance operation regularity was the same as usual. To verify the algorithm, a data acquisition device was also installed for each home appliance, thereby obtaining the operation regularity of the home appliance and providing a reference for the experiment results. The home voltage is 220 V, and the sample frequency of each acquisition device is 10 kHz. The home appliance includes one TV (TV), one refrigerator (RE), one microwave oven (MO), two air conditioners (AC-A and AC-B), two laptops (LAP-A and LAP-B) and two electric kettles (EK-A and EK-B). The second home appliance includes one TV(TV), one refrigerator (RE), one microwave oven (MO), one air conditioner (AC-C), one laptop (LAP-C), one rice cooker(RC) and one sterilizer (STE).
4.1. Verification of Signal Decomposition Effect
Figure 5 shows: the independently working current signal decomposed from complex home current signals, the actual current signal was directly collected by the acquisition device at the load side, and the error between the decomposed current signal and the actual current signal.
The two signals in
Figure 5 are almost identical and the error is small between the two signals. This means that the proposed algorithm can effectively extract the independent load working current from mix home current signal.
Table 1 shows the similarity coefficient and root mean square error between the decomposed signal and the actual signal in first home. As
Table 1 indicates, the similarity coefficients are higher than 0.9, and the average similarity coefficient is 0.9477. Furthermore, the mean square error between two signals is lower than 0.5, and the average mean square error is 0.305, proving that the proposed method effectively decomposes signals. The load characteristic is totally contained in the decomposed signal. Equations (7) and (8) calculate the similarity coefficient R and root mean square error (RMSE):
In Equations (7) and (8), and are the decomposed current and the actual current, respectively. and are the standard deviation for decomposed current and actual current, respectively.
Table 2 shows the similarity coefficient and root mean square error between the decomposed signal and the actual signal in the second home. As
Table 2 indicates, the similarity coefficients are higher than 0.96, and the mean square error between two signals is lower than 0.02, proving that the proposed method effectively decomposes signals for different homes.
4.2. Verification of Initial Load Identification
The working current signal of the switched load was decomposed and extracted, the working current signal was initially identified by the KNN algorithm. The number of
k in the KNN algorithm was set as five. The harmonic of load working current was treated as a characteristic dimension to calculate the Euclidean distance of the electrical characteristic.
Table 3 shows the first to fourth harmonics of six home appliances which are known samples in KNN.
The first to fourth harmonic current amplitudes of the unknown load, obtained by the load decomposition method, were used as the clustering center, and the Euclidean distance between the known load sample set and the cluster center was calculated to realize the initial identification.
Table 4 shows the types of home appliances involved in this paper, along with minimum and average Euclidean distance of the electrical characteristics between the decomposed load working current and five known samples.
Table 4 shows that unknown load one is the television, unknown three is the microwave oven, and the other identification results are obtainable by following the probability of the Euclidean distance.
Figure 6 shows the identification results of unknown loads two, four, six and eight, which are achieved through probability.
Figure 6 indicates unknown loads two, four, six and eight as being the electric kettle, air conditioner, notebook computer and electric kettle, respectively.
Table 5 shows the identification result. According to the
Table 5, unknown loads from one to seven were identified as RE, TV, MO, AC, LAP and RC, respectively. The unknown loads six and seven were both identified as RC, but can also be identified as SET, which produces mistakes in the identification result, the mistake will be corrected in modification step.
4.3. Verification of Modification Method
The electrical characteristic of decomposed current one is nearest to the TV, so the switched load was identified as TV, and it switched two times at 1pm. This data set trains the BP network. The trained BP network can fit the switching probability distribution curve of the TV.
Considering the mentioned example, four load switching times in the specific period are used to train the BP network. In the training process, the parameters of the BP neural network include the number of hidden layers (
Nhl), the number of neurons in each hidden layer (
Nneu1 and
Nneu2), the initial learning rate, iteration times and dropout rate in
Table 6.
Figure 7 shows the training process, while the network error decreases alongside increases in iteration times in the training process. When iterations times reach 10,000, the error of the four BP networks is close to zero, meaning the training of BP has concluded.
If the BP network finished the training, the load switching time distribution curve fitted by BP is near to the expected value. The identified result was input into the BP network to fit the switching probability distribution curve of the identified load.
Figure 8′s histogram indicates the switching times of the identified home appliances at each hour of the day. The switching probability distribution curve fitted by the BP neural network is close to the actual situation.
The current study obtained the switching probability distribution curves of identified home appliances. The abscissa is 24 h in day and the ordinate is the probability density. Therefore, the probability of the appliance used in each hour period is the area under the curve within the corresponding time period.
Firstly, one must analyze the EK-1, which is used all day, and the operation period is very regular. Normally, the hot pot is only used during a specific period of time rather than all day. Therefore, the switching probability distribution curve is not a hot pot; it can be judged that the load identification under the KNN algorithm is incorrect. According to the modification, the load should be classified as the refrigerator, because the refrigerator has the second highest probability in load identification part. Additionally, the time period of the refrigerator operation is consistent with the fitted the switching probability distribution curve.
Secondly, one must analyze the AC. According to the switching probability distribution curve, the air conditioner is mainly used at noon and at night, which is consistent with the normal operation conditions of the consumer. Indeed, the load identification result is correct.
Thirdly, the LAP was analyzed due to the switching probability distribution curve. Notebook computers are mainly used from 8 a.m. to 12 a.m. in the morning and 19 p.m. to 23 p.m. in the evening, which is consistent with the normal conditions. Accordingly, the result of load identification is correct. Finally, EK-2 requires analysis.
According to the distribution curve, the hot pot is always used in the morning, noon and evening, and is not used in other periods, which is consistent with the normal condition.
Figure 9 shows the switching probability distribution curve of two loads identified as the rice cooker in family two. It can be seen from the figure that the probability of RC-A running between 8:00 to 12:00 and 16:00 to 19:00 is relatively high, and these two periods are usually before lunch or dinner, which is consistent with the conditions of the normal user, so it can be judged that the identification result is correct. In contrast, the running time of RC-B is always from 20:00 to 23:00 p.m., which is generally after dinner, it is obviously inconsistent with the living habits, so it can be considered that the load identification has been misjudged, the identification result needs to be modified. By comparing with the data obtained by the independent data acquisition device, we can know that the load is indeed the disinfection cabinet. Therefore, the probability distribution curve of different loads can provide a reliable basis for the identification of similar load characteristics.
4.4. Comparison with Existing Algorithms
This paper uses the KNN algorithm to initially identify the unknown load. The common method only identifies loads based on various electrical characteristics, which is the final identification result. Furthermore, the BP neural network fits the switching frequency of different loads to obtain the load switching probability distribution curve. The initial identification of the load only provides a data bank for the modification step, so this paper selected the commonly used KNN algorithm to identify the load. As an important part of this paper, the process of load switching frequency, a BP neural network is used, which is not involved in traditional work. Compared with the method mentioned in this paper, the algorithm performance in reference [
21] performed Fourier transform on the current and treats harmonic as the feature to identify the load.
Figure 10 reveals a comparison of identification accuracy rate between the modified identification result and initial identification result. The accuracy of load identification was effectively improved following modification.
With the improvement of the load type number, the influence for the proposed method is less and the accuracy rate is higher. Indeed, the accuracy rate has been improved.