Non-invasive Load Identification Based on Real-Time Extraction of Multiple Steady-State Parameters and Optimization of State Coding

In demand-side management (DSM), the growth in power consumption has widened the peak-to-valley difference of the grid, pushing up the demand for load regulation on consumer side. The variety of load types, coupled with the ever-changing power demand, directly bear on the operation state of the entire power regulation system. This calls for active response and flexible regulation of user terminal load. For this purpose, it is necessary to probe deep into the consumption features and information of loads, laying the basis for effective load identification [1, 2].

Load identification is inseparable from load monitoring. The two main load monitoring methods are intrusive load monitoring (ILM) and non-intrusive load monitoring (NILM). Currently, NILM has replaced ILM as the mainstream approach, thanks to its low cost, simple communication, and ease of maintenance and promotion [3, 4]. The essence of NILM is load decomposition. Owing to the sheer number of users and complex structure of load system, the key difficulty of NILM is to effectively decompose load type from the total load and identify the load state.

Over the past decade, many scholars have explored deep into load identification based on NILM. For instance, Liang et al. [5] recognized the working and non-working states of each load by committee decision mechanisms (CDMs), and proved that the CDMs outperforms load identification methods based on single feature or single algorithm. Using heuristic algorithm and Bayesian classifier, Marchiori et al. [6] conducted non-intrusive load identification based on steady-state power and step change of loads. Chang et al. [7] combined neural network (NN) and wavelet transform (WT) into a load identification method, which effectively improves the speed of load identification, but some loads might not be detected if the wavelet basis function is improperly selected before the identification.

Considering the transient features of load power, Gao and Yang [8] identified loads by comparing the closeness of the data on load features; Based on high-frequency sampling, this load identification strategy has a high requirement on sampling and induces a huge workload of data analysis, both of which limit the application scope of the strategy. Abdullah-al-nahid et al. [9] improved Canny edge detection to recognize the pattern of low-power loads, but the improved method is slow in detecting edge mutations and not highly accurate. Qi et al. [10] developed an identification method for household loads based on Fisher’s supervised discrimination, and obtained the optimal identification features by reducing the dimensionality of characteristic load samples from eight typical electrical appliances; Nevertheless, the identification accuracy of their method depends heavily on the accuracy of sample classification.

Yang et al. [11] put forward a load identification method that fuses feature sequences; On average, the load accuracy and identification accuracy of this method are both above 90.8%; But the high accuracy is achieved at the cost of a long time and heavy computing load. Wang et al. [12] created an NILM algorithm based on voltage-current (V-I) trajectory: The V-I trajectory increment was extracted through interpolation, and the loads were identified accurately with the aid of support vector machine (SVM). In 2019, Welikala et al. [13] designed a novel NILM approach that actively identifies real-time loads in view of appliance usage patterns (AUPs), shedding new light on load identification amidst massive load monitoring data.

The above non-intrusive load identification methods each has its merits and defects. During implementation, every method is limited by some conditions. If some conditions are not satisfied, it is difficult to create a highly accurate model by the method, making the method unpromotable in practice. To solve the problem, this paper presents a non-intrusive load identification method based on real-time extraction of multiple steady-state parameters and the optimization of state coding. Specifically, the improved affinity propagation (AP) algorithm was adopted to cluster the original load data, and construct a sample set of load power characteristic parameters. A load decomposition model was established, and the objective function was optimized by genetic algorithm (GA), facilitating the decomposition and re-identification of household loads. Empirical results show that our method achieved an accuracy above 96%. Our method applies to load identification against massive data. It is widely applicable, and easy to popularize. The research results provide a technical reference for non-intrusive identification of household loads.

3.1 NILM system

As shown in Figure 4, our NILM system is designed based on smart remote load controller (SRLC). The SRLC is a product independently developed by our research team. Embedded with advanced sensors and processors, the SRLC-based system can monitor the operation and energy consumption of user loads online, and, in association with multi-function smart meters, to collect all load data, preprocess current and voltage signals, and extract and analyze characteristic load parameters. The collection of active power was taken as an example to illustrate the proposed NILM system.

4.png

Figure 4. The monitoring circuit of the SRLC-based NILM

Let L₁, L₂, L₃, ..., L_N be the power loads of a household. Then, the set of samples for load data acquisition L can be described as:

$L=({{L}_{1}},{{L}_{2}},{{L}_{3}},...{{L}_{N}})$             (1)

where, L₁, L₂, L₃, ..., L_N are vectors. The load data that correspond to the vectors can be respectively expressed as:

$\begin{align} & {{L}_{1}}(k)=\{{{l}_{1}}(0),{{l}_{1}}(1),...{{l}_{1}}(k)\}, \\ & {{L}_{2}}(k)=\{{{l}_{2}}(0),{{l}_{2}}(1),...{{l}_{2}}(k)\}, \\ & {{L}_{3}}(k)=\{{{l}_{3}}(0),{{l}_{3}}(1),...{{l}_{3}}(k)\}, \\ & ..., \\ & {{L}_{n}}(k)=\{{{l}_{N}}(0),{{l}_{N}}(1),...{{l}_{N}}(k)\} \\ \end{align}$       (2)

where, k is the number of sampling points; N is the total number of appliances; L₁(k), L₂(k), L₃(k), ..., L_N(k) are the active powers of the sampling points. The sets of samples for the acquisition of reactive power and power factor were acquired in a similar manner.

3.2 Load feature sequence

K-means clustering (KMC) and fuzzy C-means (FCM) clustering are two popular clustering methods. For both methods, the number of clusters must be configured rationally in advance. The subjective configuration will affect the clustering results. Besides, the clustering results will change with the initial cluster center. Therefore, it is necessary to find a clustering algorithm that can automatically determine the number of clusters, reduce the influence of subjective factors, and stabilize the clustering results.

Through comparison, AP algorithm was found to satisfy the above requirements [16, 17]. However, the AP algorithm has difficulty in processing massive data, not to mention converging quickly to the optimal solution. Thus, the core vector machine (CVM) was introduced to improve the AP algorithm [18, 19]. The CVM compresses and processes the original load data, creating a load feature sequence Φ. The workflow of the improved AP algorithm is explained in Figure 5 below.

5.png

Figure 5. Flow chart of improved AP algorithm

Here, the improved AP clustering algorithm is adopted to split the data from the 30 appliances into workday sub-sequence and holiday sub-sequence, with the aim to improve the operability of load decomposition and the accuracy of load identification. 

As shown in Figure 5, the original sequence was compressed by the CVM into a new sequence X_i:

${{X}_{i}}=\{{{x}_{i}}(0),{{x}_{i}}(1),{{x}_{i}}(2)\cdots ,{{x}_{i}}(l)\}$                    (3)

where, l is the length of the sequence; x_i(l) is the feature points on the sequence.

Responsibility r and availability a [20] can be respectively computed by:

$r(\beta ,\delta )\text{=}s(\beta ,\delta )-\max \{a(\beta ,j)+s(\beta ,j)\} $              (4)

$a(\beta ,\delta )=\left\{ \begin{matrix}  \min \left\{ 0,r(\delta ,\delta )+\sum\limits_{j\ne \beta ,\delta }{\max \left\{ 0,r(j,\delta ) \right\}} \right\}  \\  \sum\limits_{j\ne \delta }{\max \{0,r(j,\delta )\}}  \\\end{matrix} \right\}$.                  (5)

where, r(β,δ) is the responsibility of point β relative to point δ; a(β,δ) is the availability of point β relative to point δ.

The household load sequence Φ of N appliances can be described as:

$\Phi =({{Y}_{1}},{{Y}_{2}},{{Y}_{3}},\cdots ,{{Y}_{N}})$              (6)

where, Y₁, Y₂, Y₃, ..., Y_N are vectors. The load data that correspond to these vectors can be respectively expressed as:

$\begin{align} & {{Y}_{1}}(n)=\{y_{1}^{1}(0),y_{1}^{2}(1),\cdots ,y_{1}^{{{W}_{1}}}(n)\}, \\ & {{Y}_{2}}(n)=\{y_{2}^{1}(0),y_{2}^{2}(1),\cdots ,y_{2}^{{{W}_{2}}}(n)\}, \\ & {{Y}_{3}}(n)=\{y_{3}^{1}(0),y_{3}^{2}(1),\cdots ,y_{3}^{{{W}_{3}}}(n)\}, \\ & \cdots , \\ & {{Y}_{N}}(n)=\{y_{N}^{1}(0),y_{N}^{2}(1),\cdots ,y_{N}^{{{W}_{S}}}(n)\} \\ \end{align}$                 (7)

where, Y_N(n) is load power eigenvectors; y_N is power feature point; n is the number of power feature points; N is the total number of loads; W_S is the number of all possible working states.

3.3 Load decomposition and optimization

The characteristic load sequence was taken as the known information and matching template, with the hope that the steady-state load features are repeatable and stackable. Repeatability means that the loads can be identified effectively with the same feature index and feature extraction method. Stackability means the characteristic parameters of a single load could be extracted from the overall load data of several operating equipment for state identification. When multiple power loads work at the same time, the total load power always changes with the state of appliances. Therefore, this paper identifies the working state of appliances through load decomposition, and then identify the type of loads. Based on steady-state load features, the local decomposition model can be approximated by [20, 21]:

$\left\{ \begin{align} & {{P}_{L}}(k)=\sum\limits_{i=1}^{N}{\sum\limits_{m=1}^{M(i)}{{{P}_{i.m}}(k)}}+e(k) \\ & {{Q}_{L}}(k)=\sum\limits_{i=1}^{N}{\sum\limits_{m=1}^{M(i)}{{{Q}_{i.m}}(k)}}+e(k) \\ & F(k)={{{P}_{L}}(k)}/{\sqrt{{{P}_{L}}{{(k)}^{2}}+{{Q}_{L}}{{(k)}^{2}}}}\; \\ \end{align} \right.$              (8)

where, P_L(k) and Q_L(k) are the total active power and total reactive power at the k-th sampling point; P_i,m(k) and Q_i,m(k) are the active power and reactive power of appliance i at the k-th sampling point in working state m; e(k) is the noise or error at the k-th sampling point.

Among the various types of consumer side loads, some loads have similar features in active power. Sometimes, when the working state of a load changes, it is impossible to obtain the accurate running state vector of the load. Therefore, this paper creates a multivariate identifier based on active power, reactive power, and power factor. For each sampling point, the objective function of optimization can be expressed as:

$\begin{align} & \min F({{P}_{L}}(k),{{Q}_{L}}(k),\Phi ,{{M}_{S}}) \\ & =\left| {{\lambda }_{1}}({{P}_{L}}(k)-{{\Phi }_{P}}({{M}_{S}})) \right| \\ & +\left| (1-{{\lambda }_{1}})({{Q}_{L}}(k)-{{\Phi }_{Q}}({{M}_{S}})) \right| \\ & +\left| {{\lambda }_{2}}({{F}_{L}}(k)-{{\Phi }_{F}}({{M}_{S}})) \right| \\ \end{align}$         (9)

where, $\lambda_{1} \in[0,1], \lambda_{2} \in[0,+\infty)$ are the weight of active and reactive powers and that of power factor, respectively; P_L, Q_L, F_L are the real-time measured values of total active power, total reactive power, and power factor, respectively; Φ_P, Φ_Q, Φ_F are the total active power, total reactive power, and power factor of household load sequence Φ found in M_S; M_S is the sequence of possible working states of all loads, which consists of load states S_h ($h \in[1, N]$):

${{M}_{S}}=\left\{ \begin{matrix}  {{S}_{1}}  \\  {{S}_{\text{2}}}  \\  {{S}_{\text{3}}}  \\  \begin{matrix}  \vdots   \\  {{S}_{N}}  \\\end{matrix}  \\\end{matrix} \right\}\text{ }{{\text{S}}_{N}}\in {{N}^{*}},{{\text{S}}_{N}}\in [0,{{W}_{S}}]$                (10)

At any sampling point, the working state S_h of a load and its corresponding power are unknown. Based on the characteristic load sequence, this paper attempts to find the optimal power under the working state combination $M_{S}^{\prime}$, which has the smallest Euclidean distance from the currently sampled power, according to the total power of the working states in the state sequence M_S. The optimal power is the currently identified power of the target load.

The working states of various loads can be combined in various forms. The heavy computing load drags down the computing speed. To improve the efficiency, the GA was adopted to solve the optimization problem [22, 23]. The workflow of the GA-based solving process is explained in Figure 6 below.

6.png

Figure 6. Flow chart of GA-based solving process

Based on the principle of NILM and the load parameter detection function of SRLC, the power consumption parameters (e.g. voltage, current, power, and power factor) could be acquired by installing a monitoring device at the supply entrance of power load. There is no need to install sensors within each appliance or load, reducing the investment and maintenance cost of hardware.

The AP algorithm was improved to cluster the collected load data. On this basis, a characteristic load sequence was constructed, which contains multiple steady-state parameters, and used as the reference template for load decomposition and identification. The clustering algorithm is not sensitive to the initial value or outliers, and outputs valid and stable results.

A load decomposition model was built based on the working state and coding information of appliances, and solved by the GA. Our method could correctly recognize more than 96% of loads, especially the loads with the same power or small power. In addition, our method can be easily transplanted onto software platforms, and applied well in various fields.