Power distribution has become more efficient and reliable with the advent of smart grid. Although the distribution and supply services have been unbundled in recent years at most restructured electricity markets in smart grid systems. Having unbundled, load balancing, load forecasting and demand response management using smart meters continue to pose significant challenges to industry and research. Since the rate of power or electricity supply to customers has become competitive that needs to be tailored to the customer’s behavior based on accurate information from electricity suppliers. To do so, a deep knowledge of the customers’ behavior is required for the electricity provider not only for ensuring the uninterrupted quality of electricity but also for customer satisfaction. Again, for the sake of tariff planning, the customers’ contributions to the entire daily load pattern are gaining increased attention. Yet identifying and classifying customers effectively is very challenging. This paper elicits a general schema for analyzing electrical load supposedly used to produce effective results [1].

The basic model of a smart grid system is depicted in Figure 1, the interconnection of dependent domains used for power generation, transmission, distribution, market, operation, and customers as well [2].

The relevant consumption pattern for categorizing customers is likely to establish a suitable customer decision model on the basis of load pattern behavior.

An efficient clustering technique has been proposed in this study that employs the raw meter data collected daily from smart meter with an aim to load profiling, using clustering as well as the K-means algorithm. In doing so, the machine learning algorithm is used to time series data generated from the smart meter. This data source provides utilities with information that can be used to develop customized electricity load profiles (PCs) as well as to aid in areas such as improved load forecasting; Demand Side Management (DSM) strategies, Time of Use (ToU) tariff design and electricity settlement. Knowing the form of the power consumption might be helpful for deciding how to manage local energy generation and consumption for planning and operating energy systems. There is growing concerned about the type of electricity utilized as a result of the accessibility of cutting-edge load control technology and the expanding variety of flexible demand management options that provide incentives and rewards to the participating customers. Conceptually, electricity customers could be categorized based on their consumption. However, the following challenges are associated with the analysis of load profiling:

This paper focuses on the characteristics and habits of consumers and cites pertinent works in its discussion of these topics. The topic of evaluating the efficacy of clustering algorithms for classifying load patterns was then discussed in detail. The contracted power, supply voltage level, annual active and reactive energy (maximum, minimum, average value, and standard deviation), utilization level (defined as the energy consumption to rated power ratio), and power factor are a few examples of these external features. The usage of macro-categories, when applied to all electrical users, minimizes the number of load patterns that must be handled concurrently by the categorization methods inside each macro-category, leading to more cost-effective computation procedures.

The analysis involves two-fold steps as follows:

A sheer number of authors have already examined the load profile of providing services beyond billing using fine-grained data on electricity consumption. A lot of research has been done on non-intrusive load monitoring (NILM), which tries to figure out how much energy each appliance in a home uses. This NILM method uses measurements of power use taken at high frequencies, usually between 1 Hz and multiple times per second, and does not require the installation of extra sensors [1,4]. In order to deliver unique services, such as feedback to encourage more energy-efficient behavior among users, it is necessary to have the ability able to assign each appliance its true proportion of electricity consumption [2]. The accuracy of NILM techniques suffers in real-time contexts due to the lack of high-frequency measurements and the need for time- and data-intensive training operations [5]. In contrast to NILM methods, our focus is not on calculating the consumption of specific appliances but rather on identifying broad categories of end users. Also, we don’t need data sampled very frequently, as we may make do with data taken once an hour or even less frequently. Electricity consumption data was normally recorded at 15-, 30-, or 60-minute intervals, although a number of related methods relied on this data. Such data have even been used by a number of authors to examine how consumption patterns have changed over time [2,6,7]. Again in [8] Geoffrey et al. proposed a predictive model comparing the analysis of energy consumption among regression, decision tree, and neural networks. These frameworks aim to forecast future electricity consumption on the customer side and, by connecting these patterns across time, to support supply management on the electrical provider side [9]. According to [10], where 471 consumers were tracked down and inspected by the power company to do automatic clustering, this is a noteworthy example of the clustering techniques for grouping electrical load patterns. The authors find instances of wasteful billing methods throughout their analysis of the cluster results and current non-residential customer rates (for instance, when there is a weak association between discriminatory variables and real load patterns). Nishant et al. in [11] addressed an ANN model for accurate load forecasting to resolve the challenges imposed by conventional methods like mean and mode.The suggested methods concentrate on finding patterns in the available consumption statistics rather than tying them to particular traits of the family or commercial structure that caused the patterns to develop. Additionally, the conclusions from that approach are based on modest data sets that include traces from about 3000 households. In our research, we take into account consumption traces from more than 8760 hours of data) for a group of 24 representative facilities with a range of end-use levels. By utilizing self-organizing maps, Kushan et al. are able to discover new(commercial) consumers as well as consumption patterns that differ from "typical" behavior. There are a variety of other comparable techniques that use self-organizing maps (SOMs) to evaluate data sets that contain traces about numerous different customers[12]. A neural network-based unsupervised learning technique that can be used to automatically identify clusters from an otherwise unstructured (and unlabeled) set of data is the core of SOM. An hourly load data for a Set of 24 Facilities from residential, commercial, industrial, and various end-users has been calculated in [13]. To detect the anomalies in the power consumption in the smart grid, recognizing the load pattern is highly important [14]. However, various risks are associated while considering the smart meter data [15]. The wide ranges of cyber attacks [16] are liable to make any disruption to the meter data that might cause radical changes in load usage. Several machine learning algorithms such as classification, clustering, neural networks, and data mining are state-of-the-art techniques that can be applied to implement the precise load classification[17,18].

Load profiling in the big data era requires processing large amounts of data. With the proliferation of Advanced Metering Infrastructure (AMI) devices, through load profiling more information can be collected from the power consumption data of individual customers via the smart meter. Load profiling enhances our knowledge of electrical consumption patterns. This analysis performed load profiling among several ranges of electric users based on data from smart meters in order to understand the variance in load profiles at different time intervals.

Phase 2 includes the following steps:

Having determined the bounds clustering and unsupervised learning steps we then present the results of our comprehensive analysis of the load data and quantify the error introduced by the model.

Figure 1 shows how clients were categorized by time slot. One property is the maximum single value and the other four are the maximum electrical load. The raw dataset was reduced from 2,53,760 samples to 35,600 samples, with the most informative load profile clustering information preserved by the feature matrix. These load profiles were clustered using SOM. The total of each point’s distance to the centroid for all 5 groups is reduced iteratively by this clustering method. Most machine learning models for binary classification do not simply generate a value of 1 or 0 while making a prediction. Instead, they produce a continuous number in the [0,1] range. Values that are at or above a given threshold (for example, 0.5) are classed as 1, whereas those that are below the threshold are classified as 0. The points on the ROC curve represent the FPR and TPR for various threshold values. Classifications will change depending on the specified threshold value, which can range from 0 to 1. The "up and to the left" orientation of the ROC curve indicates superior performance when classifying the efficacy of a model.

Algorithm In this paper, we intend to implement machine learning algorithms on the Mendeley dataset of smart grid which is described in Section III. In the phase of Unsupervised training self-organizing map (SOM) is used. The comparison among classified results will be calculated by the error matrix afterward.

Tools In terms of tools, We have implemented our propounded model in MATLAB R2022a. The execution platform of both tools is on a Windows 10 PC with a configuration of an Intel Core i7 CPU (3.50GHz) and 64GB memory.

In smart grid technology, load profiling techniques largely improve the understanding of electricity usage trends. This research presents an analytical evaluation of data mining strategies for load profiling with respect to clustering methods and criteria for evaluating clustering results. The information contained in load profiling has a lot of potential use both in academia and industries not for only research but also making proper energy plan for future. In this investigation, smart meters has been employed and machine learning to do in-depth analysis for load forecasting, demand response management, and event detection. The purpose of this research is to provide a framework for using and experimentally analyzing detailed information on power use collected by smart meters. In order to reduce computational cost, dimensional, and redundancy in data during clustering, 14 features were discovered in smart meter data. In the future, the application of load profiling in demand response event can be taken into consideration, where the feature can be divided into the incentive-based and price-based programs. In the framework of massive energy data analysis, we aim to extend our work to some significant load profiling obstacles and hurdles.