Powering the Future: A Data-Driven Approach to Utility Operation & Maintenance

I. Introduction

The Industry 4.0 revolution is touching every industry, and the power sector is no exception. Power utility faces challenges viz. unexpected variable load growth, lack of systematic maintenance, repeated outages, poor asset management, unreliable power supply & pilferage of electricity. Machine learning has the potential to resolve all these problems, results reducing maintenance cost, enhance efficiency and improve planning, and better operation and control. We can addresses these issues of O&M by deploying data driven models in varied forms to mitigate the problems faced by the power utilities. Machine Learning enables predictive maintenance for identifying fault locations and analyzing the abnormality pattern and addressing them immediately.?The capacity addition of the power distribution transformer is based on peak load which can be predicted more accurately through the linear regression model. LT feeder load is manageable by implementing ML which uses proximation of grouping of LT feeder loading data. The geospatial analysis makes it more precise with respect to the location. Deep learning enabled AI applications can detect hot spots or faults in overhead lines, and an algorithm to learn and make a prediction based on a trained dataset can detect the failures before happen.

II. PROPOSED ALGORITHM & MODEL IMPLEMENTATION

A.?????Predictive Maintenance.

On the basis of past record of planned maintenance, some crucial parameters which affects lifespan of equipment of substation, have been taken viz. number of outages, duration of outages, peak loading of transformer and aging of transformer etc. After finding these crucial parameters of maintenance, through the rating and ranking method, an order-wise distribution substation list is created. A classification machine learning model is used to predict the substations most critical to maintenance work. Logistic regression has been selected with its high accuracy (94%) and implemented with the dataset. Logistic regression is most popular model of classification where it classified PM activity with the crucial parameters of maintenance. This is a powerful machine learning algorithm that uses the sigmoid function and has the ability to provide probability to classify the selection of preventive maintenance with a continuous and discrete dataset of crucial parameters.

B.?????Electricity Pilferage Detection.

In the existing system, T&D loss at LT level is calculated by comparing billed energy to the input energy of transformer. Field teams survey feeder-wise, identify the locations, and remove the possibility of theft from those locations. This is a time-consuming process.

The proposed system of detecting theft is based on data analytics and machine learning K means clustering model. K means clustering is an unsupervised learning model. By implementing this model, groups are formed, which are represented by the value of K [2].

In the high loss zone, based on the consumption analysis of total and night consumption patterns (Fig. 2), we can identify locations where the probability of theft is high. Additionally, in the same locality, K-means clustering segregates the high and low consumption pattern

(Fig. 3). As a result, we can save 60-80% of man-days.

(Yellow Colour Group - High probability of theft .

Red Colour Group - Medium probability of theft.

Green Colour Group - Low probabity of theft.)

C.?Transformer Load Forecasting.

Linear regression has been applied for forecasting the peak load of transformer of the year with respect to month and loading. Before applying train & test of the dataset, correlation has been found 84%. Equation for the linear regression is y = b0 + b1*x1+ b2*x2, where b0 is y intercept and b1 & b2 slope coefficient and y is dependent variable which provide peak load of transformer derived from the independent variable x1& x2. If we take the winter and summer peaks combined in the dataset, then non-linear regression will be applied to forecast the load growth, which will be in the form of y = ax2 + bx + c(Fig4). Based on the consitency of the load, we may plan capex and save the transformer and other equipments accordingly. [6].

D. Asset Management & Monitoring.

A significant number of equipment exists in the field, making it challenging to monitor them all effectively. However, equipment monitoring is crucial to ensure the health and proper functioning of these assets. Continuous monitoring reduces the likelihood of equipment failure.

Assets can be efficiently captured through OCR (optical character recognition), ensuring accurate data acquisition. Additionally, by obtaining geocoordinates, it becomes easier to locate, analyze, and monitor them using GeoPandas—an open-source Python library that combines the analytical capabilities of Python Pandas with geospatial functionalities.

OCR works by segmenting the text image into sections and identifying empty and non-empty regions. Pictures of equipment such as ACBs and RMUs are captured in text form, resulting in near error-free recognition

E. Inventory Management System

Many factors influence good inventory management, including financial considerations, suppliers, and effective management. However, one crucial factor is accurate planning and predictions of material consumption in projects and maintenance work. Stock prediction, based on the consumption patterns of materials using linear regression, stands as one of the best solutions. For instance, the consumption of piercing connectors is high during summer and low in winters. By analyzing the consumption pattern on a monthly basis, we can accurately predict the requirements. Similarly, the consumption of hardware fittings for LT poles and LT cables is high in winter and low in summer. Utilizing linear regression enables us to optimize the minimum stock level and improve material management.

F. ?Fault Analysis & Prediction.

Frequent tripping in the overhead HT conductor was identified as a major issue in every utility. The highest occurrence of tripping causing majorpower interruptions, revenue loss, and customer dissatisfaction. To address this problem, an exploratory data analysis (EDA) and statistical methods may be employed to analyze the location of faults section-wise and the frequency of faults month-wise using the dataset of HT tripping from the past year. This analysis enabled the identification of fault patterns through graphical representation.

Based on the insights gained from the analysis, solutions were developed and implemented in the field. As a result, there will be a significant reduction in trippings, leading to improved power reliability and customer satisfaction.

G.?LT Load Management

Transformer No load loss is constant but load loss varies as per the loading condition (Fig.6). Optimization of loading by transferring load from overloaded transformer to underloaded transformer enhance the efficiency of transformer.?

For LT feeder, loading loss is calculated by P=IR2 (Where P: Power, I: current, R: Resistance). If LT feeder is overloaded and another is under loaded as per figure. After optimization of loading into LT feeders, 50-80 % load loss can be saved(Table.1)

As per the requirement, the LT load can be shifted to another nearby LT feeder. In the next step, automating information from the LT feeder becomes possible through GeoPandas or Geopy, an open-source Python-based geospatial analytics tool. GeoPandas or Geopy comprises three fundamental components: Vector, Raster, and CRS. The Vector component represents geometries of LT feeder data, facilitating navigation from an overloaded nearby LT feeder to an underloaded one based on points, lines, or polygons, whichever is applicable. The Raster component deals with pixel data, enabling us to locate an LT feeder or transformer on the map. It resembles a Google satellite map, storing information about color, height, width, length, temperature, etc. Lastly, CRS (Coordinate Reference System) helps identify places through the latitude and longitude coordinate system. Pandas utilizes the data frame of LT loading to send analytic information about the underloading and overloading of nearby LT feeders.[5]

H. Fault Detection?

Finding the faults on overhead LT AB cables that cause leakage current in the neutral, leading to outages and technical losses, can be challenging.

The proposed solution involves an AI application based on Tensor Flow Object Detection. A dataset is generated with bound boxes of the whole image using a deep learning algorithm. After creating boundary boxes, features of faulty or burned trees are extracted and evaluated. In the final processing step, overlapping boxes representing faulty portions of cables are combined into one frame (Fig.7). Based on these visuals, a model is trained and tested using the dataset [4]."

III. Way Forward.

Eight models have been proposed and discussed based on the routine activities of utility operation and maintenance. Load forecasting plays a vital role in planning for capital expenditures and safeguarding LT and HT equipment from overloading and burning, ensuring consistency in loading.

To prevent asset theft or misplacement, an efficient asset tracking system based on machine learning is essential. Data clustering, based on consumption patterns, aids in theft detection, while implementing a smart metering management system will enable real-time theft detection in the near future.

Predictive maintenance using machine learning significantly reduces maintenance costs and allows for the identification of the most critical substation order-wise. Typically, maintenance of distribution transformer substations is based on visible faults or past problems faced.

Moreover, LT feeder load optimization and AI applications for fault detection in overhead lines further contribute to cost reduction and save man-days. The use of data analytics and data-driven decisions holds immense potential to improve operational efficiency, affordability, and transparency in the power utility sector.

IV. References

[1].?Farzana Kabir, Brandon Foggo, Student Member IEEE and Nanpeng Yu, Senior Member, IEEE, “Data Driven Predictive Maintenance of Distribution Transformers” China International Conference on Electricity Distribution, 2018.

[2]. P. Jokar, N. Arianpoo, V.C.M. Leung, "Electricity theft detection in AMI using customers? consumption patterns", IEEE Trans. Smart Grid, vol. 7, no. 1, pp. 216-226, 2016.

[3] Gopal lal Rajora, Miguel A. Sanz-Bobi, Carlos Mateo Domingo, “Application of Machine Learning Methods for Asset Management on Power Distribution Networks.

[4] Yellamma Pachipala;?M Harika;?B Aakanksha;?M Kavitha, “Object Detection using TensorFlow”, 2022 International Conference on Electronics and Renewable Systems (ICEARS)

[5] Naile Kas, Osman Bulent Tor, Mahmut Erkut Cebeci, “Theoretical and Practical Aspects of Implementing a Low Voltage Express Feeder between Nearby Distribution Transformers to Reduce Annual Losses”, ICSG Istanbul -2017.

[6] N. Amral;?C. S. Ozveren;?D. King,” Short term load forecasting using Multiple Linear Regression, 2007 42nd International Universities Power Engineering Conference.

#utilityanalytics #digitaltransformation #businessintelligence #forecasting #Disruptiveinnovation #datascience #machinelearning #energyanalytics #artificialintelligence ?#dataanalytics #deeplearning #objectdetection #tensorflow?#linearregression #classification #logisticregression #EDA #Kmeansclusterring #predictivemaintenance #loadmanagement #statisticalanalysis #powerdistribution #utility ?#OCR ?#python #geopanda?#geopy