Electricity Theft Detection & Prevention using Artificial Intelligence for African Utilities
image credit: Source: Unsplash.com
- Dec 3, 2019 6:00 am GMT
- 1341 views
Modern smart grids rely on advanced metering infrastructure (AMI) networks for monitoring and billing purposes. Unfortunately, an approach such as this does suffer from electricity theft globally.
The AMI networks rely heavily on smart meters located on the customer’s premises to frequently report their energy consumption. However, this approach has the potential to hinder traditional physical electricity theft including line hooking or meter tampering (especially in the low-income areas).
Non-Technical losses result in high financial losses for several countries such as the United States ($6 billion/year) and India ($17 billion/year). Nigeria lost almost 40% of its electricity revenue due to non-technical losses.
The first nation on the continent to successfully digitize its built infrastructure, and thereby generate the data suitable for AI/ML, will reap huge benefits in improved infrastructure provision, in better public services, and in generating whole new areas of economic activity and enterprise. An application of Artificial Intelligence (AI) that provides systems the ability to automatically learn and improve from experience without being explicitly programmed.
The process of learning begins with observations or data such as examples, direct experience or instruction, in order to look for patterns in data and make better decisions in the future based on the examples that we provide.
We recently have been working on an algorithm that is aimed at the detection and prevention of electricity theft for African Utilities. The machine learning algorithm employed trains on a customer’s historical energy consumption to be able to predict future energy usage. An irregularity between the predicted energy usage and current energy usage will then be flagged as potential electricity theft, allowing rapid alarm detection. The model developed uses a type of decision tree machine learning algorithm to train and predict customer energy consumption.
The machine learning model is trained and tested on a Smart Meter dataset provided by one of the European Utilities. This data comprises of both residential and commercial premises. However, for this research, the residential smart meter data set was considered. The data is accumulated from over 5000 houses. The data contains the information about the customer id, code for date/time, electricity consumption for every 30 minutes (in kWh). This data was recorded for over two years.
The daily profile for every customer comprises then of 48 power consumption readings. The customer ID is crucial since the utilities can use this to get the location of the different customers. The dataset also comes with post-trial survey data which gives information on the number of inhabitants at each house
The model designed used a supervised machine learning algorithm called M5P decision tree to detect electricity theft. Decision trees are generated by algorithms that split a dataset into multiple branching segments based on decision rules. These decision rules are determined by identifying a relationship between input attributes and the outputs. Decision trees empower predictive modeling with higher accuracy, better stability and provide ease of interpretation. Decision trees are well suited for this task since there is a classification algorithm.
The training and validation dataset is then created for each of the different customer data. The training data was obtained over a certain season for a certain customer in one year and the validation data was then chosen for the same season and same customer but for a different year. There are 4 seasons in a year, but these can be divided into Summer and Winter due to the similarity of Autumn with Summer and Spring with Winter. Summer was then assumed to run from May to October and Winter from November to April. For instance, for customer 1020 in the dataset, our algorithm extracted their energy usage for each day from November to April 2009 as training data. The validation data is then extracted as the daily energy used from November to April 2010. This process was done for all the customers for the different months of the year. The training dataset is used to generate the decision rules representative of normal energy consumption for each of the customers in the training dataset. Decision trees are more prone to overfitting and hence careful consideration is made to prevent overfitting. Overfitting is a phenomenon in machine learning when the algorithm is too closely fit the training data that it cannot be able to reliably predict future observations that were not in the training data.
Since we did not have any physical data where the utility had flagged customers stealing energy, we resorted to simulating typical energy theft values. Energy theft cases were modeled to illustrate real-world energy theft scenarios.
Two types of energy theft were simulated with the first being when the consumer’s smart meter reports less energy consumption than actual consumed energy. This would normally occur either because of smart meter tampering to slow down its reading or bypassing of the smart meter.
The second energy theft type is one more common in third world countries, unauthorized tapping of the electricity line. This would result in a higher than normal smart meter reading. These 2 types of energy theft were simulated by adding/subtracting a random value deviating from 0 to 1 kWh to every measurement of the energy consumption data in the validation dataset. The original data was used to learn the consumption model and simulated energy theft data was used in validating the model for an energy theft case.
With the model built and currently testing we aim to build a software which will be able to display this data and give accurate reporting For a larger dataset with more number of features, a time-series based model can be trained ( using LSTM and GRU) for fine-tuning of hyperparameters to recognize highly complex, time-variant power usage patterns and determine fraudulent customers.
After analysis of Power usage patterns of different localities with more features, the system can be used by the authorities at higher levels. With the knowledge of power usage patterns, specific power demand for the future can be predicted which can help reduce transmission losses and installation of power storage infrastructure at specific locations.
This research demonstrated the successful application of decision tree learning for detecting energy theft. The conducted experiments unveiled the ability of the machine learning model to accurately predict energy consumption values from the same month of a year, subsequent weeks, and within the same weather season.
Furthermore, the historical data were used in these experiments to generate the machine learning model and predict future energy consumption. Using the RMSE (Root Mean Square Error) it was shown that the machine learning algorithm was able to accurately predict the future values and hence detect electricity theft. In the smart grid data analytics system, it is necessary to know the real-time electricity consumption data to forecast the exact future demand for electricity and plan accordingly. The identification of power theft will also extend its support for load forecasting that permits the utilities to exactly predict the power demand for future specific to an individual customer.
We can not argue that AI essentially promises to make the generation, transmission, and delivery of electricity more efficient, helping cut power costs and maximizing the use of renewable energy. If implemented right AI could end up being the most powerful asset for all Utilities globally.