Most of the Low-Carbon Technologies (LCTs) used today (mainly EV battery charging, rooftop solar systems and data centre loads) are all non-linear loads. Ohm’s law is not applicable to such loads, meaning that the current is not proportional to the voltage and fluctuates based on the alternating load impedance. Utilities face many problems as their traditional energy flow profiles change significantly with every LCT implementation.
This transient energy flow resulting from LCT adoption, coupled with ageing assets and growing electricity demand, puts extra stress on the local distribution networks. This may negatively impact end-customers who may witness increased electricity bills to meet up with the rising network reinforcement and maintenance costs. A significant increase in electricity bills would definitely not be acceptable to energy regulators from a socio-economic standpoint.
Many Utilities have experienced that high penetration of low-voltage interconnected solar energy systems can lead to repeated voltage swells at Substations, leading to frequent tripping of their inverter relays. Furthermore, voltage sags are likely to occur if electric vehicles dominate the same installation. Thus, traditional methods for meeting integration of new loads by simply tapping to Substation Transformers with available capacity ceases to be a viable solution for Utilities with large connected LCTs.
A leading Utility company with significant LCTs integrated to its grid was experiencing voltage fluctuations and low-voltage problems at the end nodes.
Our strategic partner, OrxaGrid, won the bid proposal to develop a cohesive solution to resolve the various challenges posed by the client’s fast-growing LCT infrastructure.
OrxaGrid developed a Machine Learning (ML)-based Voltage Violation Prediction algorithm that takes the load profile information of the Substation Transformers as inputs and alerts the Utility of any upcoming voltage violations of the limits set out by the energy regulator.
To predict the voltage violation events, OrxaGrid’s intuitive algorithm seamlessly integrated the Transformers’ real-time busbar voltage data, feeder current, real and reactive power, top oil temperature, and the outdoor ambient temperature along with its historical and external influence data.
At the core of the Voltage Violation Prediction algorithm was an intricate gradient boosting regression tree model integrated with a squared error objective function that made the entire model highly customisable depending on a wide set of model-tuning hyperparameters. This meant that the trained model size could be quickly traded off against the learning rate thereby making the entire model highly scalable. The regression model was trained with cyclical input features, engineered from sensor timestamps that minimised the requirement for external data feeds such as weather feeds.
Here are two snapshots of the voltage prediction plots that were presented by the Voltage Violation Prediction algorithm:
A validation of historic data demonstrated that temporal voltage profiles could be predicted a day in advance with an average root mean square error of only 0.776 V.
Furthermore, the violation alerts generated by OrxaGrid’s solution accurately identified the locked-out Transformers’ tap changers and undeclared voltage over-generation on domestic solar energy installations.
Built and deployed on an edge-computing platform with limited computational resources, OrxaGrid’s Voltage Violation Prediction solution successfully provided the Utility deeper insights into its low-voltage networks that enabled it to profile Substation-wise future risks and plan in advance the reinforcement activities based on medium and longterm violation predictions.